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  */
  31
  32 #define DEBUG_SUBSYSTEM S_LNET
  33
  34 #include <linux/ctype.h>
  35 #include <linux/log2.h>
  36 #include <linux/ktime.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/uaccess.h>
  39 #ifdef HAVE_SCHED_HEADERS
  40 #include <linux/sched/signal.h>
  41 #endif
  42 #include <lnet/udsp.h>
  43 #include <lnet/lib-lnet.h>
  44
  45 #define D_LNI D_CONSOLE
  46
  47 /*
  48  * initialize ln_api_mutex statically, since it needs to be used in
  49  * discovery_set callback. That module parameter callback can be called
  50  * before module init completes. The mutex needs to be ready for use then.
  51  */
  52 struct lnet the_lnet = {
  53         .ln_api_mutex = __MUTEX_INITIALIZER(the_lnet.ln_api_mutex),
  54 };              /* THE state of the network */
  55 EXPORT_SYMBOL(the_lnet);
  56
  57 static char *ip2nets = "";
  58 module_param(ip2nets, charp, 0444);
  59 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  60
  61 static char *networks = "";
  62 module_param(networks, charp, 0444);
  63 MODULE_PARM_DESC(networks, "local networks");
  64
  65 static char *routes = "";
  66 module_param(routes, charp, 0444);
  67 MODULE_PARM_DESC(routes, "routes to non-local networks");
  68
  69 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  70 module_param(rnet_htable_size, int, 0444);
  71 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  72
  73 static int use_tcp_bonding;
  74 module_param(use_tcp_bonding, int, 0444);
  75 MODULE_PARM_DESC(use_tcp_bonding,
  76                  "use_tcp_bonding parameter has been removed");
  77
  78 unsigned int lnet_numa_range = 0;
  79 module_param(lnet_numa_range, uint, 0444);
  80 MODULE_PARM_DESC(lnet_numa_range,
  81                 "NUMA range to consider during Multi-Rail selection");
  82
  83 /*
  84  * lnet_health_sensitivity determines by how much we decrement the health
  85  * value on sending error. The value defaults to 100, which means health
  86  * interface health is decremented by 100 points every failure.
  87  */
  88 unsigned int lnet_health_sensitivity = 100;
  89 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
  90 #ifdef HAVE_KERNEL_PARAM_OPS
  91 static struct kernel_param_ops param_ops_health_sensitivity = {
  92         .set = sensitivity_set,
  93         .get = param_get_int,
  94 };
  95 #define param_check_health_sensitivity(name, p) \
  96                 __param_check(name, p, int)
  97 module_param(lnet_health_sensitivity, health_sensitivity, S_IRUGO|S_IWUSR);
  98 #else
  99 module_param_call(lnet_health_sensitivity, sensitivity_set, param_get_int,
 100                   &lnet_health_sensitivity, S_IRUGO|S_IWUSR);
 101 #endif
 102 MODULE_PARM_DESC(lnet_health_sensitivity,
 103                 "Value to decrement the health value by on error");
 104
 105 /*
 106  * lnet_recovery_interval determines how often we should perform recovery
 107  * on unhealthy interfaces.
 108  */
 109 unsigned int lnet_recovery_interval = 1;
 110 static int recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp);
 111 #ifdef HAVE_KERNEL_PARAM_OPS
 112 static struct kernel_param_ops param_ops_recovery_interval = {
 113         .set = recovery_interval_set,
 114         .get = param_get_int,
 115 };
 116 #define param_check_recovery_interval(name, p) \
 117                 __param_check(name, p, int)
 118 module_param(lnet_recovery_interval, recovery_interval, S_IRUGO|S_IWUSR);
 119 #else
 120 module_param_call(lnet_recovery_interval, recovery_interval_set, param_get_int,
 121                   &lnet_recovery_interval, S_IRUGO|S_IWUSR);
 122 #endif
 123 MODULE_PARM_DESC(lnet_recovery_interval,
 124                 "DEPRECATED - Interval to recover unhealthy interfaces in seconds");
 125
 126 unsigned int lnet_recovery_limit;
 127 module_param(lnet_recovery_limit, uint, 0644);
 128 MODULE_PARM_DESC(lnet_recovery_limit,
 129                  "How long to attempt recovery of unhealthy peer interfaces in seconds. Set to 0 to allow indefinite recovery");
 130
 131 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
 132 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
 133
 134 static struct kernel_param_ops param_ops_interfaces_max = {
 135         .set = intf_max_set,
 136         .get = param_get_int,
 137 };
 138
 139 #define param_check_interfaces_max(name, p) \
 140                 __param_check(name, p, int)
 141
 142 #ifdef HAVE_KERNEL_PARAM_OPS
 143 module_param(lnet_interfaces_max, interfaces_max, 0644);
 144 #else
 145 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
 146                   &param_ops_interfaces_max, 0644);
 147 #endif
 148 MODULE_PARM_DESC(lnet_interfaces_max,
 149                 "Maximum number of interfaces in a node.");
 150
 151 unsigned lnet_peer_discovery_disabled = 0;
 152 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
 153
 154 static struct kernel_param_ops param_ops_discovery_disabled = {
 155         .set = discovery_set,
 156         .get = param_get_int,
 157 };
 158
 159 #define param_check_discovery_disabled(name, p) \
 160                 __param_check(name, p, int)
 161 #ifdef HAVE_KERNEL_PARAM_OPS
 162 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
 163 #else
 164 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
 165                   &param_ops_discovery_disabled, 0644);
 166 #endif
 167 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 168                 "Set to 1 to disable peer discovery on this node.");
 169
 170 unsigned int lnet_drop_asym_route;
 171 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
 172
 173 static struct kernel_param_ops param_ops_drop_asym_route = {
 174         .set = drop_asym_route_set,
 175         .get = param_get_int,
 176 };
 177
 178 #define param_check_drop_asym_route(name, p)    \
 179         __param_check(name, p, int)
 180 #ifdef HAVE_KERNEL_PARAM_OPS
 181 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
 182 #else
 183 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
 184                   &param_ops_drop_asym_route, 0644);
 185 #endif
 186 MODULE_PARM_DESC(lnet_drop_asym_route,
 187                  "Set to 1 to drop asymmetrical route messages.");
 188
 189 #define LNET_TRANSACTION_TIMEOUT_DEFAULT 50
 190 unsigned int lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_DEFAULT;
 191 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
 192 #ifdef HAVE_KERNEL_PARAM_OPS
 193 static struct kernel_param_ops param_ops_transaction_timeout = {
 194         .set = transaction_to_set,
 195         .get = param_get_int,
 196 };
 197
 198 #define param_check_transaction_timeout(name, p) \
 199                 __param_check(name, p, int)
 200 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
 201 #else
 202 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
 203                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
 204 #endif
 205 MODULE_PARM_DESC(lnet_transaction_timeout,
 206                 "Maximum number of seconds to wait for a peer response.");
 207
 208 #define LNET_RETRY_COUNT_DEFAULT 2
 209 unsigned int lnet_retry_count = LNET_RETRY_COUNT_DEFAULT;
 210 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
 211 #ifdef HAVE_KERNEL_PARAM_OPS
 212 static struct kernel_param_ops param_ops_retry_count = {
 213         .set = retry_count_set,
 214         .get = param_get_int,
 215 };
 216
 217 #define param_check_retry_count(name, p) \
 218                 __param_check(name, p, int)
 219 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
 220 #else
 221 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
 222                   &lnet_retry_count, S_IRUGO|S_IWUSR);
 223 #endif
 224 MODULE_PARM_DESC(lnet_retry_count,
 225                  "Maximum number of times to retry transmitting a message");
 226
 227 unsigned int lnet_response_tracking = 3;
 228 static int response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp);
 229
 230 #ifdef HAVE_KERNEL_PARAM_OPS
 231 static struct kernel_param_ops param_ops_response_tracking = {
 232         .set = response_tracking_set,
 233         .get = param_get_int,
 234 };
 235
 236 #define param_check_response_tracking(name, p)  \
 237         __param_check(name, p, int)
 238 module_param(lnet_response_tracking, response_tracking, 0644);
 239 #else
 240 module_param_call(lnet_response_tracking, response_tracking_set, param_get_int,
 241                   &lnet_response_tracking, 0644);
 242 #endif
 243 MODULE_PARM_DESC(lnet_response_tracking,
 244                  "(0|1|2|3) LNet Internal Only|GET Reply only|PUT ACK only|Full Tracking (default)");
 245
 246 #define LNET_LND_TIMEOUT_DEFAULT ((LNET_TRANSACTION_TIMEOUT_DEFAULT - 1) / \
 247                                   (LNET_RETRY_COUNT_DEFAULT + 1))
 248 unsigned int lnet_lnd_timeout = LNET_LND_TIMEOUT_DEFAULT;
 249 static void lnet_set_lnd_timeout(void)
 250 {
 251         lnet_lnd_timeout = (lnet_transaction_timeout - 1) /
 252                            (lnet_retry_count + 1);
 253 }
 254
 255 /*
 256  * This sequence number keeps track of how many times DLC was used to
 257  * update the local NIs. It is incremented when a NI is added or
 258  * removed and checked when sending a message to determine if there is
 259  * a need to re-run the selection algorithm. See lnet_select_pathway()
 260  * for more details on its usage.
 261  */
 262 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
 263
 264 static int lnet_ping(struct lnet_process_id id, signed long timeout,
 265                      struct lnet_process_id __user *ids, int n_ids);
 266
 267 static int lnet_discover(struct lnet_process_id id, __u32 force,
 268                          struct lnet_process_id __user *ids, int n_ids);
 269
 270 static int
 271 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 272 {
 273         int rc;
 274         unsigned *sensitivity = (unsigned *)kp->arg;
 275         unsigned long value;
 276
 277         rc = kstrtoul(val, 0, &value);
 278         if (rc) {
 279                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
 280                 return rc;
 281         }
 282
 283         /*
 284          * The purpose of locking the api_mutex here is to ensure that
 285          * the correct value ends up stored properly.
 286          */
 287         mutex_lock(&the_lnet.ln_api_mutex);
 288
 289         if (value > LNET_MAX_HEALTH_VALUE) {
 290                 mutex_unlock(&the_lnet.ln_api_mutex);
 291                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
 292                        LNET_MAX_HEALTH_VALUE, value);
 293                 return -EINVAL;
 294         }
 295
 296         if (*sensitivity != 0 && value == 0 && lnet_retry_count != 0) {
 297                 lnet_retry_count = 0;
 298                 lnet_set_lnd_timeout();
 299         }
 300
 301         *sensitivity = value;
 302
 303         mutex_unlock(&the_lnet.ln_api_mutex);
 304
 305         return 0;
 306 }
 307
 308 static int
 309 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 310 {
 311         CWARN("'lnet_recovery_interval' has been deprecated\n");
 312
 313         return 0;
 314 }
 315
 316 static int
 317 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
 318 {
 319         int rc;
 320         unsigned *discovery_off = (unsigned *)kp->arg;
 321         unsigned long value;
 322         struct lnet_ping_buffer *pbuf;
 323
 324         rc = kstrtoul(val, 0, &value);
 325         if (rc) {
 326                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
 327                 return rc;
 328         }
 329
 330         value = (value) ? 1 : 0;
 331
 332         /*
 333          * The purpose of locking the api_mutex here is to ensure that
 334          * the correct value ends up stored properly.
 335          */
 336         mutex_lock(&the_lnet.ln_api_mutex);
 337
 338         if (value == *discovery_off) {
 339                 mutex_unlock(&the_lnet.ln_api_mutex);
 340                 return 0;
 341         }
 342
 343         /*
 344          * We still want to set the discovery value even when LNet is not
 345          * running. This is the case when LNet is being loaded and we want
 346          * the module parameters to take effect. Otherwise if we're
 347          * changing the value dynamically, we want to set it after
 348          * updating the peers
 349          */
 350         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 351                 *discovery_off = value;
 352                 mutex_unlock(&the_lnet.ln_api_mutex);
 353                 return 0;
 354         }
 355
 356         /* tell peers that discovery setting has changed */
 357         lnet_net_lock(LNET_LOCK_EX);
 358         pbuf = the_lnet.ln_ping_target;
 359         if (value)
 360                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
 361         else
 362                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
 363         lnet_net_unlock(LNET_LOCK_EX);
 364
 365         /* only send a push when we're turning off discovery */
 366         if (*discovery_off <= 0 && value > 0)
 367                 lnet_push_update_to_peers(1);
 368         *discovery_off = value;
 369
 370         mutex_unlock(&the_lnet.ln_api_mutex);
 371
 372         return 0;
 373 }
 374
 375 static int
 376 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
 377 {
 378         int rc;
 379         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
 380         unsigned long value;
 381
 382         rc = kstrtoul(val, 0, &value);
 383         if (rc) {
 384                 CERROR("Invalid module parameter value for "
 385                        "'lnet_drop_asym_route'\n");
 386                 return rc;
 387         }
 388
 389         /*
 390          * The purpose of locking the api_mutex here is to ensure that
 391          * the correct value ends up stored properly.
 392          */
 393         mutex_lock(&the_lnet.ln_api_mutex);
 394
 395         if (value == *drop_asym_route) {
 396                 mutex_unlock(&the_lnet.ln_api_mutex);
 397                 return 0;
 398         }
 399
 400         *drop_asym_route = value;
 401
 402         mutex_unlock(&the_lnet.ln_api_mutex);
 403
 404         return 0;
 405 }
 406
 407 static int
 408 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
 409 {
 410         int rc;
 411         unsigned *transaction_to = (unsigned *)kp->arg;
 412         unsigned long value;
 413
 414         rc = kstrtoul(val, 0, &value);
 415         if (rc) {
 416                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
 417                 return rc;
 418         }
 419
 420         /*
 421          * The purpose of locking the api_mutex here is to ensure that
 422          * the correct value ends up stored properly.
 423          */
 424         mutex_lock(&the_lnet.ln_api_mutex);
 425
 426         if (value <= lnet_retry_count || value == 0) {
 427                 mutex_unlock(&the_lnet.ln_api_mutex);
 428                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
 429                        "Has to be greater than lnet_retry_count (%u)\n",
 430                        value, lnet_retry_count);
 431                 return -EINVAL;
 432         }
 433
 434         if (value == *transaction_to) {
 435                 mutex_unlock(&the_lnet.ln_api_mutex);
 436                 return 0;
 437         }
 438
 439         *transaction_to = value;
 440         /* Update the lnet_lnd_timeout now that we've modified the
 441          * transaction timeout
 442          */
 443         lnet_set_lnd_timeout();
 444
 445         mutex_unlock(&the_lnet.ln_api_mutex);
 446
 447         return 0;
 448 }
 449
 450 static int
 451 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
 452 {
 453         int rc;
 454         unsigned *retry_count = (unsigned *)kp->arg;
 455         unsigned long value;
 456
 457         rc = kstrtoul(val, 0, &value);
 458         if (rc) {
 459                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
 460                 return rc;
 461         }
 462
 463         /*
 464          * The purpose of locking the api_mutex here is to ensure that
 465          * the correct value ends up stored properly.
 466          */
 467         mutex_lock(&the_lnet.ln_api_mutex);
 468
 469         if (lnet_health_sensitivity == 0 && value > 0) {
 470                 mutex_unlock(&the_lnet.ln_api_mutex);
 471                 CERROR("Can not set lnet_retry_count when health feature is turned off\n");
 472                 return -EINVAL;
 473         }
 474
 475         if (value > lnet_transaction_timeout) {
 476                 mutex_unlock(&the_lnet.ln_api_mutex);
 477                 CERROR("Invalid value for lnet_retry_count (%lu). "
 478                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
 479                        value, lnet_transaction_timeout);
 480                 return -EINVAL;
 481         }
 482
 483         *retry_count = value;
 484
 485         /* Update the lnet_lnd_timeout now that we've modified the
 486          * retry count
 487          */
 488         lnet_set_lnd_timeout();
 489
 490         mutex_unlock(&the_lnet.ln_api_mutex);
 491
 492         return 0;
 493 }
 494
 495 static int
 496 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
 497 {
 498         int value, rc;
 499
 500         rc = kstrtoint(val, 0, &value);
 501         if (rc) {
 502                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
 503                 return rc;
 504         }
 505
 506         if (value < LNET_INTERFACES_MIN) {
 507                 CWARN("max interfaces provided are too small, setting to %d\n",
 508                       LNET_INTERFACES_MAX_DEFAULT);
 509                 value = LNET_INTERFACES_MAX_DEFAULT;
 510         }
 511
 512         *(int *)kp->arg = value;
 513
 514         return 0;
 515 }
 516
 517 static int
 518 response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp)
 519 {
 520         int rc;
 521         unsigned long new_value;
 522
 523         rc = kstrtoul(val, 0, &new_value);
 524         if (rc) {
 525                 CERROR("Invalid value for 'lnet_response_tracking'\n");
 526                 return -EINVAL;
 527         }
 528
 529         if (new_value < 0 || new_value > 3) {
 530                 CWARN("Invalid value (%lu) for 'lnet_response_tracking'\n",
 531                       new_value);
 532                 return -EINVAL;
 533         }
 534
 535         lnet_response_tracking = new_value;
 536
 537         return 0;
 538 }
 539
 540 static const char *
 541 lnet_get_routes(void)
 542 {
 543         return routes;
 544 }
 545
 546 static const char *
 547 lnet_get_networks(void)
 548 {
 549         const char *nets;
 550         int rc;
 551
 552         if (*networks != 0 && *ip2nets != 0) {
 553                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 554                                    "'ip2nets' but not both at once\n");
 555                 return NULL;
 556         }
 557
 558         if (*ip2nets != 0) {
 559                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 560                 return (rc == 0) ? nets : NULL;
 561         }
 562
 563         if (*networks != 0)
 564                 return networks;
 565
 566         return "tcp";
 567 }
 568
 569 static void
 570 lnet_init_locks(void)
 571 {
 572         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 573         spin_lock_init(&the_lnet.ln_msg_resend_lock);
 574         init_completion(&the_lnet.ln_mt_wait_complete);
 575         mutex_init(&the_lnet.ln_lnd_mutex);
 576 }
 577
 578 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 579 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 580                                             *  MDs kmem_cache */
 581 struct kmem_cache *lnet_udsp_cachep;       /* udsp cache */
 582 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
 583 struct kmem_cache *lnet_msg_cachep;
 584
 585 static int
 586 lnet_slab_setup(void)
 587 {
 588         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 589          * allocated in <size-xxx> kmem_cache).
 590          */
 591         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 592                                             0, 0, NULL);
 593         if (!lnet_mes_cachep)
 594                 return -ENOMEM;
 595
 596         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 597                                                   LNET_SMALL_MD_SIZE, 0, 0,
 598                                                   NULL);
 599         if (!lnet_small_mds_cachep)
 600                 return -ENOMEM;
 601
 602         lnet_udsp_cachep = kmem_cache_create("lnet_udsp",
 603                                              sizeof(struct lnet_udsp),
 604                                              0, 0, NULL);
 605         if (!lnet_udsp_cachep)
 606                 return -ENOMEM;
 607
 608         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
 609                                             0, 0, NULL);
 610         if (!lnet_rspt_cachep)
 611                 return -ENOMEM;
 612
 613         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
 614                                             0, 0, NULL);
 615         if (!lnet_msg_cachep)
 616                 return -ENOMEM;
 617
 618         return 0;
 619 }
 620
 621 static void
 622 lnet_slab_cleanup(void)
 623 {
 624         if (lnet_msg_cachep) {
 625                 kmem_cache_destroy(lnet_msg_cachep);
 626                 lnet_msg_cachep = NULL;
 627         }
 628
 629         if (lnet_rspt_cachep) {
 630                 kmem_cache_destroy(lnet_rspt_cachep);
 631                 lnet_rspt_cachep = NULL;
 632         }
 633
 634         if (lnet_udsp_cachep) {
 635                 kmem_cache_destroy(lnet_udsp_cachep);
 636                 lnet_udsp_cachep = NULL;
 637         }
 638
 639         if (lnet_small_mds_cachep) {
 640                 kmem_cache_destroy(lnet_small_mds_cachep);
 641                 lnet_small_mds_cachep = NULL;
 642         }
 643
 644         if (lnet_mes_cachep) {
 645                 kmem_cache_destroy(lnet_mes_cachep);
 646                 lnet_mes_cachep = NULL;
 647         }
 648 }
 649
 650 static int
 651 lnet_create_remote_nets_table(void)
 652 {
 653         int               i;
 654         struct list_head *hash;
 655
 656         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 657         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 658         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
 659         if (hash == NULL) {
 660                 CERROR("Failed to create remote nets hash table\n");
 661                 return -ENOMEM;
 662         }
 663
 664         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 665                 INIT_LIST_HEAD(&hash[i]);
 666         the_lnet.ln_remote_nets_hash = hash;
 667         return 0;
 668 }
 669
 670 static void
 671 lnet_destroy_remote_nets_table(void)
 672 {
 673         int i;
 674
 675         if (the_lnet.ln_remote_nets_hash == NULL)
 676                 return;
 677
 678         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 679                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 680
 681         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
 682                            LNET_REMOTE_NETS_HASH_SIZE);
 683         the_lnet.ln_remote_nets_hash = NULL;
 684 }
 685
 686 static void
 687 lnet_destroy_locks(void)
 688 {
 689         if (the_lnet.ln_res_lock != NULL) {
 690                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 691                 the_lnet.ln_res_lock = NULL;
 692         }
 693
 694         if (the_lnet.ln_net_lock != NULL) {
 695                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 696                 the_lnet.ln_net_lock = NULL;
 697         }
 698 }
 699
 700 static int
 701 lnet_create_locks(void)
 702 {
 703         lnet_init_locks();
 704
 705         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 706         if (the_lnet.ln_res_lock == NULL)
 707                 goto failed;
 708
 709         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 710         if (the_lnet.ln_net_lock == NULL)
 711                 goto failed;
 712
 713         return 0;
 714
 715  failed:
 716         lnet_destroy_locks();
 717         return -ENOMEM;
 718 }
 719
 720 static void lnet_assert_wire_constants(void)
 721 {
 722         /* Wire protocol assertions generated by 'wirecheck'
 723          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 724          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 725          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7)
 726          */
 727
 728         /* Constants... */
 729         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
 730         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
 731         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
 732         BUILD_BUG_ON(LNET_MSG_ACK != 0);
 733         BUILD_BUG_ON(LNET_MSG_PUT != 1);
 734         BUILD_BUG_ON(LNET_MSG_GET != 2);
 735         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
 736         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
 737
 738         BUILD_BUG_ON((int)sizeof(lnet_nid_t) != 8);
 739         BUILD_BUG_ON((int)sizeof(lnet_pid_t) != 4);
 740
 741         /* Checks for struct lnet_nid */
 742         BUILD_BUG_ON((int)sizeof(struct lnet_nid) != 20);
 743         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_size) != 0);
 744         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_size) != 1);
 745         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_type) != 1);
 746         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_type) != 1);
 747         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_num) != 2);
 748         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_num) != 2);
 749         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_addr) != 4);
 750         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_addr) != 16);
 751
 752         /* Checks for struct lnet_process_id_packed */
 753         BUILD_BUG_ON((int)sizeof(struct lnet_process_id_packed) != 12);
 754         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, nid) != 0);
 755         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->nid) != 8);
 756         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, pid) != 8);
 757         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->pid) != 4);
 758
 759         /* Checks for struct lnet_handle_wire */
 760         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
 761         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 762                                    wh_interface_cookie) != 0);
 763         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
 764         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 765                                    wh_object_cookie) != 8);
 766         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
 767
 768         /* Checks for struct struct lnet_magicversion */
 769         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
 770         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
 771         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
 772         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
 773         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
 774         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
 775                                    version_minor) != 6);
 776         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
 777
 778         /* Checks for struct struct lnet_hdr */
 779         BUILD_BUG_ON((int)sizeof(struct lnet_hdr) != 72);
 780         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_nid) != 0);
 781         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_nid) != 8);
 782         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_nid) != 8);
 783         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_nid) != 8);
 784         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, dest_pid) != 16);
 785         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->dest_pid) != 4);
 786         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, src_pid) != 20);
 787         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->src_pid) != 4);
 788         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, type) != 24);
 789         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->type) != 4);
 790         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, payload_length) != 28);
 791         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->payload_length) != 4);
 792         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg) != 32);
 793         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg) != 40);
 794
 795         /* Ack */
 796         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) != 32);
 797         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) != 16);
 798         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.match_bits) != 48);
 799         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) != 8);
 800         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.ack.mlength) != 56);
 801         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) != 4);
 802
 803         /* Put */
 804         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) != 32);
 805         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) != 16);
 806         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.match_bits) != 48);
 807         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) != 8);
 808         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.hdr_data) != 56);
 809         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) != 8);
 810         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.ptl_index) != 64);
 811         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) != 4);
 812         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.put.offset) != 68);
 813         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) != 4);
 814
 815         /* Get */
 816         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.return_wmd) != 32);
 817         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) != 16);
 818         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.match_bits) != 48);
 819         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) != 8);
 820         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.ptl_index) != 56);
 821         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) != 4);
 822         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.src_offset) != 60);
 823         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) != 4);
 824         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.get.sink_length) != 64);
 825         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) != 4);
 826
 827         /* Reply */
 828         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) != 32);
 829         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) != 16);
 830
 831         /* Hello */
 832         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.incarnation) != 32);
 833         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) != 8);
 834         BUILD_BUG_ON((int)offsetof(struct lnet_hdr, msg.hello.type) != 40);
 835         BUILD_BUG_ON((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) != 4);
 836
 837         /* Checks for struct lnet_ni_status and related constants */
 838         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
 839         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
 840         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
 841
 842         /* Checks for struct lnet_ni_status */
 843         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
 844         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
 845         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
 846         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
 847         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
 848         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_unused) != 12);
 849         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) != 4);
 850
 851         /* Checks for struct lnet_ping_info and related constants */
 852         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
 853         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
 854         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
 855         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
 856         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
 857         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
 858         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
 859         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 31);
 860
 861         /* Checks for struct lnet_ping_info */
 862         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
 863         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
 864         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
 865         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
 866         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
 867         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
 868         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
 869         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
 870         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
 871         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
 872         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) != 0);
 873
 874         /* Acceptor connection request */
 875         BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 876
 877         /* Checks for struct lnet_acceptor_connreq */
 878         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq) != 16);
 879         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_magic) != 0);
 880         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_magic) != 4);
 881         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_version) != 4);
 882         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_version) != 4);
 883         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_nid) != 8);
 884         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_nid) != 8);
 885
 886         /* Checks for struct lnet_acceptor_connreq_v2 */
 887         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq_v2) != 28);
 888         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_magic) != 0);
 889         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_magic) != 4);
 890         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_version) != 4);
 891         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_version) != 4);
 892         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_nid) != 8);
 893         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_nid) != 20);
 894
 895         /* Checks for struct lnet_counters_common */
 896         BUILD_BUG_ON((int)sizeof(struct lnet_counters_common) != 60);
 897         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_alloc) != 0);
 898         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_alloc) != 4);
 899         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_max) != 4);
 900         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_max) != 4);
 901         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_errors) != 8);
 902         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_errors) != 4);
 903         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_count) != 12);
 904         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_count) != 4);
 905         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_count) != 16);
 906         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_count) != 4);
 907         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_count) != 20);
 908         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_count) != 4);
 909         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_count) != 24);
 910         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_count) != 4);
 911         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_length) != 28);
 912         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_length) != 8);
 913         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_length) != 36);
 914         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_length) != 8);
 915         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_length) != 44);
 916         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_length) != 8);
 917         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_length) != 52);
 918         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_length) != 8);
 919 }
 920
 921 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 922 {
 923         const struct lnet_lnd *lnd;
 924
 925         /* holding lnd mutex */
 926         if (type >= NUM_LNDS)
 927                 return NULL;
 928         lnd = the_lnet.ln_lnds[type];
 929         LASSERT(!lnd || lnd->lnd_type == type);
 930
 931         return lnd;
 932 }
 933
 934 unsigned int
 935 lnet_get_lnd_timeout(void)
 936 {
 937         return lnet_lnd_timeout;
 938 }
 939 EXPORT_SYMBOL(lnet_get_lnd_timeout);
 940
 941 void
 942 lnet_register_lnd(const struct lnet_lnd *lnd)
 943 {
 944         mutex_lock(&the_lnet.ln_lnd_mutex);
 945
 946         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 947         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 948
 949         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
 950
 951         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 952
 953         mutex_unlock(&the_lnet.ln_lnd_mutex);
 954 }
 955 EXPORT_SYMBOL(lnet_register_lnd);
 956
 957 void
 958 lnet_unregister_lnd(const struct lnet_lnd *lnd)
 959 {
 960         mutex_lock(&the_lnet.ln_lnd_mutex);
 961
 962         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 963
 964         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
 965         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 966
 967         mutex_unlock(&the_lnet.ln_lnd_mutex);
 968 }
 969 EXPORT_SYMBOL(lnet_unregister_lnd);
 970
 971 static void
 972 lnet_counters_get_common_locked(struct lnet_counters_common *common)
 973 {
 974         struct lnet_counters *ctr;
 975         int i;
 976
 977         /* FIXME !!! Their is no assert_lnet_net_locked() to ensure this
 978          * actually called under the protection of the lnet_net_lock.
 979          */
 980         memset(common, 0, sizeof(*common));
 981
 982         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 983                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
 984                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
 985                 common->lcc_errors       += ctr->lct_common.lcc_errors;
 986                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
 987                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
 988                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
 989                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
 990                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
 991                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
 992                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
 993                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
 994         }
 995 }
 996
 997 void
 998 lnet_counters_get_common(struct lnet_counters_common *common)
 999 {
1000         lnet_net_lock(LNET_LOCK_EX);
1001         lnet_counters_get_common_locked(common);
1002         lnet_net_unlock(LNET_LOCK_EX);
1003 }
1004 EXPORT_SYMBOL(lnet_counters_get_common);
1005
1006 int
1007 lnet_counters_get(struct lnet_counters *counters)
1008 {
1009         struct lnet_counters *ctr;
1010         struct lnet_counters_health *health = &counters->lct_health;
1011         int i, rc = 0;
1012
1013         memset(counters, 0, sizeof(*counters));
1014
1015         lnet_net_lock(LNET_LOCK_EX);
1016
1017         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1018                 GOTO(out_unlock, rc = -ENODEV);
1019
1020         lnet_counters_get_common_locked(&counters->lct_common);
1021
1022         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1023                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
1024                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
1025                 health->lch_response_timeout_count +=
1026                                 ctr->lct_health.lch_response_timeout_count;
1027                 health->lch_local_interrupt_count +=
1028                                 ctr->lct_health.lch_local_interrupt_count;
1029                 health->lch_local_dropped_count +=
1030                                 ctr->lct_health.lch_local_dropped_count;
1031                 health->lch_local_aborted_count +=
1032                                 ctr->lct_health.lch_local_aborted_count;
1033                 health->lch_local_no_route_count +=
1034                                 ctr->lct_health.lch_local_no_route_count;
1035                 health->lch_local_timeout_count +=
1036                                 ctr->lct_health.lch_local_timeout_count;
1037                 health->lch_local_error_count +=
1038                                 ctr->lct_health.lch_local_error_count;
1039                 health->lch_remote_dropped_count +=
1040                                 ctr->lct_health.lch_remote_dropped_count;
1041                 health->lch_remote_error_count +=
1042                                 ctr->lct_health.lch_remote_error_count;
1043                 health->lch_remote_timeout_count +=
1044                                 ctr->lct_health.lch_remote_timeout_count;
1045                 health->lch_network_timeout_count +=
1046                                 ctr->lct_health.lch_network_timeout_count;
1047         }
1048 out_unlock:
1049         lnet_net_unlock(LNET_LOCK_EX);
1050         return rc;
1051 }
1052 EXPORT_SYMBOL(lnet_counters_get);
1053
1054 void
1055 lnet_counters_reset(void)
1056 {
1057         struct lnet_counters *counters;
1058         int             i;
1059
1060         lnet_net_lock(LNET_LOCK_EX);
1061
1062         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1063                 goto avoid_reset;
1064
1065         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
1066                 memset(counters, 0, sizeof(struct lnet_counters));
1067 avoid_reset:
1068         lnet_net_unlock(LNET_LOCK_EX);
1069 }
1070
1071 static char *
1072 lnet_res_type2str(int type)
1073 {
1074         switch (type) {
1075         default:
1076                 LBUG();
1077         case LNET_COOKIE_TYPE_MD:
1078                 return "MD";
1079         case LNET_COOKIE_TYPE_ME:
1080                 return "ME";
1081         case LNET_COOKIE_TYPE_EQ:
1082                 return "EQ";
1083         }
1084 }
1085
1086 static void
1087 lnet_res_container_cleanup(struct lnet_res_container *rec)
1088 {
1089         int     count = 0;
1090
1091         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
1092                 return;
1093
1094         while (!list_empty(&rec->rec_active)) {
1095                 struct list_head *e = rec->rec_active.next;
1096
1097                 list_del_init(e);
1098                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
1099                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
1100
1101                 } else { /* NB: Active MEs should be attached on portals */
1102                         LBUG();
1103                 }
1104                 count++;
1105         }
1106
1107         if (count > 0) {
1108                 /* Found alive MD/ME/EQ, user really should unlink/free
1109                  * all of them before finalize LNet, but if someone didn't,
1110                  * we have to recycle garbage for him */
1111                 CERROR("%d active elements on exit of %s container\n",
1112                        count, lnet_res_type2str(rec->rec_type));
1113         }
1114
1115         if (rec->rec_lh_hash != NULL) {
1116                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
1117                 rec->rec_lh_hash = NULL;
1118         }
1119
1120         rec->rec_type = 0; /* mark it as finalized */
1121 }
1122
1123 static int
1124 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1125 {
1126         int     rc = 0;
1127         int     i;
1128
1129         LASSERT(rec->rec_type == 0);
1130
1131         rec->rec_type = type;
1132         INIT_LIST_HEAD(&rec->rec_active);
1133
1134         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1135
1136         /* Arbitrary choice of hash table size */
1137         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1138                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1139         if (rec->rec_lh_hash == NULL) {
1140                 rc = -ENOMEM;
1141                 goto out;
1142         }
1143
1144         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1145                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1146
1147         return 0;
1148
1149 out:
1150         CERROR("Failed to setup %s resource container\n",
1151                lnet_res_type2str(type));
1152         lnet_res_container_cleanup(rec);
1153         return rc;
1154 }
1155
1156 static void
1157 lnet_res_containers_destroy(struct lnet_res_container **recs)
1158 {
1159         struct lnet_res_container       *rec;
1160         int                             i;
1161
1162         cfs_percpt_for_each(rec, i, recs)
1163                 lnet_res_container_cleanup(rec);
1164
1165         cfs_percpt_free(recs);
1166 }
1167
1168 static struct lnet_res_container **
1169 lnet_res_containers_create(int type)
1170 {
1171         struct lnet_res_container       **recs;
1172         struct lnet_res_container       *rec;
1173         int                             rc;
1174         int                             i;
1175
1176         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1177         if (recs == NULL) {
1178                 CERROR("Failed to allocate %s resource containers\n",
1179                        lnet_res_type2str(type));
1180                 return NULL;
1181         }
1182
1183         cfs_percpt_for_each(rec, i, recs) {
1184                 rc = lnet_res_container_setup(rec, i, type);
1185                 if (rc != 0) {
1186                         lnet_res_containers_destroy(recs);
1187                         return NULL;
1188                 }
1189         }
1190
1191         return recs;
1192 }
1193
1194 struct lnet_libhandle *
1195 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1196 {
1197         /* ALWAYS called with lnet_res_lock held */
1198         struct list_head        *head;
1199         struct lnet_libhandle   *lh;
1200         unsigned int            hash;
1201
1202         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1203                 return NULL;
1204
1205         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1206         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1207
1208         list_for_each_entry(lh, head, lh_hash_chain) {
1209                 if (lh->lh_cookie == cookie)
1210                         return lh;
1211         }
1212
1213         return NULL;
1214 }
1215
1216 void
1217 lnet_res_lh_initialize(struct lnet_res_container *rec,
1218                        struct lnet_libhandle *lh)
1219 {
1220         /* ALWAYS called with lnet_res_lock held */
1221         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1222         unsigned int    hash;
1223
1224         lh->lh_cookie = rec->rec_lh_cookie;
1225         rec->rec_lh_cookie += 1 << ibits;
1226
1227         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1228
1229         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1230 }
1231
1232 struct list_head **
1233 lnet_create_array_of_queues(void)
1234 {
1235         struct list_head **qs;
1236         struct list_head *q;
1237         int i;
1238
1239         qs = cfs_percpt_alloc(lnet_cpt_table(),
1240                               sizeof(struct list_head));
1241         if (!qs) {
1242                 CERROR("Failed to allocate queues\n");
1243                 return NULL;
1244         }
1245
1246         cfs_percpt_for_each(q, i, qs)
1247                 INIT_LIST_HEAD(q);
1248
1249         return qs;
1250 }
1251
1252 static int lnet_unprepare(void);
1253
1254 static int
1255 lnet_prepare(lnet_pid_t requested_pid)
1256 {
1257         /* Prepare to bring up the network */
1258         struct lnet_res_container **recs;
1259         int                       rc = 0;
1260
1261         if (requested_pid == LNET_PID_ANY) {
1262                 /* Don't instantiate LNET just for me */
1263                 return -ENETDOWN;
1264         }
1265
1266         LASSERT(the_lnet.ln_refcount == 0);
1267
1268         the_lnet.ln_routing = 0;
1269
1270         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1271         the_lnet.ln_pid = requested_pid;
1272
1273         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1274         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1275         INIT_LIST_HEAD(&the_lnet.ln_nets);
1276         INIT_LIST_HEAD(&the_lnet.ln_routers);
1277         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1278         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1279         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1280         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1281         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1282         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1283         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1284         INIT_LIST_HEAD(&the_lnet.ln_udsp_list);
1285         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1286         the_lnet.ln_mt_handler = NULL;
1287         init_completion(&the_lnet.ln_started);
1288
1289         rc = lnet_slab_setup();
1290         if (rc != 0)
1291                 goto failed;
1292
1293         rc = lnet_create_remote_nets_table();
1294         if (rc != 0)
1295                 goto failed;
1296
1297         /*
1298          * NB the interface cookie in wire handles guards against delayed
1299          * replies and ACKs appearing valid after reboot.
1300          */
1301         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1302
1303         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1304                                                 sizeof(struct lnet_counters));
1305         if (the_lnet.ln_counters == NULL) {
1306                 CERROR("Failed to allocate counters for LNet\n");
1307                 rc = -ENOMEM;
1308                 goto failed;
1309         }
1310
1311         rc = lnet_peer_tables_create();
1312         if (rc != 0)
1313                 goto failed;
1314
1315         rc = lnet_msg_containers_create();
1316         if (rc != 0)
1317                 goto failed;
1318
1319         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1320                                       LNET_COOKIE_TYPE_EQ);
1321         if (rc != 0)
1322                 goto failed;
1323
1324         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1325         if (recs == NULL) {
1326                 rc = -ENOMEM;
1327                 goto failed;
1328         }
1329
1330         the_lnet.ln_md_containers = recs;
1331
1332         rc = lnet_portals_create();
1333         if (rc != 0) {
1334                 CERROR("Failed to create portals for LNet: %d\n", rc);
1335                 goto failed;
1336         }
1337
1338         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1339         if (!the_lnet.ln_mt_zombie_rstqs) {
1340                 rc = -ENOMEM;
1341                 goto failed;
1342         }
1343
1344         return 0;
1345
1346  failed:
1347         lnet_unprepare();
1348         return rc;
1349 }
1350
1351 static int
1352 lnet_unprepare (void)
1353 {
1354         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1355          * have shut down already, so it is safe to unlink and free all
1356          * descriptors, even those that appear committed to a network op (eg MD
1357          * with non-zero pending count) */
1358
1359         lnet_fail_nid(LNET_NID_ANY, 0);
1360
1361         LASSERT(the_lnet.ln_refcount == 0);
1362         LASSERT(list_empty(&the_lnet.ln_test_peers));
1363         LASSERT(list_empty(&the_lnet.ln_nets));
1364
1365         if (the_lnet.ln_mt_zombie_rstqs) {
1366                 lnet_clean_zombie_rstqs();
1367                 the_lnet.ln_mt_zombie_rstqs = NULL;
1368         }
1369
1370         lnet_assert_handler_unused(the_lnet.ln_mt_handler);
1371         the_lnet.ln_mt_handler = NULL;
1372
1373         lnet_portals_destroy();
1374
1375         if (the_lnet.ln_md_containers != NULL) {
1376                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1377                 the_lnet.ln_md_containers = NULL;
1378         }
1379
1380         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1381
1382         lnet_msg_containers_destroy();
1383         lnet_peer_uninit();
1384         lnet_rtrpools_free(0);
1385
1386         if (the_lnet.ln_counters != NULL) {
1387                 cfs_percpt_free(the_lnet.ln_counters);
1388                 the_lnet.ln_counters = NULL;
1389         }
1390         lnet_destroy_remote_nets_table();
1391         lnet_udsp_destroy(true);
1392         lnet_slab_cleanup();
1393
1394         return 0;
1395 }
1396
1397 struct lnet_ni  *
1398 lnet_net2ni_locked(__u32 net_id, int cpt)
1399 {
1400         struct lnet_ni   *ni;
1401         struct lnet_net  *net;
1402
1403         LASSERT(cpt != LNET_LOCK_EX);
1404
1405         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1406                 if (net->net_id == net_id) {
1407                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1408                                         ni_netlist);
1409                         return ni;
1410                 }
1411         }
1412
1413         return NULL;
1414 }
1415
1416 struct lnet_ni *
1417 lnet_net2ni_addref(__u32 net)
1418 {
1419         struct lnet_ni *ni;
1420
1421         lnet_net_lock(0);
1422         ni = lnet_net2ni_locked(net, 0);
1423         if (ni)
1424                 lnet_ni_addref_locked(ni, 0);
1425         lnet_net_unlock(0);
1426
1427         return ni;
1428 }
1429 EXPORT_SYMBOL(lnet_net2ni_addref);
1430
1431 struct lnet_net *
1432 lnet_get_net_locked(__u32 net_id)
1433 {
1434         struct lnet_net  *net;
1435
1436         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1437                 if (net->net_id == net_id)
1438                         return net;
1439         }
1440
1441         return NULL;
1442 }
1443
1444 void
1445 lnet_net_clr_pref_rtrs(struct lnet_net *net)
1446 {
1447         struct list_head zombies;
1448         struct lnet_nid_list *ne;
1449         struct lnet_nid_list *tmp;
1450
1451         INIT_LIST_HEAD(&zombies);
1452
1453         lnet_net_lock(LNET_LOCK_EX);
1454         list_splice_init(&net->net_rtr_pref_nids, &zombies);
1455         lnet_net_unlock(LNET_LOCK_EX);
1456
1457         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1458                 list_del_init(&ne->nl_list);
1459                 LIBCFS_FREE(ne, sizeof(*ne));
1460         }
1461 }
1462
1463 int
1464 lnet_net_add_pref_rtr(struct lnet_net *net,
1465                       struct lnet_nid *gw_nid)
1466 __must_hold(&the_lnet.ln_api_mutex)
1467 {
1468         struct lnet_nid_list *ne;
1469
1470         /* This function is called with api_mutex held. When the api_mutex
1471          * is held the list can not be modified, as it is only modified as
1472          * a result of applying a UDSP and that happens under api_mutex
1473          * lock.
1474          */
1475         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1476                 if (nid_same(&ne->nl_nid, gw_nid))
1477                         return -EEXIST;
1478         }
1479
1480         LIBCFS_ALLOC(ne, sizeof(*ne));
1481         if (!ne)
1482                 return -ENOMEM;
1483
1484         ne->nl_nid = *gw_nid;
1485
1486         /* Lock the cpt to protect against addition and checks in the
1487          * selection algorithm
1488          */
1489         lnet_net_lock(LNET_LOCK_EX);
1490         list_add(&ne->nl_list, &net->net_rtr_pref_nids);
1491         lnet_net_unlock(LNET_LOCK_EX);
1492
1493         return 0;
1494 }
1495
1496 bool
1497 lnet_net_is_pref_rtr_locked(struct lnet_net *net, struct lnet_nid *rtr_nid)
1498 {
1499         struct lnet_nid_list *ne;
1500
1501         CDEBUG(D_NET, "%s: rtr pref empty: %d\n",
1502                libcfs_net2str(net->net_id),
1503                list_empty(&net->net_rtr_pref_nids));
1504
1505         if (list_empty(&net->net_rtr_pref_nids))
1506                 return false;
1507
1508         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1509                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1510                        libcfs_nidstr(&ne->nl_nid),
1511                        libcfs_nidstr(rtr_nid));
1512                 if (nid_same(rtr_nid, &ne->nl_nid))
1513                         return true;
1514         }
1515
1516         return false;
1517 }
1518
1519 static unsigned int
1520 lnet_nid4_cpt_hash(lnet_nid_t nid, unsigned int number)
1521 {
1522         __u64           key = nid;
1523         unsigned int    val;
1524
1525         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1526
1527         if (number == 1)
1528                 return 0;
1529
1530         val = hash_long(key, LNET_CPT_BITS);
1531         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1532         if (val < number)
1533                 return val;
1534
1535         return (unsigned int)(key + val + (val >> 1)) % number;
1536 }
1537
1538 unsigned int
1539 lnet_nid_cpt_hash(struct lnet_nid *nid, unsigned int number)
1540 {
1541         unsigned int val;
1542         u32 h = 0;
1543         int i;
1544
1545         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1546
1547         if (number == 1)
1548                 return 0;
1549
1550         if (nid_is_nid4(nid))
1551                 return lnet_nid4_cpt_hash(lnet_nid_to_nid4(nid), number);
1552
1553         for (i = 0; i < 4; i++)
1554                 h = hash_32(nid->nid_addr[i]^h, 32);
1555         val = hash_32(LNET_NID_NET(nid) ^ h, LNET_CPT_BITS);
1556         if (val < number)
1557                 return val;
1558         return (unsigned int)(h + val + (val >> 1)) % number;
1559 }
1560
1561 int
1562 lnet_cpt_of_nid_locked(struct lnet_nid *nid, struct lnet_ni *ni)
1563 {
1564         struct lnet_net *net;
1565
1566         /* must called with hold of lnet_net_lock */
1567         if (LNET_CPT_NUMBER == 1)
1568                 return 0; /* the only one */
1569
1570         /*
1571          * If NI is provided then use the CPT identified in the NI cpt
1572          * list if one exists. If one doesn't exist, then that NI is
1573          * associated with all CPTs and it follows that the net it belongs
1574          * to is implicitly associated with all CPTs, so just hash the nid
1575          * and return that.
1576          */
1577         if (ni != NULL) {
1578                 if (ni->ni_cpts != NULL)
1579                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1580                                                              ni->ni_ncpts)];
1581                 else
1582                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1583         }
1584
1585         /* no NI provided so look at the net */
1586         net = lnet_get_net_locked(LNET_NID_NET(nid));
1587
1588         if (net != NULL && net->net_cpts != NULL) {
1589                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1590         }
1591
1592         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1593 }
1594
1595 int
1596 lnet_nid2cpt(struct lnet_nid *nid, struct lnet_ni *ni)
1597 {
1598         int     cpt;
1599         int     cpt2;
1600
1601         if (LNET_CPT_NUMBER == 1)
1602                 return 0; /* the only one */
1603
1604         cpt = lnet_net_lock_current();
1605
1606         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1607
1608         lnet_net_unlock(cpt);
1609
1610         return cpt2;
1611 }
1612 EXPORT_SYMBOL(lnet_nid2cpt);
1613
1614 int
1615 lnet_cpt_of_nid(lnet_nid_t nid4, struct lnet_ni *ni)
1616 {
1617         struct lnet_nid nid;
1618
1619         if (LNET_CPT_NUMBER == 1)
1620                 return 0; /* the only one */
1621
1622         lnet_nid4_to_nid(nid4, &nid);
1623         return lnet_nid2cpt(&nid, ni);
1624 }
1625 EXPORT_SYMBOL(lnet_cpt_of_nid);
1626
1627 int
1628 lnet_islocalnet_locked(__u32 net_id)
1629 {
1630         struct lnet_net *net;
1631         bool local;
1632
1633         net = lnet_get_net_locked(net_id);
1634
1635         local = net != NULL;
1636
1637         return local;
1638 }
1639
1640 int
1641 lnet_islocalnet(__u32 net_id)
1642 {
1643         int cpt;
1644         bool local;
1645
1646         cpt = lnet_net_lock_current();
1647
1648         local = lnet_islocalnet_locked(net_id);
1649
1650         lnet_net_unlock(cpt);
1651
1652         return local;
1653 }
1654
1655 struct lnet_ni  *
1656 lnet_nid_to_ni_locked(struct lnet_nid *nid, int cpt)
1657 {
1658         struct lnet_net  *net;
1659         struct lnet_ni *ni;
1660
1661         LASSERT(cpt != LNET_LOCK_EX);
1662
1663         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1664                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1665                         if (nid_same(&ni->ni_nid, nid))
1666                                 return ni;
1667                 }
1668         }
1669
1670         return NULL;
1671 }
1672
1673 struct lnet_ni  *
1674 lnet_nid2ni_locked(lnet_nid_t nid4, int cpt)
1675 {
1676         struct lnet_nid nid;
1677
1678         lnet_nid4_to_nid(nid4, &nid);
1679         return lnet_nid_to_ni_locked(&nid, cpt);
1680 }
1681
1682 struct lnet_ni *
1683 lnet_nid2ni_addref(lnet_nid_t nid4)
1684 {
1685         struct lnet_ni *ni;
1686         struct lnet_nid nid;
1687
1688         lnet_nid4_to_nid(nid4, &nid);
1689
1690         lnet_net_lock(0);
1691         ni = lnet_nid_to_ni_locked(&nid, 0);
1692         if (ni)
1693                 lnet_ni_addref_locked(ni, 0);
1694         lnet_net_unlock(0);
1695
1696         return ni;
1697 }
1698 EXPORT_SYMBOL(lnet_nid2ni_addref);
1699
1700 struct lnet_ni *
1701 lnet_nid_to_ni_addref(struct lnet_nid *nid)
1702 {
1703         struct lnet_ni *ni;
1704
1705         lnet_net_lock(0);
1706         ni = lnet_nid_to_ni_locked(nid, 0);
1707         if (ni)
1708                 lnet_ni_addref_locked(ni, 0);
1709         lnet_net_unlock(0);
1710
1711         return ni;
1712 }
1713 EXPORT_SYMBOL(lnet_nid_to_ni_addref);
1714
1715 int
1716 lnet_islocalnid4(lnet_nid_t nid)
1717 {
1718         struct lnet_ni  *ni;
1719         int             cpt;
1720
1721         cpt = lnet_net_lock_current();
1722         ni = lnet_nid2ni_locked(nid, cpt);
1723         lnet_net_unlock(cpt);
1724
1725         return ni != NULL;
1726 }
1727
1728 int
1729 lnet_islocalnid(struct lnet_nid *nid)
1730 {
1731         struct lnet_ni  *ni;
1732         int             cpt;
1733
1734         cpt = lnet_net_lock_current();
1735         ni = lnet_nid_to_ni_locked(nid, cpt);
1736         lnet_net_unlock(cpt);
1737
1738         return ni != NULL;
1739 }
1740
1741 int
1742 lnet_count_acceptor_nets(void)
1743 {
1744         /* Return the # of NIs that need the acceptor. */
1745         int              count = 0;
1746         struct lnet_net  *net;
1747         int              cpt;
1748
1749         cpt = lnet_net_lock_current();
1750         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1751                 /* all socklnd type networks should have the acceptor
1752                  * thread started */
1753                 if (net->net_lnd->lnd_accept != NULL)
1754                         count++;
1755         }
1756
1757         lnet_net_unlock(cpt);
1758
1759         return count;
1760 }
1761
1762 struct lnet_ping_buffer *
1763 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1764 {
1765         struct lnet_ping_buffer *pbuf;
1766
1767         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1768         if (pbuf) {
1769                 pbuf->pb_nnis = nnis;
1770                 pbuf->pb_needs_post = false;
1771                 atomic_set(&pbuf->pb_refcnt, 1);
1772         }
1773
1774         return pbuf;
1775 }
1776
1777 void
1778 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1779 {
1780         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1781         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1782 }
1783
1784 static struct lnet_ping_buffer *
1785 lnet_ping_target_create(int nnis)
1786 {
1787         struct lnet_ping_buffer *pbuf;
1788
1789         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1790         if (pbuf == NULL) {
1791                 CERROR("Can't allocate ping source [%d]\n", nnis);
1792                 return NULL;
1793         }
1794
1795         pbuf->pb_info.pi_nnis = nnis;
1796         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1797         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1798         pbuf->pb_info.pi_features =
1799                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1800
1801         return pbuf;
1802 }
1803
1804 static inline int
1805 lnet_get_net_ni_count_locked(struct lnet_net *net)
1806 {
1807         struct lnet_ni  *ni;
1808         int             count = 0;
1809
1810         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1811                 count++;
1812
1813         return count;
1814 }
1815
1816 static inline int
1817 lnet_get_net_ni_count_pre(struct lnet_net *net)
1818 {
1819         struct lnet_ni  *ni;
1820         int             count = 0;
1821
1822         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1823                 count++;
1824
1825         return count;
1826 }
1827
1828 static inline int
1829 lnet_get_ni_count(void)
1830 {
1831         struct lnet_ni  *ni;
1832         struct lnet_net *net;
1833         int             count = 0;
1834
1835         lnet_net_lock(0);
1836
1837         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1838                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1839                         count++;
1840         }
1841
1842         lnet_net_unlock(0);
1843
1844         return count;
1845 }
1846
1847 void
1848 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1849 {
1850         struct lnet_ni_status *stat;
1851         int nnis;
1852         int i;
1853
1854         __swab32s(&pbuf->pb_info.pi_magic);
1855         __swab32s(&pbuf->pb_info.pi_features);
1856         __swab32s(&pbuf->pb_info.pi_pid);
1857         __swab32s(&pbuf->pb_info.pi_nnis);
1858         nnis = pbuf->pb_info.pi_nnis;
1859         if (nnis > pbuf->pb_nnis)
1860                 nnis = pbuf->pb_nnis;
1861         for (i = 0; i < nnis; i++) {
1862                 stat = &pbuf->pb_info.pi_ni[i];
1863                 __swab64s(&stat->ns_nid);
1864                 __swab32s(&stat->ns_status);
1865         }
1866 }
1867
1868 int
1869 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1870 {
1871         if (!pinfo)
1872                 return -EINVAL;
1873         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1874                 return -EPROTO;
1875         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1876                 return -EPROTO;
1877         /* Loopback is guaranteed to be present */
1878         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1879                 return -ERANGE;
1880         if (LNET_PING_INFO_LONI(pinfo) != LNET_NID_LO_0)
1881                 return -EPROTO;
1882         return 0;
1883 }
1884
1885 static void
1886 lnet_ping_target_destroy(void)
1887 {
1888         struct lnet_net *net;
1889         struct lnet_ni  *ni;
1890
1891         lnet_net_lock(LNET_LOCK_EX);
1892
1893         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1894                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1895                         lnet_ni_lock(ni);
1896                         ni->ni_status = NULL;
1897                         lnet_ni_unlock(ni);
1898                 }
1899         }
1900
1901         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1902         the_lnet.ln_ping_target = NULL;
1903
1904         lnet_net_unlock(LNET_LOCK_EX);
1905 }
1906
1907 static void
1908 lnet_ping_target_event_handler(struct lnet_event *event)
1909 {
1910         struct lnet_ping_buffer *pbuf = event->md_user_ptr;
1911
1912         if (event->unlinked)
1913                 lnet_ping_buffer_decref(pbuf);
1914 }
1915
1916 static int
1917 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1918                        struct lnet_handle_md *ping_mdh,
1919                        int ni_count, bool set_eq)
1920 {
1921         struct lnet_processid id = {
1922                 .nid = LNET_ANY_NID,
1923                 .pid = LNET_PID_ANY
1924         };
1925         struct lnet_me *me;
1926         struct lnet_md md = { NULL };
1927         int rc;
1928
1929         if (set_eq)
1930                 the_lnet.ln_ping_target_handler =
1931                         lnet_ping_target_event_handler;
1932
1933         *ppbuf = lnet_ping_target_create(ni_count);
1934         if (*ppbuf == NULL) {
1935                 rc = -ENOMEM;
1936                 goto fail_free_eq;
1937         }
1938
1939         /* Ping target ME/MD */
1940         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
1941                           LNET_PROTO_PING_MATCHBITS, 0,
1942                           LNET_UNLINK, LNET_INS_AFTER);
1943         if (IS_ERR(me)) {
1944                 rc = PTR_ERR(me);
1945                 CERROR("Can't create ping target ME: %d\n", rc);
1946                 goto fail_decref_ping_buffer;
1947         }
1948
1949         /* initialize md content */
1950         md.start     = &(*ppbuf)->pb_info;
1951         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1952         md.threshold = LNET_MD_THRESH_INF;
1953         md.max_size  = 0;
1954         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1955                        LNET_MD_MANAGE_REMOTE;
1956         md.handler   = the_lnet.ln_ping_target_handler;
1957         md.user_ptr  = *ppbuf;
1958
1959         rc = LNetMDAttach(me, &md, LNET_RETAIN, ping_mdh);
1960         if (rc != 0) {
1961                 CERROR("Can't attach ping target MD: %d\n", rc);
1962                 goto fail_decref_ping_buffer;
1963         }
1964         lnet_ping_buffer_addref(*ppbuf);
1965
1966         return 0;
1967
1968 fail_decref_ping_buffer:
1969         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
1970         lnet_ping_buffer_decref(*ppbuf);
1971         *ppbuf = NULL;
1972 fail_free_eq:
1973         return rc;
1974 }
1975
1976 static void
1977 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1978                     struct lnet_handle_md *ping_mdh)
1979 {
1980         LNetMDUnlink(*ping_mdh);
1981         LNetInvalidateMDHandle(ping_mdh);
1982
1983         /* NB the MD could be busy; this just starts the unlink */
1984         wait_var_event_warning(&pbuf->pb_refcnt,
1985                                atomic_read(&pbuf->pb_refcnt) <= 1,
1986                                "Still waiting for ping data MD to unlink\n");
1987 }
1988
1989 static void
1990 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1991 {
1992         struct lnet_ni          *ni;
1993         struct lnet_net         *net;
1994         struct lnet_ni_status *ns;
1995         int                     i;
1996         int                     rc;
1997
1998         i = 0;
1999         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2000                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2001                         LASSERT(i < pbuf->pb_nnis);
2002
2003                         ns = &pbuf->pb_info.pi_ni[i];
2004
2005                         if (!nid_is_nid4(&ni->ni_nid))
2006                                 continue;
2007                         ns->ns_nid = lnet_nid_to_nid4(&ni->ni_nid);
2008
2009                         lnet_ni_lock(ni);
2010                         ns->ns_status = lnet_ni_get_status_locked(ni);
2011                         ni->ni_status = ns;
2012                         lnet_ni_unlock(ni);
2013
2014                         i++;
2015                 }
2016         }
2017         /*
2018          * We (ab)use the ns_status of the loopback interface to
2019          * transmit the sequence number. The first interface listed
2020          * must be the loopback interface.
2021          */
2022         rc = lnet_ping_info_validate(&pbuf->pb_info);
2023         if (rc) {
2024                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
2025                 LBUG();
2026         }
2027         LNET_PING_BUFFER_SEQNO(pbuf) =
2028                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
2029 }
2030
2031 static void
2032 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
2033                         struct lnet_handle_md ping_mdh)
2034 {
2035         struct lnet_ping_buffer *old_pbuf = NULL;
2036         struct lnet_handle_md old_ping_md;
2037
2038         /* switch the NIs to point to the new ping info created */
2039         lnet_net_lock(LNET_LOCK_EX);
2040
2041         if (!the_lnet.ln_routing)
2042                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
2043         if (!lnet_peer_discovery_disabled)
2044                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
2045
2046         /* Ensure only known feature bits have been set. */
2047         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
2048         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
2049
2050         lnet_ping_target_install_locked(pbuf);
2051
2052         if (the_lnet.ln_ping_target) {
2053                 old_pbuf = the_lnet.ln_ping_target;
2054                 old_ping_md = the_lnet.ln_ping_target_md;
2055         }
2056         the_lnet.ln_ping_target_md = ping_mdh;
2057         the_lnet.ln_ping_target = pbuf;
2058
2059         lnet_net_unlock(LNET_LOCK_EX);
2060
2061         if (old_pbuf) {
2062                 /* unlink and free the old ping info */
2063                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
2064                 lnet_ping_buffer_decref(old_pbuf);
2065         }
2066
2067         lnet_push_update_to_peers(0);
2068 }
2069
2070 static void
2071 lnet_ping_target_fini(void)
2072 {
2073         lnet_ping_md_unlink(the_lnet.ln_ping_target,
2074                             &the_lnet.ln_ping_target_md);
2075
2076         lnet_assert_handler_unused(the_lnet.ln_ping_target_handler);
2077         lnet_ping_target_destroy();
2078 }
2079
2080 /* Resize the push target. */
2081 int lnet_push_target_resize(void)
2082 {
2083         struct lnet_handle_md mdh;
2084         struct lnet_handle_md old_mdh;
2085         struct lnet_ping_buffer *pbuf;
2086         struct lnet_ping_buffer *old_pbuf;
2087         int nnis;
2088         int rc;
2089
2090 again:
2091         nnis = the_lnet.ln_push_target_nnis;
2092         if (nnis <= 0) {
2093                 CDEBUG(D_NET, "Invalid nnis %d\n", nnis);
2094                 return -EINVAL;
2095         }
2096
2097         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
2098          * dropped when we need to resize again (see "old_pbuf" below) or when
2099          * LNet is shutdown (see lnet_push_target_fini())
2100          */
2101         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
2102         if (!pbuf) {
2103                 CDEBUG(D_NET, "Can't allocate pbuf for nnis %d\n", nnis);
2104                 return -ENOMEM;
2105         }
2106
2107         rc = lnet_push_target_post(pbuf, &mdh);
2108         if (rc) {
2109                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
2110                 lnet_ping_buffer_decref(pbuf);
2111                 return rc;
2112         }
2113
2114         lnet_net_lock(LNET_LOCK_EX);
2115         old_pbuf = the_lnet.ln_push_target;
2116         old_mdh = the_lnet.ln_push_target_md;
2117         the_lnet.ln_push_target = pbuf;
2118         the_lnet.ln_push_target_md = mdh;
2119         lnet_net_unlock(LNET_LOCK_EX);
2120
2121         if (old_pbuf) {
2122                 LNetMDUnlink(old_mdh);
2123                 /* Drop ref set by lnet_ping_buffer_alloc() */
2124                 lnet_ping_buffer_decref(old_pbuf);
2125         }
2126
2127         /* Received another push or reply that requires a larger buffer */
2128         if (nnis < the_lnet.ln_push_target_nnis)
2129                 goto again;
2130
2131         CDEBUG(D_NET, "nnis %d success\n", nnis);
2132         return 0;
2133 }
2134
2135 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
2136                           struct lnet_handle_md *mdhp)
2137 {
2138         struct lnet_processid id = { LNET_ANY_NID, LNET_PID_ANY };
2139         struct lnet_md md = { NULL };
2140         struct lnet_me *me;
2141         int rc;
2142
2143         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
2144                           LNET_PROTO_PING_MATCHBITS, 0,
2145                           LNET_UNLINK, LNET_INS_AFTER);
2146         if (IS_ERR(me)) {
2147                 rc = PTR_ERR(me);
2148                 CERROR("Can't create push target ME: %d\n", rc);
2149                 return rc;
2150         }
2151
2152         pbuf->pb_needs_post = false;
2153
2154         /* This reference is dropped by lnet_push_target_event_handler() */
2155         lnet_ping_buffer_addref(pbuf);
2156
2157         /* initialize md content */
2158         md.start     = &pbuf->pb_info;
2159         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
2160         md.threshold = 1;
2161         md.max_size  = 0;
2162         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
2163         md.user_ptr  = pbuf;
2164         md.handler   = the_lnet.ln_push_target_handler;
2165
2166         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdhp);
2167         if (rc) {
2168                 CERROR("Can't attach push MD: %d\n", rc);
2169                 lnet_ping_buffer_decref(pbuf);
2170                 pbuf->pb_needs_post = true;
2171                 return rc;
2172         }
2173
2174         CDEBUG(D_NET, "posted push target %p\n", pbuf);
2175
2176         return 0;
2177 }
2178
2179 static void lnet_push_target_event_handler(struct lnet_event *ev)
2180 {
2181         struct lnet_ping_buffer *pbuf = ev->md_user_ptr;
2182
2183         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
2184                ev->unlinked);
2185
2186         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2187                 lnet_swap_pinginfo(pbuf);
2188
2189         if (ev->type == LNET_EVENT_UNLINK) {
2190                 /* Drop ref added by lnet_push_target_post() */
2191                 lnet_ping_buffer_decref(pbuf);
2192                 return;
2193         }
2194
2195         lnet_peer_push_event(ev);
2196         if (ev->unlinked)
2197                 /* Drop ref added by lnet_push_target_post */
2198                 lnet_ping_buffer_decref(pbuf);
2199 }
2200
2201 /* Initialize the push target. */
2202 static int lnet_push_target_init(void)
2203 {
2204         int rc;
2205
2206         if (the_lnet.ln_push_target)
2207                 return -EALREADY;
2208
2209         the_lnet.ln_push_target_handler =
2210                 lnet_push_target_event_handler;
2211
2212         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
2213         LASSERT(rc == 0);
2214
2215         /* Start at the required minimum, we'll enlarge if required. */
2216         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
2217
2218         rc = lnet_push_target_resize();
2219
2220         if (rc) {
2221                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2222                 the_lnet.ln_push_target_handler = NULL;
2223         }
2224
2225         return rc;
2226 }
2227
2228 /* Clean up the push target. */
2229 static void lnet_push_target_fini(void)
2230 {
2231         if (!the_lnet.ln_push_target)
2232                 return;
2233
2234         /* Unlink and invalidate to prevent new references. */
2235         LNetMDUnlink(the_lnet.ln_push_target_md);
2236         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2237
2238         /* Wait for the unlink to complete. */
2239         wait_var_event_warning(&the_lnet.ln_push_target->pb_refcnt,
2240                                atomic_read(&the_lnet.ln_push_target->pb_refcnt) <= 1,
2241                                "Still waiting for ping data MD to unlink\n");
2242
2243         /* Drop ref set by lnet_ping_buffer_alloc() */
2244         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2245         the_lnet.ln_push_target = NULL;
2246         the_lnet.ln_push_target_nnis = 0;
2247
2248         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2249         lnet_assert_handler_unused(the_lnet.ln_push_target_handler);
2250         the_lnet.ln_push_target_handler = NULL;
2251 }
2252
2253 static int
2254 lnet_ni_tq_credits(struct lnet_ni *ni)
2255 {
2256         int     credits;
2257
2258         LASSERT(ni->ni_ncpts >= 1);
2259
2260         if (ni->ni_ncpts == 1)
2261                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2262
2263         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2264         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2265         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2266
2267         return credits;
2268 }
2269
2270 static void
2271 lnet_ni_unlink_locked(struct lnet_ni *ni)
2272 {
2273         /* move it to zombie list and nobody can find it anymore */
2274         LASSERT(!list_empty(&ni->ni_netlist));
2275         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2276         lnet_ni_decref_locked(ni, 0);
2277 }
2278
2279 static void
2280 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2281 {
2282         int             i;
2283         int             islo;
2284         struct lnet_ni  *ni;
2285         struct list_head *zombie_list = &net->net_ni_zombie;
2286
2287         /*
2288          * Now wait for the NIs I just nuked to show up on the zombie
2289          * list and shut them down in guaranteed thread context
2290          */
2291         i = 2;
2292         while (!list_empty(zombie_list)) {
2293                 int     *ref;
2294                 int     j;
2295
2296                 ni = list_entry(zombie_list->next,
2297                                 struct lnet_ni, ni_netlist);
2298                 list_del_init(&ni->ni_netlist);
2299                 /* the ni should be in deleting state. If it's not it's
2300                  * a bug */
2301                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2302                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2303                         if (*ref == 0)
2304                                 continue;
2305                         /* still busy, add it back to zombie list */
2306                         list_add(&ni->ni_netlist, zombie_list);
2307                         break;
2308                 }
2309
2310                 if (!list_empty(&ni->ni_netlist)) {
2311                         /* Unlock mutex while waiting to allow other
2312                          * threads to read the LNet state and fall through
2313                          * to avoid deadlock
2314                          */
2315                         lnet_net_unlock(LNET_LOCK_EX);
2316                         mutex_unlock(&the_lnet.ln_api_mutex);
2317
2318                         ++i;
2319                         if ((i & (-i)) == i) {
2320                                 CDEBUG(D_WARNING,
2321                                        "Waiting for zombie LNI %s\n",
2322                                        libcfs_nidstr(&ni->ni_nid));
2323                         }
2324                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2325
2326                         mutex_lock(&the_lnet.ln_api_mutex);
2327                         lnet_net_lock(LNET_LOCK_EX);
2328                         continue;
2329                 }
2330
2331                 lnet_net_unlock(LNET_LOCK_EX);
2332
2333                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2334
2335                 LASSERT(!in_interrupt());
2336                 /* Holding the mutex makes it safe for lnd_shutdown
2337                  * to call module_put(). Module unload cannot finish
2338                  * until lnet_unregister_lnd() completes, and that
2339                  * requires the mutex.
2340                  */
2341                 mutex_lock(&the_lnet.ln_lnd_mutex);
2342                 (net->net_lnd->lnd_shutdown)(ni);
2343                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2344
2345                 if (!islo)
2346                         CDEBUG(D_LNI, "Removed LNI %s\n",
2347                               libcfs_nidstr(&ni->ni_nid));
2348
2349                 lnet_ni_free(ni);
2350                 i = 2;
2351                 lnet_net_lock(LNET_LOCK_EX);
2352         }
2353 }
2354
2355 /* shutdown down the NI and release refcount */
2356 static void
2357 lnet_shutdown_lndni(struct lnet_ni *ni)
2358 {
2359         int i;
2360         struct lnet_net *net = ni->ni_net;
2361
2362         lnet_net_lock(LNET_LOCK_EX);
2363         lnet_ni_lock(ni);
2364         ni->ni_state = LNET_NI_STATE_DELETING;
2365         lnet_ni_unlock(ni);
2366         lnet_ni_unlink_locked(ni);
2367         lnet_incr_dlc_seq();
2368         lnet_net_unlock(LNET_LOCK_EX);
2369
2370         /* clear messages for this NI on the lazy portal */
2371         for (i = 0; i < the_lnet.ln_nportals; i++)
2372                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2373
2374         lnet_net_lock(LNET_LOCK_EX);
2375         lnet_clear_zombies_nis_locked(net);
2376         lnet_net_unlock(LNET_LOCK_EX);
2377 }
2378
2379 static void
2380 lnet_shutdown_lndnet(struct lnet_net *net)
2381 {
2382         struct lnet_ni *ni;
2383
2384         lnet_net_lock(LNET_LOCK_EX);
2385
2386         list_del_init(&net->net_list);
2387
2388         while (!list_empty(&net->net_ni_list)) {
2389                 ni = list_entry(net->net_ni_list.next,
2390                                 struct lnet_ni, ni_netlist);
2391                 lnet_net_unlock(LNET_LOCK_EX);
2392                 lnet_shutdown_lndni(ni);
2393                 lnet_net_lock(LNET_LOCK_EX);
2394         }
2395
2396         lnet_net_unlock(LNET_LOCK_EX);
2397
2398         /* Do peer table cleanup for this net */
2399         lnet_peer_tables_cleanup(net);
2400
2401         lnet_net_free(net);
2402 }
2403
2404 static void
2405 lnet_shutdown_lndnets(void)
2406 {
2407         struct lnet_net *net;
2408         LIST_HEAD(resend);
2409         struct lnet_msg *msg, *tmp;
2410
2411         /* NB called holding the global mutex */
2412
2413         /* All quiet on the API front */
2414         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2415         LASSERT(the_lnet.ln_refcount == 0);
2416
2417         lnet_net_lock(LNET_LOCK_EX);
2418         the_lnet.ln_state = LNET_STATE_STOPPING;
2419
2420         /*
2421          * move the nets to the zombie list to avoid them being
2422          * picked up for new work. LONET is also included in the
2423          * Nets that will be moved to the zombie list
2424          */
2425         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2426
2427         /* Drop the cached loopback Net. */
2428         if (the_lnet.ln_loni != NULL) {
2429                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2430                 the_lnet.ln_loni = NULL;
2431         }
2432         lnet_net_unlock(LNET_LOCK_EX);
2433
2434         /* iterate through the net zombie list and delete each net */
2435         while (!list_empty(&the_lnet.ln_net_zombie)) {
2436                 net = list_entry(the_lnet.ln_net_zombie.next,
2437                                  struct lnet_net, net_list);
2438                 lnet_shutdown_lndnet(net);
2439         }
2440
2441         spin_lock(&the_lnet.ln_msg_resend_lock);
2442         list_splice(&the_lnet.ln_msg_resend, &resend);
2443         spin_unlock(&the_lnet.ln_msg_resend_lock);
2444
2445         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2446                 list_del_init(&msg->msg_list);
2447                 msg->msg_no_resend = true;
2448                 lnet_finalize(msg, -ECANCELED);
2449         }
2450
2451         lnet_net_lock(LNET_LOCK_EX);
2452         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2453         lnet_net_unlock(LNET_LOCK_EX);
2454 }
2455
2456 static int
2457 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2458 {
2459         int                     rc = -EINVAL;
2460         struct lnet_tx_queue    *tq;
2461         int                     i;
2462         struct lnet_net         *net = ni->ni_net;
2463
2464         mutex_lock(&the_lnet.ln_lnd_mutex);
2465
2466         if (tun) {
2467                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2468                 ni->ni_lnd_tunables_set = true;
2469         }
2470
2471         rc = (net->net_lnd->lnd_startup)(ni);
2472
2473         mutex_unlock(&the_lnet.ln_lnd_mutex);
2474
2475         if (rc != 0) {
2476                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2477                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2478                 goto failed0;
2479         }
2480
2481         lnet_ni_lock(ni);
2482         ni->ni_state = LNET_NI_STATE_ACTIVE;
2483         lnet_ni_unlock(ni);
2484
2485         /* We keep a reference on the loopback net through the loopback NI */
2486         if (net->net_lnd->lnd_type == LOLND) {
2487                 lnet_ni_addref(ni);
2488                 LASSERT(the_lnet.ln_loni == NULL);
2489                 the_lnet.ln_loni = ni;
2490                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2491                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2492                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2493                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2494                 return 0;
2495         }
2496
2497         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2498             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2499                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2500                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2501                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2502                                         "" : "per-peer ");
2503                 /* shutdown the NI since if we get here then it must've already
2504                  * been started
2505                  */
2506                 lnet_shutdown_lndni(ni);
2507                 return -EINVAL;
2508         }
2509
2510         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2511                 tq->tq_credits_min =
2512                 tq->tq_credits_max =
2513                 tq->tq_credits = lnet_ni_tq_credits(ni);
2514         }
2515
2516         atomic_set(&ni->ni_tx_credits,
2517                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2518         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2519
2520         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2521                 libcfs_nidstr(&ni->ni_nid),
2522                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2523                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2524                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2525                 ni->ni_net->net_tunables.lct_peer_timeout);
2526
2527         return 0;
2528 failed0:
2529         lnet_ni_free(ni);
2530         return rc;
2531 }
2532
2533 static int
2534 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2535 {
2536         struct lnet_ni *ni;
2537         struct lnet_net *net_l = NULL;
2538         LIST_HEAD(local_ni_list);
2539         int rc;
2540         int ni_count = 0;
2541         __u32 lnd_type;
2542         const struct lnet_lnd  *lnd;
2543         int peer_timeout =
2544                 net->net_tunables.lct_peer_timeout;
2545         int maxtxcredits =
2546                 net->net_tunables.lct_max_tx_credits;
2547         int peerrtrcredits =
2548                 net->net_tunables.lct_peer_rtr_credits;
2549
2550         /*
2551          * make sure that this net is unique. If it isn't then
2552          * we are adding interfaces to an already existing network, and
2553          * 'net' is just a convenient way to pass in the list.
2554          * if it is unique we need to find the LND and load it if
2555          * necessary.
2556          */
2557         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2558                 lnd_type = LNET_NETTYP(net->net_id);
2559
2560                 mutex_lock(&the_lnet.ln_lnd_mutex);
2561                 lnd = lnet_find_lnd_by_type(lnd_type);
2562
2563                 if (lnd == NULL) {
2564                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2565                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2566                         mutex_lock(&the_lnet.ln_lnd_mutex);
2567
2568                         lnd = lnet_find_lnd_by_type(lnd_type);
2569                         if (lnd == NULL) {
2570                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2571                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2572                                 libcfs_lnd2str(lnd_type),
2573                                 libcfs_lnd2modname(lnd_type), rc);
2574 #ifndef HAVE_MODULE_LOADING_SUPPORT
2575                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2576                                                 "compiled with kernel module "
2577                                                 "loading support.");
2578 #endif
2579                                 rc = -EINVAL;
2580                                 goto failed0;
2581                         }
2582                 }
2583
2584                 net->net_lnd = lnd;
2585
2586                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2587
2588                 net_l = net;
2589         }
2590
2591         /*
2592          * net_l: if the network being added is unique then net_l
2593          *        will point to that network
2594          *        if the network being added is not unique then
2595          *        net_l points to the existing network.
2596          *
2597          * When we enter the loop below, we'll pick NIs off he
2598          * network beign added and start them up, then add them to
2599          * a local ni list. Once we've successfully started all
2600          * the NIs then we join the local NI list (of started up
2601          * networks) with the net_l->net_ni_list, which should
2602          * point to the correct network to add the new ni list to
2603          *
2604          * If any of the new NIs fail to start up, then we want to
2605          * iterate through the local ni list, which should include
2606          * any NIs which were successfully started up, and shut
2607          * them down.
2608          *
2609          * After than we want to delete the network being added,
2610          * to avoid a memory leak.
2611          */
2612         while (!list_empty(&net->net_ni_added)) {
2613                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2614                                 ni_netlist);
2615                 list_del_init(&ni->ni_netlist);
2616
2617                 /* make sure that the the NI we're about to start
2618                  * up is actually unique. if it's not fail. */
2619                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2620                                         ni->ni_interface)) {
2621                         rc = -EEXIST;
2622                         goto failed1;
2623                 }
2624
2625                 /* adjust the pointer the parent network, just in case it
2626                  * the net is a duplicate */
2627                 ni->ni_net = net_l;
2628
2629                 rc = lnet_startup_lndni(ni, tun);
2630
2631                 if (rc < 0)
2632                         goto failed1;
2633
2634                 lnet_ni_addref(ni);
2635                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2636
2637                 ni_count++;
2638         }
2639
2640         lnet_net_lock(LNET_LOCK_EX);
2641         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2642         lnet_incr_dlc_seq();
2643         lnet_net_unlock(LNET_LOCK_EX);
2644
2645         /* if the network is not unique then we don't want to keep
2646          * it around after we're done. Free it. Otherwise add that
2647          * net to the global the_lnet.ln_nets */
2648         if (net_l != net && net_l != NULL) {
2649                 /*
2650                  * TODO - note. currently the tunables can not be updated
2651                  * once added
2652                  */
2653                 lnet_net_free(net);
2654         } else {
2655                 /*
2656                  * restore tunables after it has been overwitten by the
2657                  * lnd
2658                  */
2659                 if (peer_timeout != -1)
2660                         net->net_tunables.lct_peer_timeout = peer_timeout;
2661                 if (maxtxcredits != -1)
2662                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2663                 if (peerrtrcredits != -1)
2664                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2665
2666                 lnet_net_lock(LNET_LOCK_EX);
2667                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2668                 lnet_net_unlock(LNET_LOCK_EX);
2669         }
2670
2671         return ni_count;
2672
2673 failed1:
2674         /*
2675          * shutdown the new NIs that are being started up
2676          * free the NET being started
2677          */
2678         while (!list_empty(&local_ni_list)) {
2679                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2680                                 ni_netlist);
2681
2682                 lnet_shutdown_lndni(ni);
2683         }
2684
2685 failed0:
2686         lnet_net_free(net);
2687
2688         return rc;
2689 }
2690
2691 static int
2692 lnet_startup_lndnets(struct list_head *netlist)
2693 {
2694         struct lnet_net         *net;
2695         int                     rc;
2696         int                     ni_count = 0;
2697
2698         /*
2699          * Change to running state before bringing up the LNDs. This
2700          * allows lnet_shutdown_lndnets() to assert that we've passed
2701          * through here.
2702          */
2703         lnet_net_lock(LNET_LOCK_EX);
2704         the_lnet.ln_state = LNET_STATE_RUNNING;
2705         lnet_net_unlock(LNET_LOCK_EX);
2706
2707         while (!list_empty(netlist)) {
2708                 net = list_entry(netlist->next, struct lnet_net, net_list);
2709                 list_del_init(&net->net_list);
2710
2711                 rc = lnet_startup_lndnet(net, NULL);
2712
2713                 if (rc < 0)
2714                         goto failed;
2715
2716                 ni_count += rc;
2717         }
2718
2719         return ni_count;
2720 failed:
2721         lnet_shutdown_lndnets();
2722
2723         return rc;
2724 }
2725
2726 static int lnet_genl_parse_list(struct sk_buff *msg,
2727                                 const struct ln_key_list *data[], u16 idx)
2728 {
2729         const struct ln_key_list *list = data[idx];
2730         const struct ln_key_props *props;
2731         struct nlattr *node;
2732         u16 count;
2733
2734         if (!list)
2735                 return 0;
2736
2737         if (!list->lkl_maxattr)
2738                 return -ERANGE;
2739
2740         props = list->lkl_list;
2741         if (!props)
2742                 return -EINVAL;
2743
2744         node = nla_nest_start(msg, LN_SCALAR_ATTR_LIST);
2745         if (!node)
2746                 return -ENOBUFS;
2747
2748         for (count = 1; count <= list->lkl_maxattr; count++) {
2749                 struct nlattr *key = nla_nest_start(msg, count);
2750
2751                 if (count == 1)
2752                         nla_put_u16(msg, LN_SCALAR_ATTR_LIST_SIZE,
2753                                     list->lkl_maxattr);
2754
2755                 nla_put_u16(msg, LN_SCALAR_ATTR_INDEX, count);
2756                 if (props[count].lkp_values)
2757                         nla_put_string(msg, LN_SCALAR_ATTR_VALUE,
2758                                        props[count].lkp_values);
2759                 if (props[count].lkp_key_format)
2760                         nla_put_u16(msg, LN_SCALAR_ATTR_KEY_FORMAT,
2761                                     props[count].lkp_key_format);
2762                 nla_put_u16(msg, LN_SCALAR_ATTR_NLA_TYPE,
2763                             props[count].lkp_data_type);
2764                 if (props[count].lkp_data_type == NLA_NESTED) {
2765                         int rc;
2766
2767                         rc = lnet_genl_parse_list(msg, data, ++idx);
2768                         if (rc < 0)
2769                                 return rc;
2770                 }
2771
2772                 nla_nest_end(msg, key);
2773         }
2774
2775         nla_nest_end(msg, node);
2776         return 0;
2777 }
2778
2779 int lnet_genl_send_scalar_list(struct sk_buff *msg, u32 portid, u32 seq,
2780                                const struct genl_family *family, int flags,
2781                                u8 cmd, const struct ln_key_list *data[])
2782 {
2783         int rc = 0;
2784         void *hdr;
2785
2786         if (!data[0])
2787                 return -EINVAL;
2788
2789         hdr = genlmsg_put(msg, portid, seq, family, flags, cmd);
2790         if (!hdr)
2791                 GOTO(canceled, rc = -EMSGSIZE);
2792
2793         rc = lnet_genl_parse_list(msg, data, 0);
2794         if (rc < 0)
2795                 GOTO(canceled, rc);
2796
2797         genlmsg_end(msg, hdr);
2798 canceled:
2799         if (rc < 0)
2800                 genlmsg_cancel(msg, hdr);
2801         return rc;
2802 }
2803 EXPORT_SYMBOL(lnet_genl_send_scalar_list);
2804
2805 /**
2806  * Initialize LNet library.
2807  *
2808  * Automatically called at module loading time. Caller has to call
2809  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2810  * latter returned 0. It must be called exactly once.
2811  *
2812  * \retval 0 on success
2813  * \retval -ve on failures.
2814  */
2815 int lnet_lib_init(void)
2816 {
2817         int rc;
2818
2819         lnet_assert_wire_constants();
2820
2821         /* refer to global cfs_cpt_table for now */
2822         the_lnet.ln_cpt_table = cfs_cpt_tab;
2823         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2824
2825         LASSERT(the_lnet.ln_cpt_number > 0);
2826         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2827                 /* we are under risk of consuming all lh_cookie */
2828                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2829                        "please change setting of CPT-table and retry\n",
2830                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2831                 return -E2BIG;
2832         }
2833
2834         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2835                 the_lnet.ln_cpt_bits++;
2836
2837         rc = lnet_create_locks();
2838         if (rc != 0) {
2839                 CERROR("Can't create LNet global locks: %d\n", rc);
2840                 return rc;
2841         }
2842
2843         the_lnet.ln_refcount = 0;
2844         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2845         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2846
2847         /* The hash table size is the number of bits it takes to express the set
2848          * ln_num_routes, minus 1 (better to under estimate than over so we
2849          * don't waste memory). */
2850         if (rnet_htable_size <= 0)
2851                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2852         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2853                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2854         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2855                                            order_base_2(rnet_htable_size) - 1);
2856
2857         /* All LNDs apart from the LOLND are in separate modules.  They
2858          * register themselves when their module loads, and unregister
2859          * themselves when their module is unloaded. */
2860         lnet_register_lnd(&the_lolnd);
2861         return 0;
2862 }
2863
2864 /**
2865  * Finalize LNet library.
2866  *
2867  * \pre lnet_lib_init() called with success.
2868  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2869  *
2870  * As this happens at module-unload, all lnds must already be unloaded,
2871  * so they must already be unregistered.
2872  */
2873 void lnet_lib_exit(void)
2874 {
2875         int i;
2876
2877         LASSERT(the_lnet.ln_refcount == 0);
2878         lnet_unregister_lnd(&the_lolnd);
2879         for (i = 0; i < NUM_LNDS; i++)
2880                 LASSERT(!the_lnet.ln_lnds[i]);
2881         lnet_destroy_locks();
2882 }
2883
2884 /**
2885  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2886  *
2887  * Users must call this function at least once before any other functions.
2888  * For each successful call there must be a corresponding call to
2889  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2890  * ignored.
2891  *
2892  * The PID used by LNet may be different from the one requested.
2893  * See LNetGetId().
2894  *
2895  * \param requested_pid PID requested by the caller.
2896  *
2897  * \return >= 0 on success, and < 0 error code on failures.
2898  */
2899 int
2900 LNetNIInit(lnet_pid_t requested_pid)
2901 {
2902         int                     im_a_router = 0;
2903         int                     rc;
2904         int                     ni_count;
2905         struct lnet_ping_buffer *pbuf;
2906         struct lnet_handle_md   ping_mdh;
2907         LIST_HEAD(net_head);
2908         struct lnet_net         *net;
2909
2910         mutex_lock(&the_lnet.ln_api_mutex);
2911
2912         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2913
2914         if (the_lnet.ln_refcount > 0) {
2915                 rc = the_lnet.ln_refcount++;
2916                 mutex_unlock(&the_lnet.ln_api_mutex);
2917                 return rc;
2918         }
2919
2920         rc = lnet_prepare(requested_pid);
2921         if (rc != 0) {
2922                 mutex_unlock(&the_lnet.ln_api_mutex);
2923                 return rc;
2924         }
2925
2926         /* create a network for Loopback network */
2927         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2928         if (net == NULL) {
2929                 rc = -ENOMEM;
2930                 goto err_empty_list;
2931         }
2932
2933         /* Add in the loopback NI */
2934         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2935                 rc = -ENOMEM;
2936                 goto err_empty_list;
2937         }
2938
2939         if (use_tcp_bonding)
2940                 CWARN("use_tcp_bonding has been removed. Use Multi-Rail and Dynamic Discovery instead, see LU-13641\n");
2941
2942         /* If LNet is being initialized via DLC it is possible
2943          * that the user requests not to load module parameters (ones which
2944          * are supported by DLC) on initialization.  Therefore, make sure not
2945          * to load networks, routes and forwarding from module parameters
2946          * in this case.  On cleanup in case of failure only clean up
2947          * routes if it has been loaded */
2948         if (!the_lnet.ln_nis_from_mod_params) {
2949                 rc = lnet_parse_networks(&net_head, lnet_get_networks());
2950                 if (rc < 0)
2951                         goto err_empty_list;
2952         }
2953
2954         ni_count = lnet_startup_lndnets(&net_head);
2955         if (ni_count < 0) {
2956                 rc = ni_count;
2957                 goto err_empty_list;
2958         }
2959
2960         if (!the_lnet.ln_nis_from_mod_params) {
2961                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2962                 if (rc != 0)
2963                         goto err_shutdown_lndnis;
2964
2965                 rc = lnet_rtrpools_alloc(im_a_router);
2966                 if (rc != 0)
2967                         goto err_destroy_routes;
2968         }
2969
2970         rc = lnet_acceptor_start();
2971         if (rc != 0)
2972                 goto err_destroy_routes;
2973
2974         the_lnet.ln_refcount = 1;
2975         /* Now I may use my own API functions... */
2976
2977         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2978         if (rc != 0)
2979                 goto err_acceptor_stop;
2980
2981         lnet_ping_target_update(pbuf, ping_mdh);
2982
2983         the_lnet.ln_mt_handler = lnet_mt_event_handler;
2984
2985         rc = lnet_push_target_init();
2986         if (rc != 0)
2987                 goto err_stop_ping;
2988
2989         rc = lnet_peer_discovery_start();
2990         if (rc != 0)
2991                 goto err_destroy_push_target;
2992
2993         rc = lnet_monitor_thr_start();
2994         if (rc != 0)
2995                 goto err_stop_discovery_thr;
2996
2997         lnet_fault_init();
2998         lnet_router_debugfs_init();
2999
3000         mutex_unlock(&the_lnet.ln_api_mutex);
3001
3002         complete_all(&the_lnet.ln_started);
3003
3004         /* wait for all routers to start */
3005         lnet_wait_router_start();
3006
3007         return 0;
3008
3009 err_stop_discovery_thr:
3010         lnet_peer_discovery_stop();
3011 err_destroy_push_target:
3012         lnet_push_target_fini();
3013 err_stop_ping:
3014         lnet_ping_target_fini();
3015 err_acceptor_stop:
3016         the_lnet.ln_refcount = 0;
3017         lnet_acceptor_stop();
3018 err_destroy_routes:
3019         if (!the_lnet.ln_nis_from_mod_params)
3020                 lnet_destroy_routes();
3021 err_shutdown_lndnis:
3022         lnet_shutdown_lndnets();
3023 err_empty_list:
3024         lnet_unprepare();
3025         LASSERT(rc < 0);
3026         mutex_unlock(&the_lnet.ln_api_mutex);
3027         while (!list_empty(&net_head)) {
3028                 struct lnet_net *net;
3029
3030                 net = list_entry(net_head.next, struct lnet_net, net_list);
3031                 list_del_init(&net->net_list);
3032                 lnet_net_free(net);
3033         }
3034         return rc;
3035 }
3036 EXPORT_SYMBOL(LNetNIInit);
3037
3038 /**
3039  * Stop LNet interfaces, routing, and forwarding.
3040  *
3041  * Users must call this function once for each successful call to LNetNIInit().
3042  * Once the LNetNIFini() operation has been started, the results of pending
3043  * API operations are undefined.
3044  *
3045  * \return always 0 for current implementation.
3046  */
3047 int
3048 LNetNIFini(void)
3049 {
3050         mutex_lock(&the_lnet.ln_api_mutex);
3051
3052         LASSERT(the_lnet.ln_refcount > 0);
3053
3054         if (the_lnet.ln_refcount != 1) {
3055                 the_lnet.ln_refcount--;
3056         } else {
3057                 LASSERT(!the_lnet.ln_niinit_self);
3058
3059                 lnet_fault_fini();
3060
3061                 lnet_router_debugfs_fini();
3062                 lnet_monitor_thr_stop();
3063                 lnet_peer_discovery_stop();
3064                 lnet_push_target_fini();
3065                 lnet_ping_target_fini();
3066
3067                 /* Teardown fns that use my own API functions BEFORE here */
3068                 the_lnet.ln_refcount = 0;
3069
3070                 lnet_acceptor_stop();
3071                 lnet_destroy_routes();
3072                 lnet_shutdown_lndnets();
3073                 lnet_unprepare();
3074         }
3075
3076         mutex_unlock(&the_lnet.ln_api_mutex);
3077         return 0;
3078 }
3079 EXPORT_SYMBOL(LNetNIFini);
3080
3081 /**
3082  * Grabs the ni data from the ni structure and fills the out
3083  * parameters
3084  *
3085  * \param[in] ni network        interface structure
3086  * \param[out] cfg_ni           NI config information
3087  * \param[out] tun              network and LND tunables
3088  */
3089 static void
3090 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
3091                    struct lnet_ioctl_config_lnd_tunables *tun,
3092                    struct lnet_ioctl_element_stats *stats,
3093                    __u32 tun_size)
3094 {
3095         size_t min_size = 0;
3096         int i;
3097
3098         if (!ni || !cfg_ni || !tun || !nid_is_nid4(&ni->ni_nid))
3099                 return;
3100
3101         if (ni->ni_interface != NULL) {
3102                 strncpy(cfg_ni->lic_ni_intf,
3103                         ni->ni_interface,
3104                         sizeof(cfg_ni->lic_ni_intf));
3105         }
3106
3107         cfg_ni->lic_nid = lnet_nid_to_nid4(&ni->ni_nid);
3108         cfg_ni->lic_status = lnet_ni_get_status_locked(ni);
3109         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
3110
3111         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
3112
3113         if (stats) {
3114                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
3115                                                        LNET_STATS_TYPE_SEND);
3116                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
3117                                                        LNET_STATS_TYPE_RECV);
3118                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
3119                                                        LNET_STATS_TYPE_DROP);
3120         }
3121
3122         /*
3123          * tun->lt_tun will always be present, but in order to be
3124          * backwards compatible, we need to deal with the cases when
3125          * tun->lt_tun is smaller than what the kernel has, because it
3126          * comes from an older version of a userspace program, then we'll
3127          * need to copy as much information as we have available space.
3128          */
3129         min_size = tun_size - sizeof(tun->lt_cmn);
3130         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
3131
3132         /* copy over the cpts */
3133         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
3134             ni->ni_cpts == NULL)  {
3135                 for (i = 0; i < ni->ni_ncpts; i++)
3136                         cfg_ni->lic_cpts[i] = i;
3137         } else {
3138                 for (i = 0;
3139                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
3140                      i < LNET_MAX_SHOW_NUM_CPT;
3141                      i++)
3142                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
3143         }
3144         cfg_ni->lic_ncpts = ni->ni_ncpts;
3145 }
3146
3147 /**
3148  * NOTE: This is a legacy function left in the code to be backwards
3149  * compatible with older userspace programs. It should eventually be
3150  * removed.
3151  *
3152  * Grabs the ni data from the ni structure and fills the out
3153  * parameters
3154  *
3155  * \param[in] ni network        interface structure
3156  * \param[out] config           config information
3157  */
3158 static void
3159 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
3160                          struct lnet_ioctl_config_data *config)
3161 {
3162         struct lnet_ioctl_net_config *net_config;
3163         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
3164         size_t min_size, tunable_size = 0;
3165         int i;
3166
3167         if (!ni || !config || !nid_is_nid4(&ni->ni_nid))
3168                 return;
3169
3170         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
3171         if (!net_config)
3172                 return;
3173
3174         if (!ni->ni_interface)
3175                 return;
3176
3177         strncpy(net_config->ni_interface,
3178                 ni->ni_interface,
3179                 sizeof(net_config->ni_interface));
3180
3181         config->cfg_nid = lnet_nid_to_nid4(&ni->ni_nid);
3182         config->cfg_config_u.cfg_net.net_peer_timeout =
3183                 ni->ni_net->net_tunables.lct_peer_timeout;
3184         config->cfg_config_u.cfg_net.net_max_tx_credits =
3185                 ni->ni_net->net_tunables.lct_max_tx_credits;
3186         config->cfg_config_u.cfg_net.net_peer_tx_credits =
3187                 ni->ni_net->net_tunables.lct_peer_tx_credits;
3188         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
3189                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
3190
3191         net_config->ni_status = lnet_ni_get_status_locked(ni);
3192
3193         if (ni->ni_cpts) {
3194                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
3195
3196                 for (i = 0; i < num_cpts; i++)
3197                         net_config->ni_cpts[i] = ni->ni_cpts[i];
3198
3199                 config->cfg_ncpts = num_cpts;
3200         }
3201
3202         /*
3203          * See if user land tools sent in a newer and larger version
3204          * of struct lnet_tunables than what the kernel uses.
3205          */
3206         min_size = sizeof(*config) + sizeof(*net_config);
3207
3208         if (config->cfg_hdr.ioc_len > min_size)
3209                 tunable_size = config->cfg_hdr.ioc_len - min_size;
3210
3211         /* Don't copy too much data to user space */
3212         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
3213         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
3214
3215         if (lnd_cfg && min_size) {
3216                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
3217                 config->cfg_config_u.cfg_net.net_interface_count = 1;
3218
3219                 /* Tell user land that kernel side has less data */
3220                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
3221                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
3222                         config->cfg_hdr.ioc_len -= min_size;
3223                 }
3224         }
3225 }
3226
3227 struct lnet_ni *
3228 lnet_get_ni_idx_locked(int idx)
3229 {
3230         struct lnet_ni          *ni;
3231         struct lnet_net         *net;
3232
3233         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3234                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3235                         if (idx-- == 0)
3236                                 return ni;
3237                 }
3238         }
3239
3240         return NULL;
3241 }
3242
3243 int lnet_get_net_healthv_locked(struct lnet_net *net)
3244 {
3245         struct lnet_ni *ni;
3246         int best_healthv = 0;
3247         int healthv, ni_fatal;
3248
3249         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3250                 healthv = atomic_read(&ni->ni_healthv);
3251                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
3252                 if (!ni_fatal && healthv > best_healthv)
3253                         best_healthv = healthv;
3254         }
3255
3256         return best_healthv;
3257 }
3258
3259 struct lnet_ni *
3260 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
3261 {
3262         struct lnet_ni          *ni;
3263         struct lnet_net         *net = mynet;
3264
3265         /*
3266          * It is possible that the net has been cleaned out while there is
3267          * a message being sent. This function accessed the net without
3268          * checking if the list is empty
3269          */
3270         if (prev == NULL) {
3271                 if (net == NULL)
3272                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
3273                                         net_list);
3274                 if (list_empty(&net->net_ni_list))
3275                         return NULL;
3276                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3277                                 ni_netlist);
3278
3279                 return ni;
3280         }
3281
3282         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
3283                 /* if you reached the end of the ni list and the net is
3284                  * specified, then there are no more nis in that net */
3285                 if (net != NULL)
3286                         return NULL;
3287
3288                 /* we reached the end of this net ni list. move to the
3289                  * next net */
3290                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
3291                         /* no more nets and no more NIs. */
3292                         return NULL;
3293
3294                 /* get the next net */
3295                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
3296                                  net_list);
3297                 if (list_empty(&net->net_ni_list))
3298                         return NULL;
3299                 /* get the ni on it */
3300                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3301                                 ni_netlist);
3302
3303                 return ni;
3304         }
3305
3306         if (list_empty(&prev->ni_netlist))
3307                 return NULL;
3308
3309         /* there are more nis left */
3310         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
3311
3312         return ni;
3313 }
3314
3315 int
3316 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3317 {
3318         struct lnet_ni *ni;
3319         int cpt;
3320         int rc = -ENOENT;
3321         int idx = config->cfg_count;
3322
3323         cpt = lnet_net_lock_current();
3324
3325         ni = lnet_get_ni_idx_locked(idx);
3326
3327         if (ni != NULL) {
3328                 rc = 0;
3329                 lnet_ni_lock(ni);
3330                 lnet_fill_ni_info_legacy(ni, config);
3331                 lnet_ni_unlock(ni);
3332         }
3333
3334         lnet_net_unlock(cpt);
3335         return rc;
3336 }
3337
3338 int
3339 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3340                    struct lnet_ioctl_config_lnd_tunables *tun,
3341                    struct lnet_ioctl_element_stats *stats,
3342                    __u32 tun_size)
3343 {
3344         struct lnet_ni          *ni;
3345         int                     cpt;
3346         int                     rc = -ENOENT;
3347
3348         if (!cfg_ni || !tun || !stats)
3349                 return -EINVAL;
3350
3351         cpt = lnet_net_lock_current();
3352
3353         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3354
3355         if (ni) {
3356                 rc = 0;
3357                 lnet_ni_lock(ni);
3358                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3359                 lnet_ni_unlock(ni);
3360         }
3361
3362         lnet_net_unlock(cpt);
3363         return rc;
3364 }
3365
3366 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3367 {
3368         struct lnet_ni *ni;
3369         int cpt;
3370         int rc = -ENOENT;
3371
3372         if (!msg_stats)
3373                 return -EINVAL;
3374
3375         cpt = lnet_net_lock_current();
3376
3377         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3378
3379         if (ni) {
3380                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3381                 rc = 0;
3382         }
3383
3384         lnet_net_unlock(cpt);
3385
3386         return rc;
3387 }
3388
3389 static int lnet_add_net_common(struct lnet_net *net,
3390                                struct lnet_ioctl_config_lnd_tunables *tun)
3391 {
3392         struct lnet_handle_md ping_mdh;
3393         struct lnet_ping_buffer *pbuf;
3394         struct lnet_remotenet *rnet;
3395         struct lnet_ni *ni;
3396         int net_ni_count;
3397         __u32 net_id;
3398         int rc;
3399
3400         lnet_net_lock(LNET_LOCK_EX);
3401         rnet = lnet_find_rnet_locked(net->net_id);
3402         lnet_net_unlock(LNET_LOCK_EX);
3403         /*
3404          * make sure that the net added doesn't invalidate the current
3405          * configuration LNet is keeping
3406          */
3407         if (rnet) {
3408                 CERROR("Adding net %s will invalidate routing configuration\n",
3409                        libcfs_net2str(net->net_id));
3410                 lnet_net_free(net);
3411                 return -EUSERS;
3412         }
3413
3414         /*
3415          * make sure you calculate the correct number of slots in the ping
3416          * buffer. Since the ping info is a flattened list of all the NIs,
3417          * we should allocate enough slots to accomodate the number of NIs
3418          * which will be added.
3419          *
3420          * since ni hasn't been configured yet, use
3421          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3422          */
3423         net_ni_count = lnet_get_net_ni_count_pre(net);
3424
3425         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3426                                     net_ni_count + lnet_get_ni_count(),
3427                                     false);
3428         if (rc < 0) {
3429                 lnet_net_free(net);
3430                 return rc;
3431         }
3432
3433         if (tun)
3434                 memcpy(&net->net_tunables,
3435                        &tun->lt_cmn, sizeof(net->net_tunables));
3436         else
3437                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3438
3439         net_id = net->net_id;
3440
3441         rc = lnet_startup_lndnet(net,
3442                                  (tun) ? &tun->lt_tun : NULL);
3443         if (rc < 0)
3444                 goto failed;
3445
3446         lnet_net_lock(LNET_LOCK_EX);
3447         net = lnet_get_net_locked(net_id);
3448         LASSERT(net);
3449
3450         /* apply the UDSPs */
3451         rc = lnet_udsp_apply_policies_on_net(net);
3452         if (rc)
3453                 CERROR("Failed to apply UDSPs on local net %s\n",
3454                        libcfs_net2str(net->net_id));
3455
3456         /* At this point we lost track of which NI was just added, so we
3457          * just re-apply the policies on all of the NIs on this net
3458          */
3459         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3460                 rc = lnet_udsp_apply_policies_on_ni(ni);
3461                 if (rc)
3462                         CERROR("Failed to apply UDSPs on ni %s\n",
3463                                libcfs_nidstr(&ni->ni_nid));
3464         }
3465         lnet_net_unlock(LNET_LOCK_EX);
3466
3467         /*
3468          * Start the acceptor thread if this is the first network
3469          * being added that requires the thread.
3470          */
3471         if (net->net_lnd->lnd_accept) {
3472                 rc = lnet_acceptor_start();
3473                 if (rc < 0) {
3474                         /* shutdown the net that we just started */
3475                         CERROR("Failed to start up acceptor thread\n");
3476                         lnet_shutdown_lndnet(net);
3477                         goto failed;
3478                 }
3479         }
3480
3481         lnet_net_lock(LNET_LOCK_EX);
3482         lnet_peer_net_added(net);
3483         lnet_net_unlock(LNET_LOCK_EX);
3484
3485         lnet_ping_target_update(pbuf, ping_mdh);
3486
3487         return 0;
3488
3489 failed:
3490         lnet_ping_md_unlink(pbuf, &ping_mdh);
3491         lnet_ping_buffer_decref(pbuf);
3492         return rc;
3493 }
3494
3495 static void
3496 lnet_set_tune_defaults(struct lnet_ioctl_config_lnd_tunables *tun)
3497 {
3498         if (tun) {
3499                 if (!tun->lt_cmn.lct_peer_timeout)
3500                         tun->lt_cmn.lct_peer_timeout = DEFAULT_PEER_TIMEOUT;
3501                 if (!tun->lt_cmn.lct_peer_tx_credits)
3502                         tun->lt_cmn.lct_peer_tx_credits = DEFAULT_PEER_CREDITS;
3503                 if (!tun->lt_cmn.lct_max_tx_credits)
3504                         tun->lt_cmn.lct_max_tx_credits = DEFAULT_CREDITS;
3505         }
3506 }
3507
3508 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3509                                       struct lnet_ioctl_config_lnd_tunables *tun)
3510 {
3511         struct lnet_net *net;
3512         const char *nets;
3513         int rc;
3514         LIST_HEAD(net_head);
3515
3516         rc = lnet_parse_ip2nets(&nets, ip2nets);
3517         if (rc < 0)