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 unsigned int lnet_max_recovery_ping_interval = 900;
 132 unsigned int lnet_max_recovery_ping_count = 9;
 133 static int max_recovery_ping_interval_set(const char *val,
 134                                           cfs_kernel_param_arg_t *kp);
 135
 136 #define param_check_max_recovery_ping_interval(name, p) \
 137                 __param_check(name, p, int)
 138
 139 #ifdef HAVE_KERNEL_PARAM_OPS
 140 static struct kernel_param_ops param_ops_max_recovery_ping_interval = {
 141         .set = max_recovery_ping_interval_set,
 142         .get = param_get_int,
 143 };
 144 module_param(lnet_max_recovery_ping_interval, max_recovery_ping_interval, 0644);
 145 #else
 146 module_param_call(lnet_max_recovery_ping_interval, max_recovery_ping_interval,
 147                   param_get_int, &lnet_max_recovery_ping_interval, 0644);
 148 #endif
 149 MODULE_PARM_DESC(lnet_max_recovery_ping_interval,
 150                  "The max interval between LNet recovery pings, in seconds");
 151
 152 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
 153 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
 154
 155 static struct kernel_param_ops param_ops_interfaces_max = {
 156         .set = intf_max_set,
 157         .get = param_get_int,
 158 };
 159
 160 #define param_check_interfaces_max(name, p) \
 161                 __param_check(name, p, int)
 162
 163 #ifdef HAVE_KERNEL_PARAM_OPS
 164 module_param(lnet_interfaces_max, interfaces_max, 0644);
 165 #else
 166 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
 167                   &param_ops_interfaces_max, 0644);
 168 #endif
 169 MODULE_PARM_DESC(lnet_interfaces_max,
 170                 "Maximum number of interfaces in a node.");
 171
 172 unsigned lnet_peer_discovery_disabled = 0;
 173 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
 174
 175 static struct kernel_param_ops param_ops_discovery_disabled = {
 176         .set = discovery_set,
 177         .get = param_get_int,
 178 };
 179
 180 #define param_check_discovery_disabled(name, p) \
 181                 __param_check(name, p, int)
 182 #ifdef HAVE_KERNEL_PARAM_OPS
 183 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
 184 #else
 185 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
 186                   &param_ops_discovery_disabled, 0644);
 187 #endif
 188 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 189                 "Set to 1 to disable peer discovery on this node.");
 190
 191 unsigned int lnet_drop_asym_route;
 192 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
 193
 194 static struct kernel_param_ops param_ops_drop_asym_route = {
 195         .set = drop_asym_route_set,
 196         .get = param_get_int,
 197 };
 198
 199 #define param_check_drop_asym_route(name, p)    \
 200         __param_check(name, p, int)
 201 #ifdef HAVE_KERNEL_PARAM_OPS
 202 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
 203 #else
 204 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
 205                   &param_ops_drop_asym_route, 0644);
 206 #endif
 207 MODULE_PARM_DESC(lnet_drop_asym_route,
 208                  "Set to 1 to drop asymmetrical route messages.");
 209
 210 #define LNET_TRANSACTION_TIMEOUT_DEFAULT 50
 211 unsigned int lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_DEFAULT;
 212 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
 213 #ifdef HAVE_KERNEL_PARAM_OPS
 214 static struct kernel_param_ops param_ops_transaction_timeout = {
 215         .set = transaction_to_set,
 216         .get = param_get_int,
 217 };
 218
 219 #define param_check_transaction_timeout(name, p) \
 220                 __param_check(name, p, int)
 221 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
 222 #else
 223 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
 224                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
 225 #endif
 226 MODULE_PARM_DESC(lnet_transaction_timeout,
 227                 "Maximum number of seconds to wait for a peer response.");
 228
 229 #define LNET_RETRY_COUNT_DEFAULT 2
 230 unsigned int lnet_retry_count = LNET_RETRY_COUNT_DEFAULT;
 231 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
 232 #ifdef HAVE_KERNEL_PARAM_OPS
 233 static struct kernel_param_ops param_ops_retry_count = {
 234         .set = retry_count_set,
 235         .get = param_get_int,
 236 };
 237
 238 #define param_check_retry_count(name, p) \
 239                 __param_check(name, p, int)
 240 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
 241 #else
 242 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
 243                   &lnet_retry_count, S_IRUGO|S_IWUSR);
 244 #endif
 245 MODULE_PARM_DESC(lnet_retry_count,
 246                  "Maximum number of times to retry transmitting a message");
 247
 248 unsigned int lnet_response_tracking = 3;
 249 static int response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp);
 250
 251 #ifdef HAVE_KERNEL_PARAM_OPS
 252 static struct kernel_param_ops param_ops_response_tracking = {
 253         .set = response_tracking_set,
 254         .get = param_get_int,
 255 };
 256
 257 #define param_check_response_tracking(name, p)  \
 258         __param_check(name, p, int)
 259 module_param(lnet_response_tracking, response_tracking, 0644);
 260 #else
 261 module_param_call(lnet_response_tracking, response_tracking_set, param_get_int,
 262                   &lnet_response_tracking, 0644);
 263 #endif
 264 MODULE_PARM_DESC(lnet_response_tracking,
 265                  "(0|1|2|3) LNet Internal Only|GET Reply only|PUT ACK only|Full Tracking (default)");
 266
 267 #define LNET_LND_TIMEOUT_DEFAULT ((LNET_TRANSACTION_TIMEOUT_DEFAULT - 1) / \
 268                                   (LNET_RETRY_COUNT_DEFAULT + 1))
 269 unsigned int lnet_lnd_timeout = LNET_LND_TIMEOUT_DEFAULT;
 270 static void lnet_set_lnd_timeout(void)
 271 {
 272         lnet_lnd_timeout = (lnet_transaction_timeout - 1) /
 273                            (lnet_retry_count + 1);
 274 }
 275
 276 /*
 277  * This sequence number keeps track of how many times DLC was used to
 278  * update the local NIs. It is incremented when a NI is added or
 279  * removed and checked when sending a message to determine if there is
 280  * a need to re-run the selection algorithm. See lnet_select_pathway()
 281  * for more details on its usage.
 282  */
 283 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
 284
 285 static int lnet_ping(struct lnet_process_id id4, struct lnet_nid *src_nid,
 286                      signed long timeout, struct lnet_process_id __user *ids,
 287                      int n_ids);
 288
 289 static int lnet_discover(struct lnet_process_id id, __u32 force,
 290                          struct lnet_process_id __user *ids, int n_ids);
 291
 292 static int
 293 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 294 {
 295         int rc;
 296         unsigned *sensitivity = (unsigned *)kp->arg;
 297         unsigned long value;
 298
 299         rc = kstrtoul(val, 0, &value);
 300         if (rc) {
 301                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
 302                 return rc;
 303         }
 304
 305         /*
 306          * The purpose of locking the api_mutex here is to ensure that
 307          * the correct value ends up stored properly.
 308          */
 309         mutex_lock(&the_lnet.ln_api_mutex);
 310
 311         if (value > LNET_MAX_HEALTH_VALUE) {
 312                 mutex_unlock(&the_lnet.ln_api_mutex);
 313                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
 314                        LNET_MAX_HEALTH_VALUE, value);
 315                 return -EINVAL;
 316         }
 317
 318         if (*sensitivity != 0 && value == 0 && lnet_retry_count != 0) {
 319                 lnet_retry_count = 0;
 320                 lnet_set_lnd_timeout();
 321         }
 322
 323         *sensitivity = value;
 324
 325         mutex_unlock(&the_lnet.ln_api_mutex);
 326
 327         return 0;
 328 }
 329
 330 static int
 331 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 332 {
 333         CWARN("'lnet_recovery_interval' has been deprecated\n");
 334
 335         return 0;
 336 }
 337
 338 static int
 339 max_recovery_ping_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 340 {
 341         int rc;
 342         unsigned long value;
 343
 344         rc = kstrtoul(val, 0, &value);
 345         if (rc) {
 346                 CERROR("Invalid module parameter value for 'lnet_max_recovery_ping_interval'\n");
 347                 return rc;
 348         }
 349
 350         if (!value) {
 351                 CERROR("Invalid max ping timeout. Must be strictly positive\n");
 352                 return -EINVAL;
 353         }
 354
 355         /* The purpose of locking the api_mutex here is to ensure that
 356          * the correct value ends up stored properly.
 357          */
 358         mutex_lock(&the_lnet.ln_api_mutex);
 359         lnet_max_recovery_ping_interval = value;
 360         lnet_max_recovery_ping_count = 0;
 361         value >>= 1;
 362         while (value) {
 363                 lnet_max_recovery_ping_count++;
 364                 value >>= 1;
 365         }
 366         mutex_unlock(&the_lnet.ln_api_mutex);
 367
 368         return 0;
 369 }
 370
 371 static int
 372 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
 373 {
 374         int rc;
 375         unsigned *discovery_off = (unsigned *)kp->arg;
 376         unsigned long value;
 377         struct lnet_ping_buffer *pbuf;
 378
 379         rc = kstrtoul(val, 0, &value);
 380         if (rc) {
 381                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
 382                 return rc;
 383         }
 384
 385         value = (value) ? 1 : 0;
 386
 387         /*
 388          * The purpose of locking the api_mutex here is to ensure that
 389          * the correct value ends up stored properly.
 390          */
 391         mutex_lock(&the_lnet.ln_api_mutex);
 392
 393         if (value == *discovery_off) {
 394                 mutex_unlock(&the_lnet.ln_api_mutex);
 395                 return 0;
 396         }
 397
 398         /*
 399          * We still want to set the discovery value even when LNet is not
 400          * running. This is the case when LNet is being loaded and we want
 401          * the module parameters to take effect. Otherwise if we're
 402          * changing the value dynamically, we want to set it after
 403          * updating the peers
 404          */
 405         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 406                 *discovery_off = value;
 407                 mutex_unlock(&the_lnet.ln_api_mutex);
 408                 return 0;
 409         }
 410
 411         /* tell peers that discovery setting has changed */
 412         lnet_net_lock(LNET_LOCK_EX);
 413         pbuf = the_lnet.ln_ping_target;
 414         if (value)
 415                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
 416         else
 417                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
 418         lnet_net_unlock(LNET_LOCK_EX);
 419
 420         /* only send a push when we're turning off discovery */
 421         if (*discovery_off <= 0 && value > 0)
 422                 lnet_push_update_to_peers(1);
 423         *discovery_off = value;
 424
 425         mutex_unlock(&the_lnet.ln_api_mutex);
 426
 427         return 0;
 428 }
 429
 430 static int
 431 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
 432 {
 433         int rc;
 434         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
 435         unsigned long value;
 436
 437         rc = kstrtoul(val, 0, &value);
 438         if (rc) {
 439                 CERROR("Invalid module parameter value for "
 440                        "'lnet_drop_asym_route'\n");
 441                 return rc;
 442         }
 443
 444         /*
 445          * The purpose of locking the api_mutex here is to ensure that
 446          * the correct value ends up stored properly.
 447          */
 448         mutex_lock(&the_lnet.ln_api_mutex);
 449
 450         if (value == *drop_asym_route) {
 451                 mutex_unlock(&the_lnet.ln_api_mutex);
 452                 return 0;
 453         }
 454
 455         *drop_asym_route = value;
 456
 457         mutex_unlock(&the_lnet.ln_api_mutex);
 458
 459         return 0;
 460 }
 461
 462 static int
 463 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
 464 {
 465         int rc;
 466         unsigned *transaction_to = (unsigned *)kp->arg;
 467         unsigned long value;
 468
 469         rc = kstrtoul(val, 0, &value);
 470         if (rc) {
 471                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
 472                 return rc;
 473         }
 474
 475         /*
 476          * The purpose of locking the api_mutex here is to ensure that
 477          * the correct value ends up stored properly.
 478          */
 479         mutex_lock(&the_lnet.ln_api_mutex);
 480
 481         if (value <= lnet_retry_count || value == 0) {
 482                 mutex_unlock(&the_lnet.ln_api_mutex);
 483                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
 484                        "Has to be greater than lnet_retry_count (%u)\n",
 485                        value, lnet_retry_count);
 486                 return -EINVAL;
 487         }
 488
 489         if (value == *transaction_to) {
 490                 mutex_unlock(&the_lnet.ln_api_mutex);
 491                 return 0;
 492         }
 493
 494         *transaction_to = value;
 495         /* Update the lnet_lnd_timeout now that we've modified the
 496          * transaction timeout
 497          */
 498         lnet_set_lnd_timeout();
 499
 500         mutex_unlock(&the_lnet.ln_api_mutex);
 501
 502         return 0;
 503 }
 504
 505 static int
 506 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
 507 {
 508         int rc;
 509         unsigned *retry_count = (unsigned *)kp->arg;
 510         unsigned long value;
 511
 512         rc = kstrtoul(val, 0, &value);
 513         if (rc) {
 514                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
 515                 return rc;
 516         }
 517
 518         /*
 519          * The purpose of locking the api_mutex here is to ensure that
 520          * the correct value ends up stored properly.
 521          */
 522         mutex_lock(&the_lnet.ln_api_mutex);
 523
 524         if (lnet_health_sensitivity == 0 && value > 0) {
 525                 mutex_unlock(&the_lnet.ln_api_mutex);
 526                 CERROR("Can not set lnet_retry_count when health feature is turned off\n");
 527                 return -EINVAL;
 528         }
 529
 530         if (value > lnet_transaction_timeout) {
 531                 mutex_unlock(&the_lnet.ln_api_mutex);
 532                 CERROR("Invalid value for lnet_retry_count (%lu). "
 533                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
 534                        value, lnet_transaction_timeout);
 535                 return -EINVAL;
 536         }
 537
 538         *retry_count = value;
 539
 540         /* Update the lnet_lnd_timeout now that we've modified the
 541          * retry count
 542          */
 543         lnet_set_lnd_timeout();
 544
 545         mutex_unlock(&the_lnet.ln_api_mutex);
 546
 547         return 0;
 548 }
 549
 550 static int
 551 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
 552 {
 553         int value, rc;
 554
 555         rc = kstrtoint(val, 0, &value);
 556         if (rc) {
 557                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
 558                 return rc;
 559         }
 560
 561         if (value < LNET_INTERFACES_MIN) {
 562                 CWARN("max interfaces provided are too small, setting to %d\n",
 563                       LNET_INTERFACES_MAX_DEFAULT);
 564                 value = LNET_INTERFACES_MAX_DEFAULT;
 565         }
 566
 567         *(int *)kp->arg = value;
 568
 569         return 0;
 570 }
 571
 572 static int
 573 response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp)
 574 {
 575         int rc;
 576         unsigned long new_value;
 577
 578         rc = kstrtoul(val, 0, &new_value);
 579         if (rc) {
 580                 CERROR("Invalid value for 'lnet_response_tracking'\n");
 581                 return -EINVAL;
 582         }
 583
 584         if (new_value < 0 || new_value > 3) {
 585                 CWARN("Invalid value (%lu) for 'lnet_response_tracking'\n",
 586                       new_value);
 587                 return -EINVAL;
 588         }
 589
 590         lnet_response_tracking = new_value;
 591
 592         return 0;
 593 }
 594
 595 static const char *
 596 lnet_get_routes(void)
 597 {
 598         return routes;
 599 }
 600
 601 static const char *
 602 lnet_get_networks(void)
 603 {
 604         const char *nets;
 605         int rc;
 606
 607         if (*networks != 0 && *ip2nets != 0) {
 608                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 609                                    "'ip2nets' but not both at once\n");
 610                 return NULL;
 611         }
 612
 613         if (*ip2nets != 0) {
 614                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 615                 return (rc == 0) ? nets : NULL;
 616         }
 617
 618         if (*networks != 0)
 619                 return networks;
 620
 621         return "tcp";
 622 }
 623
 624 static void
 625 lnet_init_locks(void)
 626 {
 627         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 628         spin_lock_init(&the_lnet.ln_msg_resend_lock);
 629         init_completion(&the_lnet.ln_mt_wait_complete);
 630         mutex_init(&the_lnet.ln_lnd_mutex);
 631 }
 632
 633 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 634 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 635                                             *  MDs kmem_cache */
 636 struct kmem_cache *lnet_udsp_cachep;       /* udsp cache */
 637 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
 638 struct kmem_cache *lnet_msg_cachep;
 639
 640 static int
 641 lnet_slab_setup(void)
 642 {
 643         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 644          * allocated in <size-xxx> kmem_cache).
 645          */
 646         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 647                                             0, 0, NULL);
 648         if (!lnet_mes_cachep)
 649                 return -ENOMEM;
 650
 651         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 652                                                   LNET_SMALL_MD_SIZE, 0, 0,
 653                                                   NULL);
 654         if (!lnet_small_mds_cachep)
 655                 return -ENOMEM;
 656
 657         lnet_udsp_cachep = kmem_cache_create("lnet_udsp",
 658                                              sizeof(struct lnet_udsp),
 659                                              0, 0, NULL);
 660         if (!lnet_udsp_cachep)
 661                 return -ENOMEM;
 662
 663         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
 664                                             0, 0, NULL);
 665         if (!lnet_rspt_cachep)
 666                 return -ENOMEM;
 667
 668         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
 669                                             0, 0, NULL);
 670         if (!lnet_msg_cachep)
 671                 return -ENOMEM;
 672
 673         return 0;
 674 }
 675
 676 static void
 677 lnet_slab_cleanup(void)
 678 {
 679         if (lnet_msg_cachep) {
 680                 kmem_cache_destroy(lnet_msg_cachep);
 681                 lnet_msg_cachep = NULL;
 682         }
 683
 684         if (lnet_rspt_cachep) {
 685                 kmem_cache_destroy(lnet_rspt_cachep);
 686                 lnet_rspt_cachep = NULL;
 687         }
 688
 689         if (lnet_udsp_cachep) {
 690                 kmem_cache_destroy(lnet_udsp_cachep);
 691                 lnet_udsp_cachep = NULL;
 692         }
 693
 694         if (lnet_small_mds_cachep) {
 695                 kmem_cache_destroy(lnet_small_mds_cachep);
 696                 lnet_small_mds_cachep = NULL;
 697         }
 698
 699         if (lnet_mes_cachep) {
 700                 kmem_cache_destroy(lnet_mes_cachep);
 701                 lnet_mes_cachep = NULL;
 702         }
 703 }
 704
 705 static int
 706 lnet_create_remote_nets_table(void)
 707 {
 708         int               i;
 709         struct list_head *hash;
 710
 711         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 712         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 713         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
 714         if (hash == NULL) {
 715                 CERROR("Failed to create remote nets hash table\n");
 716                 return -ENOMEM;
 717         }
 718
 719         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 720                 INIT_LIST_HEAD(&hash[i]);
 721         the_lnet.ln_remote_nets_hash = hash;
 722         return 0;
 723 }
 724
 725 static void
 726 lnet_destroy_remote_nets_table(void)
 727 {
 728         int i;
 729
 730         if (the_lnet.ln_remote_nets_hash == NULL)
 731                 return;
 732
 733         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 734                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 735
 736         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
 737                            LNET_REMOTE_NETS_HASH_SIZE);
 738         the_lnet.ln_remote_nets_hash = NULL;
 739 }
 740
 741 static void
 742 lnet_destroy_locks(void)
 743 {
 744         if (the_lnet.ln_res_lock != NULL) {
 745                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 746                 the_lnet.ln_res_lock = NULL;
 747         }
 748
 749         if (the_lnet.ln_net_lock != NULL) {
 750                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 751                 the_lnet.ln_net_lock = NULL;
 752         }
 753 }
 754
 755 static int
 756 lnet_create_locks(void)
 757 {
 758         lnet_init_locks();
 759
 760         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 761         if (the_lnet.ln_res_lock == NULL)
 762                 goto failed;
 763
 764         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 765         if (the_lnet.ln_net_lock == NULL)
 766                 goto failed;
 767
 768         return 0;
 769
 770  failed:
 771         lnet_destroy_locks();
 772         return -ENOMEM;
 773 }
 774
 775 static void lnet_assert_wire_constants(void)
 776 {
 777         /* Wire protocol assertions generated by 'wirecheck'
 778          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 779          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 780          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7)
 781          */
 782
 783         /* Constants... */
 784         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
 785         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
 786         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
 787         BUILD_BUG_ON(LNET_MSG_ACK != 0);
 788         BUILD_BUG_ON(LNET_MSG_PUT != 1);
 789         BUILD_BUG_ON(LNET_MSG_GET != 2);
 790         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
 791         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
 792
 793         BUILD_BUG_ON((int)sizeof(lnet_nid_t) != 8);
 794         BUILD_BUG_ON((int)sizeof(lnet_pid_t) != 4);
 795
 796         /* Checks for struct lnet_nid */
 797         BUILD_BUG_ON((int)sizeof(struct lnet_nid) != 20);
 798         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_size) != 0);
 799         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_size) != 1);
 800         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_type) != 1);
 801         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_type) != 1);
 802         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_num) != 2);
 803         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_num) != 2);
 804         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_addr) != 4);
 805         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_addr) != 16);
 806
 807         /* Checks for struct lnet_process_id_packed */
 808         BUILD_BUG_ON((int)sizeof(struct lnet_process_id_packed) != 12);
 809         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, nid) != 0);
 810         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->nid) != 8);
 811         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, pid) != 8);
 812         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->pid) != 4);
 813
 814         /* Checks for struct lnet_handle_wire */
 815         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
 816         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 817                                    wh_interface_cookie) != 0);
 818         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
 819         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 820                                    wh_object_cookie) != 8);
 821         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
 822
 823         /* Checks for struct struct lnet_magicversion */
 824         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
 825         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
 826         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
 827         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
 828         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
 829         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
 830                                    version_minor) != 6);
 831         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
 832
 833         /* Checks for struct _lnet_hdr_nid4 */
 834         BUILD_BUG_ON((int)sizeof(struct _lnet_hdr_nid4) != 72);
 835         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, dest_nid) != 0);
 836         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->dest_nid) != 8);
 837         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, src_nid) != 8);
 838         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->src_nid) != 8);
 839         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, dest_pid) != 16);
 840         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->dest_pid) != 4);
 841         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, src_pid) != 20);
 842         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->src_pid) != 4);
 843         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, type) != 24);
 844         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->type) != 4);
 845         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, payload_length) != 28);
 846         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->payload_length) != 4);
 847         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg) != 32);
 848         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg) != 40);
 849
 850         /* Ack */
 851         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.dst_wmd) != 32);
 852         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.dst_wmd) != 16);
 853         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.match_bits) != 48);
 854         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.match_bits) != 8);
 855         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.mlength) != 56);
 856         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.mlength) != 4);
 857
 858         /* Put */
 859         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.ack_wmd) != 32);
 860         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.ack_wmd) != 16);
 861         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.match_bits) != 48);
 862         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.match_bits) != 8);
 863         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.hdr_data) != 56);
 864         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.hdr_data) != 8);
 865         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.ptl_index) != 64);
 866         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.ptl_index) != 4);
 867         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.offset) != 68);
 868         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.offset) != 4);
 869
 870         /* Get */
 871         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.return_wmd) != 32);
 872         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.return_wmd) != 16);
 873         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.match_bits) != 48);
 874         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.match_bits) != 8);
 875         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.ptl_index) != 56);
 876         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.ptl_index) != 4);
 877         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.src_offset) != 60);
 878         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.src_offset) != 4);
 879         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.sink_length) != 64);
 880         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.sink_length) != 4);
 881
 882         /* Reply */
 883         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.reply.dst_wmd) != 32);
 884         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.reply.dst_wmd) != 16);
 885
 886         /* Hello */
 887         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.hello.incarnation) != 32);
 888         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.hello.incarnation) != 8);
 889         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.hello.type) != 40);
 890         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.hello.type) != 4);
 891
 892         /* Checks for struct lnet_ni_status and related constants */
 893         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
 894         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
 895         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
 896
 897         /* Checks for struct lnet_ni_status */
 898         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
 899         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
 900         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
 901         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
 902         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
 903         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_unused) != 12);
 904         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) != 4);
 905
 906         /* Checks for struct lnet_ping_info and related constants */
 907         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
 908         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
 909         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
 910         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
 911         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
 912         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
 913         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
 914         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 31);
 915
 916         /* Checks for struct lnet_ping_info */
 917         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
 918         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
 919         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
 920         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
 921         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
 922         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
 923         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
 924         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
 925         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
 926         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
 927         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) != 0);
 928
 929         /* Acceptor connection request */
 930         BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 931
 932         /* Checks for struct lnet_acceptor_connreq */
 933         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq) != 16);
 934         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_magic) != 0);
 935         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_magic) != 4);
 936         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_version) != 4);
 937         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_version) != 4);
 938         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_nid) != 8);
 939         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_nid) != 8);
 940
 941         /* Checks for struct lnet_acceptor_connreq_v2 */
 942         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq_v2) != 28);
 943         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_magic) != 0);
 944         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_magic) != 4);
 945         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_version) != 4);
 946         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_version) != 4);
 947         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_nid) != 8);
 948         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_nid) != 20);
 949
 950         /* Checks for struct lnet_counters_common */
 951         BUILD_BUG_ON((int)sizeof(struct lnet_counters_common) != 60);
 952         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_alloc) != 0);
 953         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_alloc) != 4);
 954         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_max) != 4);
 955         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_max) != 4);
 956         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_errors) != 8);
 957         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_errors) != 4);
 958         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_count) != 12);
 959         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_count) != 4);
 960         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_count) != 16);
 961         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_count) != 4);
 962         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_count) != 20);
 963         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_count) != 4);
 964         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_count) != 24);
 965         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_count) != 4);
 966         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_length) != 28);
 967         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_length) != 8);
 968         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_length) != 36);
 969         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_length) != 8);
 970         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_length) != 44);
 971         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_length) != 8);
 972         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_length) != 52);
 973         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_length) != 8);
 974 }
 975
 976 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 977 {
 978         const struct lnet_lnd *lnd;
 979
 980         /* holding lnd mutex */
 981         if (type >= NUM_LNDS)
 982                 return NULL;
 983         lnd = the_lnet.ln_lnds[type];
 984         LASSERT(!lnd || lnd->lnd_type == type);
 985
 986         return lnd;
 987 }
 988
 989 unsigned int
 990 lnet_get_lnd_timeout(void)
 991 {
 992         return lnet_lnd_timeout;
 993 }
 994 EXPORT_SYMBOL(lnet_get_lnd_timeout);
 995
 996 void
 997 lnet_register_lnd(const struct lnet_lnd *lnd)
 998 {
 999         mutex_lock(&the_lnet.ln_lnd_mutex);
1000
1001         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
1002         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
1003
1004         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
1005
1006         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
1007
1008         mutex_unlock(&the_lnet.ln_lnd_mutex);
1009 }
1010 EXPORT_SYMBOL(lnet_register_lnd);
1011
1012 void
1013 lnet_unregister_lnd(const struct lnet_lnd *lnd)
1014 {
1015         mutex_lock(&the_lnet.ln_lnd_mutex);
1016
1017         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
1018
1019         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
1020         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
1021
1022         mutex_unlock(&the_lnet.ln_lnd_mutex);
1023 }
1024 EXPORT_SYMBOL(lnet_unregister_lnd);
1025
1026 static void
1027 lnet_counters_get_common_locked(struct lnet_counters_common *common)
1028 {
1029         struct lnet_counters *ctr;
1030         int i;
1031
1032         /* FIXME !!! Their is no assert_lnet_net_locked() to ensure this
1033          * actually called under the protection of the lnet_net_lock.
1034          */
1035         memset(common, 0, sizeof(*common));
1036
1037         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1038                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
1039                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
1040                 common->lcc_errors       += ctr->lct_common.lcc_errors;
1041                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
1042                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
1043                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
1044                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
1045                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
1046                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
1047                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
1048                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
1049         }
1050 }
1051
1052 void
1053 lnet_counters_get_common(struct lnet_counters_common *common)
1054 {
1055         lnet_net_lock(LNET_LOCK_EX);
1056         lnet_counters_get_common_locked(common);
1057         lnet_net_unlock(LNET_LOCK_EX);
1058 }
1059 EXPORT_SYMBOL(lnet_counters_get_common);
1060
1061 int
1062 lnet_counters_get(struct lnet_counters *counters)
1063 {
1064         struct lnet_counters *ctr;
1065         struct lnet_counters_health *health = &counters->lct_health;
1066         int i, rc = 0;
1067
1068         memset(counters, 0, sizeof(*counters));
1069
1070         lnet_net_lock(LNET_LOCK_EX);
1071
1072         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1073                 GOTO(out_unlock, rc = -ENODEV);
1074
1075         lnet_counters_get_common_locked(&counters->lct_common);
1076
1077         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1078                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
1079                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
1080                 health->lch_response_timeout_count +=
1081                                 ctr->lct_health.lch_response_timeout_count;
1082                 health->lch_local_interrupt_count +=
1083                                 ctr->lct_health.lch_local_interrupt_count;
1084                 health->lch_local_dropped_count +=
1085                                 ctr->lct_health.lch_local_dropped_count;
1086                 health->lch_local_aborted_count +=
1087                                 ctr->lct_health.lch_local_aborted_count;
1088                 health->lch_local_no_route_count +=
1089                                 ctr->lct_health.lch_local_no_route_count;
1090                 health->lch_local_timeout_count +=
1091                                 ctr->lct_health.lch_local_timeout_count;
1092                 health->lch_local_error_count +=
1093                                 ctr->lct_health.lch_local_error_count;
1094                 health->lch_remote_dropped_count +=
1095                                 ctr->lct_health.lch_remote_dropped_count;
1096                 health->lch_remote_error_count +=
1097                                 ctr->lct_health.lch_remote_error_count;
1098                 health->lch_remote_timeout_count +=
1099                                 ctr->lct_health.lch_remote_timeout_count;
1100                 health->lch_network_timeout_count +=
1101                                 ctr->lct_health.lch_network_timeout_count;
1102         }
1103 out_unlock:
1104         lnet_net_unlock(LNET_LOCK_EX);
1105         return rc;
1106 }
1107 EXPORT_SYMBOL(lnet_counters_get);
1108
1109 void
1110 lnet_counters_reset(void)
1111 {
1112         struct lnet_counters *counters;
1113         int             i;
1114
1115         lnet_net_lock(LNET_LOCK_EX);
1116
1117         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1118                 goto avoid_reset;
1119
1120         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
1121                 memset(counters, 0, sizeof(struct lnet_counters));
1122 avoid_reset:
1123         lnet_net_unlock(LNET_LOCK_EX);
1124 }
1125
1126 static char *
1127 lnet_res_type2str(int type)
1128 {
1129         switch (type) {
1130         default:
1131                 LBUG();
1132         case LNET_COOKIE_TYPE_MD:
1133                 return "MD";
1134         case LNET_COOKIE_TYPE_ME:
1135                 return "ME";
1136         case LNET_COOKIE_TYPE_EQ:
1137                 return "EQ";
1138         }
1139 }
1140
1141 static void
1142 lnet_res_container_cleanup(struct lnet_res_container *rec)
1143 {
1144         int     count = 0;
1145
1146         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
1147                 return;
1148
1149         while (!list_empty(&rec->rec_active)) {
1150                 struct list_head *e = rec->rec_active.next;
1151
1152                 list_del_init(e);
1153                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
1154                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
1155
1156                 } else { /* NB: Active MEs should be attached on portals */
1157                         LBUG();
1158                 }
1159                 count++;
1160         }
1161
1162         if (count > 0) {
1163                 /* Found alive MD/ME/EQ, user really should unlink/free
1164                  * all of them before finalize LNet, but if someone didn't,
1165                  * we have to recycle garbage for him */
1166                 CERROR("%d active elements on exit of %s container\n",
1167                        count, lnet_res_type2str(rec->rec_type));
1168         }
1169
1170         if (rec->rec_lh_hash != NULL) {
1171                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
1172                 rec->rec_lh_hash = NULL;
1173         }
1174
1175         rec->rec_type = 0; /* mark it as finalized */
1176 }
1177
1178 static int
1179 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1180 {
1181         int     rc = 0;
1182         int     i;
1183
1184         LASSERT(rec->rec_type == 0);
1185
1186         rec->rec_type = type;
1187         INIT_LIST_HEAD(&rec->rec_active);
1188
1189         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1190
1191         /* Arbitrary choice of hash table size */
1192         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1193                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1194         if (rec->rec_lh_hash == NULL) {
1195                 rc = -ENOMEM;
1196                 goto out;
1197         }
1198
1199         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1200                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1201
1202         return 0;
1203
1204 out:
1205         CERROR("Failed to setup %s resource container\n",
1206                lnet_res_type2str(type));
1207         lnet_res_container_cleanup(rec);
1208         return rc;
1209 }
1210
1211 static void
1212 lnet_res_containers_destroy(struct lnet_res_container **recs)
1213 {
1214         struct lnet_res_container       *rec;
1215         int                             i;
1216
1217         cfs_percpt_for_each(rec, i, recs)
1218                 lnet_res_container_cleanup(rec);
1219
1220         cfs_percpt_free(recs);
1221 }
1222
1223 static struct lnet_res_container **
1224 lnet_res_containers_create(int type)
1225 {
1226         struct lnet_res_container       **recs;
1227         struct lnet_res_container       *rec;
1228         int                             rc;
1229         int                             i;
1230
1231         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1232         if (recs == NULL) {
1233                 CERROR("Failed to allocate %s resource containers\n",
1234                        lnet_res_type2str(type));
1235                 return NULL;
1236         }
1237
1238         cfs_percpt_for_each(rec, i, recs) {
1239                 rc = lnet_res_container_setup(rec, i, type);
1240                 if (rc != 0) {
1241                         lnet_res_containers_destroy(recs);
1242                         return NULL;
1243                 }
1244         }
1245
1246         return recs;
1247 }
1248
1249 struct lnet_libhandle *
1250 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1251 {
1252         /* ALWAYS called with lnet_res_lock held */
1253         struct list_head        *head;
1254         struct lnet_libhandle   *lh;
1255         unsigned int            hash;
1256
1257         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1258                 return NULL;
1259
1260         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1261         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1262
1263         list_for_each_entry(lh, head, lh_hash_chain) {
1264                 if (lh->lh_cookie == cookie)
1265                         return lh;
1266         }
1267
1268         return NULL;
1269 }
1270
1271 void
1272 lnet_res_lh_initialize(struct lnet_res_container *rec,
1273                        struct lnet_libhandle *lh)
1274 {
1275         /* ALWAYS called with lnet_res_lock held */
1276         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1277         unsigned int    hash;
1278
1279         lh->lh_cookie = rec->rec_lh_cookie;
1280         rec->rec_lh_cookie += 1 << ibits;
1281
1282         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1283
1284         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1285 }
1286
1287 struct list_head **
1288 lnet_create_array_of_queues(void)
1289 {
1290         struct list_head **qs;
1291         struct list_head *q;
1292         int i;
1293
1294         qs = cfs_percpt_alloc(lnet_cpt_table(),
1295                               sizeof(struct list_head));
1296         if (!qs) {
1297                 CERROR("Failed to allocate queues\n");
1298                 return NULL;
1299         }
1300
1301         cfs_percpt_for_each(q, i, qs)
1302                 INIT_LIST_HEAD(q);
1303
1304         return qs;
1305 }
1306
1307 static int lnet_unprepare(void);
1308
1309 static int
1310 lnet_prepare(lnet_pid_t requested_pid)
1311 {
1312         /* Prepare to bring up the network */
1313         struct lnet_res_container **recs;
1314         int                       rc = 0;
1315
1316         if (requested_pid == LNET_PID_ANY) {
1317                 /* Don't instantiate LNET just for me */
1318                 return -ENETDOWN;
1319         }
1320
1321         LASSERT(the_lnet.ln_refcount == 0);
1322
1323         the_lnet.ln_routing = 0;
1324
1325         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1326         the_lnet.ln_pid = requested_pid;
1327
1328         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1329         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1330         INIT_LIST_HEAD(&the_lnet.ln_nets);
1331         INIT_LIST_HEAD(&the_lnet.ln_routers);
1332         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1333         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1334         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1335         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1336         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1337         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1338         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1339         INIT_LIST_HEAD(&the_lnet.ln_udsp_list);
1340         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1341         the_lnet.ln_mt_handler = NULL;
1342         init_completion(&the_lnet.ln_started);
1343
1344         rc = lnet_slab_setup();
1345         if (rc != 0)
1346                 goto failed;
1347
1348         rc = lnet_create_remote_nets_table();
1349         if (rc != 0)
1350                 goto failed;
1351
1352         /*
1353          * NB the interface cookie in wire handles guards against delayed
1354          * replies and ACKs appearing valid after reboot.
1355          */
1356         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1357
1358         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1359                                                 sizeof(struct lnet_counters));
1360         if (the_lnet.ln_counters == NULL) {
1361                 CERROR("Failed to allocate counters for LNet\n");
1362                 rc = -ENOMEM;
1363                 goto failed;
1364         }
1365
1366         rc = lnet_peer_tables_create();
1367         if (rc != 0)
1368                 goto failed;
1369
1370         rc = lnet_msg_containers_create();
1371         if (rc != 0)
1372                 goto failed;
1373
1374         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1375                                       LNET_COOKIE_TYPE_EQ);
1376         if (rc != 0)
1377                 goto failed;
1378
1379         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1380         if (recs == NULL) {
1381                 rc = -ENOMEM;
1382                 goto failed;
1383         }
1384
1385         the_lnet.ln_md_containers = recs;
1386
1387         rc = lnet_portals_create();
1388         if (rc != 0) {
1389                 CERROR("Failed to create portals for LNet: %d\n", rc);
1390                 goto failed;
1391         }
1392
1393         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1394         if (!the_lnet.ln_mt_zombie_rstqs) {
1395                 rc = -ENOMEM;
1396                 goto failed;
1397         }
1398
1399         return 0;
1400
1401  failed:
1402         lnet_unprepare();
1403         return rc;
1404 }
1405
1406 static int
1407 lnet_unprepare(void)
1408 {
1409         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1410          * have shut down already, so it is safe to unlink and free all
1411          * descriptors, even those that appear committed to a network op (eg MD
1412          * with non-zero pending count) */
1413
1414         lnet_fail_nid(LNET_NID_ANY, 0);
1415
1416         LASSERT(the_lnet.ln_refcount == 0);
1417         LASSERT(list_empty(&the_lnet.ln_test_peers));
1418         LASSERT(list_empty(&the_lnet.ln_nets));
1419
1420         if (the_lnet.ln_mt_zombie_rstqs) {
1421                 lnet_clean_zombie_rstqs();
1422                 the_lnet.ln_mt_zombie_rstqs = NULL;
1423         }
1424
1425         lnet_assert_handler_unused(the_lnet.ln_mt_handler);
1426         the_lnet.ln_mt_handler = NULL;
1427
1428         lnet_portals_destroy();
1429
1430         if (the_lnet.ln_md_containers != NULL) {
1431                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1432                 the_lnet.ln_md_containers = NULL;
1433         }
1434
1435         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1436
1437         lnet_msg_containers_destroy();
1438         lnet_peer_uninit();
1439         lnet_rtrpools_free(0);
1440
1441         if (the_lnet.ln_counters != NULL) {
1442                 cfs_percpt_free(the_lnet.ln_counters);
1443                 the_lnet.ln_counters = NULL;
1444         }
1445         lnet_destroy_remote_nets_table();
1446         lnet_udsp_destroy(true);
1447         lnet_slab_cleanup();
1448
1449         return 0;
1450 }
1451
1452 struct lnet_ni  *
1453 lnet_net2ni_locked(__u32 net_id, int cpt)
1454 {
1455         struct lnet_ni   *ni;
1456         struct lnet_net  *net;
1457
1458         LASSERT(cpt != LNET_LOCK_EX);
1459
1460         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1461                 if (net->net_id == net_id) {
1462                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1463                                         ni_netlist);
1464                         return ni;
1465                 }
1466         }
1467
1468         return NULL;
1469 }
1470
1471 struct lnet_ni *
1472 lnet_net2ni_addref(__u32 net)
1473 {
1474         struct lnet_ni *ni;
1475
1476         lnet_net_lock(0);
1477         ni = lnet_net2ni_locked(net, 0);
1478         if (ni)
1479                 lnet_ni_addref_locked(ni, 0);
1480         lnet_net_unlock(0);
1481
1482         return ni;
1483 }
1484 EXPORT_SYMBOL(lnet_net2ni_addref);
1485
1486 struct lnet_net *
1487 lnet_get_net_locked(__u32 net_id)
1488 {
1489         struct lnet_net  *net;
1490
1491         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1492                 if (net->net_id == net_id)
1493                         return net;
1494         }
1495
1496         return NULL;
1497 }
1498
1499 void
1500 lnet_net_clr_pref_rtrs(struct lnet_net *net)
1501 {
1502         struct list_head zombies;
1503         struct lnet_nid_list *ne;
1504         struct lnet_nid_list *tmp;
1505
1506         INIT_LIST_HEAD(&zombies);
1507
1508         lnet_net_lock(LNET_LOCK_EX);
1509         list_splice_init(&net->net_rtr_pref_nids, &zombies);
1510         lnet_net_unlock(LNET_LOCK_EX);
1511
1512         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1513                 list_del_init(&ne->nl_list);
1514                 LIBCFS_FREE(ne, sizeof(*ne));
1515         }
1516 }
1517
1518 int
1519 lnet_net_add_pref_rtr(struct lnet_net *net,
1520                       struct lnet_nid *gw_nid)
1521 __must_hold(&the_lnet.ln_api_mutex)
1522 {
1523         struct lnet_nid_list *ne;
1524
1525         /* This function is called with api_mutex held. When the api_mutex
1526          * is held the list can not be modified, as it is only modified as
1527          * a result of applying a UDSP and that happens under api_mutex
1528          * lock.
1529          */
1530         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1531                 if (nid_same(&ne->nl_nid, gw_nid))
1532                         return -EEXIST;
1533         }
1534
1535         LIBCFS_ALLOC(ne, sizeof(*ne));
1536         if (!ne)
1537                 return -ENOMEM;
1538
1539         ne->nl_nid = *gw_nid;
1540
1541         /* Lock the cpt to protect against addition and checks in the
1542          * selection algorithm
1543          */
1544         lnet_net_lock(LNET_LOCK_EX);
1545         list_add(&ne->nl_list, &net->net_rtr_pref_nids);
1546         lnet_net_unlock(LNET_LOCK_EX);
1547
1548         return 0;
1549 }
1550
1551 bool
1552 lnet_net_is_pref_rtr_locked(struct lnet_net *net, struct lnet_nid *rtr_nid)
1553 {
1554         struct lnet_nid_list *ne;
1555
1556         CDEBUG(D_NET, "%s: rtr pref empty: %d\n",
1557                libcfs_net2str(net->net_id),
1558                list_empty(&net->net_rtr_pref_nids));
1559
1560         if (list_empty(&net->net_rtr_pref_nids))
1561                 return false;
1562
1563         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1564                 CDEBUG(D_NET, "Comparing pref %s with gw %s\n",
1565                        libcfs_nidstr(&ne->nl_nid),
1566                        libcfs_nidstr(rtr_nid));
1567                 if (nid_same(rtr_nid, &ne->nl_nid))
1568                         return true;
1569         }
1570
1571         return false;
1572 }
1573
1574 static unsigned int
1575 lnet_nid4_cpt_hash(lnet_nid_t nid, unsigned int number)
1576 {
1577         __u64 key = nid;
1578         __u64 pair_bits = 0x0001000100010001LLU;
1579         __u64 mask = pair_bits * 0xFF;
1580         __u64 pair_sum;
1581
1582         /* Use (sum-by-multiplication of nid bytes) mod (number of CPTs)
1583          * to match nid to a CPT.
1584          */
1585         pair_sum = (key & mask) + ((key >> 8) & mask);
1586         pair_sum = (pair_sum * pair_bits) >> 48;
1587
1588         CDEBUG(D_NET, "Match nid %s to cpt %u\n",
1589                libcfs_nid2str(nid), (unsigned int)(pair_sum) % number);
1590
1591         return (unsigned int)(pair_sum) % number;
1592 }
1593
1594 unsigned int
1595 lnet_nid_cpt_hash(struct lnet_nid *nid, unsigned int number)
1596 {
1597         unsigned int val;
1598         u32 h = 0;
1599         int i;
1600
1601         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1602
1603         if (number == 1)
1604                 return 0;
1605
1606         if (nid_is_nid4(nid))
1607                 return lnet_nid4_cpt_hash(lnet_nid_to_nid4(nid), number);
1608
1609         for (i = 0; i < 4; i++)
1610                 h = hash_32(nid->nid_addr[i]^h, 32);
1611         val = hash_32(LNET_NID_NET(nid) ^ h, LNET_CPT_BITS);
1612         if (val < number)
1613                 return val;
1614         return (unsigned int)(h + val + (val >> 1)) % number;
1615 }
1616
1617 int
1618 lnet_cpt_of_nid_locked(struct lnet_nid *nid, struct lnet_ni *ni)
1619 {
1620         struct lnet_net *net;
1621
1622         /* must called with hold of lnet_net_lock */
1623         if (LNET_CPT_NUMBER == 1)
1624                 return 0; /* the only one */
1625
1626         /*
1627          * If NI is provided then use the CPT identified in the NI cpt
1628          * list if one exists. If one doesn't exist, then that NI is
1629          * associated with all CPTs and it follows that the net it belongs
1630          * to is implicitly associated with all CPTs, so just hash the nid
1631          * and return that.
1632          */
1633         if (ni != NULL) {
1634                 if (ni->ni_cpts != NULL)
1635                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1636                                                              ni->ni_ncpts)];
1637                 else
1638                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1639         }
1640
1641         /* no NI provided so look at the net */
1642         net = lnet_get_net_locked(LNET_NID_NET(nid));
1643
1644         if (net != NULL && net->net_cpts != NULL) {
1645                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1646         }
1647
1648         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1649 }
1650
1651 int
1652 lnet_nid2cpt(struct lnet_nid *nid, struct lnet_ni *ni)
1653 {
1654         int     cpt;
1655         int     cpt2;
1656
1657         if (LNET_CPT_NUMBER == 1)
1658                 return 0; /* the only one */
1659
1660         cpt = lnet_net_lock_current();
1661
1662         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1663
1664         lnet_net_unlock(cpt);
1665
1666         return cpt2;
1667 }
1668 EXPORT_SYMBOL(lnet_nid2cpt);
1669
1670 int
1671 lnet_cpt_of_nid(lnet_nid_t nid4, struct lnet_ni *ni)
1672 {
1673         struct lnet_nid nid;
1674
1675         if (LNET_CPT_NUMBER == 1)
1676                 return 0; /* the only one */
1677
1678         lnet_nid4_to_nid(nid4, &nid);
1679         return lnet_nid2cpt(&nid, ni);
1680 }
1681 EXPORT_SYMBOL(lnet_cpt_of_nid);
1682
1683 int
1684 lnet_islocalnet_locked(__u32 net_id)
1685 {
1686         struct lnet_net *net;
1687         bool local;
1688
1689         net = lnet_get_net_locked(net_id);
1690
1691         local = net != NULL;
1692
1693         return local;
1694 }
1695
1696 int
1697 lnet_islocalnet(__u32 net_id)
1698 {
1699         int cpt;
1700         bool local;
1701
1702         cpt = lnet_net_lock_current();
1703
1704         local = lnet_islocalnet_locked(net_id);
1705
1706         lnet_net_unlock(cpt);
1707
1708         return local;
1709 }
1710
1711 struct lnet_ni  *
1712 lnet_nid_to_ni_locked(struct lnet_nid *nid, int cpt)
1713 {
1714         struct lnet_net  *net;
1715         struct lnet_ni *ni;
1716
1717         LASSERT(cpt != LNET_LOCK_EX);
1718
1719         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1720                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1721                         if (nid_same(&ni->ni_nid, nid))
1722                                 return ni;
1723                 }
1724         }
1725
1726         return NULL;
1727 }
1728
1729 struct lnet_ni  *
1730 lnet_nid2ni_locked(lnet_nid_t nid4, int cpt)
1731 {
1732         struct lnet_nid nid;
1733
1734         lnet_nid4_to_nid(nid4, &nid);
1735         return lnet_nid_to_ni_locked(&nid, cpt);
1736 }
1737
1738 struct lnet_ni *
1739 lnet_nid2ni_addref(lnet_nid_t nid4)
1740 {
1741         struct lnet_ni *ni;
1742         struct lnet_nid nid;
1743
1744         lnet_nid4_to_nid(nid4, &nid);
1745
1746         lnet_net_lock(0);
1747         ni = lnet_nid_to_ni_locked(&nid, 0);
1748         if (ni)
1749                 lnet_ni_addref_locked(ni, 0);
1750         lnet_net_unlock(0);
1751
1752         return ni;
1753 }
1754 EXPORT_SYMBOL(lnet_nid2ni_addref);
1755
1756 struct lnet_ni *
1757 lnet_nid_to_ni_addref(struct lnet_nid *nid)
1758 {
1759         struct lnet_ni *ni;
1760
1761         lnet_net_lock(0);
1762         ni = lnet_nid_to_ni_locked(nid, 0);
1763         if (ni)
1764                 lnet_ni_addref_locked(ni, 0);
1765         lnet_net_unlock(0);
1766
1767         return ni;
1768 }
1769 EXPORT_SYMBOL(lnet_nid_to_ni_addref);
1770
1771 int
1772 lnet_islocalnid(struct lnet_nid *nid)
1773 {
1774         struct lnet_ni  *ni;
1775         int             cpt;
1776
1777         cpt = lnet_net_lock_current();
1778         ni = lnet_nid_to_ni_locked(nid, cpt);
1779         lnet_net_unlock(cpt);
1780
1781         return ni != NULL;
1782 }
1783
1784 int
1785 lnet_count_acceptor_nets(void)
1786 {
1787         /* Return the # of NIs that need the acceptor. */
1788         int              count = 0;
1789         struct lnet_net  *net;
1790         int              cpt;
1791
1792         cpt = lnet_net_lock_current();
1793         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1794                 /* all socklnd type networks should have the acceptor
1795                  * thread started */
1796                 if (net->net_lnd->lnd_accept != NULL)
1797                         count++;
1798         }
1799
1800         lnet_net_unlock(cpt);
1801
1802         return count;
1803 }
1804
1805 struct lnet_ping_buffer *
1806 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1807 {
1808         struct lnet_ping_buffer *pbuf;
1809
1810         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1811         if (pbuf) {
1812                 pbuf->pb_nnis = nnis;
1813                 pbuf->pb_needs_post = false;
1814                 atomic_set(&pbuf->pb_refcnt, 1);
1815         }
1816
1817         return pbuf;
1818 }
1819
1820 void
1821 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1822 {
1823         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1824         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1825 }
1826
1827 static struct lnet_ping_buffer *
1828 lnet_ping_target_create(int nnis)
1829 {
1830         struct lnet_ping_buffer *pbuf;
1831
1832         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1833         if (pbuf == NULL) {
1834                 CERROR("Can't allocate ping source [%d]\n", nnis);
1835                 return NULL;
1836         }
1837
1838         pbuf->pb_info.pi_nnis = nnis;
1839         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1840         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1841         pbuf->pb_info.pi_features =
1842                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1843
1844         return pbuf;
1845 }
1846
1847 static inline int
1848 lnet_get_net_ni_count_locked(struct lnet_net *net)
1849 {
1850         struct lnet_ni  *ni;
1851         int             count = 0;
1852
1853         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1854                 count++;
1855
1856         return count;
1857 }
1858
1859 static inline int
1860 lnet_get_net_ni_count_pre(struct lnet_net *net)
1861 {
1862         struct lnet_ni  *ni;
1863         int             count = 0;
1864
1865         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1866                 count++;
1867
1868         return count;
1869 }
1870
1871 static inline int
1872 lnet_get_ni_count(void)
1873 {
1874         struct lnet_ni  *ni;
1875         struct lnet_net *net;
1876         int             count = 0;
1877
1878         lnet_net_lock(0);
1879
1880         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1881                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1882                         count++;
1883         }
1884
1885         lnet_net_unlock(0);
1886
1887         return count;
1888 }
1889
1890 void
1891 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1892 {
1893         struct lnet_ni_status *stat;
1894         int nnis;
1895         int i;
1896
1897         __swab32s(&pbuf->pb_info.pi_magic);
1898         __swab32s(&pbuf->pb_info.pi_features);
1899         __swab32s(&pbuf->pb_info.pi_pid);
1900         __swab32s(&pbuf->pb_info.pi_nnis);
1901         nnis = pbuf->pb_info.pi_nnis;
1902         if (nnis > pbuf->pb_nnis)
1903                 nnis = pbuf->pb_nnis;
1904         for (i = 0; i < nnis; i++) {
1905                 stat = &pbuf->pb_info.pi_ni[i];
1906                 __swab64s(&stat->ns_nid);
1907                 __swab32s(&stat->ns_status);
1908         }
1909 }
1910
1911 int
1912 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1913 {
1914         if (!pinfo)
1915                 return -EINVAL;
1916         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1917                 return -EPROTO;
1918         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1919                 return -EPROTO;
1920         /* Loopback is guaranteed to be present */
1921         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1922                 return -ERANGE;
1923         if (LNET_PING_INFO_LONI(pinfo) != LNET_NID_LO_0)
1924                 return -EPROTO;
1925         return 0;
1926 }
1927
1928 static void
1929 lnet_ping_target_destroy(void)
1930 {
1931         struct lnet_net *net;
1932         struct lnet_ni  *ni;
1933
1934         lnet_net_lock(LNET_LOCK_EX);
1935
1936         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1937                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1938                         lnet_ni_lock(ni);
1939                         ni->ni_status = NULL;
1940                         lnet_ni_unlock(ni);
1941                 }
1942         }
1943
1944         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1945         the_lnet.ln_ping_target = NULL;
1946
1947         lnet_net_unlock(LNET_LOCK_EX);
1948 }
1949
1950 static void
1951 lnet_ping_target_event_handler(struct lnet_event *event)
1952 {
1953         struct lnet_ping_buffer *pbuf = event->md_user_ptr;
1954
1955         if (event->unlinked)
1956                 lnet_ping_buffer_decref(pbuf);
1957 }
1958
1959 static int
1960 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1961                        struct lnet_handle_md *ping_mdh,
1962                        int ni_count, bool set_eq)
1963 {
1964         struct lnet_processid id = {
1965                 .nid = LNET_ANY_NID,
1966                 .pid = LNET_PID_ANY
1967         };
1968         struct lnet_me *me;
1969         struct lnet_md md = { NULL };
1970         int rc;
1971
1972         if (set_eq)
1973                 the_lnet.ln_ping_target_handler =
1974                         lnet_ping_target_event_handler;
1975
1976         *ppbuf = lnet_ping_target_create(ni_count);
1977         if (*ppbuf == NULL) {
1978                 rc = -ENOMEM;
1979                 goto fail_free_eq;
1980         }
1981
1982         /* Ping target ME/MD */
1983         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
1984                           LNET_PROTO_PING_MATCHBITS, 0,
1985                           LNET_UNLINK, LNET_INS_AFTER);
1986         if (IS_ERR(me)) {
1987                 rc = PTR_ERR(me);
1988                 CERROR("Can't create ping target ME: %d\n", rc);
1989                 goto fail_decref_ping_buffer;
1990         }
1991
1992         /* initialize md content */
1993         md.start     = &(*ppbuf)->pb_info;
1994         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1995         md.threshold = LNET_MD_THRESH_INF;
1996         md.max_size  = 0;
1997         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1998                        LNET_MD_MANAGE_REMOTE;
1999         md.handler   = the_lnet.ln_ping_target_handler;
2000         md.user_ptr  = *ppbuf;
2001
2002         rc = LNetMDAttach(me, &md, LNET_RETAIN, ping_mdh);
2003         if (rc != 0) {
2004                 CERROR("Can't attach ping target MD: %d\n", rc);
2005                 goto fail_decref_ping_buffer;
2006         }
2007         lnet_ping_buffer_addref(*ppbuf);
2008
2009         return 0;
2010
2011 fail_decref_ping_buffer:
2012         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
2013         lnet_ping_buffer_decref(*ppbuf);
2014         *ppbuf = NULL;
2015 fail_free_eq:
2016         return rc;
2017 }
2018
2019 static void
2020 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
2021                     struct lnet_handle_md *ping_mdh)
2022 {
2023         LNetMDUnlink(*ping_mdh);
2024         LNetInvalidateMDHandle(ping_mdh);
2025
2026         /* NB the MD could be busy; this just starts the unlink */
2027         wait_var_event_warning(&pbuf->pb_refcnt,
2028                                atomic_read(&pbuf->pb_refcnt) <= 1,
2029                                "Still waiting for ping data MD to unlink\n");
2030 }
2031
2032 static void
2033 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
2034 {
2035         struct lnet_ni          *ni;
2036         struct lnet_net         *net;
2037         struct lnet_ni_status *ns;
2038         int                     i;
2039         int                     rc;
2040
2041         i = 0;
2042         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2043                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2044                         LASSERT(i < pbuf->pb_nnis);
2045
2046                         ns = &pbuf->pb_info.pi_ni[i];
2047
2048                         if (!nid_is_nid4(&ni->ni_nid))
2049                                 continue;
2050                         ns->ns_nid = lnet_nid_to_nid4(&ni->ni_nid);
2051
2052                         lnet_ni_lock(ni);
2053                         ns->ns_status = lnet_ni_get_status_locked(ni);
2054                         ni->ni_status = ns;
2055                         lnet_ni_unlock(ni);
2056
2057                         i++;
2058                 }
2059         }
2060         /*
2061          * We (ab)use the ns_status of the loopback interface to
2062          * transmit the sequence number. The first interface listed
2063          * must be the loopback interface.
2064          */
2065         rc = lnet_ping_info_validate(&pbuf->pb_info);
2066         if (rc) {
2067                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
2068                 LBUG();
2069         }
2070         LNET_PING_BUFFER_SEQNO(pbuf) =
2071                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
2072 }
2073
2074 static void
2075 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
2076                         struct lnet_handle_md ping_mdh)
2077 {
2078         struct lnet_ping_buffer *old_pbuf = NULL;
2079         struct lnet_handle_md old_ping_md;
2080
2081         /* switch the NIs to point to the new ping info created */
2082         lnet_net_lock(LNET_LOCK_EX);
2083
2084         if (!the_lnet.ln_routing)
2085                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
2086         if (!lnet_peer_discovery_disabled)
2087                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
2088
2089         /* Ensure only known feature bits have been set. */
2090         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
2091         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
2092
2093         lnet_ping_target_install_locked(pbuf);
2094
2095         if (the_lnet.ln_ping_target) {
2096                 old_pbuf = the_lnet.ln_ping_target;
2097                 old_ping_md = the_lnet.ln_ping_target_md;
2098         }
2099         the_lnet.ln_ping_target_md = ping_mdh;
2100         the_lnet.ln_ping_target = pbuf;
2101
2102         lnet_net_unlock(LNET_LOCK_EX);
2103
2104         if (old_pbuf) {
2105                 /* unlink and free the old ping info */
2106                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
2107                 lnet_ping_buffer_decref(old_pbuf);
2108         }
2109
2110         lnet_push_update_to_peers(0);
2111 }
2112
2113 static void
2114 lnet_ping_target_fini(void)
2115 {
2116         lnet_ping_md_unlink(the_lnet.ln_ping_target,
2117                             &the_lnet.ln_ping_target_md);
2118
2119         lnet_assert_handler_unused(the_lnet.ln_ping_target_handler);
2120         lnet_ping_target_destroy();
2121 }
2122
2123 /* Resize the push target. */
2124 int lnet_push_target_resize(void)
2125 {
2126         struct lnet_handle_md mdh;
2127         struct lnet_handle_md old_mdh;
2128         struct lnet_ping_buffer *pbuf;
2129         struct lnet_ping_buffer *old_pbuf;
2130         int nnis;
2131         int rc;
2132
2133 again:
2134         nnis = the_lnet.ln_push_target_nnis;
2135         if (nnis <= 0) {
2136                 CDEBUG(D_NET, "Invalid nnis %d\n", nnis);
2137                 return -EINVAL;
2138         }
2139
2140         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
2141          * dropped when we need to resize again (see "old_pbuf" below) or when
2142          * LNet is shutdown (see lnet_push_target_fini())
2143          */
2144         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
2145         if (!pbuf) {
2146                 CDEBUG(D_NET, "Can't allocate pbuf for nnis %d\n", nnis);
2147                 return -ENOMEM;
2148         }
2149
2150         rc = lnet_push_target_post(pbuf, &mdh);
2151         if (rc) {
2152                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
2153                 lnet_ping_buffer_decref(pbuf);
2154                 return rc;
2155         }
2156
2157         lnet_net_lock(LNET_LOCK_EX);
2158         old_pbuf = the_lnet.ln_push_target;
2159         old_mdh = the_lnet.ln_push_target_md;
2160         the_lnet.ln_push_target = pbuf;
2161         the_lnet.ln_push_target_md = mdh;
2162         lnet_net_unlock(LNET_LOCK_EX);
2163
2164         if (old_pbuf) {
2165                 LNetMDUnlink(old_mdh);
2166                 /* Drop ref set by lnet_ping_buffer_alloc() */
2167                 lnet_ping_buffer_decref(old_pbuf);
2168         }
2169
2170         /* Received another push or reply that requires a larger buffer */
2171         if (nnis < the_lnet.ln_push_target_nnis)
2172                 goto again;
2173
2174         CDEBUG(D_NET, "nnis %d success\n", nnis);
2175         return 0;
2176 }
2177
2178 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
2179                           struct lnet_handle_md *mdhp)
2180 {
2181         struct lnet_processid id = { LNET_ANY_NID, LNET_PID_ANY };
2182         struct lnet_md md = { NULL };
2183         struct lnet_me *me;
2184         int rc;
2185
2186         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
2187                           LNET_PROTO_PING_MATCHBITS, 0,
2188                           LNET_UNLINK, LNET_INS_AFTER);
2189         if (IS_ERR(me)) {
2190                 rc = PTR_ERR(me);
2191                 CERROR("Can't create push target ME: %d\n", rc);
2192                 return rc;
2193         }
2194
2195         pbuf->pb_needs_post = false;
2196
2197         /* This reference is dropped by lnet_push_target_event_handler() */
2198         lnet_ping_buffer_addref(pbuf);
2199
2200         /* initialize md content */
2201         md.start     = &pbuf->pb_info;
2202         md.length    = LNET_PING_INFO_SIZE(pbuf->pb_nnis);
2203         md.threshold = 1;
2204         md.max_size  = 0;
2205         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
2206         md.user_ptr  = pbuf;
2207         md.handler   = the_lnet.ln_push_target_handler;
2208
2209         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdhp);
2210         if (rc) {
2211                 CERROR("Can't attach push MD: %d\n", rc);
2212                 lnet_ping_buffer_decref(pbuf);
2213                 pbuf->pb_needs_post = true;
2214                 return rc;
2215         }
2216
2217         CDEBUG(D_NET, "posted push target %p\n", pbuf);
2218
2219         return 0;
2220 }
2221
2222 static void lnet_push_target_event_handler(struct lnet_event *ev)
2223 {
2224         struct lnet_ping_buffer *pbuf = ev->md_user_ptr;
2225
2226         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
2227                ev->unlinked);
2228
2229         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2230                 lnet_swap_pinginfo(pbuf);
2231
2232         if (ev->type == LNET_EVENT_UNLINK) {
2233                 /* Drop ref added by lnet_push_target_post() */
2234                 lnet_ping_buffer_decref(pbuf);
2235                 return;
2236         }
2237
2238         lnet_peer_push_event(ev);
2239         if (ev->unlinked)
2240                 /* Drop ref added by lnet_push_target_post */
2241                 lnet_ping_buffer_decref(pbuf);
2242 }
2243
2244 /* Initialize the push target. */
2245 static int lnet_push_target_init(void)
2246 {
2247         int rc;
2248
2249         if (the_lnet.ln_push_target)
2250                 return -EALREADY;
2251
2252         the_lnet.ln_push_target_handler =
2253                 lnet_push_target_event_handler;
2254
2255         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
2256         LASSERT(rc == 0);
2257
2258         /* Start at the required minimum, we'll enlarge if required. */
2259         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
2260
2261         rc = lnet_push_target_resize();
2262
2263         if (rc) {
2264                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2265                 the_lnet.ln_push_target_handler = NULL;
2266         }
2267
2268         return rc;
2269 }
2270
2271 /* Clean up the push target. */
2272 static void lnet_push_target_fini(void)
2273 {
2274         if (!the_lnet.ln_push_target)
2275                 return;
2276
2277         /* Unlink and invalidate to prevent new references. */
2278         LNetMDUnlink(the_lnet.ln_push_target_md);
2279         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2280
2281         /* Wait for the unlink to complete. */
2282         wait_var_event_warning(&the_lnet.ln_push_target->pb_refcnt,
2283                                atomic_read(&the_lnet.ln_push_target->pb_refcnt) <= 1,
2284                                "Still waiting for ping data MD to unlink\n");
2285
2286         /* Drop ref set by lnet_ping_buffer_alloc() */
2287         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2288         the_lnet.ln_push_target = NULL;
2289         the_lnet.ln_push_target_nnis = 0;
2290
2291         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2292         lnet_assert_handler_unused(the_lnet.ln_push_target_handler);
2293         the_lnet.ln_push_target_handler = NULL;
2294 }
2295
2296 static int
2297 lnet_ni_tq_credits(struct lnet_ni *ni)
2298 {
2299         int     credits;
2300
2301         LASSERT(ni->ni_ncpts >= 1);
2302
2303         if (ni->ni_ncpts == 1)
2304                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2305
2306         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2307         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2308         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2309
2310         return credits;
2311 }
2312
2313 static void
2314 lnet_ni_unlink_locked(struct lnet_ni *ni)
2315 {
2316         /* move it to zombie list and nobody can find it anymore */
2317         LASSERT(!list_empty(&ni->ni_netlist));
2318         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2319         lnet_ni_decref_locked(ni, 0);
2320 }
2321
2322 static void
2323 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2324 {
2325         int             i;
2326         int             islo;
2327         struct lnet_ni  *ni;
2328         struct list_head *zombie_list = &net->net_ni_zombie;
2329
2330         /*
2331          * Now wait for the NIs I just nuked to show up on the zombie
2332          * list and shut them down in guaranteed thread context
2333          */
2334         i = 2;
2335         while (!list_empty(zombie_list)) {
2336                 int     *ref;
2337                 int     j;
2338
2339                 ni = list_entry(zombie_list->next,
2340                                 struct lnet_ni, ni_netlist);
2341                 list_del_init(&ni->ni_netlist);
2342                 /* the ni should be in deleting state. If it's not it's
2343                  * a bug */
2344                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2345                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2346                         if (*ref == 0)
2347                                 continue;
2348                         /* still busy, add it back to zombie list */
2349                         list_add(&ni->ni_netlist, zombie_list);
2350                         break;
2351                 }
2352
2353                 if (!list_empty(&ni->ni_netlist)) {
2354                         /* Unlock mutex while waiting to allow other
2355                          * threads to read the LNet state and fall through
2356                          * to avoid deadlock
2357                          */
2358                         lnet_net_unlock(LNET_LOCK_EX);
2359                         mutex_unlock(&the_lnet.ln_api_mutex);
2360
2361                         ++i;
2362                         if ((i & (-i)) == i) {
2363                                 CDEBUG(D_WARNING,
2364                                        "Waiting for zombie LNI %s\n",
2365                                        libcfs_nidstr(&ni->ni_nid));
2366                         }
2367                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2368
2369                         mutex_lock(&the_lnet.ln_api_mutex);
2370                         lnet_net_lock(LNET_LOCK_EX);
2371                         continue;
2372                 }
2373
2374                 lnet_net_unlock(LNET_LOCK_EX);
2375
2376                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2377
2378                 LASSERT(!in_interrupt());
2379                 /* Holding the LND mutex makes it safe for lnd_shutdown
2380                  * to call module_put(). Module unload cannot finish
2381                  * until lnet_unregister_lnd() completes, and that
2382                  * requires the LND mutex.
2383                  */
2384                 mutex_unlock(&the_lnet.ln_api_mutex);
2385                 mutex_lock(&the_lnet.ln_lnd_mutex);
2386                 (net->net_lnd->lnd_shutdown)(ni);
2387                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2388                 mutex_lock(&the_lnet.ln_api_mutex);
2389
2390                 if (!islo)
2391                         CDEBUG(D_LNI, "Removed LNI %s\n",
2392                               libcfs_nidstr(&ni->ni_nid));
2393
2394                 lnet_ni_free(ni);
2395                 i = 2;
2396                 lnet_net_lock(LNET_LOCK_EX);
2397         }
2398 }
2399
2400 /* shutdown down the NI and release refcount */
2401 static void
2402 lnet_shutdown_lndni(struct lnet_ni *ni)
2403 {
2404         int i;
2405         struct lnet_net *net = ni->ni_net;
2406
2407         lnet_net_lock(LNET_LOCK_EX);
2408         lnet_ni_lock(ni);
2409         ni->ni_state = LNET_NI_STATE_DELETING;
2410         lnet_ni_unlock(ni);
2411         lnet_ni_unlink_locked(ni);
2412         lnet_incr_dlc_seq();
2413         lnet_net_unlock(LNET_LOCK_EX);
2414
2415         /* clear messages for this NI on the lazy portal */
2416         for (i = 0; i < the_lnet.ln_nportals; i++)
2417                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2418
2419         lnet_net_lock(LNET_LOCK_EX);
2420         lnet_clear_zombies_nis_locked(net);
2421         lnet_net_unlock(LNET_LOCK_EX);
2422 }
2423
2424 static void
2425 lnet_shutdown_lndnet(struct lnet_net *net)
2426 {
2427         struct lnet_ni *ni;
2428
2429         lnet_net_lock(LNET_LOCK_EX);
2430
2431         list_del_init(&net->net_list);
2432
2433         while (!list_empty(&net->net_ni_list)) {
2434                 ni = list_entry(net->net_ni_list.next,
2435                                 struct lnet_ni, ni_netlist);
2436                 lnet_net_unlock(LNET_LOCK_EX);
2437                 lnet_shutdown_lndni(ni);
2438                 lnet_net_lock(LNET_LOCK_EX);
2439         }
2440
2441         lnet_net_unlock(LNET_LOCK_EX);
2442
2443         /* Do peer table cleanup for this net */
2444         lnet_peer_tables_cleanup(net);
2445
2446         lnet_net_free(net);
2447 }
2448
2449 static void
2450 lnet_shutdown_lndnets(void)
2451 {
2452         struct lnet_net *net;
2453         LIST_HEAD(resend);
2454         struct lnet_msg *msg, *tmp;
2455
2456         /* NB called holding the global mutex */
2457
2458         /* All quiet on the API front */
2459         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING ||
2460                 the_lnet.ln_state == LNET_STATE_STOPPING);
2461         LASSERT(the_lnet.ln_refcount == 0);
2462
2463         lnet_net_lock(LNET_LOCK_EX);
2464         the_lnet.ln_state = LNET_STATE_STOPPING;
2465
2466         /*
2467          * move the nets to the zombie list to avoid them being
2468          * picked up for new work. LONET is also included in the
2469          * Nets that will be moved to the zombie list
2470          */
2471         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2472
2473         /* Drop the cached loopback Net. */
2474         if (the_lnet.ln_loni != NULL) {
2475                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2476                 the_lnet.ln_loni = NULL;
2477         }
2478         lnet_net_unlock(LNET_LOCK_EX);
2479
2480         /* iterate through the net zombie list and delete each net */
2481         while (!list_empty(&the_lnet.ln_net_zombie)) {
2482                 net = list_entry(the_lnet.ln_net_zombie.next,
2483                                  struct lnet_net, net_list);
2484                 lnet_shutdown_lndnet(net);
2485         }
2486
2487         spin_lock(&the_lnet.ln_msg_resend_lock);
2488         list_splice(&the_lnet.ln_msg_resend, &resend);
2489         spin_unlock(&the_lnet.ln_msg_resend_lock);
2490
2491         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2492                 list_del_init(&msg->msg_list);
2493                 msg->msg_no_resend = true;
2494                 lnet_finalize(msg, -ECANCELED);
2495         }
2496
2497         lnet_net_lock(LNET_LOCK_EX);
2498         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2499         lnet_net_unlock(LNET_LOCK_EX);
2500 }
2501
2502 static int
2503 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2504 {
2505         int                     rc = -EINVAL;
2506         struct lnet_tx_queue    *tq;
2507         int                     i;
2508         struct lnet_net         *net = ni->ni_net;
2509
2510         mutex_lock(&the_lnet.ln_lnd_mutex);
2511
2512         if (tun) {
2513                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2514                 ni->ni_lnd_tunables_set = true;
2515         }
2516
2517         rc = (net->net_lnd->lnd_startup)(ni);
2518
2519         mutex_unlock(&the_lnet.ln_lnd_mutex);
2520
2521         if (rc != 0) {
2522                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2523                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2524                 goto failed0;
2525         }
2526
2527         lnet_ni_lock(ni);
2528         ni->ni_state = LNET_NI_STATE_ACTIVE;
2529         lnet_ni_unlock(ni);
2530
2531         /* We keep a reference on the loopback net through the loopback NI */
2532         if (net->net_lnd->lnd_type == LOLND) {
2533                 lnet_ni_addref(ni);
2534                 LASSERT(the_lnet.ln_loni == NULL);
2535                 the_lnet.ln_loni = ni;
2536                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2537                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2538                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2539                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2540                 return 0;
2541         }
2542
2543         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2544             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2545                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2546                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2547                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2548                                         "" : "per-peer ");
2549                 /* shutdown the NI since if we get here then it must've already
2550                  * been started
2551                  */
2552                 lnet_shutdown_lndni(ni);
2553                 return -EINVAL;
2554         }
2555
2556         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2557                 tq->tq_credits_min =
2558                 tq->tq_credits_max =
2559                 tq->tq_credits = lnet_ni_tq_credits(ni);
2560         }
2561
2562         atomic_set(&ni->ni_tx_credits,
2563                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2564         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2565
2566         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2567                 libcfs_nidstr(&ni->ni_nid),
2568                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2569                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2570                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2571                 ni->ni_net->net_tunables.lct_peer_timeout);
2572
2573         return 0;
2574 failed0:
2575         lnet_ni_free(ni);
2576         return rc;
2577 }
2578
2579 static int
2580 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2581 {
2582         struct lnet_ni *ni;
2583         struct lnet_net *net_l = NULL;
2584         LIST_HEAD(local_ni_list);
2585         int rc;
2586         int ni_count = 0;
2587         __u32 lnd_type;
2588         const struct lnet_lnd  *lnd;
2589         int peer_timeout =
2590                 net->net_tunables.lct_peer_timeout;
2591         int maxtxcredits =
2592                 net->net_tunables.lct_max_tx_credits;
2593         int peerrtrcredits =
2594                 net->net_tunables.lct_peer_rtr_credits;
2595
2596         /*
2597          * make sure that this net is unique. If it isn't then
2598          * we are adding interfaces to an already existing network, and
2599          * 'net' is just a convenient way to pass in the list.
2600          * if it is unique we need to find the LND and load it if
2601          * necessary.
2602          */
2603         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2604                 lnd_type = LNET_NETTYP(net->net_id);
2605
2606                 mutex_lock(&the_lnet.ln_lnd_mutex);
2607                 lnd = lnet_find_lnd_by_type(lnd_type);
2608
2609                 if (lnd == NULL) {
2610                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2611                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2612                         mutex_lock(&the_lnet.ln_lnd_mutex);
2613
2614                         lnd = lnet_find_lnd_by_type(lnd_type);
2615                         if (lnd == NULL) {
2616                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2617                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2618                                 libcfs_lnd2str(lnd_type),
2619                                 libcfs_lnd2modname(lnd_type), rc);
2620 #ifndef HAVE_MODULE_LOADING_SUPPORT
2621                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2622                                                 "compiled with kernel module "
2623                                                 "loading support.");
2624 #endif
2625                                 rc = -EINVAL;
2626                                 goto failed0;
2627                         }
2628                 }
2629
2630                 net->net_lnd = lnd;
2631
2632                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2633
2634                 net_l = net;
2635         }
2636
2637         /*
2638          * net_l: if the network being added is unique then net_l
2639          *        will point to that network
2640          *        if the network being added is not unique then
2641          *        net_l points to the existing network.
2642          *
2643          * When we enter the loop below, we'll pick NIs off he
2644          * network beign added and start them up, then add them to
2645          * a local ni list. Once we've successfully started all
2646          * the NIs then we join the local NI list (of started up
2647          * networks) with the net_l->net_ni_list, which should
2648          * point to the correct network to add the new ni list to
2649          *
2650          * If any of the new NIs fail to start up, then we want to
2651          * iterate through the local ni list, which should include
2652          * any NIs which were successfully started up, and shut
2653          * them down.
2654          *
2655          * After than we want to delete the network being added,
2656          * to avoid a memory leak.
2657          */
2658         while (!list_empty(&net->net_ni_added)) {
2659                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2660                                 ni_netlist);
2661                 list_del_init(&ni->ni_netlist);
2662
2663                 /* make sure that the the NI we're about to start
2664                  * up is actually unique. if it's not fail. */
2665                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2666                                         ni->ni_interface)) {
2667                         rc = -EEXIST;
2668                         goto failed1;
2669                 }
2670
2671                 /* adjust the pointer the parent network, just in case it
2672                  * the net is a duplicate */
2673                 ni->ni_net = net_l;
2674
2675                 rc = lnet_startup_lndni(ni, tun);
2676
2677                 if (rc < 0)
2678                         goto failed1;
2679
2680                 lnet_ni_addref(ni);
2681                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2682
2683                 ni_count++;
2684         }
2685
2686         lnet_net_lock(LNET_LOCK_EX);
2687         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2688         lnet_incr_dlc_seq();
2689         lnet_net_unlock(LNET_LOCK_EX);
2690
2691         /* if the network is not unique then we don't want to keep
2692          * it around after we're done. Free it. Otherwise add that
2693          * net to the global the_lnet.ln_nets */
2694         if (net_l != net && net_l != NULL) {
2695                 /*
2696                  * TODO - note. currently the tunables can not be updated
2697                  * once added
2698                  */
2699                 lnet_net_free(net);
2700         } else {
2701                 /*
2702                  * restore tunables after it has been overwitten by the
2703                  * lnd
2704                  */
2705                 if (peer_timeout != -1)
2706                         net->net_tunables.lct_peer_timeout = peer_timeout;
2707                 if (maxtxcredits != -1)
2708                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2709                 if (peerrtrcredits != -1)
2710                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2711
2712                 lnet_net_lock(LNET_LOCK_EX);
2713                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2714                 lnet_net_unlock(LNET_LOCK_EX);
2715         }
2716
2717         return ni_count;
2718
2719 failed1:
2720         /*
2721          * shutdown the new NIs that are being started up
2722          * free the NET being started
2723          */
2724         while (!list_empty(&local_ni_list)) {
2725                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2726                                 ni_netlist);
2727
2728                 lnet_shutdown_lndni(ni);
2729         }
2730
2731 failed0:
2732         lnet_net_free(net);
2733
2734         return rc;
2735 }
2736
2737 static int
2738 lnet_startup_lndnets(struct list_head *netlist)
2739 {
2740         struct lnet_net         *net;
2741         int                     rc;
2742         int                     ni_count = 0;
2743
2744         /*
2745          * Change to running state before bringing up the LNDs. This
2746          * allows lnet_shutdown_lndnets() to assert that we've passed
2747          * through here.
2748          */
2749         lnet_net_lock(LNET_LOCK_EX);
2750         the_lnet.ln_state = LNET_STATE_RUNNING;
2751         lnet_net_unlock(LNET_LOCK_EX);
2752
2753         while (!list_empty(netlist)) {
2754                 net = list_entry(netlist->next, struct lnet_net, net_list);
2755                 list_del_init(&net->net_list);
2756
2757                 rc = lnet_startup_lndnet(net, NULL);
2758
2759                 if (rc < 0)
2760                         goto failed;
2761
2762                 ni_count += rc;
2763         }
2764
2765         return ni_count;
2766 failed:
2767         lnet_shutdown_lndnets();
2768
2769         return rc;
2770 }
2771
2772 static int lnet_genl_parse_list(struct sk_buff *msg,
2773                                 const struct ln_key_list *data[], u16 idx)
2774 {
2775         const struct ln_key_list *list = data[idx];
2776         const struct ln_key_props *props;
2777         struct nlattr *node;
2778         u16 count;
2779
2780         if (!list)
2781                 return 0;
2782
2783         if (!list->lkl_maxattr)
2784                 return -ERANGE;
2785
2786         props = list->lkl_list;
2787         if (!props)
2788                 return -EINVAL;
2789
2790         node = nla_nest_start(msg, LN_SCALAR_ATTR_LIST);
2791         if (!node)
2792                 return -ENOBUFS;
2793
2794         for (count = 1; count <= list->lkl_maxattr; count++) {
2795                 struct nlattr *key = nla_nest_start(msg, count);
2796
2797                 if (count == 1)
2798                         nla_put_u16(msg, LN_SCALAR_ATTR_LIST_SIZE,
2799                                     list->lkl_maxattr);
2800
2801                 nla_put_u16(msg, LN_SCALAR_ATTR_INDEX, count);
2802                 if (props[count].lkp_value)
2803                         nla_put_string(msg, LN_SCALAR_ATTR_VALUE,
2804                                        props[count].lkp_value);
2805                 if (props[count].lkp_key_format)
2806                         nla_put_u16(msg, LN_SCALAR_ATTR_KEY_FORMAT,
2807                                     props[count].lkp_key_format);
2808                 nla_put_u16(msg, LN_SCALAR_ATTR_NLA_TYPE,
2809                             props[count].lkp_data_type);
2810                 if (props[count].lkp_data_type == NLA_NESTED) {
2811                         int rc;
2812
2813                         rc = lnet_genl_parse_list(msg, data, ++idx);
2814                         if (rc < 0)
2815                                 return rc;
2816                         idx = rc;
2817                 }
2818
2819                 nla_nest_end(msg, key);
2820         }
2821
2822         nla_nest_end(msg, node);
2823         return idx;
2824 }
2825
2826 int lnet_genl_send_scalar_list(struct sk_buff *msg, u32 portid, u32 seq,
2827                                const struct genl_family *family, int flags,
2828                                u8 cmd, const struct ln_key_list *data[])
2829 {
2830         int rc = 0;
2831         void *hdr;
2832
2833         if (!data[0])
2834                 return -EINVAL;
2835
2836         hdr = genlmsg_put(msg, portid, seq, family, flags, cmd);
2837         if (!hdr)
2838                 GOTO(canceled, rc = -EMSGSIZE);
2839
2840         rc = lnet_genl_parse_list(msg, data, 0);
2841         if (rc < 0)
2842                 GOTO(canceled, rc);
2843
2844         genlmsg_end(msg, hdr);
2845 canceled:
2846         if (rc < 0)
2847                 genlmsg_cancel(msg, hdr);
2848         return rc > 0 ? 0 : rc;
2849 }
2850 EXPORT_SYMBOL(lnet_genl_send_scalar_list);
2851
2852 /**
2853  * Initialize LNet library.
2854  *
2855  * Automatically called at module loading time. Caller has to call
2856  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2857  * latter returned 0. It must be called exactly once.
2858  *
2859  * \retval 0 on success
2860  * \retval -ve on failures.
2861  */
2862 int lnet_lib_init(void)
2863 {
2864         int rc;
2865
2866         lnet_assert_wire_constants();
2867
2868         /* refer to global cfs_cpt_table for now */
2869         the_lnet.ln_cpt_table = cfs_cpt_tab;
2870         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2871
2872         LASSERT(the_lnet.ln_cpt_number > 0);
2873         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2874                 /* we are under risk of consuming all lh_cookie */
2875                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2876                        "please change setting of CPT-table and retry\n",
2877                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2878                 return -E2BIG;
2879         }
2880
2881         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2882                 the_lnet.ln_cpt_bits++;
2883
2884         rc = lnet_create_locks();
2885         if (rc != 0) {
2886                 CERROR("Can't create LNet global locks: %d\n", rc);
2887                 return rc;
2888         }
2889
2890         the_lnet.ln_refcount = 0;
2891         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2892         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2893
2894         /* The hash table size is the number of bits it takes to express the set
2895          * ln_num_routes, minus 1 (better to under estimate than over so we
2896          * don't waste memory). */
2897         if (rnet_htable_size <= 0)
2898                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2899         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2900                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2901         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2902                                            order_base_2(rnet_htable_size) - 1);
2903
2904         /* All LNDs apart from the LOLND are in separate modules.  They
2905          * register themselves when their module loads, and unregister
2906          * themselves when their module is unloaded. */
2907         lnet_register_lnd(&the_lolnd);
2908         return 0;
2909 }
2910
2911 /**
2912  * Finalize LNet library.
2913  *
2914  * \pre lnet_lib_init() called with success.
2915  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2916  *
2917  * As this happens at module-unload, all lnds must already be unloaded,
2918  * so they must already be unregistered.
2919  */
2920 void lnet_lib_exit(void)
2921 {
2922         int i;
2923
2924         LASSERT(the_lnet.ln_refcount == 0);
2925         lnet_unregister_lnd(&the_lolnd);
2926         for (i = 0; i < NUM_LNDS; i++)
2927                 LASSERT(!the_lnet.ln_lnds[i]);
2928         lnet_destroy_locks();
2929 }
2930
2931 /**
2932  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2933  *
2934  * Users must call this function at least once before any other functions.
2935  * For each successful call there must be a corresponding call to
2936  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2937  * ignored.
2938  *
2939  * The PID used by LNet may be different from the one requested.
2940  * See LNetGetId().
2941  *
2942  * \param requested_pid PID requested by the caller.
2943  *
2944  * \return >= 0 on success, and < 0 error code on failures.
2945  */
2946 int
2947 LNetNIInit(lnet_pid_t requested_pid)
2948 {
2949         int                     im_a_router = 0;
2950         int                     rc;
2951         int                     ni_count;
2952         struct lnet_ping_buffer *pbuf;
2953         struct lnet_handle_md   ping_mdh;
2954         LIST_HEAD(net_head);
2955         struct lnet_net         *net;
2956
2957         mutex_lock(&the_lnet.ln_api_mutex);
2958
2959         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2960
2961         if (the_lnet.ln_state == LNET_STATE_STOPPING) {
2962                 mutex_unlock(&the_lnet.ln_api_mutex);
2963                 return -ESHUTDOWN;
2964         }
2965
2966         if (the_lnet.ln_refcount > 0) {
2967                 rc = the_lnet.ln_refcount++;
2968                 mutex_unlock(&the_lnet.ln_api_mutex);
2969                 return rc;
2970         }
2971
2972         rc = lnet_prepare(requested_pid);
2973         if (rc != 0) {
2974                 mutex_unlock(&the_lnet.ln_api_mutex);
2975                 return rc;
2976         }
2977
2978         /* create a network for Loopback network */
2979         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2980         if (net == NULL) {
2981                 rc = -ENOMEM;
2982                 goto err_empty_list;
2983         }
2984
2985         /* Add in the loopback NI */
2986         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2987                 rc = -ENOMEM;
2988                 goto err_empty_list;
2989         }
2990
2991         if (use_tcp_bonding)
2992                 CWARN("use_tcp_bonding has been removed. Use Multi-Rail and Dynamic Discovery instead, see LU-13641\n");
2993
2994         /* If LNet is being initialized via DLC it is possible
2995          * that the user requests not to load module parameters (ones which
2996          * are supported by DLC) on initialization.  Therefore, make sure not
2997          * to load networks, routes and forwarding from module parameters
2998          * in this case.  On cleanup in case of failure only clean up
2999          * routes if it has been loaded */
3000         if (!the_lnet.ln_nis_from_mod_params) {
3001                 rc = lnet_parse_networks(&net_head, lnet_get_networks());
3002                 if (rc < 0)
3003                         goto err_empty_list;
3004         }
3005
3006         ni_count = lnet_startup_lndnets(&net_head);
3007         if (ni_count < 0) {
3008                 rc = ni_count;
3009                 goto err_empty_list;
3010         }
3011
3012         if (!the_lnet.ln_nis_from_mod_params) {
3013                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
3014                 if (rc != 0)
3015                         goto err_shutdown_lndnis;
3016
3017                 rc = lnet_rtrpools_alloc(im_a_router);
3018                 if (rc != 0)
3019                         goto err_destroy_routes;
3020         }
3021
3022         rc = lnet_acceptor_start();
3023         if (rc != 0)
3024                 goto err_destroy_routes;
3025
3026         the_lnet.ln_refcount = 1;
3027         /* Now I may use my own API functions... */
3028
3029         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
3030         if (rc != 0)
3031                 goto err_acceptor_stop;
3032
3033         lnet_ping_target_update(pbuf, ping_mdh);
3034
3035         the_lnet.ln_mt_handler = lnet_mt_event_handler;
3036
3037         rc = lnet_push_target_init();
3038         if (rc != 0)
3039                 goto err_stop_ping;
3040
3041         rc = lnet_peer_discovery_start();
3042         if (rc != 0)
3043                 goto err_destroy_push_target;
3044
3045         rc = lnet_monitor_thr_start();
3046         if (rc != 0)
3047                 goto err_stop_discovery_thr;
3048
3049         lnet_fault_init();
3050         lnet_router_debugfs_init();
3051
3052         mutex_unlock(&the_lnet.ln_api_mutex);
3053
3054         complete_all(&the_lnet.ln_started);
3055
3056         /* wait for all routers to start */
3057         lnet_wait_router_start();
3058
3059         return 0;
3060
3061 err_stop_discovery_thr:
3062         lnet_peer_discovery_stop();
3063 err_destroy_push_target:
3064         lnet_push_target_fini();
3065 err_stop_ping:
3066         lnet_ping_target_fini();
3067 err_acceptor_stop:
3068         the_lnet.ln_refcount = 0;
3069         lnet_acceptor_stop();
3070 err_destroy_routes:
3071         if (!the_lnet.ln_nis_from_mod_params)
3072                 lnet_destroy_routes();
3073 err_shutdown_lndnis:
3074         lnet_shutdown_lndnets();
3075 err_empty_list:
3076         lnet_unprepare();
3077         LASSERT(rc < 0);
3078         mutex_unlock(&the_lnet.ln_api_mutex);
3079         while (!list_empty(&net_head)) {
3080                 struct lnet_net *net;
3081
3082                 net = list_entry(net_head.next, struct lnet_net, net_list);
3083                 list_del_init(&net->net_list);
3084                 lnet_net_free(net);
3085         }
3086         return rc;
3087 }
3088 EXPORT_SYMBOL(LNetNIInit);
3089
3090 /**
3091  * Stop LNet interfaces, routing, and forwarding.
3092  *
3093  * Users must call this function once for each successful call to LNetNIInit().
3094  * Once the LNetNIFini() operation has been started, the results of pending
3095  * API operations are undefined.
3096  *
3097  * \return always 0 for current implementation.
3098  */
3099 int
3100 LNetNIFini(void)
3101 {
3102         mutex_lock(&the_lnet.ln_api_mutex);
3103
3104         LASSERT(the_lnet.ln_refcount > 0);
3105
3106         if (the_lnet.ln_refcount != 1) {
3107                 the_lnet.ln_refcount--;
3108         } else {
3109                 LASSERT(!the_lnet.ln_niinit_self);
3110
3111                 lnet_net_lock(LNET_LOCK_EX);
3112                 the_lnet.ln_state = LNET_STATE_STOPPING;
3113                 lnet_net_unlock(LNET_LOCK_EX);
3114
3115                 lnet_fault_fini();
3116
3117                 lnet_router_debugfs_fini();
3118                 lnet_monitor_thr_stop();
3119                 lnet_peer_discovery_stop();
3120                 lnet_push_target_fini();
3121                 lnet_ping_target_fini();
3122
3123                 /* Teardown fns that use my own API functions BEFORE here */
3124                 the_lnet.ln_refcount = 0;
3125
3126                 lnet_acceptor_stop();
3127                 lnet_destroy_routes();
3128                 lnet_shutdown_lndnets();
3129                 lnet_unprepare();
3130         }
3131
3132         mutex_unlock(&the_lnet.ln_api_mutex);
3133         return 0;
3134 }
3135 EXPORT_SYMBOL(LNetNIFini);
3136
3137 /**
3138  * Grabs the ni data from the ni structure and fills the out
3139  * parameters
3140  *
3141  * \param[in] ni network        interface structure
3142  * \param[out] cfg_ni           NI config information
3143  * \param[out] tun              network and LND tunables
3144  */
3145 static void
3146 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
3147                    struct lnet_ioctl_config_lnd_tunables *tun,
3148                    struct lnet_ioctl_element_stats *stats,
3149                    __u32 tun_size)
3150 {
3151         size_t min_size = 0;
3152         int i;
3153
3154         if (!ni || !cfg_ni || !tun || !nid_is_nid4(&ni->ni_nid))
3155                 return;
3156
3157         if (ni->ni_interface != NULL) {
3158                 strncpy(cfg_ni->lic_ni_intf,
3159                         ni->ni_interface,
3160                         sizeof(cfg_ni->lic_ni_intf));
3161         }
3162
3163         cfg_ni->lic_nid = lnet_nid_to_nid4(&ni->ni_nid);
3164         cfg_ni->lic_status = lnet_ni_get_status_locked(ni);
3165         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
3166
3167         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
3168
3169         if (stats) {
3170                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
3171                                                        LNET_STATS_TYPE_SEND);
3172                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
3173                                                        LNET_STATS_TYPE_RECV);
3174                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
3175                                                        LNET_STATS_TYPE_DROP);
3176         }
3177
3178         /*
3179          * tun->lt_tun will always be present, but in order to be
3180          * backwards compatible, we need to deal with the cases when
3181          * tun->lt_tun is smaller than what the kernel has, because it
3182          * comes from an older version of a userspace program, then we'll
3183          * need to copy as much information as we have available space.
3184          */
3185         min_size = tun_size - sizeof(tun->lt_cmn);
3186         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
3187
3188         /* copy over the cpts */
3189         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
3190             ni->ni_cpts == NULL)  {
3191                 for (i = 0; i < ni->ni_ncpts; i++)
3192                         cfg_ni->lic_cpts[i] = i;
3193         } else {
3194                 for (i = 0;
3195                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
3196                      i < LNET_MAX_SHOW_NUM_CPT;
3197                      i++)
3198                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
3199         }
3200         cfg_ni->lic_ncpts = ni->ni_ncpts;
3201 }
3202
3203 /**
3204  * NOTE: This is a legacy function left in the code to be backwards
3205  * compatible with older userspace programs. It should eventually be
3206  * removed.
3207  *
3208  * Grabs the ni data from the ni structure and fills the out
3209  * parameters
3210  *
3211  * \param[in] ni network        interface structure
3212  * \param[out] config           config information
3213  */
3214 static void
3215 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
3216                          struct lnet_ioctl_config_data *config)
3217 {
3218         struct lnet_ioctl_net_config *net_config;
3219         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
3220         size_t min_size, tunable_size = 0;
3221         int i;
3222
3223         if (!ni || !config || !nid_is_nid4(&ni->ni_nid))
3224                 return;
3225
3226         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
3227         if (!net_config)
3228                 return;
3229
3230         if (!ni->ni_interface)
3231                 return;
3232
3233         strncpy(net_config->ni_interface,
3234                 ni->ni_interface,
3235                 sizeof(net_config->ni_interface));
3236
3237         config->cfg_nid = lnet_nid_to_nid4(&ni->ni_nid);
3238         config->cfg_config_u.cfg_net.net_peer_timeout =
3239                 ni->ni_net->net_tunables.lct_peer_timeout;
3240         config->cfg_config_u.cfg_net.net_max_tx_credits =
3241                 ni->ni_net->net_tunables.lct_max_tx_credits;
3242         config->cfg_config_u.cfg_net.net_peer_tx_credits =
3243                 ni->ni_net->net_tunables.lct_peer_tx_credits;
3244         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
3245                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
3246
3247         net_config->ni_status = lnet_ni_get_status_locked(ni);
3248
3249         if (ni->ni_cpts) {
3250                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
3251
3252                 for (i = 0; i < num_cpts; i++)
3253                         net_config->ni_cpts[i] = ni->ni_cpts[i];
3254
3255                 config->cfg_ncpts = num_cpts;
3256         }
3257
3258         /*
3259          * See if user land tools sent in a newer and larger version
3260          * of struct lnet_tunables than what the kernel uses.
3261          */
3262         min_size = sizeof(*config) + sizeof(*net_config);
3263
3264         if (config->cfg_hdr.ioc_len > min_size)
3265                 tunable_size = config->cfg_hdr.ioc_len - min_size;
3266
3267         /* Don't copy too much data to user space */
3268         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
3269         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
3270
3271         if (lnd_cfg && min_size) {
3272                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
3273                 config->cfg_config_u.cfg_net.net_interface_count = 1;
3274
3275                 /* Tell user land that kernel side has less data */
3276                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
3277                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
3278                         config->cfg_hdr.ioc_len -= min_size;
3279                 }
3280         }
3281 }
3282
3283 struct lnet_ni *
3284 lnet_get_ni_idx_locked(int idx)
3285 {
3286         struct lnet_ni          *ni;
3287         struct lnet_net         *net;
3288
3289         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3290                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3291                         if (idx-- == 0)
3292                                 return ni;
3293                 }
3294         }
3295
3296         return NULL;
3297 }
3298
3299 int lnet_get_net_healthv_locked(struct lnet_net *net)
3300 {
3301         struct lnet_ni *ni;
3302         int best_healthv = 0;
3303         int healthv, ni_fatal;
3304
3305         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3306                 healthv = atomic_read(&ni->ni_healthv);
3307                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
3308                 if (!ni_fatal && healthv > best_healthv)
3309                         best_healthv = healthv;
3310         }
3311
3312         return best_healthv;
3313 }
3314
3315 struct lnet_ni *
3316 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
3317 {
3318         struct lnet_ni          *ni;
3319         struct lnet_net         *net = mynet;
3320
3321         /*
3322          * It is possible that the net has been cleaned out while there is
3323          * a message being sent. This function accessed the net without
3324          * checking if the list is empty
3325          */
3326         if (prev == NULL) {
3327                 if (net == NULL)
3328                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
3329                                         net_list);
3330                 if (list_empty(&net->net_ni_list))
3331                         return NULL;
3332                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3333                                 ni_netlist);
3334
3335                 return ni;
3336         }
3337
3338         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
3339                 /* if you reached the end of the ni list and the net is
3340                  * specified, then there are no more nis in that net */
3341                 if (net != NULL)
3342                         return NULL;
3343
3344                 /* we reached the end of this net ni list. move to the
3345                  * next net */
3346                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
3347                         /* no more nets and no more NIs. */
3348                         return NULL;
3349
3350                 /* get the next net */
3351                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
3352                                  net_list);
3353                 if (list_empty(&net->net_ni_list))
3354                         return NULL;
3355                 /* get the ni on it */
3356                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
3357                                 ni_netlist);
3358
3359                 return ni;
3360         }
3361
3362         if (list_empty(&prev->ni_netlist))
3363                 return NULL;
3364
3365         /* there are more nis left */
3366         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
3367
3368         return ni;
3369 }
3370
3371 int
3372 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3373 {
3374         struct lnet_ni *ni;
3375         int cpt;
3376         int rc = -ENOENT;
3377         int idx = config->cfg_count;
3378
3379         cpt = lnet_net_lock_current();
3380
3381         ni = lnet_get_ni_idx_locked(idx);
3382
3383         if (ni != NULL) {
3384                 rc = 0;
3385                 lnet_ni_lock(ni);
3386                 lnet_fill_ni_info_legacy(ni, config);
3387                 lnet_ni_unlock(ni);
3388         }
3389
3390         lnet_net_unlock(cpt);
3391         return rc;
3392 }
3393
3394 int
3395 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3396                    struct lnet_ioctl_config_lnd_tunables *tun,
3397                    struct lnet_ioctl_element_stats *stats,
3398                    __u32 tun_size)
3399 {
3400         struct lnet_ni          *ni;
3401         int                     cpt;
3402         int                     rc = -ENOENT;
3403
3404         if (!cfg_ni || !tun || !stats)
3405                 return -EINVAL;
3406
3407         cpt = lnet_net_lock_current();
3408
3409         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3410
3411         if (ni) {
3412                 rc = 0;
3413                 lnet_ni_lock(ni);
3414                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3415                 lnet_ni_unlock(ni);
3416         }
3417
3418         lnet_net_unlock(cpt);
3419         return rc;
3420 }
3421
3422 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3423 {
3424         struct lnet_ni *ni;
3425         int cpt;
3426         int rc = -ENOENT;
3427
3428         if (!msg_stats)
3429                 return -EINVAL;
3430
3431         cpt = lnet_net_lock_current();
3432
3433         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3434
3435         if (ni) {
3436                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3437                 rc = 0;
3438         }
3439
3440         lnet_net_unlock(cpt);
3441
3442         return rc;
3443 }
3444
3445 static int lnet_add_net_common(struct lnet_net *net,
3446                                struct lnet_ioctl_config_lnd_tunables *tun)
3447 {
3448         struct lnet_handle_md ping_mdh;
3449         struct lnet_ping_buffer *pbuf;
3450         struct lnet_remotenet *rnet;
3451         struct lnet_ni *ni;
3452         int net_ni_count;
3453         __u32 net_id;
3454         int rc;
3455
3456         lnet_net_lock(LNET_LOCK_EX);
3457         rnet = lnet_find_rnet_locked(net->net_id);
3458         lnet_net_unlock(LNET_LOCK_EX);
3459         /*
3460          * make sure that the net added doesn't invalidate the current
3461          * configuration LNet is keeping
3462          */
3463         if (rnet) {
3464                 CERROR("Adding net %s will invalidate routing configuration\n",
3465                        libcfs_net2str(net->net_id));
3466                 lnet_net_free(net);
3467                 return -EUSERS;
3468         }
3469
3470         /*
3471          * make sure you calculate the correct number of slots in the ping
3472          * buffer. Since the ping info is a flattened list of all the NIs,
3473          * we should allocate enough slots to accomodate the number of NIs
3474          * which will be added.
3475          *
3476          * since ni hasn't been configured yet, use
3477          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3478          */
3479         net_ni_count = lnet_get_net_ni_count_pre(net);
3480
3481         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3482                                     net_ni_count + lnet_get_ni_count(),
3483                                     false);
3484         if (rc < 0) {
3485                 lnet_net_free(net);
3486                 return rc;
3487         }
3488
3489         if (tun)
3490                 memcpy(&net->net_tunables,
3491                        &tun->lt_cmn, sizeof(net->net_tunables));
3492         else
3493                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3494
3495         net_id = net->net_id;
3496
3497         rc = lnet_startup_lndnet(net,
3498                                  (tun) ? &tun->lt_tun : NULL);
3499         if (rc < 0)
3500                 goto failed;
3501
3502         lnet_net_lock(LNET_LOCK_EX);
3503         net = lnet_get_net_locked(net_id);
3504         LASSERT(net);
3505
3506         /* apply the UDSPs */
3507         rc = lnet_udsp_apply_policies_on_net(net);
3508         if (rc)
3509                 CERROR("Failed to apply UDSPs on local net %s\n",
3510                        libcfs_net2str(net->net_id));
3511
3512         /* At this point we lost track of which NI was just added, so we
3513          * just re-apply the policies on all of the NIs on this net
3514          */
3515         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3516                 rc = lnet_udsp_apply_policies_on_ni(ni);
3517                 if (rc)
3518                         CERROR("Failed to apply UDSPs on ni %s\n",
3519                                libcfs_nidstr(&ni->ni_nid));
3520         }
3521         lnet_net_unlock(LNET_LOCK_EX);
3522
3523         /*
3524          * Start the acceptor thread if this is the first network
3525          * being added that requires the thread.
3526          */
3527         if (net->net_lnd->lnd_accept) {
3528                 rc = lnet_acceptor_start();
3529                 if (rc < 0) {
3530                         /* shutdown the net that we just started */
3531                         CERROR("Failed to start up acceptor thread\n");
3532                         lnet_shutdown_lndnet(net);
3533                         goto failed;
3534                 }
3535         }
3536
3537         lnet_net_lock(LNET_LOCK_EX);
3538         lnet_peer_net_added(net);
3539         lnet_net_unlock(LNET_LOCK_EX);
3540
3541         lnet_ping_target_update(pbuf, ping_mdh);
3542
3543         return 0;
3544
3545 failed:
3546         lnet_ping_md_unlink(pbuf, &ping_mdh);
3547         lnet_ping_buffer_decref(pbuf);
3548         return rc;
3549 }
3550
3551 static void
3552 lnet_set_tune_defaults(struct lnet_ioctl_config_lnd_tunables *tun)
3553 {
3554         if (tun) {
3555                 if (!tun->lt_cmn.lct_peer_timeout)
3556                         tun->lt_cmn.lct_peer_timeout = DEFAULT_PEER_TIMEOUT;
3557                 if (!tun->lt_cmn.lct_peer_tx_credits)
3558                         tun->lt_cmn.lct_peer_tx_credits = DEFAULT_PEER_CREDITS;
3559                 if (!tun->lt_cmn.lct_max_tx_credits)
3560                         tun->lt_cmn.lct_max_tx_credits = DEFAULT_CREDITS;
3561         }
3562 }
3563
3564 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3565                                       struct lnet_ioctl_config_lnd_tunables *tun)
3566 {
3567         struct lnet_net *net;
3568         const char *nets;
3569         int rc;
3570         LIST_HEAD(net_head);
3571
3572         rc = lnet_parse_ip2nets(&nets, ip2nets);
3573         if (rc < 0)
3574                 return rc;
3575
3576         rc = lnet_parse_networks(&net_head, nets);
3577         if (rc < 0)
3578                 return rc;
3579
3580         lnet_set_tune_defaults(tun);
3581
3582         mutex_lock(&the_lnet.ln_api_mutex);
3583         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3584                 rc = -ESHUTDOWN;
3585                 goto out;
3586         }
3587         while (!list_empty(&net_head)) {
3588                 net = list_entry(net_head.next, struct lnet_net, net_list);
3589                 list_del_init(&net->net_list);
3590                 rc = lnet_add_net_common(net, tun);
3591                 if (rc < 0)
3592                         goto out;
3593         }
3594
3595 out:
3596         mutex_unlock(&the_lnet.ln_api_mutex);
3597
3598         while (!list_empty(&net_head)) {
3599                 net = list_entry(net_head.next, struct lnet_net, net_list);
3600                 list_del_init(&net->net_list);
3601                 lnet_net_free(net);
3602         }
3603         return rc;
3604 }
3605
3606 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3607 {
3608         struct lnet_net *net;
3609         struct lnet_ni *ni;
3610         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3611         int rc, i;
3612         __u32 net_id, lnd_type;
3613
3614         /* get the tunables if they are available */
3615         if (conf->lic_cfg_hdr.ioc_len >=
3616             sizeof(*conf) + sizeof(*tun))
3617                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3618                         conf->lic_bulk;
3619
3620         /* handle legacy ip2nets from DLC */
3621         if (conf->lic_legacy_ip2nets[0] != '\0')
3622                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3623                                                   tun);
3624
3625         net_id = LNET_NIDNET(conf->lic_nid);
3626         lnd_type = LNET_NETTYP(net_id);
3627
3628         if (!libcfs_isknown_lnd(lnd_type)) {
3629                 CERROR("No valid net and lnd information provided\n");
3630                 return -EINVAL;
3631         }
3632
3633         net = lnet_net_alloc(net_id, NULL);
3634         if (!net)
3635                 return -ENOMEM;
3636
3637         for (i = 0; i < conf->lic_ncpts; i++) {
3638                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3639                         return -EINVAL;
3640         }
3641
3642         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3643                                        conf->lic_ni_intf);
3644         if (!ni)
3645                 return -ENOMEM;
3646
3647         lnet_set_tune_defaults(tun);
3648
3649         mutex_lock(&the_lnet.ln_api_mutex);
3650         if (the_lnet.ln_state != LNET_STATE_RUNNING)
3651                 rc = -ESHUTDOWN;
3652         else
3653                 rc = lnet_add_net_common(net, tun);
3654
3655         mutex_unlock(&the_lnet.ln_api_mutex);
3656
3657         return rc;
3658 }
3659
3660 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3661 {
3662         struct lnet_net  *net;
3663         struct lnet_ni *ni;
3664         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3665         struct lnet_ping_buffer *pbuf;
3666         struct lnet_handle_md  ping_mdh;
3667         int               rc;
3668         int               net_count;
3669         __u32             addr;
3670
3671         /* don't allow userspace to shutdown the LOLND */
3672         if (LNET_NETTYP(net_id) == LOLND)
3673                 return -EINVAL;
3674
3675         mutex_lock(&the_lnet.ln_api_mutex);
3676         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3677                 rc = -ESHUTDOWN;
3678                 goto unlock_api_mutex;
3679         }
3680
3681         lnet_net_lock(0);
3682
3683         net = lnet_get_net_locked(net_id);
3684         if (!net) {
3685                 CERROR("net %s not found\n",
3686                        libcfs_net2str(net_id));
3687                 rc = -ENOENT;
3688                 goto unlock_net;
3689         }
3690
3691         addr = LNET_NIDADDR(conf->lic_nid);
3692         if (addr == 0) {
3693                 /* remove the entire net */
3694                 net_count = lnet_get_net_ni_count_locked(net);
3695
3696                 lnet_net_unlock(0);
3697
3698                 /* create and link a new ping info, before removing the old one */
3699                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3700                                         lnet_get_ni_count() - net_count,
3701                                         false);
3702                 if (rc != 0)
3703                         goto unlock_api_mutex;
3704
3705                 lnet_shutdown_lndnet(net);
3706
3707                 lnet_acceptor_stop();
3708
3709                 lnet_ping_target_update(pbuf, ping_mdh);
3710
3711                 goto unlock_api_mutex;
3712         }
3713
3714         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3715         if (!ni) {
3716                 CERROR("nid %s not found\n",
3717                        libcfs_nid2str(conf->lic_nid));
3718                 rc = -ENOENT;
3719                 goto unlock_net;
3720         }
3721
3722         net_count = lnet_get_net_ni_count_locked(net);
3723
3724         lnet_net_unlock(0);
3725
3726         /* create and link a new ping info, before removing the old one */
3727         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3728                                   lnet_get_ni_count() - 1, false);
3729         if (rc != 0)
3730                 goto unlock_api_mutex;
3731
3732         lnet_shutdown_lndni(ni);
3733
3734         lnet_acceptor_stop();
3735
3736         lnet_ping_target_update(pbuf, ping_mdh);
3737
3738         /* check if the net is empty and remove it if it is */
3739         if (net_count == 1)
3740                 lnet_shutdown_lndnet(net);
3741
3742         goto unlock_api_mutex;
3743
3744 unlock_net:
3745         lnet_net_unlock(0);
3746 unlock_api_mutex:
3747         mutex_unlock(&the_lnet.ln_api_mutex);
3748
3749         return rc;
3750 }
3751
3752 /*
3753  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3754  * They are only expected to be called for unique networks.
3755  * That can be as a result of older DLC library
3756  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3757  */
3758 int
3759 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3760 {
3761         struct lnet_net *net;
3762         LIST_HEAD(net_head);
3763         int rc;
3764         struct lnet_ioctl_config_lnd_tunables tun;
3765         const char *nets = conf->cfg_config_u.cfg_net.net_intf;
3766
3767         /* Create a net/ni structures for the network string */
3768         rc = lnet_parse_networks(&net_head, nets);
3769         if (rc <= 0)
3770                 return rc == 0 ? -EINVAL : rc;
3771
3772         mutex_lock(&the_lnet.ln_api_mutex);
3773         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3774                 rc = -ESHUTDOWN;
3775                 goto out_unlock_clean;
3776         }
3777
3778         if (rc > 1) {
3779                 rc = -EINVAL; /* only add one network per call */
3780                 goto out_unlock_clean;
3781         }
3782
3783         net = list_entry(net_head.next, struct lnet_net, net_list);
3784         list_del_init(&net->net_list);
3785
3786         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3787
3788         memset(&tun, 0, sizeof(tun));
3789
3790         tun.lt_cmn.lct_peer_timeout =
3791           (!conf->cfg_config_u.cfg_net.net_peer_timeout) ? DEFAULT_PEER_TIMEOUT :
3792                 conf->cfg_config_u.cfg_net.net_peer_timeout;
3793         tun.lt_cmn.lct_peer_tx_credits =
3794           (!conf->cfg_config_u.cfg_net.net_peer_tx_credits) ? DEFAULT_PEER_CREDITS :
3795                 conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3796         tun.lt_cmn.lct_peer_rtr_credits =
3797           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3798         tun.lt_cmn.lct_max_tx_credits =
3799           (!conf->cfg_config_u.cfg_net.net_max_tx_credits) ? DEFAULT_CREDITS :
3800                 conf->cfg_config_u.cfg_net.net_max_tx_credits;
3801
3802         rc = lnet_add_net_common(net, &tun);
3803
3804 out_unlock_clean:
3805         mutex_unlock(&the_lnet.ln_api_mutex);
3806         while (!list_empty(&net_head)) {
3807                 /* net_head list is empty in success case */
3808                 net = list_entry(net_head.next, struct lnet_net, net_list);
3809                 list_del_init(&net->net_list);
3810                 lnet_net_free(net);
3811         }
3812         return rc;
3813 }
3814
3815 int
3816 lnet_dyn_del_net(__u32 net_id)
3817 {
3818         struct lnet_net  *net;
3819         struct lnet_ping_buffer *pbuf;
3820         struct lnet_handle_md ping_mdh;
3821         int               rc;
3822         int               net_ni_count;
3823
3824         /* don't allow userspace to shutdown the LOLND */
3825         if (LNET_NETTYP(net_id) == LOLND)
3826                 return -EINVAL;
3827
3828         mutex_lock(&the_lnet.ln_api_mutex);
3829         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3830                 rc = -ESHUTDOWN;
3831                 goto out;
3832         }
3833
3834         lnet_net_lock(0);
3835
3836         net = lnet_get_net_locked(net_id);
3837         if (net == NULL) {
3838                 lnet_net_unlock(0);
3839                 rc = -EINVAL;
3840                 goto out;
3841         }
3842
3843         net_ni_count = lnet_get_net_ni_count_locked(net);
3844
3845         lnet_net_unlock(0);
3846
3847         /* create and link a new ping info, before removing the old one */
3848         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3849                                     lnet_get_ni_count() - net_ni_count, false);
3850         if (rc != 0)
3851                 goto out;
3852
3853         lnet_shutdown_lndnet(net);
3854
3855         lnet_acceptor_stop();
3856
3857         lnet_ping_target_update(pbuf, ping_mdh);
3858
3859 out:
3860         mutex_unlock(&the_lnet.ln_api_mutex);
3861
3862         return rc;
3863 }
3864
3865 void lnet_incr_dlc_seq(void)
3866 {
3867         atomic_inc(&lnet_dlc_seq_no);
3868 }
3869
3870 __u32 lnet_get_dlc_seq_locked(void)
3871 {
3872         return atomic_read(&lnet_dlc_seq_no);
3873 }
3874
3875 static void
3876 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3877 {
3878         struct lnet_net *net;
3879         struct lnet_ni *ni;
3880
3881         lnet_net_lock(LNET_LOCK_EX);
3882         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3883                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3884                         if (all || (nid_is_nid4(&ni->ni_nid) &&
3885                                     lnet_nid_to_nid4(&ni->ni_nid) == nid)) {
3886                                 atomic_set(&ni->ni_healthv, value);
3887                                 if (list_empty(&ni->ni_recovery) &&
3888                                     value < LNET_MAX_HEALTH_VALUE) {
3889                                         CERROR("manually adding local NI %s to recovery\n",
3890                                                libcfs_nidstr(&ni->ni_nid));
3891                                         list_add_tail(&ni->ni_recovery,
3892                                                       &the_lnet.ln_mt_localNIRecovq);
3893                                         lnet_ni_addref_locked(ni, 0);
3894                                 }
3895                                 if (!all) {
3896                                         lnet_net_unlock(LNET_LOCK_EX);
3897                                         return;
3898                                 }
3899                         }
3900                 }
3901         }
3902         lnet_net_unlock(LNET_LOCK_EX);
3903 }
3904
3905 static void
3906 lnet_ni_set_conns_per_peer(lnet_nid_t nid, int value, bool all)
3907 {
3908         struct lnet_net *net;
3909         struct lnet_ni *ni;
3910
3911         lnet_net_lock(LNET_LOCK_EX);
3912         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3913                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3914                         if (lnet_nid_to_nid4(&ni->ni_nid) != nid && !all)
3915                                 continue;
3916                         if (LNET_NETTYP(net->net_id) == SOCKLND)
3917                                 ni->ni_lnd_tunables.lnd_tun_u.lnd_sock.lnd_conns_per_peer = value;
3918                         else if (LNET_NETTYP(net->net_id) == O2IBLND)
3919                                 ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = value;
3920                         if (!all) {
3921                                 lnet_net_unlock(LNET_LOCK_EX);
3922                                 return;
3923                         }
3924                 }
3925         }
3926         lnet_net_unlock(LNET_LOCK_EX);
3927 }
3928
3929 static int
3930 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3931 {
3932         int cpt, rc = 0;
3933         struct lnet_ni *ni;
3934         lnet_nid_t nid = stats->hlni_nid;
3935
3936         cpt = lnet_net_lock_current();
3937         ni = lnet_nid2ni_locked(nid, cpt);
3938
3939         if (!ni) {
3940                 rc = -ENOENT;
3941                 goto unlock;
3942         }
3943
3944         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3945         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3946         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3947         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3948         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3949         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3950         stats->hlni_fatal_error = atomic_read(&ni->ni_fatal_error_on);
3951         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3952         stats->hlni_ping_count = ni->ni_ping_count;
3953         stats->hlni_next_ping = ni->ni_next_ping;
3954
3955 unlock:
3956         lnet_net_unlock(cpt);
3957
3958         return rc;
3959 }
3960
3961 static int
3962 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3963 {
3964         struct lnet_ni *ni;
3965         int i = 0;
3966
3967         lnet_net_lock(LNET_LOCK_EX);
3968         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
3969                 if (!nid_is_nid4(&ni->ni_nid))
3970                         continue;
3971                 list->rlst_nid_array[i] = lnet_nid_to_nid4(&ni->ni_nid);
3972                 i++;
3973                 if (i >= LNET_MAX_SHOW_NUM_NID)
3974                         break;
3975         }
3976         lnet_net_unlock(LNET_LOCK_EX);
3977         list->rlst_num_nids = i;
3978
3979         return 0;
3980 }
3981
3982 static int
3983 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3984 {
3985         struct lnet_peer_ni *lpni;
3986         int i = 0;
3987
3988         lnet_net_lock(LNET_LOCK_EX);
3989         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
3990                 list->rlst_nid_array[i] = lnet_nid_to_nid4(&lpni->lpni_nid);
3991                 i++;
3992                 if (i >= LNET_MAX_SHOW_NUM_NID)
3993                         break;
3994         }
3995         lnet_net_unlock(LNET_LOCK_EX);
3996         list->rlst_num_nids = i;
3997
3998         return 0;
3999 }
4000
4001 /**
4002  * LNet ioctl handler.
4003  *
4004  */
4005 int
4006 LNetCtl(unsigned int cmd, void *arg)
4007 {
4008         struct libcfs_ioctl_data *data = arg;
4009         struct lnet_ioctl_config_data *config;
4010         struct lnet_process_id    id4 = {};
4011         struct lnet_processid     id = {};
4012         struct lnet_ni           *ni;
4013         struct lnet_nid           nid;
4014         int                       rc;
4015
4016         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
4017                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
4018
4019         switch (cmd) {
4020         case IOC_LIBCFS_GET_NI:
4021                 rc = LNetGetId(data->ioc_count, &id);
4022                 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
4023                 return rc;
4024
4025         case IOC_LIBCFS_FAIL_NID:
4026                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
4027
4028         case IOC_LIBCFS_ADD_ROUTE: {
4029                 /* default router sensitivity to 1 */
4030                 unsigned int sensitivity = 1;
4031                 config = arg;
4032
4033                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4034                         return -EINVAL;
4035
4036                 if (config->cfg_config_u.cfg_route.rtr_sensitivity) {
4037                         sensitivity =
4038                           config->cfg_config_u.cfg_route.rtr_sensitivity;
4039                 }
4040
4041                 lnet_nid4_to_nid(config->cfg_nid, &nid);
4042                 mutex_lock(&the_lnet.ln_api_mutex);
4043                 rc = lnet_add_route(config->cfg_net,
4044                                     config->cfg_config_u.cfg_route.rtr_hop,
4045                                     &nid,
4046                                     config->cfg_config_u.cfg_route.
4047                                         rtr_priority, sensitivity);
4048                 mutex_unlock(&the_lnet.ln_api_mutex);
4049                 return rc;
4050         }
4051
4052         case IOC_LIBCFS_DEL_ROUTE:
4053                 config = arg;
4054
4055                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4056                         return -EINVAL;
4057
4058                 lnet_nid4_to_nid(config->cfg_nid, &nid);
4059                 mutex_lock(&the_lnet.ln_api_mutex);
4060                 rc = lnet_del_route(config->cfg_net, &nid);
4061                 mutex_unlock(&the_lnet.ln_api_mutex);
4062                 return rc;
4063
4064         case IOC_LIBCFS_GET_ROUTE:
4065                 config = arg;
4066
4067                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4068                         return -EINVAL;
4069
4070                 mutex_lock(&the_lnet.ln_api_mutex);
4071                 rc = lnet_get_route(config->cfg_count,
4072                                     &config->cfg_net,
4073                                     &config->cfg_config_u.cfg_route.rtr_hop,
4074                                     &config->cfg_nid,
4075                                     &config->cfg_config_u.cfg_route.rtr_flags,
4076                                     &config->cfg_config_u.cfg_route.
4077                                         rtr_priority,
4078                                     &config->cfg_config_u.cfg_route.
4079                                         rtr_sensitivity);
4080                 mutex_unlock(&the_lnet.ln_api_mutex);
4081                 return rc;
4082
4083         case IOC_LIBCFS_GET_LOCAL_NI: {
4084                 struct lnet_ioctl_config_ni *cfg_ni;
4085                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
4086                 struct lnet_ioctl_element_stats *stats;
4087                 __u32 tun_size;
4088
4089                 cfg_ni = arg;
4090
4091                 /* get the tunables if they are available */
4092                 if (cfg_ni->lic_cfg_hdr.ioc_len <
4093                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
4094                         return -EINVAL;
4095
4096                 stats = (struct lnet_ioctl_element_stats *)
4097                         cfg_ni->lic_bulk;
4098                 tun = (struct lnet_ioctl_config_lnd_tunables *)
4099                                 (cfg_ni->lic_bulk + sizeof(*stats));
4100
4101                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
4102                         sizeof(*stats);
4103
4104                 mutex_lock(&the_lnet.ln_api_mutex);
4105                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
4106                 mutex_unlock(&the_lnet.ln_api_mutex);
4107                 return rc;
4108         }
4109
4110         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
4111                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
4112
4113                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
4114                         return -EINVAL;
4115
4116                 mutex_lock(&the_lnet.ln_api_mutex);
4117                 rc = lnet_get_ni_stats(msg_stats);
4118                 mutex_unlock(&the_lnet.ln_api_mutex);
4119
4120                 return rc;
4121         }
4122
4123         case IOC_LIBCFS_GET_NET: {
4124                 size_t total = sizeof(*config) +
4125                                sizeof(struct lnet_ioctl_net_config);
4126                 config = arg;
4127
4128                 if (config->cfg_hdr.ioc_len < total)
4129                         return -EINVAL;
4130
4131                 mutex_lock(&the_lnet.ln_api_mutex);
4132                 rc = lnet_get_net_config(config);
4133                 mutex_unlock(&the_lnet.ln_api_mutex);
4134                 return rc;
4135         }
4136
4137         case IOC_LIBCFS_GET_LNET_STATS:
4138         {
4139                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
4140
4141                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
4142                         return -EINVAL;
4143
4144                 mutex_lock(&the_lnet.ln_api_mutex);
4145                 rc = lnet_counters_get(&lnet_stats->st_cntrs);
4146                 mutex_unlock(&the_lnet.ln_api_mutex);
4147                 return rc;
4148         }
4149
4150         case IOC_LIBCFS_RESET_LNET_STATS:
4151         {
4152                 mutex_lock(&the_lnet.ln_api_mutex);
4153                 lnet_counters_reset();
4154                 mutex_unlock(&the_lnet.ln_api_mutex);
4155                 return 0;
4156         }
4157
4158         case IOC_LIBCFS_CONFIG_RTR:
4159                 config = arg;
4160
4161                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4162                         return -EINVAL;
4163
4164                 mutex_lock(&the_lnet.ln_api_mutex);
4165                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
4166                         rc = lnet_rtrpools_enable();
4167                         mutex_unlock(&the_lnet.ln_api_mutex);
4168                         return rc;
4169                 }
4170                 lnet_rtrpools_disable();
4171                 mutex_unlock(&the_lnet.ln_api_mutex);
4172                 return 0;
4173
4174         case IOC_LIBCFS_ADD_BUF:
4175                 config = arg;
4176
4177                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4178                         return -EINVAL;
4179
4180                 mutex_lock(&the_lnet.ln_api_mutex);
4181                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
4182                                                 buf_tiny,
4183                                           config->cfg_config_u.cfg_buffers.
4184                                                 buf_small,
4185                                           config->cfg_config_u.cfg_buffers.
4186                                                 buf_large);
4187                 mutex_unlock(&the_lnet.ln_api_mutex);
4188                 return rc;
4189
4190         case IOC_LIBCFS_SET_NUMA_RANGE: {
4191                 struct lnet_ioctl_set_value *numa;
4192                 numa = arg;
4193                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
4194                         return -EINVAL;
4195                 lnet_net_lock(LNET_LOCK_EX);
4196                 lnet_numa_range = numa->sv_value;
4197                 lnet_net_unlock(LNET_LOCK_EX);
4198                 return 0;
4199         }
4200
4201         case IOC_LIBCFS_GET_NUMA_RANGE: {
4202                 struct lnet_ioctl_set_value *numa;
4203                 numa = arg;
4204                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
4205                         return -EINVAL;
4206                 numa->sv_value = lnet_numa_range;
4207                 return 0;
4208         }
4209
4210         case IOC_LIBCFS_GET_BUF: {
4211                 struct lnet_ioctl_pool_cfg *pool_cfg;
4212                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
4213
4214                 config = arg;
4215
4216                 if (config->cfg_hdr.ioc_len < total)
4217                         return -EINVAL;
4218
4219                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
4220
4221                 mutex_lock(&the_lnet.ln_api_mutex);
4222                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
4223                 mutex_unlock(&the_lnet.ln_api_mutex);
4224                 return rc;
4225         }
4226
4227         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
4228                 struct lnet_ioctl_local_ni_hstats *stats = arg;
4229
4230                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
4231                         return -EINVAL;
4232
4233                 mutex_lock(&the_lnet.ln_api_mutex);
4234                 rc = lnet_get_local_ni_hstats(stats);
4235                 mutex_unlock(&the_lnet.ln_api_mutex);
4236
4237                 return rc;
4238         }
4239
4240         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
4241                 struct lnet_ioctl_recovery_list *list = arg;
4242                 if (list->rlst_hdr.ioc_len < sizeof(*list))
4243                         return -EINVAL;
4244
4245                 mutex_lock(&the_lnet.ln_api_mutex);
4246                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
4247                         rc = lnet_get_local_ni_recovery_list(list);
4248                 else
4249                         rc = lnet_get_peer_ni_recovery_list(list);
4250                 mutex_unlock(&the_lnet.ln_api_mutex);
4251                 return rc;
4252         }
4253
4254         case IOC_LIBCFS_ADD_PEER_NI: {
4255                 struct lnet_ioctl_peer_cfg *cfg = arg;
4256
4257                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4258                         return -EINVAL;
4259
4260                 mutex_lock(&the_lnet.ln_api_mutex);
4261                 rc = lnet_add_peer_ni(cfg->prcfg_prim_nid,
4262                                       cfg->prcfg_cfg_nid,
4263                                       cfg->prcfg_mr, false);
4264                 mutex_unlock(&the_lnet.ln_api_mutex);
4265                 return rc;
4266         }
4267
4268         case IOC_LIBCFS_DEL_PEER_NI: {
4269                 struct lnet_ioctl_peer_cfg *cfg = arg;
4270
4271                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4272                         return -EINVAL;
4273
4274                 mutex_lock(&the_lnet.ln_api_mutex);
4275                 rc = lnet_del_peer_ni(cfg->prcfg_prim_nid,
4276                                       cfg->prcfg_cfg_nid);
4277                 mutex_unlock(&the_lnet.ln_api_mutex);
4278                 return rc;
4279         }
4280
4281         case IOC_LIBCFS_GET_PEER_INFO: {
4282                 struct lnet_ioctl_peer *peer_info = arg;
4283
4284                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
4285                         return -EINVAL;
4286
4287                 mutex_lock(&the_lnet.ln_api_mutex);
4288                 rc = lnet_get_peer_ni_info(
4289                    peer_info->pr_count,
4290                    &peer_info->pr_nid,
4291                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
4292                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
4293                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
4294                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
4295                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
4296                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
4297                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
4298                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
4299                 mutex_unlock(&the_lnet.ln_api_mutex);
4300                 return rc;
4301         }
4302
4303         case IOC_LIBCFS_GET_PEER_NI: {
4304                 struct lnet_ioctl_peer_cfg *cfg = arg;
4305
4306                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4307                         return -EINVAL;
4308
4309                 mutex_lock(&the_lnet.ln_api_mutex);
4310                 rc = lnet_get_peer_info(cfg,
4311                                         (void __user *)cfg->prcfg_bulk);
4312                 mutex_unlock(&the_lnet.ln_api_mutex);
4313                 return rc;
4314         }
4315
4316         case IOC_LIBCFS_GET_PEER_LIST: {
4317                 struct lnet_ioctl_peer_cfg *cfg = arg;
4318
4319                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4320                         return -EINVAL;
4321
4322                 mutex_lock(&the_lnet.ln_api_mutex);
4323                 rc = lnet_get_peer_list(&cfg->prcfg_count, &cfg->prcfg_size,
4324                                 (struct lnet_process_id __user *)cfg->prcfg_bulk);
4325                 mutex_unlock(&the_lnet.ln_api_mutex);
4326                 return rc;
4327         }
4328
4329         case IOC_LIBCFS_SET_HEALHV: {
4330                 struct lnet_ioctl_reset_health_cfg *cfg = arg;
4331                 int value;
4332                 if (cfg->rh_hdr.ioc_len < sizeof(*cfg))
4333                         return -EINVAL;
4334                 if (cfg->rh_value < 0 ||
4335                     cfg->rh_value > LNET_MAX_HEALTH_VALUE)
4336                         value = LNET_MAX_HEALTH_VALUE;
4337                 else
4338                         value = cfg->rh_value;
4339                 CDEBUG(D_NET, "Manually setting healthv to %d for %s:%s. all = %d\n",
4340                        value, (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) ?
4341                        "local" : "peer", libcfs_nid2str(cfg->rh_nid), cfg->rh_all);
4342                 mutex_lock(&the_lnet.ln_api_mutex);
4343                 if (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI)
4344                         lnet_ni_set_healthv(cfg->rh_nid, value,
4345                                              cfg->rh_all);
4346                 else
4347                         lnet_peer_ni_set_healthv(cfg->rh_nid, value,
4348                                                   cfg->rh_all);
4349                 mutex_unlock(&the_lnet.ln_api_mutex);
4350                 return 0;
4351         }
4352
4353         case IOC_LIBCFS_SET_CONNS_PER_PEER: {
4354                 struct lnet_ioctl_reset_conns_per_peer_cfg *cfg = arg;
4355                 int value;
4356
4357                 if (cfg->rcpp_hdr.ioc_len < sizeof(*cfg))
4358                         return -EINVAL;
4359                 if (cfg->rcpp_value < 0)
4360                         value = 1;
4361                 else
4362                         value = cfg->rcpp_value;
4363                 CDEBUG(D_NET,
4364                        "Setting conns_per_peer to %d for %s. all = %d\n",
4365                        value, libcfs_nid2str(cfg->rcpp_nid), cfg->rcpp_all);
4366                 mutex_lock(&the_lnet.ln_api_mutex);
4367                 lnet_ni_set_conns_per_peer(cfg->rcpp_nid, value, cfg->rcpp_all);
4368                 mutex_unlock(&the_lnet.ln_api_mutex);
4369                 return 0;
4370         }
4371
4372         case IOC_LIBCFS_NOTIFY_ROUTER: {
4373                 time64_t deadline = ktime_get_real_seconds() - data->ioc_u64[0];
4374
4375                 /* The deadline passed in by the user should be some time in
4376                  * seconds in the future since the UNIX epoch. We have to map
4377                  * that deadline to the wall clock.
4378                  */
4379                 deadline += ktime_get_seconds();
4380                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags, false,
4381                                    deadline);
4382         }
4383
4384         case IOC_LIBCFS_LNET_DIST:
4385                 lnet_nid4_to_nid(data->ioc_nid, &nid);
4386                 rc = LNetDist(&nid, &nid, &data->ioc_u32[1]);
4387                 if (rc < 0 && rc != -EHOSTUNREACH)
4388                         return rc;
4389
4390                 data->ioc_nid = lnet_nid_to_nid4(&nid);
4391                 data->ioc_u32[0] = rc;
4392                 return 0;
4393
4394         case IOC_LIBCFS_TESTPROTOCOMPAT:
4395                 the_lnet.ln_testprotocompat = data->ioc_flags;
4396                 return 0;
4397
4398         case IOC_LIBCFS_LNET_FAULT:
4399                 return lnet_fault_ctl(data->ioc_flags, data);
4400
4401         case IOC_LIBCFS_PING: {
4402                 signed long timeout;
4403
4404                 id4.nid = data->ioc_nid;
4405                 id4.pid = data->ioc_u32[0];
4406
4407                 /* If timeout is negative then set default of 3 minutes */
4408                 if (((s32)data->ioc_u32[1] <= 0) ||
4409                     data->ioc_u32[1] > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
4410                         timeout = cfs_time_seconds(DEFAULT_PEER_TIMEOUT);
4411                 else
4412                         timeout = nsecs_to_jiffies(data->ioc_u32[1] * NSEC_PER_MSEC);
4413
4414                 rc = lnet_ping(id4, &LNET_ANY_NID, timeout, data->ioc_pbuf1,
4415                                data->ioc_plen1 / sizeof(struct lnet_process_id));
4416
4417                 if (rc < 0)
4418                         return rc;
4419
4420                 data->ioc_count = rc;
4421                 return 0;
4422         }
4423
4424         case IOC_LIBCFS_PING_PEER: {
4425                 struct lnet_ioctl_ping_data *ping = arg;
4426                 struct lnet_nid src_nid = LNET_ANY_NID;
4427                 struct lnet_peer *lp;
4428                 signed long timeout;
4429
4430                 /* Check if the supplied ping data supports source nid
4431                  * NB: This check is sufficient if lnet_ioctl_ping_data has
4432                  * additional fields added, but if they are re-ordered or
4433                  * fields removed then this will break. It is expected that
4434                  * these ioctls will be replaced with netlink implementation, so
4435                  * it is probably not worth coming up with a more robust version
4436                  * compatibility scheme.
4437                  */
4438                 if (ping->ping_hdr.ioc_len >= sizeof(struct lnet_ioctl_ping_data))
4439                         lnet_nid4_to_nid(ping->ping_src, &src_nid);
4440
4441                 /* If timeout is negative then set default of 3 minutes */
4442                 if (((s32)ping->op_param) <= 0 ||
4443                     ping->op_param > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
4444                         timeout = cfs_time_seconds(DEFAULT_PEER_TIMEOUT);
4445                 else
4446                         timeout = nsecs_to_jiffies(ping->op_param * NSEC_PER_MSEC);
4447
4448                 rc = lnet_ping(ping->ping_id, &src_nid, timeout,
4449                                ping->ping_buf,
4450                                ping->ping_count);
4451                 if (rc < 0)
4452                         return rc;
4453
4454                 mutex_lock(&the_lnet.ln_api_mutex);
4455                 lp = lnet_find_peer4(ping->ping_id.nid);
4456                 if (lp) {
4457                         ping->ping_id.nid =
4458                                 lnet_nid_to_nid4(&lp->lp_primary_nid);
4459                         ping->mr_info = lnet_peer_is_multi_rail(lp);
4460                         lnet_peer_decref_locked(lp);
4461                 }
4462                 mutex_unlock(&the_lnet.ln_api_mutex);
4463
4464                 ping->ping_count = rc;
4465                 return 0;
4466         }
4467
4468         case IOC_LIBCFS_DISCOVER: {
4469                 struct lnet_ioctl_ping_data *discover = arg;
4470                 struct lnet_peer *lp;
4471
4472                 rc = lnet_discover(discover->ping_id, discover->op_param,
4473                                    discover->ping_buf,
4474                                    discover->ping_count);
4475                 if (rc < 0)
4476                         return rc;
4477
4478                 mutex_lock(&the_lnet.ln_api_mutex);
4479                 lp = lnet_find_peer4(discover->ping_id.nid);
4480                 if (lp) {
4481                         discover->ping_id.nid =
4482                                 lnet_nid_to_nid4(&lp->lp_primary_nid);
4483                         discover->mr_info = lnet_peer_is_multi_rail(lp);
4484                         lnet_peer_decref_locked(lp);
4485                 }
4486                 mutex_unlock(&the_lnet.ln_api_mutex);
4487
4488                 discover->ping_count = rc;
4489                 return 0;
4490         }
4491
4492         case IOC_LIBCFS_ADD_UDSP: {
4493                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4494                 __u32 bulk_size = ioc_udsp->iou_hdr.ioc_len;
4495
4496                 mutex_lock(&the_lnet.ln_api_mutex);
4497                 rc = lnet_udsp_demarshal_add(arg, bulk_size);
4498                 if (!rc) {
4499                         rc = lnet_udsp_apply_policies(NULL, false);
4500                         CDEBUG(D_NET, "policy application returned %d\n", rc);
4501                         rc = 0;
4502                 }
4503                 mutex_unlock(&the_lnet.ln_api_mutex);
4504
4505                 return rc;
4506         }
4507
4508         case IOC_LIBCFS_DEL_UDSP: {
4509                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4510                 int idx = ioc_udsp->iou_idx;
4511
4512                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4513                         return -EINVAL;
4514
4515                 mutex_lock(&the_lnet.ln_api_mutex);
4516                 rc = lnet_udsp_del_policy(idx);
4517                 if (!rc) {
4518                         rc = lnet_udsp_apply_policies(NULL, false);
4519                         CDEBUG(D_NET, "policy re-application returned %d\n",
4520                                rc);
4521                         rc = 0;
4522                 }
4523                 mutex_unlock(&the_lnet.ln_api_mutex);
4524
4525                 return rc;
4526         }
4527
4528         case IOC_LIBCFS_GET_UDSP_SIZE: {
4529                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4530                 struct lnet_udsp *udsp;
4531
4532                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4533                         return -EINVAL;
4534
4535                 rc = 0;
4536
4537                 mutex_lock(&the_lnet.ln_api_mutex);
4538                 udsp = lnet_udsp_get_policy(ioc_udsp->iou_idx);
4539                 if (!udsp) {
4540                         rc = -ENOENT;
4541                 } else {
4542                         /* coming in iou_idx will hold the idx of the udsp
4543                          * to get the size of. going out the iou_idx will
4544                          * hold the size of the UDSP found at the passed
4545                          * in index.
4546                          */
4547                         ioc_udsp->iou_idx = lnet_get_udsp_size(udsp);
4548                         if (ioc_udsp->iou_idx < 0)
4549                                 rc = -EINVAL;
4550                 }
4551                 mutex_unlock(&the_lnet.ln_api_mutex);
4552
4553                 return rc;
4554         }
4555
4556         case IOC_LIBCFS_GET_UDSP: {
4557                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4558                 struct lnet_udsp *udsp;
4559
4560                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4561                         return -EINVAL;
4562
4563                 rc = 0;
4564
4565                 mutex_lock(&the_lnet.ln_api_mutex);
4566                 udsp = lnet_udsp_get_policy(ioc_udsp->iou_idx);
4567                 if (!udsp)
4568                         rc = -ENOENT;
4569                 else
4570                         rc = lnet_udsp_marshal(udsp, ioc_udsp);
4571                 mutex_unlock(&the_lnet.ln_api_mutex);
4572
4573                 return rc;
4574         }
4575
4576         case IOC_LIBCFS_GET_CONST_UDSP_INFO: {
4577                 struct lnet_ioctl_construct_udsp_info *info = arg;
4578
4579                 if (info->cud_hdr.ioc_len < sizeof(*info))
4580                         return -EINVAL;
4581
4582                 CDEBUG(D_NET, "GET_UDSP_INFO for %s\n",
4583                        libcfs_nid2str(info->cud_nid));
4584
4585                 mutex_lock(&the_lnet.ln_api_mutex);
4586                 lnet_udsp_get_construct_info(info);
4587                 mutex_unlock(&the_lnet.ln_api_mutex);
4588
4589                 return 0;
4590         }
4591
4592         default:
4593                 ni = lnet_net2ni_addref(data->ioc_net);
4594                 if (ni == NULL)
4595                         return -EINVAL;
4596
4597                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
4598                         rc = -EINVAL;
4599                 else
4600                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
4601
4602                 lnet_ni_decref(ni);
4603                 return rc;
4604         }
4605         /* not reached */
4606 }
4607 EXPORT_SYMBOL(LNetCtl);
4608
4609 void LNetDebugPeer(struct lnet_processid *id)
4610 {
4611         lnet_debug_peer(lnet_nid_to_nid4(&id->nid));
4612 }
4613 EXPORT_SYMBOL(LNetDebugPeer);
4614
4615 /**
4616  * Determine if the specified peer \a nid is on the local node.
4617  *
4618  * \param nid   peer nid to check
4619  *
4620  * \retval true         If peer NID is on the local node.
4621  * \retval false        If peer NID is not on the local node.
4622  */
4623 bool LNetIsPeerLocal(struct lnet_nid *nid)
4624 {
4625         struct lnet_net *net;
4626         struct lnet_ni *ni;
4627         int cpt;
4628
4629         cpt = lnet_net_lock_current();
4630         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
4631                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
4632                         if (nid_same(&ni->ni_nid, nid)) {
4633                                 lnet_net_unlock(cpt);
4634                                 return true;
4635                         }
4636                 }
4637         }
4638         lnet_net_unlock(cpt);
4639
4640         return false;
4641 }
4642 EXPORT_SYMBOL(LNetIsPeerLocal);
4643
4644 /**
4645  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
4646  * Note that all interfaces share a same PID, as requested by LNetNIInit().
4647  *
4648  * \param index Index of the interface to look up.
4649  * \param id On successful return, this location will hold the
4650  * struct lnet_process_id ID of the interface.
4651  *
4652  * \retval 0 If an interface exists at \a index.
4653  * \retval -ENOENT If no interface has been found.
4654  */
4655 int
4656 LNetGetId(unsigned int index, struct lnet_processid *id)
4657 {
4658         struct lnet_ni   *ni;
4659         struct lnet_net  *net;
4660         int               cpt;
4661         int               rc = -ENOENT;
4662
4663         LASSERT(the_lnet.ln_refcount > 0);
4664
4665         cpt = lnet_net_lock_current();
4666
4667         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
4668                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
4669                         if (!nid_is_nid4(&ni->ni_nid))
4670                                 /* FIXME this needs to be handled */
4671                                 continue;
4672                         if (index-- != 0)
4673                                 continue;
4674
4675                         id->nid = ni->ni_nid;
4676                         id->pid = the_lnet.ln_pid;
4677                         rc = 0;
4678                         break;
4679                 }
4680         }
4681
4682         lnet_net_unlock(cpt);
4683         return rc;
4684 }
4685 EXPORT_SYMBOL(LNetGetId);
4686
4687 struct ping_data {
4688         int rc;
4689         int replied;
4690         struct lnet_handle_md mdh;
4691         struct completion completion;
4692 };
4693
4694 static void
4695 lnet_ping_event_handler(struct lnet_event *event)
4696 {
4697         struct ping_data *pd = event->md_user_ptr;
4698
4699         CDEBUG(D_NET, "ping event (%d %d)%s\n",
4700                event->type, event->status,
4701                event->unlinked ? " unlinked" : "");
4702
4703         if (event->status) {
4704                 if (!pd->rc)
4705                         pd->rc = event->status;
4706         } else if (event->type == LNET_EVENT_REPLY) {
4707                 pd->replied = 1;
4708                 pd->rc = event->mlength;
4709         }
4710         if (event->unlinked)
4711                 complete(&pd->completion);
4712 }
4713
4714 static int lnet_ping(struct lnet_process_id id4, struct lnet_nid *src_nid,
4715                      signed long timeout, struct lnet_process_id __user *ids,
4716                      int n_ids)
4717 {
4718         struct lnet_md md = { NULL };
4719         struct ping_data pd = { 0 };
4720         struct lnet_ping_buffer *pbuf;
4721         struct lnet_process_id tmpid;
4722         struct lnet_processid id;
4723         int i;
4724         int nob;
4725         int rc;
4726         int rc2;
4727
4728         /* n_ids limit is arbitrary */
4729         if (n_ids <= 0 || id4.nid == LNET_NID_ANY)
4730                 return -EINVAL;
4731
4732         /*
4733          * if the user buffer has more space than the lnet_interfaces_max
4734          * then only fill it up to lnet_interfaces_max
4735          */
4736         if (n_ids > lnet_interfaces_max)
4737                 n_ids = lnet_interfaces_max;
4738
4739         if (id4.pid == LNET_PID_ANY)
4740                 id4.pid = LNET_PID_LUSTRE;
4741
4742         pbuf = lnet_ping_buffer_alloc(n_ids, GFP_NOFS);
4743         if (!pbuf)
4744                 return -ENOMEM;
4745
4746         /* initialize md content */
4747         md.start     = &pbuf->pb_info;
4748         md.length    = LNET_PING_INFO_SIZE(n_ids);
4749         md.threshold = 2; /* GET/REPLY */
4750         md.max_size  = 0;
4751         md.options   = LNET_MD_TRUNCATE;
4752         md.user_ptr  = &pd;
4753         md.handler   = lnet_ping_event_handler;
4754
4755         init_completion(&pd.completion);
4756
4757         rc = LNetMDBind(&md, LNET_UNLINK, &pd.mdh);
4758         if (rc != 0) {
4759                 CERROR("Can't bind MD: %d\n", rc);
4760                 goto fail_ping_buffer_decref;
4761         }
4762
4763         lnet_pid4_to_pid(id4, &id);
4764         rc = LNetGet(src_nid, pd.mdh, &id, LNET_RESERVED_PORTAL,
4765                      LNET_PROTO_PING_MATCHBITS, 0, false);
4766
4767         if (rc != 0) {
4768                 /* Don't CERROR; this could be deliberate! */
4769                 rc2 = LNetMDUnlink(pd.mdh);
4770                 LASSERT(rc2 == 0);
4771
4772                 /* NB must wait for the UNLINK event below... */
4773         }
4774
4775         if (wait_for_completion_timeout(&pd.completion, timeout) == 0) {
4776                 /* Ensure completion in finite time... */
4777                 LNetMDUnlink(pd.mdh);
4778                 wait_for_completion(&pd.completion);
4779         }
4780         if (!pd.replied) {
4781                 rc = -EIO;
4782                 goto fail_ping_buffer_decref;
4783         }
4784
4785         nob = pd.rc;
4786         LASSERT(nob >= 0 && nob <= LNET_PING_INFO_SIZE(n_ids));
4787
4788         rc = -EPROTO;           /* if I can't parse... */
4789
4790         if (nob < 8) {
4791                 CERROR("%s: ping info too short %d\n",
4792                        libcfs_id2str(id4), nob);
4793                 goto fail_ping_buffer_decref;
4794         }
4795
4796         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
4797                 lnet_swap_pinginfo(pbuf);
4798         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
4799                 CERROR("%s: Unexpected magic %08x\n",
4800                        libcfs_id2str(id4), pbuf->pb_info.pi_magic);
4801                 goto fail_ping_buffer_decref;
4802         }
4803
4804         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
4805                 CERROR("%s: ping w/o NI status: 0x%x\n",
4806                        libcfs_id2str(id4), pbuf->pb_info.pi_features);
4807                 goto fail_ping_buffer_decref;
4808         }
4809
4810         if (nob < LNET_PING_INFO_SIZE(0)) {
4811                 CERROR("%s: Short reply %d(%d min)\n",
4812                        libcfs_id2str(id4),
4813                        nob, (int)LNET_PING_INFO_SIZE(0));
4814                 goto fail_ping_buffer_decref;
4815         }
4816
4817         if (pbuf->pb_info.pi_nnis < n_ids)
4818                 n_ids = pbuf->pb_info.pi_nnis;
4819
4820         if (nob < LNET_PING_INFO_SIZE(n_ids)) {
4821                 CERROR("%s: Short reply %d(%d expected)\n",
4822                        libcfs_id2str(id4),
4823                        nob, (int)LNET_PING_INFO_SIZE(n_ids));
4824                 goto fail_ping_buffer_decref;
4825         }
4826
4827         rc = -EFAULT;           /* if I segv in copy_to_user()... */
4828
4829         memset(&tmpid, 0, sizeof(tmpid));
4830         for (i = 0; i < n_ids; i++) {
4831                 tmpid.pid = pbuf->pb_info.pi_pid;
4832                 tmpid.nid = pbuf->pb_info.pi_ni[i].ns_nid;
4833                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
4834                         goto fail_ping_buffer_decref;
4835         }
4836         rc = pbuf->pb_info.pi_nnis;
4837
4838  fail_ping_buffer_decref:
4839         lnet_ping_buffer_decref(pbuf);
4840         return rc;
4841 }
4842
4843 static int
4844 lnet_discover(struct lnet_process_id id, __u32 force,
4845               struct lnet_process_id __user *ids, int n_ids)
4846 {
4847         struct lnet_peer_ni *lpni;
4848         struct lnet_peer_ni *p;
4849         struct lnet_peer *lp;
4850         struct lnet_process_id *buf;
4851         int cpt;
4852         int i;
4853         int rc;
4854
4855         if (n_ids <= 0 ||
4856             id.nid == LNET_NID_ANY)
4857                 return -EINVAL;
4858
4859         if (id.pid == LNET_PID_ANY)
4860                 id.pid = LNET_PID_LUSTRE;
4861
4862         /*
4863          * If the user buffer has more space than the lnet_interfaces_max,
4864          * then only fill it up to lnet_interfaces_max.
4865          */
4866         if (n_ids > lnet_interfaces_max)
4867                 n_ids = lnet_interfaces_max;
4868
4869         CFS_ALLOC_PTR_ARRAY(buf, n_ids);
4870         if (!buf)
4871                 return -ENOMEM;
4872
4873         cpt = lnet_net_lock_current();
4874         lpni = lnet_nid2peerni_locked(id.nid, LNET_NID_ANY, cpt);
4875         if (IS_ERR(lpni)) {
4876                 rc = PTR_ERR(lpni);
4877                 goto out;
4878         }
4879
4880         /*
4881          * Clearing the NIDS_UPTODATE flag ensures the peer will
4882          * be discovered, provided discovery has not been disabled.
4883          */
4884         lp = lpni->lpni_peer_net->lpn_peer;
4885         spin_lock(&lp->lp_lock);
4886         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
4887         /* If the force flag is set, force a PING and PUSH as well. */
4888         if (force)
4889                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
4890         spin_unlock(&lp->lp_lock);
4891         rc = lnet_discover_peer_locked(lpni, cpt, true);
4892         if (rc)
4893                 goto out_decref;
4894
4895         /* The lpni (or lp) for this NID may have changed and our ref is
4896          * the only thing keeping the old one around. Release the ref
4897          * and lookup the lpni again
4898          */
4899         lnet_peer_ni_decref_locked(lpni);
4900         lpni = lnet_find_peer_ni_locked(id.nid);
4901         if (!lpni) {
4902                 rc = -ENOENT;
4903                 goto out;
4904         }
4905         lp = lpni->lpni_peer_net->lpn_peer;
4906
4907         i = 0;
4908         p = NULL;
4909         while ((p = lnet_get_next_peer_ni_locked(lp, NULL, p)) != NULL) {
4910                 buf[i].pid = id.pid;
4911                 buf[i].nid = lnet_nid_to_nid4(&p->lpni_nid);
4912                 if (++i >= n_ids)
4913                         break;
4914         }
4915         rc = i;
4916
4917 out_decref:
4918         lnet_peer_ni_decref_locked(lpni);
4919 out:
4920         lnet_net_unlock(cpt);
4921
4922         if (rc >= 0)
4923                 if (copy_to_user(ids, buf, rc * sizeof(*buf)))
4924                         rc = -EFAULT;
4925         CFS_FREE_PTR_ARRAY(buf, n_ids);
4926
4927         return rc;
4928 }
4929
4930 /**
4931  * Retrieve peer discovery status.
4932  *
4933  * \retval 1 if lnet_peer_discovery_disabled is 0
4934  * \retval 0 if lnet_peer_discovery_disabled is 1
4935  */
4936 int
4937 LNetGetPeerDiscoveryStatus(void)
4938 {
4939         return !lnet_peer_discovery_disabled;
4940 }
4941 EXPORT_SYMBOL(LNetGetPeerDiscoveryStatus);