lnet/lnet/api-ni.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   4  * Use is subject to license terms.
   5  *
   6  * Copyright (c) 2011, 2017, Intel Corporation.
   7  */
   8
   9 /* This file is part of Lustre, http://www.lustre.org/ */
  10
  11 #define DEBUG_SUBSYSTEM S_LNET
  12
  13 #include <linux/ctype.h>
  14 #include <linux/generic-radix-tree.h>
  15 #include <linux/log2.h>
  16 #include <linux/ktime.h>
  17 #include <linux/moduleparam.h>
  18 #include <linux/uaccess.h>
  19 #ifdef HAVE_SCHED_HEADERS
  20 #include <linux/sched/signal.h>
  21 #endif
  22 #include <net/genetlink.h>
  23
  24 #include <libcfs/linux/linux-net.h>
  25 #include <lnet/udsp.h>
  26 #include <lnet/lib-lnet.h>
  27
  28 #define D_LNI D_CONSOLE
  29
  30 /*
  31  * initialize ln_api_mutex statically, since it needs to be used in
  32  * discovery_set callback. That module parameter callback can be called
  33  * before module init completes. The mutex needs to be ready for use then.
  34  */
  35 struct lnet the_lnet = {
  36         .ln_api_mutex = __MUTEX_INITIALIZER(the_lnet.ln_api_mutex),
  37 };              /* THE state of the network */
  38 EXPORT_SYMBOL(the_lnet);
  39
  40 static char *ip2nets = "";
  41 module_param(ip2nets, charp, 0444);
  42 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  43
  44 static char *networks = "";
  45 module_param(networks, charp, 0444);
  46 MODULE_PARM_DESC(networks, "local networks");
  47
  48 static char *routes = "";
  49 module_param(routes, charp, 0444);
  50 MODULE_PARM_DESC(routes, "routes to non-local networks");
  51
  52 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  53 module_param(rnet_htable_size, int, 0444);
  54 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  55
  56 static int use_tcp_bonding;
  57 module_param(use_tcp_bonding, int, 0444);
  58 MODULE_PARM_DESC(use_tcp_bonding,
  59                  "use_tcp_bonding parameter has been removed");
  60
  61 unsigned int lnet_numa_range = 0;
  62 module_param(lnet_numa_range, uint, 0444);
  63 MODULE_PARM_DESC(lnet_numa_range,
  64                 "NUMA range to consider during Multi-Rail selection");
  65
  66 /*
  67  * lnet_health_sensitivity determines by how much we decrement the health
  68  * value on sending error. The value defaults to 100, which means health
  69  * interface health is decremented by 100 points every failure.
  70  */
  71 unsigned int lnet_health_sensitivity = 100;
  72 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
  73 #ifdef HAVE_KERNEL_PARAM_OPS
  74 static struct kernel_param_ops param_ops_health_sensitivity = {
  75         .set = sensitivity_set,
  76         .get = param_get_int,
  77 };
  78 #define param_check_health_sensitivity(name, p) \
  79                 __param_check(name, p, int)
  80 module_param(lnet_health_sensitivity, health_sensitivity, S_IRUGO|S_IWUSR);
  81 #else
  82 module_param_call(lnet_health_sensitivity, sensitivity_set, param_get_int,
  83                   &lnet_health_sensitivity, S_IRUGO|S_IWUSR);
  84 #endif
  85 MODULE_PARM_DESC(lnet_health_sensitivity,
  86                 "Value to decrement the health value by on error");
  87
  88 /*
  89  * lnet_recovery_interval determines how often we should perform recovery
  90  * on unhealthy interfaces.
  91  */
  92 unsigned int lnet_recovery_interval = 1;
  93 static int recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp);
  94 #ifdef HAVE_KERNEL_PARAM_OPS
  95 static struct kernel_param_ops param_ops_recovery_interval = {
  96         .set = recovery_interval_set,
  97         .get = param_get_int,
  98 };
  99 #define param_check_recovery_interval(name, p) \
 100                 __param_check(name, p, int)
 101 module_param(lnet_recovery_interval, recovery_interval, S_IRUGO|S_IWUSR);
 102 #else
 103 module_param_call(lnet_recovery_interval, recovery_interval_set, param_get_int,
 104                   &lnet_recovery_interval, S_IRUGO|S_IWUSR);
 105 #endif
 106 MODULE_PARM_DESC(lnet_recovery_interval,
 107                 "DEPRECATED - Interval to recover unhealthy interfaces in seconds");
 108
 109 unsigned int lnet_recovery_limit;
 110 module_param(lnet_recovery_limit, uint, 0644);
 111 MODULE_PARM_DESC(lnet_recovery_limit,
 112                  "How long to attempt recovery of unhealthy peer interfaces in seconds. Set to 0 to allow indefinite recovery");
 113
 114 unsigned int lnet_max_recovery_ping_interval = 900;
 115 unsigned int lnet_max_recovery_ping_count = 9;
 116 static int max_recovery_ping_interval_set(const char *val,
 117                                           cfs_kernel_param_arg_t *kp);
 118
 119 #define param_check_max_recovery_ping_interval(name, p) \
 120                 __param_check(name, p, int)
 121
 122 #ifdef HAVE_KERNEL_PARAM_OPS
 123 static struct kernel_param_ops param_ops_max_recovery_ping_interval = {
 124         .set = max_recovery_ping_interval_set,
 125         .get = param_get_int,
 126 };
 127 module_param(lnet_max_recovery_ping_interval, max_recovery_ping_interval, 0644);
 128 #else
 129 module_param_call(lnet_max_recovery_ping_interval, max_recovery_ping_interval,
 130                   param_get_int, &lnet_max_recovery_ping_interval, 0644);
 131 #endif
 132 MODULE_PARM_DESC(lnet_max_recovery_ping_interval,
 133                  "The max interval between LNet recovery pings, in seconds");
 134
 135 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
 136 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
 137
 138 static struct kernel_param_ops param_ops_interfaces_max = {
 139         .set = intf_max_set,
 140         .get = param_get_int,
 141 };
 142
 143 #define param_check_interfaces_max(name, p) \
 144                 __param_check(name, p, int)
 145
 146 #ifdef HAVE_KERNEL_PARAM_OPS
 147 module_param(lnet_interfaces_max, interfaces_max, 0644);
 148 #else
 149 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
 150                   &param_ops_interfaces_max, 0644);
 151 #endif
 152 MODULE_PARM_DESC(lnet_interfaces_max,
 153                 "Maximum number of interfaces in a node.");
 154
 155 unsigned lnet_peer_discovery_disabled = 0;
 156 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
 157
 158 static struct kernel_param_ops param_ops_discovery_disabled = {
 159         .set = discovery_set,
 160         .get = param_get_int,
 161 };
 162
 163 #define param_check_discovery_disabled(name, p) \
 164                 __param_check(name, p, int)
 165 #ifdef HAVE_KERNEL_PARAM_OPS
 166 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
 167 #else
 168 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
 169                   &param_ops_discovery_disabled, 0644);
 170 #endif
 171 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 172                 "Set to 1 to disable peer discovery on this node.");
 173
 174 unsigned int lnet_drop_asym_route;
 175 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
 176
 177 static struct kernel_param_ops param_ops_drop_asym_route = {
 178         .set = drop_asym_route_set,
 179         .get = param_get_int,
 180 };
 181
 182 #define param_check_drop_asym_route(name, p)    \
 183         __param_check(name, p, int)
 184 #ifdef HAVE_KERNEL_PARAM_OPS
 185 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
 186 #else
 187 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
 188                   &param_ops_drop_asym_route, 0644);
 189 #endif
 190 MODULE_PARM_DESC(lnet_drop_asym_route,
 191                  "Set to 1 to drop asymmetrical route messages.");
 192
 193 #define LNET_TRANSACTION_TIMEOUT_DEFAULT 150
 194 unsigned int lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_DEFAULT;
 195 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
 196 #ifdef HAVE_KERNEL_PARAM_OPS
 197 static struct kernel_param_ops param_ops_transaction_timeout = {
 198         .set = transaction_to_set,
 199         .get = param_get_int,
 200 };
 201
 202 #define param_check_transaction_timeout(name, p) \
 203                 __param_check(name, p, int)
 204 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
 205 #else
 206 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
 207                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
 208 #endif
 209 MODULE_PARM_DESC(lnet_transaction_timeout,
 210                 "Maximum number of seconds to wait for a peer response.");
 211
 212 #define LNET_RETRY_COUNT_DEFAULT 2
 213 unsigned int lnet_retry_count = LNET_RETRY_COUNT_DEFAULT;
 214 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
 215 #ifdef HAVE_KERNEL_PARAM_OPS
 216 static struct kernel_param_ops param_ops_retry_count = {
 217         .set = retry_count_set,
 218         .get = param_get_int,
 219 };
 220
 221 #define param_check_retry_count(name, p) \
 222                 __param_check(name, p, int)
 223 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
 224 #else
 225 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
 226                   &lnet_retry_count, S_IRUGO|S_IWUSR);
 227 #endif
 228 MODULE_PARM_DESC(lnet_retry_count,
 229                  "Maximum number of times to retry transmitting a message");
 230
 231 unsigned int lnet_response_tracking = 3;
 232 static int response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp);
 233
 234 #ifdef HAVE_KERNEL_PARAM_OPS
 235 static struct kernel_param_ops param_ops_response_tracking = {
 236         .set = response_tracking_set,
 237         .get = param_get_int,
 238 };
 239
 240 #define param_check_response_tracking(name, p)  \
 241         __param_check(name, p, int)
 242 module_param(lnet_response_tracking, response_tracking, 0644);
 243 #else
 244 module_param_call(lnet_response_tracking, response_tracking_set, param_get_int,
 245                   &lnet_response_tracking, 0644);
 246 #endif
 247 MODULE_PARM_DESC(lnet_response_tracking,
 248                  "(0|1|2|3) LNet Internal Only|GET Reply only|PUT ACK only|Full Tracking (default)");
 249
 250 int lock_prim_nid = 1;
 251 module_param(lock_prim_nid, int, 0444);
 252 MODULE_PARM_DESC(lock_prim_nid,
 253                  "Whether nid passed down by Lustre is locked as primary");
 254
 255 #define LNET_LND_TIMEOUT_DEFAULT ((LNET_TRANSACTION_TIMEOUT_DEFAULT - 1) / \
 256                                   (LNET_RETRY_COUNT_DEFAULT + 1))
 257 unsigned int lnet_lnd_timeout = LNET_LND_TIMEOUT_DEFAULT;
 258 static void lnet_set_lnd_timeout(void)
 259 {
 260         lnet_lnd_timeout = max((lnet_transaction_timeout - 1) /
 261                                (lnet_retry_count + 1), 1U);
 262 }
 263
 264 /*
 265  * This sequence number keeps track of how many times DLC was used to
 266  * update the local NIs. It is incremented when a NI is added or
 267  * removed and checked when sending a message to determine if there is
 268  * a need to re-run the selection algorithm. See lnet_select_pathway()
 269  * for more details on its usage.
 270  */
 271 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
 272
 273 struct lnet_fail_ping {
 274         struct lnet_processid           lfp_id;
 275         int                             lfp_errno;
 276         char                            lfp_msg[256];
 277 };
 278
 279 struct lnet_genl_ping_list {
 280         unsigned int                    lgpl_index;
 281         unsigned int                    lgpl_list_count;
 282         unsigned int                    lgpl_failed_count;
 283         signed long                     lgpl_timeout;
 284         struct lnet_nid                 lgpl_src_nid;
 285         GENRADIX(struct lnet_fail_ping) lgpl_failed;
 286         GENRADIX(struct lnet_processid) lgpl_list;
 287 };
 288
 289 static int lnet_ping(struct lnet_processid *id, struct lnet_nid *src_nid,
 290                      signed long timeout, struct lnet_genl_ping_list *plist,
 291                      int n_ids);
 292
 293 static int lnet_discover(struct lnet_processid *id, u32 force,
 294                          struct lnet_genl_ping_list *dlists);
 295
 296 static int
 297 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 298 {
 299         int rc;
 300         unsigned *sensitivity = (unsigned *)kp->arg;
 301         unsigned long value;
 302
 303         rc = kstrtoul(val, 0, &value);
 304         if (rc) {
 305                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
 306                 return rc;
 307         }
 308
 309         /*
 310          * The purpose of locking the api_mutex here is to ensure that
 311          * the correct value ends up stored properly.
 312          */
 313         mutex_lock(&the_lnet.ln_api_mutex);
 314
 315         if (value > LNET_MAX_HEALTH_VALUE) {
 316                 mutex_unlock(&the_lnet.ln_api_mutex);
 317                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
 318                        LNET_MAX_HEALTH_VALUE, value);
 319                 return -EINVAL;
 320         }
 321
 322         if (*sensitivity != 0 && value == 0 && lnet_retry_count != 0) {
 323                 lnet_retry_count = 0;
 324                 lnet_set_lnd_timeout();
 325         }
 326
 327         *sensitivity = value;
 328
 329         mutex_unlock(&the_lnet.ln_api_mutex);
 330
 331         return 0;
 332 }
 333
 334 static int
 335 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 336 {
 337         CWARN("'lnet_recovery_interval' has been deprecated\n");
 338
 339         return 0;
 340 }
 341
 342 static int
 343 max_recovery_ping_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 344 {
 345         int rc;
 346         unsigned long value;
 347
 348         rc = kstrtoul(val, 0, &value);
 349         if (rc) {
 350                 CERROR("Invalid module parameter value for 'lnet_max_recovery_ping_interval'\n");
 351                 return rc;
 352         }
 353
 354         if (!value) {
 355                 CERROR("Invalid max ping timeout. Must be strictly positive\n");
 356                 return -EINVAL;
 357         }
 358
 359         /* The purpose of locking the api_mutex here is to ensure that
 360          * the correct value ends up stored properly.
 361          */
 362         mutex_lock(&the_lnet.ln_api_mutex);
 363         lnet_max_recovery_ping_interval = value;
 364         lnet_max_recovery_ping_count = 0;
 365         value >>= 1;
 366         while (value) {
 367                 lnet_max_recovery_ping_count++;
 368                 value >>= 1;
 369         }
 370         mutex_unlock(&the_lnet.ln_api_mutex);
 371
 372         return 0;
 373 }
 374
 375 static int
 376 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
 377 {
 378         int rc;
 379         unsigned *discovery_off = (unsigned *)kp->arg;
 380         unsigned long value;
 381         struct lnet_ping_buffer *pbuf;
 382
 383         rc = kstrtoul(val, 0, &value);
 384         if (rc) {
 385                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
 386                 return rc;
 387         }
 388
 389         value = (value) ? 1 : 0;
 390
 391         /*
 392          * The purpose of locking the api_mutex here is to ensure that
 393          * the correct value ends up stored properly.
 394          */
 395         mutex_lock(&the_lnet.ln_api_mutex);
 396
 397         if (value == *discovery_off) {
 398                 mutex_unlock(&the_lnet.ln_api_mutex);
 399                 return 0;
 400         }
 401
 402         /*
 403          * We still want to set the discovery value even when LNet is not
 404          * running. This is the case when LNet is being loaded and we want
 405          * the module parameters to take effect. Otherwise if we're
 406          * changing the value dynamically, we want to set it after
 407          * updating the peers
 408          */
 409         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 410                 *discovery_off = value;
 411                 mutex_unlock(&the_lnet.ln_api_mutex);
 412                 return 0;
 413         }
 414
 415         /* tell peers that discovery setting has changed */
 416         lnet_net_lock(LNET_LOCK_EX);
 417         pbuf = the_lnet.ln_ping_target;
 418         if (value)
 419                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
 420         else
 421                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
 422         lnet_net_unlock(LNET_LOCK_EX);
 423
 424         /* only send a push when we're turning off discovery */
 425         if (*discovery_off <= 0 && value > 0)
 426                 lnet_push_update_to_peers(1);
 427         *discovery_off = value;
 428
 429         mutex_unlock(&the_lnet.ln_api_mutex);
 430
 431         return 0;
 432 }
 433
 434 static int
 435 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
 436 {
 437         int rc;
 438         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
 439         unsigned long value;
 440
 441         rc = kstrtoul(val, 0, &value);
 442         if (rc) {
 443                 CERROR("Invalid module parameter value for "
 444                        "'lnet_drop_asym_route'\n");
 445                 return rc;
 446         }
 447
 448         /*
 449          * The purpose of locking the api_mutex here is to ensure that
 450          * the correct value ends up stored properly.
 451          */
 452         mutex_lock(&the_lnet.ln_api_mutex);
 453
 454         if (value == *drop_asym_route) {
 455                 mutex_unlock(&the_lnet.ln_api_mutex);
 456                 return 0;
 457         }
 458
 459         *drop_asym_route = value;
 460
 461         mutex_unlock(&the_lnet.ln_api_mutex);
 462
 463         return 0;
 464 }
 465
 466 static int
 467 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
 468 {
 469         int rc;
 470         unsigned *transaction_to = (unsigned *)kp->arg;
 471         unsigned long value;
 472
 473         rc = kstrtoul(val, 0, &value);
 474         if (rc) {
 475                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
 476                 return rc;
 477         }
 478
 479         /*
 480          * The purpose of locking the api_mutex here is to ensure that
 481          * the correct value ends up stored properly.
 482          */
 483         mutex_lock(&the_lnet.ln_api_mutex);
 484
 485         if (value <= lnet_retry_count || value == 0) {
 486                 mutex_unlock(&the_lnet.ln_api_mutex);
 487                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
 488                        "Has to be greater than lnet_retry_count (%u)\n",
 489                        value, lnet_retry_count);
 490                 return -EINVAL;
 491         }
 492
 493         if (value == *transaction_to) {
 494                 mutex_unlock(&the_lnet.ln_api_mutex);
 495                 return 0;
 496         }
 497
 498         *transaction_to = value;
 499         /* Update the lnet_lnd_timeout now that we've modified the
 500          * transaction timeout
 501          */
 502         lnet_set_lnd_timeout();
 503
 504         mutex_unlock(&the_lnet.ln_api_mutex);
 505
 506         return 0;
 507 }
 508
 509 static int
 510 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
 511 {
 512         int rc;
 513         unsigned *retry_count = (unsigned *)kp->arg;
 514         unsigned long value;
 515
 516         rc = kstrtoul(val, 0, &value);
 517         if (rc) {
 518                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
 519                 return rc;
 520         }
 521
 522         /*
 523          * The purpose of locking the api_mutex here is to ensure that
 524          * the correct value ends up stored properly.
 525          */
 526         mutex_lock(&the_lnet.ln_api_mutex);
 527
 528         if (lnet_health_sensitivity == 0 && value > 0) {
 529                 mutex_unlock(&the_lnet.ln_api_mutex);
 530                 CERROR("Can not set lnet_retry_count when health feature is turned off\n");
 531                 return -EINVAL;
 532         }
 533
 534         if (value > lnet_transaction_timeout) {
 535                 mutex_unlock(&the_lnet.ln_api_mutex);
 536                 CERROR("Invalid value for lnet_retry_count (%lu). "
 537                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
 538                        value, lnet_transaction_timeout);
 539                 return -EINVAL;
 540         }
 541
 542         *retry_count = value;
 543
 544         /* Update the lnet_lnd_timeout now that we've modified the
 545          * retry count
 546          */
 547         lnet_set_lnd_timeout();
 548
 549         mutex_unlock(&the_lnet.ln_api_mutex);
 550
 551         return 0;
 552 }
 553
 554 static int
 555 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
 556 {
 557         int value, rc;
 558
 559         rc = kstrtoint(val, 0, &value);
 560         if (rc) {
 561                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
 562                 return rc;
 563         }
 564
 565         if (value < LNET_INTERFACES_MIN) {
 566                 CWARN("max interfaces provided are too small, setting to %d\n",
 567                       LNET_INTERFACES_MAX_DEFAULT);
 568                 value = LNET_INTERFACES_MAX_DEFAULT;
 569         }
 570
 571         *(int *)kp->arg = value;
 572
 573         return 0;
 574 }
 575
 576 static int
 577 response_tracking_set(const char *val, cfs_kernel_param_arg_t *kp)
 578 {
 579         int rc;
 580         unsigned long new_value;
 581
 582         rc = kstrtoul(val, 0, &new_value);
 583         if (rc) {
 584                 CERROR("Invalid value for 'lnet_response_tracking'\n");
 585                 return -EINVAL;
 586         }
 587
 588         if (new_value < 0 || new_value > 3) {
 589                 CWARN("Invalid value (%lu) for 'lnet_response_tracking'\n",
 590                       new_value);
 591                 return -EINVAL;
 592         }
 593
 594         lnet_response_tracking = new_value;
 595
 596         return 0;
 597 }
 598
 599 static const char *
 600 lnet_get_routes(void)
 601 {
 602         return routes;
 603 }
 604
 605 static const char *
 606 lnet_get_networks(void)
 607 {
 608         const char *nets;
 609         int rc;
 610
 611         if (*networks != 0 && *ip2nets != 0) {
 612                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 613                                    "'ip2nets' but not both at once\n");
 614                 return NULL;
 615         }
 616
 617         if (*ip2nets != 0) {
 618                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 619                 return (rc == 0) ? nets : NULL;
 620         }
 621
 622         if (*networks != 0)
 623                 return networks;
 624
 625         return "tcp";
 626 }
 627
 628 static void
 629 lnet_init_locks(void)
 630 {
 631         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 632         spin_lock_init(&the_lnet.ln_msg_resend_lock);
 633         init_completion(&the_lnet.ln_mt_wait_complete);
 634         mutex_init(&the_lnet.ln_lnd_mutex);
 635 }
 636
 637 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 638 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 639                                             *  MDs kmem_cache */
 640 struct kmem_cache *lnet_udsp_cachep;       /* udsp cache */
 641 struct kmem_cache *lnet_rspt_cachep;       /* response tracker cache */
 642 struct kmem_cache *lnet_msg_cachep;
 643
 644 static int
 645 lnet_slab_setup(void)
 646 {
 647         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 648          * allocated in <size-xxx> kmem_cache).
 649          */
 650         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 651                                             0, 0, NULL);
 652         if (!lnet_mes_cachep)
 653                 return -ENOMEM;
 654
 655         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 656                                                   LNET_SMALL_MD_SIZE, 0, 0,
 657                                                   NULL);
 658         if (!lnet_small_mds_cachep)
 659                 return -ENOMEM;
 660
 661         lnet_udsp_cachep = kmem_cache_create("lnet_udsp",
 662                                              sizeof(struct lnet_udsp),
 663                                              0, 0, NULL);
 664         if (!lnet_udsp_cachep)
 665                 return -ENOMEM;
 666
 667         lnet_rspt_cachep = kmem_cache_create("lnet_rspt", sizeof(struct lnet_rsp_tracker),
 668                                             0, 0, NULL);
 669         if (!lnet_rspt_cachep)
 670                 return -ENOMEM;
 671
 672         lnet_msg_cachep = kmem_cache_create("lnet_msg", sizeof(struct lnet_msg),
 673                                             0, 0, NULL);
 674         if (!lnet_msg_cachep)
 675                 return -ENOMEM;
 676
 677         return 0;
 678 }
 679
 680 static void
 681 lnet_slab_cleanup(void)
 682 {
 683         if (lnet_msg_cachep) {
 684                 kmem_cache_destroy(lnet_msg_cachep);
 685                 lnet_msg_cachep = NULL;
 686         }
 687
 688         if (lnet_rspt_cachep) {
 689                 kmem_cache_destroy(lnet_rspt_cachep);
 690                 lnet_rspt_cachep = NULL;
 691         }
 692
 693         if (lnet_udsp_cachep) {
 694                 kmem_cache_destroy(lnet_udsp_cachep);
 695                 lnet_udsp_cachep = NULL;
 696         }
 697
 698         if (lnet_small_mds_cachep) {
 699                 kmem_cache_destroy(lnet_small_mds_cachep);
 700                 lnet_small_mds_cachep = NULL;
 701         }
 702
 703         if (lnet_mes_cachep) {
 704                 kmem_cache_destroy(lnet_mes_cachep);
 705                 lnet_mes_cachep = NULL;
 706         }
 707 }
 708
 709 static int
 710 lnet_create_remote_nets_table(void)
 711 {
 712         int               i;
 713         struct list_head *hash;
 714
 715         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 716         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 717         CFS_ALLOC_PTR_ARRAY(hash, LNET_REMOTE_NETS_HASH_SIZE);
 718         if (hash == NULL) {
 719                 CERROR("Failed to create remote nets hash table\n");
 720                 return -ENOMEM;
 721         }
 722
 723         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 724                 INIT_LIST_HEAD(&hash[i]);
 725         the_lnet.ln_remote_nets_hash = hash;
 726         return 0;
 727 }
 728
 729 static void
 730 lnet_destroy_remote_nets_table(void)
 731 {
 732         int i;
 733
 734         if (the_lnet.ln_remote_nets_hash == NULL)
 735                 return;
 736
 737         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 738                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 739
 740         CFS_FREE_PTR_ARRAY(the_lnet.ln_remote_nets_hash,
 741                            LNET_REMOTE_NETS_HASH_SIZE);
 742         the_lnet.ln_remote_nets_hash = NULL;
 743 }
 744
 745 static void
 746 lnet_destroy_locks(void)
 747 {
 748         if (the_lnet.ln_res_lock != NULL) {
 749                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 750                 the_lnet.ln_res_lock = NULL;
 751         }
 752
 753         if (the_lnet.ln_net_lock != NULL) {
 754                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 755                 the_lnet.ln_net_lock = NULL;
 756         }
 757 }
 758
 759 static int
 760 lnet_create_locks(void)
 761 {
 762         lnet_init_locks();
 763
 764         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 765         if (the_lnet.ln_res_lock == NULL)
 766                 goto failed;
 767
 768         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 769         if (the_lnet.ln_net_lock == NULL)
 770                 goto failed;
 771
 772         return 0;
 773
 774  failed:
 775         lnet_destroy_locks();
 776         return -ENOMEM;
 777 }
 778
 779 static void lnet_assert_wire_constants(void)
 780 {
 781         /* Wire protocol assertions generated by 'wirecheck'
 782          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 783          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 784          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7)
 785          */
 786
 787         /* Constants... */
 788         BUILD_BUG_ON(LNET_PROTO_TCP_MAGIC != 0xeebc0ded);
 789         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MAJOR != 1);
 790         BUILD_BUG_ON(LNET_PROTO_TCP_VERSION_MINOR != 0);
 791         BUILD_BUG_ON(LNET_MSG_ACK != 0);
 792         BUILD_BUG_ON(LNET_MSG_PUT != 1);
 793         BUILD_BUG_ON(LNET_MSG_GET != 2);
 794         BUILD_BUG_ON(LNET_MSG_REPLY != 3);
 795         BUILD_BUG_ON(LNET_MSG_HELLO != 4);
 796
 797         BUILD_BUG_ON((int)sizeof(lnet_nid_t) != 8);
 798         BUILD_BUG_ON((int)sizeof(lnet_pid_t) != 4);
 799
 800         /* Checks for struct lnet_nid */
 801         BUILD_BUG_ON((int)sizeof(struct lnet_nid) != 20);
 802         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_size) != 0);
 803         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_size) != 1);
 804         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_type) != 1);
 805         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_type) != 1);
 806         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_num) != 2);
 807         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_num) != 2);
 808         BUILD_BUG_ON((int)offsetof(struct lnet_nid, nid_addr) != 4);
 809         BUILD_BUG_ON((int)sizeof(((struct lnet_nid *)0)->nid_addr) != 16);
 810
 811         /* Checks for struct lnet_process_id_packed */
 812         BUILD_BUG_ON((int)sizeof(struct lnet_process_id_packed) != 12);
 813         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, nid) != 0);
 814         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->nid) != 8);
 815         BUILD_BUG_ON((int)offsetof(struct lnet_process_id_packed, pid) != 8);
 816         BUILD_BUG_ON((int)sizeof(((struct lnet_process_id_packed *)0)->pid) != 4);
 817
 818         /* Checks for struct lnet_handle_wire */
 819         BUILD_BUG_ON((int)sizeof(struct lnet_handle_wire) != 16);
 820         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 821                                    wh_interface_cookie) != 0);
 822         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) != 8);
 823         BUILD_BUG_ON((int)offsetof(struct lnet_handle_wire,
 824                                    wh_object_cookie) != 8);
 825         BUILD_BUG_ON((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) != 8);
 826
 827         /* Checks for struct struct lnet_magicversion */
 828         BUILD_BUG_ON((int)sizeof(struct lnet_magicversion) != 8);
 829         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, magic) != 0);
 830         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->magic) != 4);
 831         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion, version_major) != 4);
 832         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_major) != 2);
 833         BUILD_BUG_ON((int)offsetof(struct lnet_magicversion,
 834                                    version_minor) != 6);
 835         BUILD_BUG_ON((int)sizeof(((struct lnet_magicversion *)0)->version_minor) != 2);
 836
 837         /* Checks for struct _lnet_hdr_nid4 */
 838         BUILD_BUG_ON((int)sizeof(struct _lnet_hdr_nid4) != 72);
 839         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, dest_nid) != 0);
 840         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->dest_nid) != 8);
 841         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, src_nid) != 8);
 842         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->src_nid) != 8);
 843         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, dest_pid) != 16);
 844         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->dest_pid) != 4);
 845         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, src_pid) != 20);
 846         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->src_pid) != 4);
 847         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, type) != 24);
 848         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->type) != 4);
 849         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, payload_length) != 28);
 850         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->payload_length) != 4);
 851         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg) != 32);
 852         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg) != 40);
 853
 854         /* Ack */
 855         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.dst_wmd) != 32);
 856         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.dst_wmd) != 16);
 857         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.match_bits) != 48);
 858         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.match_bits) != 8);
 859         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.ack.mlength) != 56);
 860         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.ack.mlength) != 4);
 861
 862         /* Put */
 863         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.ack_wmd) != 32);
 864         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.ack_wmd) != 16);
 865         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.match_bits) != 48);
 866         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.match_bits) != 8);
 867         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.hdr_data) != 56);
 868         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.hdr_data) != 8);
 869         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.ptl_index) != 64);
 870         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.ptl_index) != 4);
 871         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.put.offset) != 68);
 872         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.put.offset) != 4);
 873
 874         /* Get */
 875         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.return_wmd) != 32);
 876         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.return_wmd) != 16);
 877         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.match_bits) != 48);
 878         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.match_bits) != 8);
 879         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.ptl_index) != 56);
 880         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.ptl_index) != 4);
 881         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.src_offset) != 60);
 882         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.src_offset) != 4);
 883         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.get.sink_length) != 64);
 884         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.get.sink_length) != 4);
 885
 886         /* Reply */
 887         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.reply.dst_wmd) != 32);
 888         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.reply.dst_wmd) != 16);
 889
 890         /* Hello */
 891         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.hello.incarnation) != 32);
 892         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.hello.incarnation) != 8);
 893         BUILD_BUG_ON((int)offsetof(struct _lnet_hdr_nid4, msg.hello.type) != 40);
 894         BUILD_BUG_ON((int)sizeof(((struct _lnet_hdr_nid4 *)0)->msg.hello.type) != 4);
 895
 896         /* Checks for struct lnet_ni_status and related constants */
 897         BUILD_BUG_ON(LNET_NI_STATUS_INVALID != 0x00000000);
 898         BUILD_BUG_ON(LNET_NI_STATUS_UP != 0x15aac0de);
 899         BUILD_BUG_ON(LNET_NI_STATUS_DOWN != 0xdeadface);
 900
 901         /* Checks for struct lnet_ni_status */
 902         BUILD_BUG_ON((int)sizeof(struct lnet_ni_status) != 16);
 903         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_nid) != 0);
 904         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) != 8);
 905         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_status) != 8);
 906         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_status) != 4);
 907         BUILD_BUG_ON((int)offsetof(struct lnet_ni_status, ns_msg_size) != 12);
 908         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_status *)0)->ns_msg_size) != 4);
 909
 910         /* Checks for struct lnet_ni_large_status */
 911         BUILD_BUG_ON((int)sizeof(struct lnet_ni_large_status) != 24);
 912         BUILD_BUG_ON((int)offsetof(struct lnet_ni_large_status, ns_status) != 0);
 913         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_large_status *)0)->ns_status) != 4);
 914         BUILD_BUG_ON((int)offsetof(struct lnet_ni_large_status, ns_nid) != 4);
 915         BUILD_BUG_ON((int)sizeof(((struct lnet_ni_large_status *)0)->ns_nid) != 20);
 916
 917         /* Checks for struct lnet_ping_info and related constants */
 918         BUILD_BUG_ON(LNET_PROTO_PING_MAGIC != 0x70696E67);
 919         BUILD_BUG_ON(LNET_PING_FEAT_INVAL != 0);
 920         BUILD_BUG_ON(LNET_PING_FEAT_BASE != 1);
 921         BUILD_BUG_ON(LNET_PING_FEAT_NI_STATUS != 2);
 922         BUILD_BUG_ON(LNET_PING_FEAT_RTE_DISABLED != 4);
 923         BUILD_BUG_ON(LNET_PING_FEAT_MULTI_RAIL != 8);
 924         BUILD_BUG_ON(LNET_PING_FEAT_DISCOVERY != 16);
 925         BUILD_BUG_ON(LNET_PING_FEAT_LARGE_ADDR != 32);
 926         BUILD_BUG_ON(LNET_PING_FEAT_PRIMARY_LARGE != 64);
 927         BUILD_BUG_ON(LNET_PING_FEAT_BITS != 127);
 928
 929         /* Checks for struct lnet_ping_info */
 930         BUILD_BUG_ON((int)sizeof(struct lnet_ping_info) != 16);
 931         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_magic) != 0);
 932         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) != 4);
 933         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_features) != 4);
 934         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_features) != 4);
 935         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_pid) != 8);
 936         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) != 4);
 937         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_nnis) != 12);
 938         BUILD_BUG_ON((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) != 4);
 939         BUILD_BUG_ON((int)offsetof(struct lnet_ping_info, pi_ni) != 16);
 940         BUILD_BUG_ON(offsetof(struct lnet_ping_info, pi_ni) != sizeof(struct lnet_ping_info));
 941
 942         /* Acceptor connection request */
 943         BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 944
 945         /* Checks for struct lnet_acceptor_connreq */
 946         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq) != 16);
 947         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_magic) != 0);
 948         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_magic) != 4);
 949         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_version) != 4);
 950         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_version) != 4);
 951         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq, acr_nid) != 8);
 952         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq *)0)->acr_nid) != 8);
 953
 954         /* Checks for struct lnet_acceptor_connreq_v2 */
 955         BUILD_BUG_ON((int)sizeof(struct lnet_acceptor_connreq_v2) != 28);
 956         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_magic) != 0);
 957         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_magic) != 4);
 958         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_version) != 4);
 959         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_version) != 4);
 960         BUILD_BUG_ON((int)offsetof(struct lnet_acceptor_connreq_v2, acr_nid) != 8);
 961         BUILD_BUG_ON((int)sizeof(((struct lnet_acceptor_connreq_v2 *)0)->acr_nid) != 20);
 962
 963         /* Checks for struct lnet_counters_common */
 964         BUILD_BUG_ON((int)sizeof(struct lnet_counters_common) != 60);
 965         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_alloc) != 0);
 966         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_alloc) != 4);
 967         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_msgs_max) != 4);
 968         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_msgs_max) != 4);
 969         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_errors) != 8);
 970         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_errors) != 4);
 971         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_count) != 12);
 972         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_count) != 4);
 973         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_count) != 16);
 974         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_count) != 4);
 975         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_count) != 20);
 976         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_count) != 4);
 977         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_count) != 24);
 978         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_count) != 4);
 979         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_send_length) != 28);
 980         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_send_length) != 8);
 981         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_recv_length) != 36);
 982         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_recv_length) != 8);
 983         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_route_length) != 44);
 984         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_route_length) != 8);
 985         BUILD_BUG_ON((int)offsetof(struct lnet_counters_common, lcc_drop_length) != 52);
 986         BUILD_BUG_ON((int)sizeof(((struct lnet_counters_common *)0)->lcc_drop_length) != 8);
 987 }
 988
 989 static const struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 990 {
 991         const struct lnet_lnd *lnd;
 992
 993         /* holding lnd mutex */
 994         if (type >= NUM_LNDS)
 995                 return NULL;
 996         lnd = the_lnet.ln_lnds[type];
 997         LASSERT(!lnd || lnd->lnd_type == type);
 998
 999         return lnd;
1000 }
1001
1002 unsigned int
1003 lnet_get_lnd_timeout(void)
1004 {
1005         return lnet_lnd_timeout;
1006 }
1007 EXPORT_SYMBOL(lnet_get_lnd_timeout);
1008
1009 void
1010 lnet_register_lnd(const struct lnet_lnd *lnd)
1011 {
1012         mutex_lock(&the_lnet.ln_lnd_mutex);
1013
1014         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
1015         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
1016
1017         the_lnet.ln_lnds[lnd->lnd_type] = lnd;
1018
1019         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
1020
1021         mutex_unlock(&the_lnet.ln_lnd_mutex);
1022 }
1023 EXPORT_SYMBOL(lnet_register_lnd);
1024
1025 void
1026 lnet_unregister_lnd(const struct lnet_lnd *lnd)
1027 {
1028         mutex_lock(&the_lnet.ln_lnd_mutex);
1029
1030         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
1031
1032         the_lnet.ln_lnds[lnd->lnd_type] = NULL;
1033         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
1034
1035         mutex_unlock(&the_lnet.ln_lnd_mutex);
1036 }
1037 EXPORT_SYMBOL(lnet_unregister_lnd);
1038
1039 static void
1040 lnet_counters_get_common_locked(struct lnet_counters_common *common)
1041 {
1042         struct lnet_counters *ctr;
1043         int i;
1044
1045         /* FIXME !!! Their is no assert_lnet_net_locked() to ensure this
1046          * actually called under the protection of the lnet_net_lock.
1047          */
1048         memset(common, 0, sizeof(*common));
1049
1050         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1051                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
1052                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
1053                 common->lcc_errors       += ctr->lct_common.lcc_errors;
1054                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
1055                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
1056                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
1057                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
1058                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
1059                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
1060                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
1061                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
1062         }
1063 }
1064
1065 void
1066 lnet_counters_get_common(struct lnet_counters_common *common)
1067 {
1068         lnet_net_lock(LNET_LOCK_EX);
1069         lnet_counters_get_common_locked(common);
1070         lnet_net_unlock(LNET_LOCK_EX);
1071 }
1072 EXPORT_SYMBOL(lnet_counters_get_common);
1073
1074 int
1075 lnet_counters_get(struct lnet_counters *counters)
1076 {
1077         struct lnet_counters *ctr;
1078         struct lnet_counters_health *health = &counters->lct_health;
1079         int i, rc = 0;
1080
1081         memset(counters, 0, sizeof(*counters));
1082
1083         lnet_net_lock(LNET_LOCK_EX);
1084
1085         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1086                 GOTO(out_unlock, rc = -ENODEV);
1087
1088         lnet_counters_get_common_locked(&counters->lct_common);
1089
1090         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
1091                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
1092                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
1093                 health->lch_response_timeout_count +=
1094                                 ctr->lct_health.lch_response_timeout_count;
1095                 health->lch_local_interrupt_count +=
1096                                 ctr->lct_health.lch_local_interrupt_count;
1097                 health->lch_local_dropped_count +=
1098                                 ctr->lct_health.lch_local_dropped_count;
1099                 health->lch_local_aborted_count +=
1100                                 ctr->lct_health.lch_local_aborted_count;
1101                 health->lch_local_no_route_count +=
1102                                 ctr->lct_health.lch_local_no_route_count;
1103                 health->lch_local_timeout_count +=
1104                                 ctr->lct_health.lch_local_timeout_count;
1105                 health->lch_local_error_count +=
1106                                 ctr->lct_health.lch_local_error_count;
1107                 health->lch_remote_dropped_count +=
1108                                 ctr->lct_health.lch_remote_dropped_count;
1109                 health->lch_remote_error_count +=
1110                                 ctr->lct_health.lch_remote_error_count;
1111                 health->lch_remote_timeout_count +=
1112                                 ctr->lct_health.lch_remote_timeout_count;
1113                 health->lch_network_timeout_count +=
1114                                 ctr->lct_health.lch_network_timeout_count;
1115         }
1116 out_unlock:
1117         lnet_net_unlock(LNET_LOCK_EX);
1118         return rc;
1119 }
1120 EXPORT_SYMBOL(lnet_counters_get);
1121
1122 void
1123 lnet_counters_reset(void)
1124 {
1125         struct lnet_counters *counters;
1126         int             i;
1127
1128         lnet_net_lock(LNET_LOCK_EX);
1129
1130         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1131                 goto avoid_reset;
1132
1133         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
1134                 memset(counters, 0, sizeof(struct lnet_counters));
1135 avoid_reset:
1136         lnet_net_unlock(LNET_LOCK_EX);
1137 }
1138
1139 static char *
1140 lnet_res_type2str(int type)
1141 {
1142         switch (type) {
1143         default:
1144                 LBUG();
1145         case LNET_COOKIE_TYPE_MD:
1146                 return "MD";
1147         case LNET_COOKIE_TYPE_ME:
1148                 return "ME";
1149         case LNET_COOKIE_TYPE_EQ:
1150                 return "EQ";
1151         }
1152 }
1153
1154 static void
1155 lnet_res_container_cleanup(struct lnet_res_container *rec)
1156 {
1157         int     count = 0;
1158
1159         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
1160                 return;
1161
1162         while (!list_empty(&rec->rec_active)) {
1163                 struct list_head *e = rec->rec_active.next;
1164
1165                 list_del_init(e);
1166                 if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
1167                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
1168
1169                 } else { /* NB: Active MEs should be attached on portals */
1170                         LBUG();
1171                 }
1172                 count++;
1173         }
1174
1175         if (count > 0) {
1176                 /* Found alive MD/ME/EQ, user really should unlink/free
1177                  * all of them before finalize LNet, but if someone didn't,
1178                  * we have to recycle garbage for him */
1179                 CERROR("%d active elements on exit of %s container\n",
1180                        count, lnet_res_type2str(rec->rec_type));
1181         }
1182
1183         if (rec->rec_lh_hash != NULL) {
1184                 CFS_FREE_PTR_ARRAY(rec->rec_lh_hash, LNET_LH_HASH_SIZE);
1185                 rec->rec_lh_hash = NULL;
1186         }
1187
1188         rec->rec_type = 0; /* mark it as finalized */
1189 }
1190
1191 static int
1192 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
1193 {
1194         int     rc = 0;
1195         int     i;
1196
1197         LASSERT(rec->rec_type == 0);
1198
1199         rec->rec_type = type;
1200         INIT_LIST_HEAD(&rec->rec_active);
1201
1202         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1203
1204         /* Arbitrary choice of hash table size */
1205         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1206                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1207         if (rec->rec_lh_hash == NULL) {
1208                 rc = -ENOMEM;
1209                 goto out;
1210         }
1211
1212         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1213                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1214
1215         return 0;
1216
1217 out:
1218         CERROR("Failed to setup %s resource container\n",
1219                lnet_res_type2str(type));
1220         lnet_res_container_cleanup(rec);
1221         return rc;
1222 }
1223
1224 static void
1225 lnet_res_containers_destroy(struct lnet_res_container **recs)
1226 {
1227         struct lnet_res_container       *rec;
1228         int                             i;
1229
1230         cfs_percpt_for_each(rec, i, recs)
1231                 lnet_res_container_cleanup(rec);
1232
1233         cfs_percpt_free(recs);
1234 }
1235
1236 static struct lnet_res_container **
1237 lnet_res_containers_create(int type)
1238 {
1239         struct lnet_res_container       **recs;
1240         struct lnet_res_container       *rec;
1241         int                             rc;
1242         int                             i;
1243
1244         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1245         if (recs == NULL) {
1246                 CERROR("Failed to allocate %s resource containers\n",
1247                        lnet_res_type2str(type));
1248                 return NULL;
1249         }
1250
1251         cfs_percpt_for_each(rec, i, recs) {
1252                 rc = lnet_res_container_setup(rec, i, type);
1253                 if (rc != 0) {
1254                         lnet_res_containers_destroy(recs);
1255                         return NULL;
1256                 }
1257         }
1258
1259         return recs;
1260 }
1261
1262 struct lnet_libhandle *
1263 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1264 {
1265         /* ALWAYS called with lnet_res_lock held */
1266         struct list_head        *head;
1267         struct lnet_libhandle   *lh;
1268         unsigned int            hash;
1269
1270         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1271                 return NULL;
1272
1273         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1274         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1275
1276         list_for_each_entry(lh, head, lh_hash_chain) {
1277                 if (lh->lh_cookie == cookie)
1278                         return lh;
1279         }
1280
1281         return NULL;
1282 }
1283
1284 void
1285 lnet_res_lh_initialize(struct lnet_res_container *rec,
1286                        struct lnet_libhandle *lh)
1287 {
1288         /* ALWAYS called with lnet_res_lock held */
1289         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1290         unsigned int    hash;
1291
1292         lh->lh_cookie = rec->rec_lh_cookie;
1293         rec->rec_lh_cookie += 1 << ibits;
1294
1295         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1296
1297         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1298 }
1299
1300 struct list_head **
1301 lnet_create_array_of_queues(void)
1302 {
1303         struct list_head **qs;
1304         struct list_head *q;
1305         int i;
1306
1307         qs = cfs_percpt_alloc(lnet_cpt_table(),
1308                               sizeof(struct list_head));
1309         if (!qs) {
1310                 CERROR("Failed to allocate queues\n");
1311                 return NULL;
1312         }
1313
1314         cfs_percpt_for_each(q, i, qs)
1315                 INIT_LIST_HEAD(q);
1316
1317         return qs;
1318 }
1319
1320 static int lnet_unprepare(void);
1321
1322 static int
1323 lnet_prepare(lnet_pid_t requested_pid)
1324 {
1325         /* Prepare to bring up the network */
1326         struct lnet_res_container **recs;
1327         int                       rc = 0;
1328
1329         if (requested_pid == LNET_PID_ANY) {
1330                 /* Don't instantiate LNET just for me */
1331                 return -ENETDOWN;
1332         }
1333
1334         LASSERT(the_lnet.ln_refcount == 0);
1335
1336         the_lnet.ln_routing = 0;
1337
1338         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1339         the_lnet.ln_pid = requested_pid;
1340
1341         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1342         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1343         INIT_LIST_HEAD(&the_lnet.ln_nets);
1344         INIT_LIST_HEAD(&the_lnet.ln_routers);
1345         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1346         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1347         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1348         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1349         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1350         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1351         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1352         INIT_LIST_HEAD(&the_lnet.ln_udsp_list);
1353         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1354         the_lnet.ln_mt_handler = NULL;
1355         init_completion(&the_lnet.ln_started);
1356         atomic_set(&the_lnet.ln_late_msg_count, 0);
1357         atomic64_set(&the_lnet.ln_late_msg_nsecs, 0);
1358
1359         rc = lnet_slab_setup();
1360         if (rc != 0)
1361                 goto failed;
1362
1363         rc = lnet_create_remote_nets_table();
1364         if (rc != 0)
1365                 goto failed;
1366
1367         /*
1368          * NB the interface cookie in wire handles guards against delayed
1369          * replies and ACKs appearing valid after reboot.
1370          */
1371         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1372
1373         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1374                                                 sizeof(struct lnet_counters));
1375         if (the_lnet.ln_counters == NULL) {
1376                 CERROR("Failed to allocate counters for LNet\n");
1377                 rc = -ENOMEM;
1378                 goto failed;
1379         }
1380
1381         rc = lnet_peer_tables_create();
1382         if (rc != 0)
1383                 goto failed;
1384
1385         rc = lnet_msg_containers_create();
1386         if (rc != 0)
1387                 goto failed;
1388
1389         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1390                                       LNET_COOKIE_TYPE_EQ);
1391         if (rc != 0)
1392                 goto failed;
1393
1394         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1395         if (recs == NULL) {
1396                 rc = -ENOMEM;
1397                 goto failed;
1398         }
1399
1400         the_lnet.ln_md_containers = recs;
1401
1402         rc = lnet_portals_create();
1403         if (rc != 0) {
1404                 CERROR("Failed to create portals for LNet: %d\n", rc);
1405                 goto failed;
1406         }
1407
1408         the_lnet.ln_mt_zombie_rstqs = lnet_create_array_of_queues();
1409         if (!the_lnet.ln_mt_zombie_rstqs) {
1410                 rc = -ENOMEM;
1411                 goto failed;
1412         }
1413
1414         return 0;
1415
1416  failed:
1417         lnet_unprepare();
1418         return rc;
1419 }
1420
1421 static int
1422 lnet_unprepare(void)
1423 {
1424         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1425          * have shut down already, so it is safe to unlink and free all
1426          * descriptors, even those that appear committed to a network op (eg MD
1427          * with non-zero pending count) */
1428
1429         lnet_fail_nid(LNET_NID_ANY, 0);
1430
1431         LASSERT(the_lnet.ln_refcount == 0);
1432         LASSERT(list_empty(&the_lnet.ln_test_peers));
1433         LASSERT(list_empty(&the_lnet.ln_nets));
1434
1435         if (the_lnet.ln_mt_zombie_rstqs) {
1436                 lnet_clean_zombie_rstqs();
1437                 the_lnet.ln_mt_zombie_rstqs = NULL;
1438         }
1439
1440         lnet_assert_handler_unused(the_lnet.ln_mt_handler, true);
1441         the_lnet.ln_mt_handler = NULL;
1442
1443         lnet_portals_destroy();
1444
1445         if (the_lnet.ln_md_containers != NULL) {
1446                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1447                 the_lnet.ln_md_containers = NULL;
1448         }
1449
1450         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1451
1452         lnet_msg_containers_destroy();
1453         lnet_peer_uninit();
1454         lnet_rtrpools_free(0);
1455
1456         if (the_lnet.ln_counters != NULL) {
1457                 cfs_percpt_free(the_lnet.ln_counters);
1458                 the_lnet.ln_counters = NULL;
1459         }
1460         lnet_destroy_remote_nets_table();
1461         lnet_udsp_destroy(true);
1462         lnet_slab_cleanup();
1463
1464         return 0;
1465 }
1466
1467 struct lnet_ni  *
1468 lnet_net2ni_locked(__u32 net_id, int cpt)
1469 {
1470         struct lnet_ni   *ni;
1471         struct lnet_net  *net;
1472
1473         LASSERT(cpt != LNET_LOCK_EX);
1474
1475         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1476                 if (net->net_id == net_id) {
1477                         ni = list_first_entry(&net->net_ni_list, struct lnet_ni,
1478                                               ni_netlist);
1479                         return ni;
1480                 }
1481         }
1482
1483         return NULL;
1484 }
1485
1486 struct lnet_ni *
1487 lnet_net2ni_addref(__u32 net)
1488 {
1489         struct lnet_ni *ni;
1490
1491         lnet_net_lock(0);
1492         ni = lnet_net2ni_locked(net, 0);
1493         if (ni)
1494                 lnet_ni_addref_locked(ni, 0);
1495         lnet_net_unlock(0);
1496
1497         return ni;
1498 }
1499 EXPORT_SYMBOL(lnet_net2ni_addref);
1500
1501 struct lnet_net *
1502 lnet_get_net_locked(__u32 net_id)
1503 {
1504         struct lnet_net  *net;
1505
1506         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1507                 if (net->net_id == net_id)
1508                         return net;
1509         }
1510
1511         return NULL;
1512 }
1513
1514 void
1515 lnet_net_clr_pref_rtrs(struct lnet_net *net)
1516 {
1517         struct list_head zombies;
1518         struct lnet_nid_list *ne;
1519         struct lnet_nid_list *tmp;
1520
1521         INIT_LIST_HEAD(&zombies);
1522
1523         lnet_net_lock(LNET_LOCK_EX);
1524         list_splice_init(&net->net_rtr_pref_nids, &zombies);
1525         lnet_net_unlock(LNET_LOCK_EX);
1526
1527         list_for_each_entry_safe(ne, tmp, &zombies, nl_list) {
1528                 list_del_init(&ne->nl_list);
1529                 LIBCFS_FREE(ne, sizeof(*ne));
1530         }
1531 }
1532
1533 int
1534 lnet_net_add_pref_rtr(struct lnet_net *net,
1535                       struct lnet_nid *gw_nid)
1536 __must_hold(&the_lnet.ln_api_mutex)
1537 {
1538         struct lnet_nid_list *ne;
1539
1540         /* This function is called with api_mutex held. When the api_mutex
1541          * is held the list can not be modified, as it is only modified as
1542          * a result of applying a UDSP and that happens under api_mutex
1543          * lock.
1544          */
1545         list_for_each_entry(ne, &net->net_rtr_pref_nids, nl_list) {
1546                 if (nid_same(&ne->nl_nid, gw_nid))
1547                         return -EEXIST;
1548         }
1549
1550         LIBCFS_ALLOC(ne, sizeof(*ne));
1551         if (!ne)
1552                 return -ENOMEM;
1553
1554         ne->nl_nid = *gw_nid;
1555
1556         /* Lock the cpt to protect against addition and checks in the
1557          * selection algorithm
1558          */
1559         lnet_net_lock(LNET_LOCK_EX);
1560         list_add(&ne->nl_list, &net->net_rtr_pref_nids);
1561         lnet_net_unlock(LNET_LOCK_EX);
1562
1563         return 0;
1564 }
1565
1566 static unsigned int
1567 lnet_nid4_cpt_hash(lnet_nid_t nid, unsigned int number)
1568 {
1569         __u64 key = nid;
1570         __u16 lnd = LNET_NETTYP(LNET_NIDNET(nid));
1571         unsigned int cpt;
1572
1573         if (lnd == KFILND || lnd == GNILND) {
1574                 cpt = hash_long(key, LNET_CPT_BITS);
1575
1576                 /* NB: The number of CPTs needn't be a power of 2 */
1577                 if (cpt >= number)
1578                         cpt = (key + cpt + (cpt >> 1)) % number;
1579         } else {
1580                 __u64 pair_bits = 0x0001000100010001LLU;
1581                 __u64 mask = pair_bits * 0xFF;
1582                 __u64 pair_sum;
1583                 /* For ipv4 NIDs, use (sum-by-multiplication of nid bytes) mod
1584                  * (number of CPTs) to match nid to a CPT.
1585                  */
1586                 pair_sum = (key & mask) + ((key >> 8) & mask);
1587                 pair_sum = (pair_sum * pair_bits) >> 48;
1588                 cpt = (unsigned int)(pair_sum) % number;
1589         }
1590
1591         CDEBUG(D_NET, "Match nid %s to cpt %u\n",
1592                libcfs_nid2str(nid), cpt);
1593
1594         return cpt;
1595 }
1596
1597 unsigned int
1598 lnet_nid_cpt_hash(struct lnet_nid *nid, unsigned int number)
1599 {
1600         unsigned int val;
1601         u32 h = 0;
1602         int i;
1603
1604         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1605
1606         if (number == 1)
1607                 return 0;
1608
1609         if (nid_is_nid4(nid))
1610                 return lnet_nid4_cpt_hash(lnet_nid_to_nid4(nid), number);
1611
1612         for (i = 0; i < 4; i++)
1613                 h = cfs_hash_32(nid->nid_addr[i]^h, 32);
1614         val = cfs_hash_32(LNET_NID_NET(nid) ^ h, LNET_CPT_BITS);
1615         if (val < number)
1616                 return val;
1617         return (unsigned int)(h + val + (val >> 1)) % number;
1618 }
1619
1620 int
1621 lnet_cpt_of_nid_locked(struct lnet_nid *nid, struct lnet_ni *ni)
1622 {
1623         struct lnet_net *net;
1624
1625         /* must called with hold of lnet_net_lock */
1626         if (LNET_CPT_NUMBER == 1)
1627                 return 0; /* the only one */
1628
1629         /*
1630          * If NI is provided then use the CPT identified in the NI cpt
1631          * list if one exists. If one doesn't exist, then that NI is
1632          * associated with all CPTs and it follows that the net it belongs
1633          * to is implicitly associated with all CPTs, so just hash the nid
1634          * and return that.
1635          */
1636         if (ni != NULL) {
1637                 if (ni->ni_cpts != NULL)
1638                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1639                                                              ni->ni_ncpts)];
1640                 else
1641                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1642         }
1643
1644         /* no NI provided so look at the net */
1645         net = lnet_get_net_locked(LNET_NID_NET(nid));
1646
1647         if (net != NULL && net->net_cpts != NULL) {
1648                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1649         }
1650
1651         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1652 }
1653
1654 int
1655 lnet_nid2cpt(struct lnet_nid *nid, struct lnet_ni *ni)
1656 {
1657         int     cpt;
1658         int     cpt2;
1659
1660         if (LNET_CPT_NUMBER == 1)
1661                 return 0; /* the only one */
1662
1663         cpt = lnet_net_lock_current();
1664
1665         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1666
1667         lnet_net_unlock(cpt);
1668
1669         return cpt2;
1670 }
1671 EXPORT_SYMBOL(lnet_nid2cpt);
1672
1673 int
1674 lnet_cpt_of_nid(lnet_nid_t nid4, struct lnet_ni *ni)
1675 {
1676         struct lnet_nid nid;
1677
1678         if (LNET_CPT_NUMBER == 1)
1679                 return 0; /* the only one */
1680
1681         lnet_nid4_to_nid(nid4, &nid);
1682         return lnet_nid2cpt(&nid, ni);
1683 }
1684 EXPORT_SYMBOL(lnet_cpt_of_nid);
1685
1686 int
1687 lnet_islocalnet_locked(__u32 net_id)
1688 {
1689         struct lnet_net *net;
1690         bool local;
1691
1692         net = lnet_get_net_locked(net_id);
1693
1694         local = net != NULL;
1695
1696         return local;
1697 }
1698
1699 int
1700 lnet_islocalnet(__u32 net_id)
1701 {
1702         int cpt;
1703         bool local;
1704
1705         cpt = lnet_net_lock_current();
1706
1707         local = lnet_islocalnet_locked(net_id);
1708
1709         lnet_net_unlock(cpt);
1710
1711         return local;
1712 }
1713
1714 struct lnet_ni  *
1715 lnet_nid_to_ni_locked(struct lnet_nid *nid, int cpt)
1716 {
1717         struct lnet_net  *net;
1718         struct lnet_ni *ni;
1719
1720         LASSERT(cpt != LNET_LOCK_EX);
1721
1722         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1723                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1724                         if (nid_same(&ni->ni_nid, nid))
1725                                 return ni;
1726                 }
1727         }
1728
1729         return NULL;
1730 }
1731
1732 struct lnet_ni *
1733 lnet_nid_to_ni_addref(struct lnet_nid *nid)
1734 {
1735         struct lnet_ni *ni;
1736
1737         lnet_net_lock(0);
1738         ni = lnet_nid_to_ni_locked(nid, 0);
1739         if (ni)
1740                 lnet_ni_addref_locked(ni, 0);
1741         lnet_net_unlock(0);
1742
1743         return ni;
1744 }
1745 EXPORT_SYMBOL(lnet_nid_to_ni_addref);
1746
1747 int
1748 lnet_islocalnid(struct lnet_nid *nid)
1749 {
1750         struct lnet_ni  *ni;
1751         int             cpt;
1752
1753         cpt = lnet_net_lock_current();
1754         ni = lnet_nid_to_ni_locked(nid, cpt);
1755         lnet_net_unlock(cpt);
1756
1757         return ni != NULL;
1758 }
1759
1760 int
1761 lnet_count_acceptor_nets(void)
1762 {
1763         /* Return the # of NIs that need the acceptor. */
1764         int              count = 0;
1765         struct lnet_net  *net;
1766         int              cpt;
1767
1768         cpt = lnet_net_lock_current();
1769         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1770                 /* all socklnd type networks should have the acceptor
1771                  * thread started */
1772                 if (net->net_lnd->lnd_accept != NULL)
1773                         count++;
1774         }
1775
1776         lnet_net_unlock(cpt);
1777
1778         return count;
1779 }
1780
1781 struct lnet_ping_buffer *
1782 lnet_ping_buffer_alloc(int nbytes, gfp_t gfp)
1783 {
1784         struct lnet_ping_buffer *pbuf;
1785
1786         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nbytes), gfp);
1787         if (pbuf) {
1788                 pbuf->pb_nbytes = nbytes;       /* sizeof of pb_info */
1789                 pbuf->pb_needs_post = false;
1790                 atomic_set(&pbuf->pb_refcnt, 1);
1791         }
1792
1793         return pbuf;
1794 }
1795
1796 void
1797 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1798 {
1799         LASSERT(atomic_read(&pbuf->pb_refcnt) == 0);
1800         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nbytes));
1801 }
1802
1803 static struct lnet_ping_buffer *
1804 lnet_ping_target_create(int nbytes)
1805 {
1806         struct lnet_ping_buffer *pbuf;
1807
1808         pbuf = lnet_ping_buffer_alloc(nbytes, GFP_NOFS);
1809         if (pbuf == NULL) {
1810                 CERROR("Can't allocate ping source [%d]\n", nbytes);
1811                 return NULL;
1812         }
1813
1814         pbuf->pb_info.pi_nnis = 0;
1815         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1816         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1817         pbuf->pb_info.pi_features =
1818                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1819
1820         return pbuf;
1821 }
1822
1823 static inline int
1824 lnet_get_net_ni_bytes_locked(struct lnet_net *net)
1825 {
1826         struct lnet_ni *ni;
1827         int bytes = 0;
1828
1829         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1830                 bytes += lnet_ping_sts_size(&ni->ni_nid);
1831
1832         return bytes;
1833 }
1834
1835 static inline int
1836 lnet_get_ni_bytes(void)
1837 {
1838         struct lnet_ni *ni;
1839         struct lnet_net *net;
1840         int bytes = 0;
1841
1842         lnet_net_lock(0);
1843
1844         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1845                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1846                         bytes += lnet_ping_sts_size(&ni->ni_nid);
1847         }
1848
1849         lnet_net_unlock(0);
1850
1851         return bytes;
1852 }
1853
1854 void
1855 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1856 {
1857         struct lnet_ni_large_status *lstat, *lend;
1858         struct lnet_ni_status *stat, *end;
1859         int nnis;
1860         int i;
1861
1862         __swab32s(&pbuf->pb_info.pi_magic);
1863         __swab32s(&pbuf->pb_info.pi_features);
1864         __swab32s(&pbuf->pb_info.pi_pid);
1865         __swab32s(&pbuf->pb_info.pi_nnis);
1866         nnis = pbuf->pb_info.pi_nnis;
1867         stat = &pbuf->pb_info.pi_ni[0];
1868         end = (void *)&pbuf->pb_info + pbuf->pb_nbytes;
1869         for (i = 0; i < nnis && stat + 1 <= end; i++, stat++) {
1870                 __swab64s(&stat->ns_nid);
1871                 __swab32s(&stat->ns_status);
1872                 if (i == 0)
1873                         /* Might be total size */
1874                         __swab32s(&stat->ns_msg_size);
1875         }
1876         if (!(pbuf->pb_info.pi_features & LNET_PING_FEAT_LARGE_ADDR))
1877                 return;
1878
1879         lstat = (struct lnet_ni_large_status *)stat;
1880         lend = (void *)end;
1881         while (lstat + 1 <= lend) {
1882                 __swab32s(&lstat->ns_status);
1883                 /* struct lnet_nid never needs to be swabed */
1884                 lstat = lnet_ping_sts_next(lstat);
1885         }
1886 }
1887
1888 int
1889 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1890 {
1891         if (!pinfo)
1892                 return -EINVAL;
1893         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1894                 return -EPROTO;
1895         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1896                 return -EPROTO;
1897         /* Loopback is guaranteed to be present */
1898         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1899                 return -ERANGE;
1900         if (LNET_PING_INFO_LONI(pinfo) != LNET_NID_LO_0)
1901                 return -EPROTO;
1902         return 0;
1903 }
1904
1905 static void
1906 lnet_ping_target_destroy(void)
1907 {
1908         struct lnet_net *net;
1909         struct lnet_ni  *ni;
1910
1911         lnet_net_lock(LNET_LOCK_EX);
1912
1913         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1914                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1915                         lnet_ni_lock(ni);
1916                         ni->ni_status = NULL;
1917                         lnet_ni_unlock(ni);
1918                 }
1919         }
1920
1921         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1922         the_lnet.ln_ping_target = NULL;
1923
1924         lnet_net_unlock(LNET_LOCK_EX);
1925 }
1926
1927 static void
1928 lnet_ping_target_event_handler(struct lnet_event *event)
1929 {
1930         struct lnet_ping_buffer *pbuf = event->md_user_ptr;
1931
1932         if (event->unlinked)
1933                 lnet_ping_buffer_decref(pbuf);
1934 }
1935
1936 static int
1937 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1938                        struct lnet_handle_md *ping_mdh,
1939                        int ni_bytes, bool set_eq)
1940 {
1941         struct lnet_processid id = {
1942                 .nid = LNET_ANY_NID,
1943                 .pid = LNET_PID_ANY
1944         };
1945         struct lnet_me *me;
1946         struct lnet_md md = { NULL };
1947         int rc;
1948
1949         if (set_eq)
1950                 the_lnet.ln_ping_target_handler =
1951                         lnet_ping_target_event_handler;
1952
1953         *ppbuf = lnet_ping_target_create(ni_bytes);
1954         if (*ppbuf == NULL) {
1955                 rc = -ENOMEM;
1956                 goto fail_free_eq;
1957         }
1958
1959         /* Ping target ME/MD */
1960         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
1961                           LNET_PROTO_PING_MATCHBITS, 0,
1962                           LNET_UNLINK, LNET_INS_AFTER);
1963         if (IS_ERR(me)) {
1964                 rc = PTR_ERR(me);
1965                 CERROR("Can't create ping target ME: %d\n", rc);
1966                 goto fail_decref_ping_buffer;
1967         }
1968
1969         /* initialize md content */
1970         md.start     = &(*ppbuf)->pb_info;
1971         md.length    = (*ppbuf)->pb_nbytes;
1972         md.threshold = LNET_MD_THRESH_INF;
1973         md.max_size  = 0;
1974         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1975                        LNET_MD_MANAGE_REMOTE;
1976         md.handler   = the_lnet.ln_ping_target_handler;
1977         md.user_ptr  = *ppbuf;
1978
1979         rc = LNetMDAttach(me, &md, LNET_RETAIN, ping_mdh);
1980         if (rc != 0) {
1981                 CERROR("Can't attach ping target MD: %d\n", rc);
1982                 goto fail_decref_ping_buffer;
1983         }
1984         lnet_ping_buffer_addref(*ppbuf);
1985
1986         return 0;
1987
1988 fail_decref_ping_buffer:
1989         LASSERT(atomic_read(&(*ppbuf)->pb_refcnt) == 1);
1990         lnet_ping_buffer_decref(*ppbuf);
1991         *ppbuf = NULL;
1992 fail_free_eq:
1993         return rc;
1994 }
1995
1996 static void
1997 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1998                     struct lnet_handle_md *ping_mdh)
1999 {
2000         LNetMDUnlink(*ping_mdh);
2001         LNetInvalidateMDHandle(ping_mdh);
2002
2003         /* NB the MD could be busy; this just starts the unlink */
2004         wait_var_event_warning(&pbuf->pb_refcnt,
2005                                atomic_read(&pbuf->pb_refcnt) <= 1,
2006                                "Still waiting for ping data MD to unlink\n");
2007 }
2008
2009 static void
2010 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
2011 {
2012         struct lnet_ni *ni;
2013         struct lnet_net *net;
2014         struct lnet_ni_status *ns, *end;
2015         struct lnet_ni_large_status *lns, *lend;
2016         int rc;
2017
2018         pbuf->pb_info.pi_nnis = 0;
2019         ns = &pbuf->pb_info.pi_ni[0];
2020         end = (void *)&pbuf->pb_info + pbuf->pb_nbytes;
2021         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2022                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2023                         if (!nid_is_nid4(&ni->ni_nid)) {
2024                                 if (ns == &pbuf->pb_info.pi_ni[1]) {
2025                                         /* This is primary, and it is long */
2026                                         pbuf->pb_info.pi_features |=
2027                                                 LNET_PING_FEAT_PRIMARY_LARGE;
2028                                 }
2029                                 continue;
2030                         }
2031                         LASSERT(ns + 1 <= end);
2032                         ns->ns_nid = lnet_nid_to_nid4(&ni->ni_nid);
2033
2034                         lnet_ni_lock(ni);
2035                         ns->ns_status = lnet_ni_get_status_locked(ni);
2036                         ni->ni_status = &ns->ns_status;
2037                         lnet_ni_unlock(ni);
2038
2039                         pbuf->pb_info.pi_nnis++;
2040                         ns++;
2041                 }
2042         }
2043
2044         lns = (void *)ns;
2045         lend = (void *)end;
2046         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2047                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2048                         if (nid_is_nid4(&ni->ni_nid))
2049                                 continue;
2050                         LASSERT(lns + 1 <= lend);
2051
2052                         lns->ns_nid = ni->ni_nid;
2053
2054                         lnet_ni_lock(ni);
2055                         lns->ns_status = lnet_ni_get_status_locked(ni);
2056                         ni->ni_status = &lns->ns_status;
2057                         lnet_ni_unlock(ni);
2058
2059                         lns = lnet_ping_sts_next(lns);
2060                 }
2061         }
2062         if ((void *)lns > (void *)ns) {
2063                 /* Record total info size */
2064                 pbuf->pb_info.pi_ni[0].ns_msg_size =
2065                         (void *)lns - (void *)&pbuf->pb_info;
2066                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_LARGE_ADDR;
2067         }
2068
2069         /* We (ab)use the ns_status of the loopback interface to
2070          * transmit the sequence number. The first interface listed
2071          * must be the loopback interface.
2072          */
2073         rc = lnet_ping_info_validate(&pbuf->pb_info);
2074         if (rc) {
2075                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
2076                 LBUG();
2077         }
2078         LNET_PING_BUFFER_SEQNO(pbuf) =
2079                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
2080 }
2081
2082 static void
2083 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
2084                         struct lnet_handle_md ping_mdh)
2085 __must_hold(&the_lnet.ln_api_mutex)
2086 {
2087         struct lnet_ping_buffer *old_pbuf = NULL;
2088         struct lnet_handle_md old_ping_md;
2089
2090         /* switch the NIs to point to the new ping info created */
2091         lnet_net_lock(LNET_LOCK_EX);
2092
2093         if (!the_lnet.ln_routing)
2094                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
2095         if (!lnet_peer_discovery_disabled)
2096                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
2097
2098         /* Ensure only known feature bits have been set. */
2099         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
2100         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
2101
2102         lnet_ping_target_install_locked(pbuf);
2103
2104         if (the_lnet.ln_ping_target) {
2105                 old_pbuf = the_lnet.ln_ping_target;
2106                 old_ping_md = the_lnet.ln_ping_target_md;
2107         }
2108         the_lnet.ln_ping_target_md = ping_mdh;
2109         the_lnet.ln_ping_target = pbuf;
2110
2111         lnet_net_unlock(LNET_LOCK_EX);
2112
2113         if (old_pbuf) {
2114                 /* unlink and free the old ping info.
2115                  * There may be outstanding traffic on this MD, and
2116                  * ln_api_mutex may be required to finalize that
2117                  * traffic. Release ln_api_mutex while we wait for
2118                  * refs on this ping buffer to drop
2119                  */
2120                 mutex_unlock(&the_lnet.ln_api_mutex);
2121                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
2122                 mutex_lock(&the_lnet.ln_api_mutex);
2123                 lnet_ping_buffer_decref(old_pbuf);
2124         }
2125
2126         lnet_push_update_to_peers(0);
2127 }
2128
2129 static void
2130 lnet_ping_target_fini(void)
2131 {
2132         lnet_ping_md_unlink(the_lnet.ln_ping_target,
2133                             &the_lnet.ln_ping_target_md);
2134
2135         lnet_assert_handler_unused(the_lnet.ln_ping_target_handler, true);
2136         lnet_ping_target_destroy();
2137 }
2138
2139 /* Resize the push target. */
2140 int lnet_push_target_resize(void)
2141 {
2142         struct lnet_handle_md mdh;
2143         struct lnet_handle_md old_mdh;
2144         struct lnet_ping_buffer *pbuf;
2145         struct lnet_ping_buffer *old_pbuf;
2146         int nbytes;
2147         int rc;
2148
2149 again:
2150         nbytes = the_lnet.ln_push_target_nbytes;
2151         if (nbytes <= 0) {
2152                 CDEBUG(D_NET, "Invalid nbytes %d\n", nbytes);
2153                 return -EINVAL;
2154         }
2155
2156         /* NB: lnet_ping_buffer_alloc() sets pbuf refcount to 1. That ref is
2157          * dropped when we need to resize again (see "old_pbuf" below) or when
2158          * LNet is shutdown (see lnet_push_target_fini())
2159          */
2160         pbuf = lnet_ping_buffer_alloc(nbytes, GFP_NOFS);
2161         if (!pbuf) {
2162                 CDEBUG(D_NET, "Can't allocate pbuf for nbytes %d\n", nbytes);
2163                 return -ENOMEM;
2164         }
2165
2166         rc = lnet_push_target_post(pbuf, &mdh);
2167         if (rc) {
2168                 CDEBUG(D_NET, "Failed to post push target: %d\n", rc);
2169                 lnet_ping_buffer_decref(pbuf);
2170                 return rc;
2171         }
2172
2173         lnet_net_lock(LNET_LOCK_EX);
2174         old_pbuf = the_lnet.ln_push_target;
2175         old_mdh = the_lnet.ln_push_target_md;
2176         the_lnet.ln_push_target = pbuf;
2177         the_lnet.ln_push_target_md = mdh;
2178         lnet_net_unlock(LNET_LOCK_EX);
2179
2180         if (old_pbuf) {
2181                 LNetMDUnlink(old_mdh);
2182                 /* Drop ref set by lnet_ping_buffer_alloc() */
2183                 lnet_ping_buffer_decref(old_pbuf);
2184         }
2185
2186         /* Received another push or reply that requires a larger buffer */
2187         if (nbytes < the_lnet.ln_push_target_nbytes)
2188                 goto again;
2189
2190         CDEBUG(D_NET, "nbytes %d success\n", nbytes);
2191         return 0;
2192 }
2193
2194 int lnet_push_target_post(struct lnet_ping_buffer *pbuf,
2195                           struct lnet_handle_md *mdhp)
2196 {
2197         struct lnet_processid id = { LNET_ANY_NID, LNET_PID_ANY };
2198         struct lnet_md md = { NULL };
2199         struct lnet_me *me;
2200         int rc;
2201
2202         me = LNetMEAttach(LNET_RESERVED_PORTAL, &id,
2203                           LNET_PROTO_PING_MATCHBITS, 0,
2204                           LNET_UNLINK, LNET_INS_AFTER);
2205         if (IS_ERR(me)) {
2206                 rc = PTR_ERR(me);
2207                 CERROR("Can't create push target ME: %d\n", rc);
2208                 return rc;
2209         }
2210
2211         pbuf->pb_needs_post = false;
2212
2213         /* This reference is dropped by lnet_push_target_event_handler() */
2214         lnet_ping_buffer_addref(pbuf);
2215
2216         /* initialize md content */
2217         md.start     = &pbuf->pb_info;
2218         md.length    = pbuf->pb_nbytes;
2219         md.threshold = 1;
2220         md.max_size  = 0;
2221         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE;
2222         md.user_ptr  = pbuf;
2223         md.handler   = the_lnet.ln_push_target_handler;
2224
2225         rc = LNetMDAttach(me, &md, LNET_UNLINK, mdhp);
2226         if (rc) {
2227                 CERROR("Can't attach push MD: %d\n", rc);
2228                 lnet_ping_buffer_decref(pbuf);
2229                 pbuf->pb_needs_post = true;
2230                 return rc;
2231         }
2232
2233         CDEBUG(D_NET, "posted push target %p\n", pbuf);
2234
2235         return 0;
2236 }
2237
2238 static void lnet_push_target_event_handler(struct lnet_event *ev)
2239 {
2240         struct lnet_ping_buffer *pbuf = ev->md_user_ptr;
2241
2242         CDEBUG(D_NET, "type %d status %d unlinked %d\n", ev->type, ev->status,
2243                ev->unlinked);
2244
2245         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
2246                 lnet_swap_pinginfo(pbuf);
2247
2248         if (ev->type == LNET_EVENT_UNLINK) {
2249                 /* Drop ref added by lnet_push_target_post() */
2250                 lnet_ping_buffer_decref(pbuf);
2251                 return;
2252         }
2253
2254         lnet_peer_push_event(ev);
2255         if (ev->unlinked)
2256                 /* Drop ref added by lnet_push_target_post */
2257                 lnet_ping_buffer_decref(pbuf);
2258 }
2259
2260 /* Initialize the push target. */
2261 static int lnet_push_target_init(void)
2262 {
2263         int rc;
2264
2265         if (the_lnet.ln_push_target)
2266                 return -EALREADY;
2267
2268         the_lnet.ln_push_target_handler =
2269                 lnet_push_target_event_handler;
2270
2271         rc = LNetSetLazyPortal(LNET_RESERVED_PORTAL);
2272         LASSERT(rc == 0);
2273
2274         /* Start at the required minimum, we'll enlarge if required. */
2275         the_lnet.ln_push_target_nbytes = LNET_PING_INFO_MIN_SIZE;
2276
2277         rc = lnet_push_target_resize();
2278         if (rc) {
2279                 LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2280                 the_lnet.ln_push_target_handler = NULL;
2281         }
2282
2283         return rc;
2284 }
2285
2286 /* Clean up the push target. */
2287 static void lnet_push_target_fini(void)
2288 {
2289         if (!the_lnet.ln_push_target)
2290                 return;
2291
2292         /* Unlink and invalidate to prevent new references. */
2293         LNetMDUnlink(the_lnet.ln_push_target_md);
2294         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
2295
2296         /* Wait for the unlink to complete. */
2297         wait_var_event_warning(&the_lnet.ln_push_target->pb_refcnt,
2298                                atomic_read(&the_lnet.ln_push_target->pb_refcnt) <= 1,
2299                                "Still waiting for ping data MD to unlink\n");
2300
2301         /* Drop ref set by lnet_ping_buffer_alloc() */
2302         lnet_ping_buffer_decref(the_lnet.ln_push_target);
2303         the_lnet.ln_push_target = NULL;
2304         the_lnet.ln_push_target_nbytes = 0;
2305
2306         LNetClearLazyPortal(LNET_RESERVED_PORTAL);
2307         lnet_assert_handler_unused(the_lnet.ln_push_target_handler, true);
2308         the_lnet.ln_push_target_handler = NULL;
2309 }
2310
2311 static int
2312 lnet_ni_tq_credits(struct lnet_ni *ni)
2313 {
2314         int     credits;
2315
2316         LASSERT(ni->ni_ncpts >= 1);
2317
2318         if (ni->ni_ncpts == 1)
2319                 return ni->ni_net->net_tunables.lct_max_tx_credits;
2320
2321         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
2322         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
2323         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
2324
2325         return credits;
2326 }
2327
2328 static void
2329 lnet_ni_unlink_locked(struct lnet_ni *ni)
2330 {
2331         /* move it to zombie list and nobody can find it anymore */
2332         LASSERT(!list_empty(&ni->ni_netlist));
2333         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
2334         lnet_ni_decref_locked(ni, 0);
2335 }
2336
2337 static void
2338 lnet_clear_zombies_nis_locked(struct lnet_net *net)
2339 {
2340         int             i;
2341         int             islo;
2342         struct lnet_ni  *ni;
2343         struct list_head *zombie_list = &net->net_ni_zombie;
2344
2345         /*
2346          * Now wait for the NIs I just nuked to show up on the zombie
2347          * list and shut them down in guaranteed thread context
2348          */
2349         i = 2;
2350         while ((ni = list_first_entry_or_null(zombie_list,
2351                                               struct lnet_ni,
2352                                               ni_netlist)) != NULL) {
2353                 int *ref;
2354                 int j;
2355
2356                 list_del_init(&ni->ni_netlist);
2357                 /* the ni should be in deleting state. If it's not it's
2358                  * a bug */
2359                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
2360                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
2361                         if (*ref == 0)
2362                                 continue;
2363                         /* still busy, add it back to zombie list */
2364                         list_add(&ni->ni_netlist, zombie_list);
2365                         break;
2366                 }
2367
2368                 if (!list_empty(&ni->ni_netlist)) {
2369                         /* Unlock mutex while waiting to allow other
2370                          * threads to read the LNet state and fall through
2371                          * to avoid deadlock
2372                          */
2373                         lnet_net_unlock(LNET_LOCK_EX);
2374                         mutex_unlock(&the_lnet.ln_api_mutex);
2375
2376                         ++i;
2377                         if ((i & (-i)) == i) {
2378                                 CDEBUG(D_WARNING,
2379                                        "Waiting for zombie LNI %s\n",
2380                                        libcfs_nidstr(&ni->ni_nid));
2381                         }
2382                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2383
2384                         mutex_lock(&the_lnet.ln_api_mutex);
2385                         lnet_net_lock(LNET_LOCK_EX);
2386                         continue;
2387                 }
2388
2389                 lnet_net_unlock(LNET_LOCK_EX);
2390
2391                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2392
2393                 LASSERT(!in_interrupt());
2394                 /* Holding the LND mutex makes it safe for lnd_shutdown
2395                  * to call module_put(). Module unload cannot finish
2396                  * until lnet_unregister_lnd() completes, and that
2397                  * requires the LND mutex.
2398                  */
2399                 mutex_unlock(&the_lnet.ln_api_mutex);
2400                 mutex_lock(&the_lnet.ln_lnd_mutex);
2401                 (net->net_lnd->lnd_shutdown)(ni);
2402                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2403                 mutex_lock(&the_lnet.ln_api_mutex);
2404
2405                 if (!islo)
2406                         CDEBUG(D_LNI, "Removed LNI %s\n",
2407                               libcfs_nidstr(&ni->ni_nid));
2408
2409                 lnet_ni_free(ni);
2410                 i = 2;
2411                 lnet_net_lock(LNET_LOCK_EX);
2412         }
2413 }
2414
2415 /* shutdown down the NI and release refcount */
2416 static void
2417 lnet_shutdown_lndni(struct lnet_ni *ni)
2418 {
2419         int i;
2420         struct lnet_net *net = ni->ni_net;
2421
2422         lnet_net_lock(LNET_LOCK_EX);
2423         lnet_ni_lock(ni);
2424         ni->ni_state = LNET_NI_STATE_DELETING;
2425         lnet_ni_unlock(ni);
2426         lnet_ni_unlink_locked(ni);
2427         lnet_incr_dlc_seq();
2428         lnet_net_unlock(LNET_LOCK_EX);
2429
2430         /* clear messages for this NI on the lazy portal */
2431         for (i = 0; i < the_lnet.ln_nportals; i++)
2432                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2433
2434         lnet_net_lock(LNET_LOCK_EX);
2435         lnet_clear_zombies_nis_locked(net);
2436         lnet_net_unlock(LNET_LOCK_EX);
2437 }
2438
2439 static void
2440 lnet_shutdown_lndnet(struct lnet_net *net)
2441 {
2442         struct lnet_ni *ni;
2443
2444         lnet_net_lock(LNET_LOCK_EX);
2445
2446         list_del_init(&net->net_list);
2447
2448         while ((ni = list_first_entry_or_null(&net->net_ni_list,
2449                                               struct lnet_ni,
2450                                               ni_netlist)) != NULL) {
2451                 lnet_net_unlock(LNET_LOCK_EX);
2452                 lnet_shutdown_lndni(ni);
2453                 lnet_net_lock(LNET_LOCK_EX);
2454         }
2455
2456         lnet_net_unlock(LNET_LOCK_EX);
2457
2458         /* Do peer table cleanup for this net */
2459         lnet_peer_tables_cleanup(net);
2460
2461         lnet_net_free(net);
2462 }
2463
2464 static void
2465 lnet_shutdown_lndnets(void)
2466 {
2467         struct lnet_net *net;
2468         LIST_HEAD(resend);
2469         struct lnet_msg *msg, *tmp;
2470
2471         /* NB called holding the global mutex */
2472
2473         /* All quiet on the API front */
2474         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING ||
2475                 the_lnet.ln_state == LNET_STATE_STOPPING);
2476         LASSERT(the_lnet.ln_refcount == 0);
2477
2478         lnet_net_lock(LNET_LOCK_EX);
2479         the_lnet.ln_state = LNET_STATE_STOPPING;
2480
2481         /*
2482          * move the nets to the zombie list to avoid them being
2483          * picked up for new work. LONET is also included in the
2484          * Nets that will be moved to the zombie list
2485          */
2486         list_splice_init(&the_lnet.ln_nets, &the_lnet.ln_net_zombie);
2487
2488         /* Drop the cached loopback Net. */
2489         if (the_lnet.ln_loni != NULL) {
2490                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2491                 the_lnet.ln_loni = NULL;
2492         }
2493         lnet_net_unlock(LNET_LOCK_EX);
2494
2495         /* iterate through the net zombie list and delete each net */
2496         while ((net = list_first_entry_or_null(&the_lnet.ln_net_zombie,
2497                                                struct lnet_net,
2498                                                net_list)) != NULL)
2499                 lnet_shutdown_lndnet(net);
2500
2501         spin_lock(&the_lnet.ln_msg_resend_lock);
2502         list_splice(&the_lnet.ln_msg_resend, &resend);
2503         spin_unlock(&the_lnet.ln_msg_resend_lock);
2504
2505         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2506                 list_del_init(&msg->msg_list);
2507                 msg->msg_no_resend = true;
2508                 lnet_finalize(msg, -ECANCELED);
2509         }
2510
2511         lnet_net_lock(LNET_LOCK_EX);
2512         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2513         lnet_net_unlock(LNET_LOCK_EX);
2514 }
2515
2516 static int
2517 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2518 {
2519         int                     rc = -EINVAL;
2520         struct lnet_tx_queue    *tq;
2521         int                     i;
2522         struct lnet_net         *net = ni->ni_net;
2523
2524         mutex_lock(&the_lnet.ln_lnd_mutex);
2525
2526         if (tun) {
2527                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2528                 ni->ni_lnd_tunables_set = true;
2529         }
2530
2531         rc = (net->net_lnd->lnd_startup)(ni);
2532
2533         mutex_unlock(&the_lnet.ln_lnd_mutex);
2534
2535         if (rc != 0) {
2536                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2537                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2538                 goto failed0;
2539         }
2540
2541         /* We keep a reference on the loopback net through the loopback NI */
2542         if (net->net_lnd->lnd_type == LOLND) {
2543                 lnet_ni_addref(ni);
2544                 LASSERT(the_lnet.ln_loni == NULL);
2545                 the_lnet.ln_loni = ni;
2546                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2547                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2548                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2549                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2550                 return 0;
2551         }
2552
2553         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2554             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2555                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2556                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2557                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2558                                         "" : "per-peer ");
2559                 /* shutdown the NI since if we get here then it must've already
2560                  * been started
2561                  */
2562                 lnet_shutdown_lndni(ni);
2563                 return -EINVAL;
2564         }
2565
2566         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2567                 tq->tq_credits_min =
2568                 tq->tq_credits_max =
2569                 tq->tq_credits = lnet_ni_tq_credits(ni);
2570         }
2571
2572         atomic_set(&ni->ni_tx_credits,
2573                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2574         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2575
2576         /* Nodes with small feet have little entropy. The NID for this
2577          * node gives the most entropy in the low bits.
2578          */
2579         add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
2580
2581         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2582                 libcfs_nidstr(&ni->ni_nid),
2583                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2584                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2585                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2586                 ni->ni_net->net_tunables.lct_peer_timeout);
2587
2588         return 0;
2589 failed0:
2590         lnet_ni_free(ni);
2591         return rc;
2592 }
2593
2594 static const struct lnet_lnd *lnet_load_lnd(u32 lnd_type)
2595 {
2596         const struct lnet_lnd *lnd;
2597         int rc = 0;
2598
2599         mutex_lock(&the_lnet.ln_lnd_mutex);
2600         lnd = lnet_find_lnd_by_type(lnd_type);
2601         if (!lnd) {
2602                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2603                 rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2604                 mutex_lock(&the_lnet.ln_lnd_mutex);
2605
2606                 lnd = lnet_find_lnd_by_type(lnd_type);
2607                 if (!lnd) {
2608                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2609                         CERROR("Can't load LND %s, module %s, rc=%d\n",
2610                         libcfs_lnd2str(lnd_type),
2611                         libcfs_lnd2modname(lnd_type), rc);
2612 #ifndef HAVE_MODULE_LOADING_SUPPORT
2613                         LCONSOLE_ERROR_MSG(0x104,
2614                                            "Your kernel must be compiled with kernel module loading support.");
2615 #endif
2616                         return ERR_PTR(-EINVAL);
2617                 }
2618         }
2619         mutex_unlock(&the_lnet.ln_lnd_mutex);
2620
2621         return lnd;
2622 }
2623
2624 static int
2625 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2626 {
2627         struct lnet_ni *ni;
2628         struct lnet_net *net_l = NULL;
2629         LIST_HEAD(local_ni_list);
2630         int rc;
2631         int ni_count = 0;
2632         __u32 lnd_type;
2633         const struct lnet_lnd  *lnd;
2634         int peer_timeout =
2635                 net->net_tunables.lct_peer_timeout;
2636         int maxtxcredits =
2637                 net->net_tunables.lct_max_tx_credits;
2638         int peerrtrcredits =
2639                 net->net_tunables.lct_peer_rtr_credits;
2640
2641         /*
2642          * make sure that this net is unique. If it isn't then
2643          * we are adding interfaces to an already existing network, and
2644          * 'net' is just a convenient way to pass in the list.
2645          * if it is unique we need to find the LND and load it if
2646          * necessary.
2647          */
2648         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2649                 lnd_type = LNET_NETTYP(net->net_id);
2650
2651                 lnd = lnet_load_lnd(lnd_type);
2652                 if (IS_ERR(lnd)) {
2653                         rc = PTR_ERR(lnd);
2654                         goto failed0;
2655                 }
2656
2657                 mutex_lock(&the_lnet.ln_lnd_mutex);
2658                 net->net_lnd = lnd;
2659                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2660
2661                 net_l = net;
2662         }
2663
2664         /*
2665          * net_l: if the network being added is unique then net_l
2666          *        will point to that network
2667          *        if the network being added is not unique then
2668          *        net_l points to the existing network.
2669          *
2670          * When we enter the loop below, we'll pick NIs off he
2671          * network beign added and start them up, then add them to
2672          * a local ni list. Once we've successfully started all
2673          * the NIs then we join the local NI list (of started up
2674          * networks) with the net_l->net_ni_list, which should
2675          * point to the correct network to add the new ni list to
2676          *
2677          * If any of the new NIs fail to start up, then we want to
2678          * iterate through the local ni list, which should include
2679          * any NIs which were successfully started up, and shut
2680          * them down.
2681          *
2682          * After than we want to delete the network being added,
2683          * to avoid a memory leak.
2684          */
2685         while ((ni = list_first_entry_or_null(&net->net_ni_added,
2686                                               struct lnet_ni,
2687                                               ni_netlist)) != NULL) {
2688                 list_del_init(&ni->ni_netlist);
2689
2690                 /* make sure that the the NI we're about to start
2691                  * up is actually unique. if it's not fail. */
2692                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2693                                         ni->ni_interface)) {
2694                         rc = -EEXIST;
2695                         goto failed1;
2696                 }
2697
2698                 /* adjust the pointer the parent network, just in case it
2699                  * the net is a duplicate */
2700                 ni->ni_net = net_l;
2701
2702                 rc = lnet_startup_lndni(ni, tun);
2703
2704                 if (rc != 0)
2705                         goto failed1;
2706
2707                 lnet_ni_addref(ni);
2708                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2709
2710                 ni_count++;
2711         }
2712
2713         lnet_net_lock(LNET_LOCK_EX);
2714         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2715         lnet_incr_dlc_seq();
2716
2717         list_for_each_entry(ni, &net_l->net_ni_list, ni_netlist) {
2718                 if (!ni)
2719                         break;
2720                 lnet_ni_lock(ni);
2721                 ni->ni_state = LNET_NI_STATE_ACTIVE;
2722                 lnet_ni_unlock(ni);
2723         }
2724         lnet_net_unlock(LNET_LOCK_EX);
2725
2726         /* if the network is not unique then we don't want to keep
2727          * it around after we're done. Free it. Otherwise add that
2728          * net to the global the_lnet.ln_nets */
2729         if (net_l != net && net_l != NULL) {
2730                 /*
2731                  * TODO - note. currently the tunables can not be updated
2732                  * once added
2733                  */
2734                 lnet_net_free(net);
2735         } else {
2736                 /*
2737                  * restore tunables after it has been overwitten by the
2738                  * lnd
2739                  */
2740                 if (peer_timeout != -1)
2741                         net->net_tunables.lct_peer_timeout = peer_timeout;
2742                 if (maxtxcredits != -1)
2743                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2744                 if (peerrtrcredits != -1)
2745                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2746
2747                 lnet_net_lock(LNET_LOCK_EX);
2748                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2749                 lnet_net_unlock(LNET_LOCK_EX);
2750         }
2751
2752         return ni_count;
2753
2754 failed1:
2755         /*
2756          * shutdown the new NIs that are being started up
2757          * free the NET being started
2758          */
2759         while ((ni = list_first_entry_or_null(&local_ni_list,
2760                                               struct lnet_ni,
2761                                               ni_netlist)) != NULL)
2762                 lnet_shutdown_lndni(ni);
2763
2764 failed0:
2765         lnet_net_free(net);
2766
2767         return rc;
2768 }
2769
2770 static int
2771 lnet_startup_lndnets(struct list_head *netlist)
2772 {
2773         struct lnet_net         *net;
2774         int                     rc;
2775         int                     ni_count = 0;
2776
2777         /*
2778          * Change to running state before bringing up the LNDs. This
2779          * allows lnet_shutdown_lndnets() to assert that we've passed
2780          * through here.
2781          */
2782         lnet_net_lock(LNET_LOCK_EX);
2783         the_lnet.ln_state = LNET_STATE_RUNNING;
2784         lnet_net_unlock(LNET_LOCK_EX);
2785
2786         while ((net = list_first_entry_or_null(netlist,
2787                                                struct lnet_net,
2788                                                net_list)) != NULL) {
2789                 list_del_init(&net->net_list);
2790
2791                 rc = lnet_startup_lndnet(net, NULL);
2792
2793                 if (rc < 0)
2794                         goto failed;
2795
2796                 ni_count += rc;
2797         }
2798
2799         return ni_count;
2800 failed:
2801         lnet_shutdown_lndnets();
2802
2803         return rc;
2804 }
2805
2806 static int lnet_genl_parse_list(struct sk_buff *msg,
2807                                 const struct ln_key_list *data[], u16 idx)
2808 {
2809         const struct ln_key_list *list = data[idx];
2810         const struct ln_key_props *props;
2811         struct nlattr *node;
2812         u16 count;
2813
2814         if (!list)
2815                 return 0;
2816
2817         if (!list->lkl_maxattr)
2818                 return -ERANGE;
2819
2820         props = list->lkl_list;
2821         if (!props)
2822                 return -EINVAL;
2823
2824         node = nla_nest_start(msg, LN_SCALAR_ATTR_LIST);
2825         if (!node)
2826                 return -ENOBUFS;
2827
2828         for (count = 1; count <= list->lkl_maxattr; count++) {
2829                 struct nlattr *key = nla_nest_start(msg, count);
2830
2831                 if (!key)
2832                         return -EMSGSIZE;
2833
2834                 if (count == 1)
2835                         nla_put_u16(msg, LN_SCALAR_ATTR_LIST_SIZE,
2836                                     list->lkl_maxattr);
2837
2838                 nla_put_u16(msg, LN_SCALAR_ATTR_INDEX, count);
2839                 if (props[count].lkp_value)
2840                         nla_put_string(msg, LN_SCALAR_ATTR_VALUE,
2841                                        props[count].lkp_value);
2842                 if (props[count].lkp_key_format)
2843                         nla_put_u16(msg, LN_SCALAR_ATTR_KEY_FORMAT,
2844                                     props[count].lkp_key_format);
2845                 nla_put_u16(msg, LN_SCALAR_ATTR_NLA_TYPE,
2846                             props[count].lkp_data_type);
2847                 if (props[count].lkp_data_type == NLA_NESTED) {
2848                         int rc;
2849
2850                         rc = lnet_genl_parse_list(msg, data, ++idx);
2851                         if (rc < 0)
2852                                 return rc;
2853                         idx = rc;
2854                 }
2855
2856                 nla_nest_end(msg, key);
2857         }
2858
2859         nla_nest_end(msg, node);
2860         return idx;
2861 }
2862
2863 int lnet_genl_send_scalar_list(struct sk_buff *msg, u32 portid, u32 seq,
2864                                const struct genl_family *family, int flags,
2865                                u8 cmd, const struct ln_key_list *data[])
2866 {
2867         int rc = 0;
2868         void *hdr;
2869
2870         if (!data[0])
2871                 return -EINVAL;
2872
2873         hdr = genlmsg_put(msg, portid, seq, family, flags, cmd);
2874         if (!hdr)
2875                 GOTO(canceled, rc = -EMSGSIZE);
2876
2877         rc = lnet_genl_parse_list(msg, data, 0);
2878         if (rc < 0)
2879                 GOTO(canceled, rc);
2880
2881         genlmsg_end(msg, hdr);
2882 canceled:
2883         if (rc < 0)
2884                 genlmsg_cancel(msg, hdr);
2885         return rc > 0 ? 0 : rc;
2886 }
2887 EXPORT_SYMBOL(lnet_genl_send_scalar_list);
2888
2889 static struct genl_family lnet_family;
2890
2891 /**
2892  * Initialize LNet library.
2893  *
2894  * Automatically called at module loading time. Caller has to call
2895  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2896  * latter returned 0. It must be called exactly once.
2897  *
2898  * \retval 0 on success
2899  * \retval -ve on failures.
2900  */
2901 int lnet_lib_init(void)
2902 {
2903         int rc;
2904
2905         lnet_assert_wire_constants();
2906
2907         /* refer to global cfs_cpt_table for now */
2908         the_lnet.ln_cpt_table = cfs_cpt_tab;
2909         the_lnet.ln_cpt_number = cfs_cpt_number(cfs_cpt_tab);
2910
2911         LASSERT(the_lnet.ln_cpt_number > 0);
2912         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2913                 /* we are under risk of consuming all lh_cookie */
2914                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2915                        "please change setting of CPT-table and retry\n",
2916                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2917                 return -E2BIG;
2918         }
2919
2920         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2921                 the_lnet.ln_cpt_bits++;
2922
2923         rc = lnet_create_locks();
2924         if (rc != 0) {
2925                 CERROR("Can't create LNet global locks: %d\n", rc);
2926                 return rc;
2927         }
2928
2929         rc = genl_register_family(&lnet_family);
2930         if (rc != 0) {
2931                 lnet_destroy_locks();
2932                 CERROR("Can't register LNet netlink family: %d\n", rc);
2933                 return rc;
2934         }
2935
2936         the_lnet.ln_refcount = 0;
2937         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2938         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2939
2940         /* The hash table size is the number of bits it takes to express the set
2941          * ln_num_routes, minus 1 (better to under estimate than over so we
2942          * don't waste memory). */
2943         if (rnet_htable_size <= 0)
2944                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2945         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2946                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2947         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2948                                            order_base_2(rnet_htable_size) - 1);
2949
2950         /* All LNDs apart from the LOLND are in separate modules.  They
2951          * register themselves when their module loads, and unregister
2952          * themselves when their module is unloaded. */
2953         lnet_register_lnd(&the_lolnd);
2954         return 0;
2955 }
2956
2957 /**
2958  * Finalize LNet library.
2959  *
2960  * \pre lnet_lib_init() called with success.
2961  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2962  *
2963  * As this happens at module-unload, all lnds must already be unloaded,
2964  * so they must already be unregistered.
2965  */
2966 void lnet_lib_exit(void)
2967 {
2968         int i;
2969
2970         LASSERT(the_lnet.ln_refcount == 0);
2971         lnet_unregister_lnd(&the_lolnd);
2972         for (i = 0; i < NUM_LNDS; i++)
2973                 LASSERT(!the_lnet.ln_lnds[i]);
2974         lnet_destroy_locks();
2975         genl_unregister_family(&lnet_family);
2976 }
2977
2978 /**
2979  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2980  *
2981  * Users must call this function at least once before any other functions.
2982  * For each successful call there must be a corresponding call to
2983  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2984  * ignored.
2985  *
2986  * The PID used by LNet may be different from the one requested.
2987  * See LNetGetId().
2988  *
2989  * \param requested_pid PID requested by the caller.
2990  *
2991  * \return >= 0 on success, and < 0 error code on failures.
2992  */
2993 int
2994 LNetNIInit(lnet_pid_t requested_pid)
2995 {
2996         int im_a_router = 0;
2997         int rc;
2998         int ni_bytes;
2999         struct lnet_ping_buffer *pbuf;
3000         struct lnet_handle_md ping_mdh;
3001         LIST_HEAD(net_head);
3002         struct lnet_net *net;
3003
3004         mutex_lock(&the_lnet.ln_api_mutex);
3005
3006         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
3007
3008         if (the_lnet.ln_state == LNET_STATE_STOPPING) {
3009                 mutex_unlock(&the_lnet.ln_api_mutex);
3010                 return -ESHUTDOWN;
3011         }
3012
3013         if (the_lnet.ln_refcount > 0) {
3014                 rc = the_lnet.ln_refcount++;
3015                 mutex_unlock(&the_lnet.ln_api_mutex);
3016                 return rc;
3017         }
3018
3019         rc = lnet_prepare(requested_pid);
3020         if (rc != 0) {
3021                 mutex_unlock(&the_lnet.ln_api_mutex);
3022                 return rc;
3023         }
3024
3025         /* create a network for Loopback network */
3026         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
3027         if (net == NULL) {
3028                 rc = -ENOMEM;
3029                 goto err_empty_list;
3030         }
3031
3032         /* Add in the loopback NI */
3033         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
3034                 rc = -ENOMEM;
3035                 goto err_empty_list;
3036         }
3037
3038         if (use_tcp_bonding)
3039                 CWARN("use_tcp_bonding has been removed. Use Multi-Rail and Dynamic Discovery instead, see LU-13641\n");
3040
3041         /* If LNet is being initialized via DLC it is possible
3042          * that the user requests not to load module parameters (ones which
3043          * are supported by DLC) on initialization.  Therefore, make sure not
3044          * to load networks, routes and forwarding from module parameters
3045          * in this case.  On cleanup in case of failure only clean up
3046          * routes if it has been loaded */
3047         if (!the_lnet.ln_nis_from_mod_params) {
3048                 rc = lnet_parse_networks(&net_head, lnet_get_networks());
3049                 if (rc < 0)
3050                         goto err_empty_list;
3051         }
3052
3053         rc = lnet_startup_lndnets(&net_head);
3054         if (rc < 0)
3055                 goto err_empty_list;
3056
3057         if (!the_lnet.ln_nis_from_mod_params) {
3058                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
3059                 if (rc != 0)
3060                         goto err_shutdown_lndnis;
3061
3062                 rc = lnet_rtrpools_alloc(im_a_router);
3063                 if (rc != 0)
3064                         goto err_destroy_routes;
3065         }
3066
3067         rc = lnet_acceptor_start();
3068         if (rc != 0)
3069                 goto err_destroy_routes;
3070
3071         the_lnet.ln_refcount = 1;
3072         /* Now I may use my own API functions... */
3073
3074         ni_bytes = LNET_PING_INFO_HDR_SIZE;
3075         list_for_each_entry(net, &the_lnet.ln_nets, net_list)
3076                 ni_bytes += lnet_get_net_ni_bytes_locked(net);
3077
3078         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_bytes, true);
3079         if (rc != 0)
3080                 goto err_acceptor_stop;
3081
3082         lnet_ping_target_update(pbuf, ping_mdh);
3083
3084         the_lnet.ln_mt_handler = lnet_mt_event_handler;
3085
3086         rc = lnet_push_target_init();
3087         if (rc != 0)
3088                 goto err_stop_ping;
3089
3090         rc = lnet_monitor_thr_start();
3091         if (rc != 0)
3092                 goto err_destroy_push_target;
3093
3094         rc = lnet_peer_discovery_start();
3095         if (rc != 0)
3096                 goto err_stop_monitor_thr;
3097
3098         lnet_fault_init();
3099         lnet_router_debugfs_init();
3100
3101         mutex_unlock(&the_lnet.ln_api_mutex);
3102
3103         complete_all(&the_lnet.ln_started);
3104
3105         /* wait for all routers to start */
3106         lnet_wait_router_start();
3107
3108         return 0;
3109
3110 err_stop_monitor_thr:
3111         lnet_monitor_thr_stop();
3112 err_destroy_push_target:
3113         lnet_push_target_fini();
3114 err_stop_ping:
3115         lnet_ping_target_fini();
3116 err_acceptor_stop:
3117         the_lnet.ln_refcount = 0;
3118         lnet_acceptor_stop();
3119 err_destroy_routes:
3120         if (!the_lnet.ln_nis_from_mod_params)
3121                 lnet_destroy_routes();
3122 err_shutdown_lndnis:
3123         lnet_shutdown_lndnets();
3124 err_empty_list:
3125         lnet_unprepare();
3126         LASSERT(rc < 0);
3127         mutex_unlock(&the_lnet.ln_api_mutex);
3128         while ((net = list_first_entry_or_null(&net_head,
3129                                                struct lnet_net,
3130                                                net_list)) != NULL) {
3131                 list_del_init(&net->net_list);
3132                 lnet_net_free(net);
3133         }
3134         return rc;
3135 }
3136 EXPORT_SYMBOL(LNetNIInit);
3137
3138 /**
3139  * Stop LNet interfaces, routing, and forwarding.
3140  *
3141  * Users must call this function once for each successful call to LNetNIInit().
3142  * Once the LNetNIFini() operation has been started, the results of pending
3143  * API operations are undefined.
3144  *
3145  * \return always 0 for current implementation.
3146  */
3147 int
3148 LNetNIFini(void)
3149 {
3150         mutex_lock(&the_lnet.ln_api_mutex);
3151
3152         LASSERT(the_lnet.ln_refcount > 0);
3153
3154         if (the_lnet.ln_refcount != 1) {
3155                 the_lnet.ln_refcount--;
3156         } else {
3157                 LASSERT(!the_lnet.ln_niinit_self);
3158
3159                 lnet_net_lock(LNET_LOCK_EX);
3160                 the_lnet.ln_state = LNET_STATE_STOPPING;
3161                 lnet_net_unlock(LNET_LOCK_EX);
3162
3163                 lnet_fault_fini();
3164
3165                 lnet_router_debugfs_fini();
3166                 lnet_peer_discovery_stop();
3167                 lnet_monitor_thr_stop();
3168                 lnet_push_target_fini();
3169                 lnet_ping_target_fini();
3170
3171                 /* Teardown fns that use my own API functions BEFORE here */
3172                 the_lnet.ln_refcount = 0;
3173
3174                 lnet_acceptor_stop();
3175                 lnet_destroy_routes();
3176                 lnet_shutdown_lndnets();
3177                 lnet_unprepare();
3178         }
3179
3180         mutex_unlock(&the_lnet.ln_api_mutex);
3181         return 0;
3182 }
3183 EXPORT_SYMBOL(LNetNIFini);
3184
3185 /**
3186  * Grabs the ni data from the ni structure and fills the out
3187  * parameters
3188  *
3189  * \param[in] ni network        interface structure
3190  * \param[out] cfg_ni           NI config information
3191  * \param[out] tun              network and LND tunables
3192  */
3193 static void
3194 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
3195                    struct lnet_ioctl_config_lnd_tunables *tun,
3196                    struct lnet_ioctl_element_stats *stats,
3197                    __u32 tun_size)
3198 {
3199         size_t min_size = 0;
3200         int i;
3201
3202         if (!ni || !cfg_ni || !tun || !nid_is_nid4(&ni->ni_nid))
3203                 return;
3204
3205         if (ni->ni_interface != NULL) {
3206                 strncpy(cfg_ni->lic_ni_intf,
3207                         ni->ni_interface,
3208                         sizeof(cfg_ni->lic_ni_intf));
3209         }
3210
3211         cfg_ni->lic_nid = lnet_nid_to_nid4(&ni->ni_nid);
3212         cfg_ni->lic_status = lnet_ni_get_status_locked(ni);
3213         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
3214
3215         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
3216
3217         if (stats) {
3218                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
3219                                                        LNET_STATS_TYPE_SEND);
3220                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
3221                                                        LNET_STATS_TYPE_RECV);
3222                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
3223                                                        LNET_STATS_TYPE_DROP);
3224         }
3225
3226         /*
3227          * tun->lt_tun will always be present, but in order to be
3228          * backwards compatible, we need to deal with the cases when
3229          * tun->lt_tun is smaller than what the kernel has, because it
3230          * comes from an older version of a userspace program, then we'll
3231          * need to copy as much information as we have available space.
3232          */
3233         min_size = tun_size - sizeof(tun->lt_cmn);
3234         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
3235
3236         /* copy over the cpts */
3237         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
3238             ni->ni_cpts == NULL)  {
3239                 for (i = 0; i < ni->ni_ncpts; i++)
3240                         cfg_ni->lic_cpts[i] = i;
3241         } else {
3242                 for (i = 0;
3243                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
3244                      i < LNET_MAX_SHOW_NUM_CPT;
3245                      i++)
3246                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
3247         }
3248         cfg_ni->lic_ncpts = ni->ni_ncpts;
3249 }
3250
3251 /**
3252  * NOTE: This is a legacy function left in the code to be backwards
3253  * compatible with older userspace programs. It should eventually be
3254  * removed.
3255  *
3256  * Grabs the ni data from the ni structure and fills the out
3257  * parameters
3258  *
3259  * \param[in] ni network        interface structure
3260  * \param[out] config           config information
3261  */
3262 static void
3263 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
3264                          struct lnet_ioctl_config_data *config)
3265 {
3266         struct lnet_ioctl_net_config *net_config;
3267         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
3268         size_t min_size, tunable_size = 0;
3269         int i;
3270
3271         if (!ni || !config || !nid_is_nid4(&ni->ni_nid))
3272                 return;
3273
3274         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
3275         if (!net_config)
3276                 return;
3277
3278         if (!ni->ni_interface)
3279                 return;
3280
3281         strncpy(net_config->ni_interface,
3282                 ni->ni_interface,
3283                 sizeof(net_config->ni_interface));
3284
3285         config->cfg_nid = lnet_nid_to_nid4(&ni->ni_nid);
3286         config->cfg_config_u.cfg_net.net_peer_timeout =
3287                 ni->ni_net->net_tunables.lct_peer_timeout;
3288         config->cfg_config_u.cfg_net.net_max_tx_credits =
3289                 ni->ni_net->net_tunables.lct_max_tx_credits;
3290         config->cfg_config_u.cfg_net.net_peer_tx_credits =
3291                 ni->ni_net->net_tunables.lct_peer_tx_credits;
3292         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
3293                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
3294
3295         net_config->ni_status = lnet_ni_get_status_locked(ni);
3296
3297         if (ni->ni_cpts) {
3298                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
3299
3300                 for (i = 0; i < num_cpts; i++)
3301                         net_config->ni_cpts[i] = ni->ni_cpts[i];
3302
3303                 config->cfg_ncpts = num_cpts;
3304         }
3305
3306         /*
3307          * See if user land tools sent in a newer and larger version
3308          * of struct lnet_tunables than what the kernel uses.
3309          */
3310         min_size = sizeof(*config) + sizeof(*net_config);
3311
3312         if (config->cfg_hdr.ioc_len > min_size)
3313                 tunable_size = config->cfg_hdr.ioc_len - min_size;
3314
3315         /* Don't copy too much data to user space */
3316         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
3317         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
3318
3319         if (lnd_cfg && min_size) {
3320                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
3321                 config->cfg_config_u.cfg_net.net_interface_count = 1;
3322
3323                 /* Tell user land that kernel side has less data */
3324                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
3325                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
3326                         config->cfg_hdr.ioc_len -= min_size;
3327                 }
3328         }
3329 }
3330
3331 struct lnet_ni *
3332 lnet_get_ni_idx_locked(int idx)
3333 {
3334         struct lnet_ni          *ni;
3335         struct lnet_net         *net;
3336
3337         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3338                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3339                         if (idx-- == 0)
3340                                 return ni;
3341                 }
3342         }
3343
3344         return NULL;
3345 }
3346
3347 int lnet_get_net_healthv_locked(struct lnet_net *net)
3348 {
3349         struct lnet_ni *ni;
3350         int best_healthv = 0;
3351         int healthv, ni_fatal;
3352
3353         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3354                 healthv = atomic_read(&ni->ni_healthv);
3355                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
3356                 if (!ni_fatal && healthv > best_healthv)
3357                         best_healthv = healthv;
3358         }
3359
3360         return best_healthv;
3361 }
3362
3363 struct lnet_ni *
3364 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
3365 {
3366         struct lnet_ni          *ni;
3367         struct lnet_net         *net = mynet;
3368
3369         /*
3370          * It is possible that the net has been cleaned out while there is
3371          * a message being sent. This function accessed the net without
3372          * checking if the list is empty
3373          */
3374         if (!prev) {
3375                 if (!net)
3376                         net = list_first_entry(&the_lnet.ln_nets,
3377                                                struct lnet_net,
3378                                                net_list);
3379                 if (list_empty(&net->net_ni_list))
3380                         return NULL;
3381                 ni = list_first_entry(&net->net_ni_list, struct lnet_ni,
3382                                       ni_netlist);
3383
3384                 return ni;
3385         }
3386
3387         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
3388                 /* if you reached the end of the ni list and the net is
3389                  * specified, then there are no more nis in that net */
3390                 if (net != NULL)
3391                         return NULL;
3392
3393                 /* we reached the end of this net ni list. move to the
3394                  * next net */
3395                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
3396                         /* no more nets and no more NIs. */
3397                         return NULL;
3398
3399                 /* get the next net */
3400                 net = list_first_entry(&prev->ni_net->net_list, struct lnet_net,
3401                                        net_list);
3402                 if (list_empty(&net->net_ni_list))
3403                         return NULL;
3404                 /* get the ni on it */
3405                 ni = list_first_entry(&net->net_ni_list, struct lnet_ni,
3406                                       ni_netlist);
3407
3408                 return ni;
3409         }
3410
3411         if (list_empty(&prev->ni_netlist))
3412                 return NULL;
3413
3414         /* there are more nis left */
3415         ni = list_first_entry(&prev->ni_netlist, struct lnet_ni, ni_netlist);
3416
3417         return ni;
3418 }
3419
3420 static int
3421 lnet_get_net_config(struct lnet_ioctl_config_data *config)
3422 {
3423         struct lnet_ni *ni;
3424         int cpt;
3425         int rc = -ENOENT;
3426         int idx = config->cfg_count;
3427
3428         cpt = lnet_net_lock_current();
3429
3430         ni = lnet_get_ni_idx_locked(idx);
3431
3432         if (ni != NULL) {
3433                 rc = 0;
3434                 lnet_ni_lock(ni);
3435                 lnet_fill_ni_info_legacy(ni, config);
3436                 lnet_ni_unlock(ni);
3437         }
3438
3439         lnet_net_unlock(cpt);
3440         return rc;
3441 }
3442
3443 static int
3444 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
3445                    struct lnet_ioctl_config_lnd_tunables *tun,
3446                    struct lnet_ioctl_element_stats *stats,
3447                    __u32 tun_size)
3448 {
3449         struct lnet_ni          *ni;
3450         int                     cpt;
3451         int                     rc = -ENOENT;
3452
3453         if (!cfg_ni || !tun || !stats)
3454                 return -EINVAL;
3455
3456         cpt = lnet_net_lock_current();
3457
3458         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
3459
3460         if (ni) {
3461                 rc = 0;
3462                 lnet_ni_lock(ni);
3463                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
3464                 lnet_ni_unlock(ni);
3465         }
3466
3467         lnet_net_unlock(cpt);
3468         return rc;
3469 }
3470
3471 static int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
3472 {
3473         struct lnet_ni *ni;
3474         int rc = -ENOENT;
3475
3476         if (!msg_stats)
3477                 return -EINVAL;
3478
3479         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3480
3481         if (ni) {
3482                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3483                 rc = 0;
3484         }
3485
3486         return rc;
3487 }
3488
3489 static int lnet_add_net_common(struct lnet_net *net,
3490                                struct lnet_ioctl_config_lnd_tunables *tun)
3491 {
3492         struct lnet_handle_md ping_mdh;
3493         struct lnet_ping_buffer *pbuf;
3494         struct lnet_remotenet *rnet;
3495         struct lnet_ni *ni;
3496         u32 net_id;
3497         int rc;
3498
3499         lnet_net_lock(LNET_LOCK_EX);
3500         rnet = lnet_find_rnet_locked(net->net_id);
3501         lnet_net_unlock(LNET_LOCK_EX);
3502         /*
3503          * make sure that the net added doesn't invalidate the current
3504          * configuration LNet is keeping
3505          */
3506         if (rnet) {
3507                 CERROR("Adding net %s will invalidate routing configuration\n",
3508                        libcfs_net2str(net->net_id));
3509                 lnet_net_free(net);
3510                 return -EUSERS;
3511         }
3512
3513         if (tun)
3514                 memcpy(&net->net_tunables,
3515                        &tun->lt_cmn, sizeof(net->net_tunables));
3516         else
3517                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3518
3519         net_id = net->net_id;
3520
3521         rc = lnet_startup_lndnet(net,
3522                                  (tun) ? &tun->lt_tun : NULL);
3523         if (rc < 0)
3524                 return rc;
3525
3526         /* make sure you calculate the correct number of slots in the ping
3527          * buffer. Since the ping info is a flattened list of all the NIs,
3528          * we should allocate enough slots to accomodate the number of NIs
3529          * which will be added.
3530          */
3531         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3532                                     LNET_PING_INFO_HDR_SIZE +
3533                                     lnet_get_ni_bytes(),
3534                                     false);
3535         if (rc < 0) {
3536                 lnet_shutdown_lndnet(net);
3537                 return rc;
3538         }
3539
3540         lnet_net_lock(LNET_LOCK_EX);
3541         net = lnet_get_net_locked(net_id);
3542         LASSERT(net);
3543
3544         /* apply the UDSPs */
3545         rc = lnet_udsp_apply_policies_on_net(net);
3546         if (rc)
3547                 CERROR("Failed to apply UDSPs on local net %s\n",
3548                        libcfs_net2str(net->net_id));
3549
3550         /* At this point we lost track of which NI was just added, so we
3551          * just re-apply the policies on all of the NIs on this net
3552          */
3553         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3554                 rc = lnet_udsp_apply_policies_on_ni(ni);
3555                 if (rc)
3556                         CERROR("Failed to apply UDSPs on ni %s\n",
3557                                libcfs_nidstr(&ni->ni_nid));
3558         }
3559         lnet_net_unlock(LNET_LOCK_EX);
3560
3561         /*
3562          * Start the acceptor thread if this is the first network
3563          * being added that requires the thread.
3564          */
3565         if (net->net_lnd->lnd_accept) {
3566                 rc = lnet_acceptor_start();
3567                 if (rc < 0) {
3568                         /* shutdown the net that we just started */
3569                         CERROR("Failed to start up acceptor thread\n");
3570                         lnet_shutdown_lndnet(net);
3571                         goto failed;
3572                 }
3573         }
3574
3575         lnet_net_lock(LNET_LOCK_EX);
3576         lnet_peer_net_added(net);
3577         lnet_net_unlock(LNET_LOCK_EX);
3578
3579         lnet_ping_target_update(pbuf, ping_mdh);
3580
3581         return 0;
3582
3583 failed:
3584         lnet_ping_md_unlink(pbuf, &ping_mdh);
3585         lnet_ping_buffer_decref(pbuf);
3586         return rc;
3587 }
3588
3589 static void
3590 lnet_set_tune_defaults(struct lnet_ioctl_config_lnd_tunables *tun)
3591 {
3592         if (tun) {
3593                 if (tun->lt_cmn.lct_peer_timeout < 0)
3594                         tun->lt_cmn.lct_peer_timeout = DEFAULT_PEER_TIMEOUT;
3595                 if (!tun->lt_cmn.lct_peer_tx_credits)
3596                         tun->lt_cmn.lct_peer_tx_credits = DEFAULT_PEER_CREDITS;
3597                 if (!tun->lt_cmn.lct_max_tx_credits)
3598                         tun->lt_cmn.lct_max_tx_credits = DEFAULT_CREDITS;
3599         }
3600 }
3601
3602 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3603                                       struct lnet_ioctl_config_lnd_tunables *tun)
3604 {
3605         struct lnet_net *net;
3606         const char *nets;
3607         int rc;
3608         LIST_HEAD(net_head);
3609
3610         rc = lnet_parse_ip2nets(&nets, ip2nets);
3611         if (rc < 0)
3612                 return rc;
3613
3614         rc = lnet_parse_networks(&net_head, nets);
3615         if (rc < 0)
3616                 return rc;
3617
3618         lnet_set_tune_defaults(tun);
3619
3620         mutex_lock(&the_lnet.ln_api_mutex);
3621         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3622                 rc = -ESHUTDOWN;
3623                 goto out;
3624         }
3625
3626         while ((net = list_first_entry_or_null(&net_head,
3627                                                struct lnet_net,
3628                                                net_list)) != NULL) {
3629                 list_del_init(&net->net_list);
3630                 rc = lnet_add_net_common(net, tun);
3631                 if (rc < 0)
3632                         goto out;
3633         }
3634
3635 out:
3636         mutex_unlock(&the_lnet.ln_api_mutex);
3637
3638         while ((net = list_first_entry_or_null(&net_head,
3639                                                struct lnet_net,
3640                                                net_list)) != NULL) {
3641                 list_del_init(&net->net_list);
3642                 lnet_net_free(net);
3643         }
3644         return rc;
3645 }
3646
3647 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf, u32 net_id,
3648                     struct lnet_nid *nid,
3649                     struct lnet_ioctl_config_lnd_tunables *tun)
3650 {
3651         struct lnet_net *net;
3652         struct lnet_ni *ni;
3653         int rc, i;
3654         u32 lnd_type;
3655
3656         /* handle legacy ip2nets from DLC */
3657         if (conf->lic_legacy_ip2nets[0] != '\0')
3658                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3659                                                   tun);
3660
3661         lnd_type = LNET_NETTYP(net_id);
3662
3663         if (!libcfs_isknown_lnd(lnd_type)) {
3664                 CERROR("No valid net and lnd information provided\n");
3665                 return -ENOENT;
3666         }
3667
3668         net = lnet_net_alloc(net_id, NULL);
3669         if (!net)
3670                 return -ENOMEM;
3671
3672         for (i = 0; i < conf->lic_ncpts; i++) {
3673                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER) {
3674                         lnet_net_free(net);
3675                         return -ERANGE;
3676                 }
3677         }
3678
3679         ni = lnet_ni_alloc_w_cpt_array(net, nid, conf->lic_cpts,
3680                                        conf->lic_ncpts, conf->lic_ni_intf);
3681         if (!ni) {
3682                 lnet_net_free(net);
3683                 return -ENOMEM;
3684         }
3685
3686         lnet_set_tune_defaults(tun);
3687
3688         mutex_lock(&the_lnet.ln_api_mutex);
3689         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3690                 lnet_net_free(net);
3691                 rc = -ESHUTDOWN;
3692         } else {
3693                 rc = lnet_add_net_common(net, tun);
3694         }
3695
3696         mutex_unlock(&the_lnet.ln_api_mutex);
3697
3698         /* If NI already exist delete this new unused copy */
3699         if (rc == -EEXIST)
3700                 lnet_ni_free(ni);
3701
3702         return rc;
3703 }
3704
3705 int lnet_dyn_del_ni(struct lnet_nid *nid)
3706 {
3707         struct lnet_net *net;
3708         struct lnet_ni *ni;
3709         u32 net_id = LNET_NID_NET(nid);
3710         struct lnet_ping_buffer *pbuf;
3711         struct lnet_handle_md ping_mdh;
3712         int net_bytes, rc;
3713         bool net_empty;
3714
3715         /* don't allow userspace to shutdown the LOLND */
3716         if (LNET_NETTYP(net_id) == LOLND)
3717                 return -EINVAL;
3718
3719         mutex_lock(&the_lnet.ln_api_mutex);
3720         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3721                 rc = -ESHUTDOWN;
3722                 goto unlock_api_mutex;
3723         }
3724
3725         lnet_net_lock(0);
3726
3727         net = lnet_get_net_locked(net_id);
3728         if (!net) {
3729                 CERROR("net %s not found\n",
3730                        libcfs_net2str(net_id));
3731                 rc = -ENOENT;
3732                 goto unlock_net;
3733         }
3734
3735         if (!nid_addr_is_set(nid)) {
3736                 /* remove the entire net */
3737                 net_bytes = lnet_get_net_ni_bytes_locked(net);
3738
3739                 lnet_net_unlock(0);
3740
3741                 /* create and link a new ping info, before removing the old one */
3742                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3743                                             LNET_PING_INFO_HDR_SIZE +
3744                                             lnet_get_ni_bytes() - net_bytes,
3745                                             false);
3746                 if (rc != 0)
3747                         goto unlock_api_mutex;
3748
3749                 lnet_shutdown_lndnet(net);
3750
3751                 lnet_acceptor_stop();
3752
3753                 lnet_ping_target_update(pbuf, ping_mdh);
3754
3755                 goto unlock_api_mutex;
3756         }
3757
3758         ni = lnet_nid_to_ni_locked(nid, 0);
3759         if (!ni) {
3760                 CERROR("nid %s not found\n", libcfs_nidstr(nid));
3761                 rc = -ENOENT;
3762                 goto unlock_net;
3763         }
3764
3765         net_bytes = lnet_get_net_ni_bytes_locked(net);
3766         net_empty = list_is_singular(&net->net_ni_list);
3767
3768         lnet_net_unlock(0);
3769
3770         /* create and link a new ping info, before removing the old one */
3771         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3772                                     (LNET_PING_INFO_HDR_SIZE +
3773                                      lnet_get_ni_bytes() -
3774                                      lnet_ping_sts_size(&ni->ni_nid)),
3775                                     false);
3776         if (rc != 0)
3777                 goto unlock_api_mutex;
3778
3779         lnet_shutdown_lndni(ni);
3780
3781         lnet_acceptor_stop();
3782
3783         lnet_ping_target_update(pbuf, ping_mdh);
3784
3785         /* check if the net is empty and remove it if it is */
3786         if (net_empty)
3787                 lnet_shutdown_lndnet(net);
3788
3789         goto unlock_api_mutex;
3790
3791 unlock_net:
3792         lnet_net_unlock(0);
3793 unlock_api_mutex:
3794         mutex_unlock(&the_lnet.ln_api_mutex);
3795
3796         return rc;
3797 }
3798
3799 /*
3800  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3801  * They are only expected to be called for unique networks.
3802  * That can be as a result of older DLC library
3803  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3804  */
3805 int
3806 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3807 {
3808         struct lnet_net *net;
3809         LIST_HEAD(net_head);
3810         int rc;
3811         struct lnet_ioctl_config_lnd_tunables tun;
3812         const char *nets = conf->cfg_config_u.cfg_net.net_intf;
3813
3814         /* Create a net/ni structures for the network string */
3815         rc = lnet_parse_networks(&net_head, nets);
3816         if (rc <= 0)
3817                 return rc == 0 ? -EINVAL : rc;
3818
3819         mutex_lock(&the_lnet.ln_api_mutex);
3820         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3821                 rc = -ESHUTDOWN;
3822                 goto out_unlock_clean;
3823         }
3824
3825         if (rc > 1) {
3826                 rc = -EINVAL; /* only add one network per call */
3827                 goto out_unlock_clean;
3828         }
3829
3830         net = list_first_entry(&net_head, struct lnet_net, net_list);
3831         list_del_init(&net->net_list);
3832
3833         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3834
3835         memset(&tun, 0, sizeof(tun));
3836
3837         tun.lt_cmn.lct_peer_timeout =
3838           (!conf->cfg_config_u.cfg_net.net_peer_timeout) ? DEFAULT_PEER_TIMEOUT :
3839                 conf->cfg_config_u.cfg_net.net_peer_timeout;
3840         tun.lt_cmn.lct_peer_tx_credits =
3841           (!conf->cfg_config_u.cfg_net.net_peer_tx_credits) ? DEFAULT_PEER_CREDITS :
3842                 conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3843         tun.lt_cmn.lct_peer_rtr_credits =
3844           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3845         tun.lt_cmn.lct_max_tx_credits =
3846           (!conf->cfg_config_u.cfg_net.net_max_tx_credits) ? DEFAULT_CREDITS :
3847                 conf->cfg_config_u.cfg_net.net_max_tx_credits;
3848
3849         rc = lnet_add_net_common(net, &tun);
3850
3851 out_unlock_clean:
3852         mutex_unlock(&the_lnet.ln_api_mutex);
3853         /* net_head list is empty in success case */
3854         while ((net = list_first_entry_or_null(&net_head,
3855                                                struct lnet_net,
3856                                                net_list)) != NULL) {
3857                 list_del_init(&net->net_list);
3858                 lnet_net_free(net);
3859         }
3860         return rc;
3861 }
3862
3863 int
3864 lnet_dyn_del_net(u32 net_id)
3865 {
3866         struct lnet_net *net;
3867         struct lnet_ping_buffer *pbuf;
3868         struct lnet_handle_md ping_mdh;
3869         int net_ni_bytes, rc;
3870
3871         /* don't allow userspace to shutdown the LOLND */
3872         if (LNET_NETTYP(net_id) == LOLND)
3873                 return -EINVAL;
3874
3875         mutex_lock(&the_lnet.ln_api_mutex);
3876         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
3877                 rc = -ESHUTDOWN;
3878                 goto out;
3879         }
3880
3881         lnet_net_lock(0);
3882
3883         net = lnet_get_net_locked(net_id);
3884         if (net == NULL) {
3885                 lnet_net_unlock(0);
3886                 rc = -EINVAL;
3887                 goto out;
3888         }
3889
3890         net_ni_bytes = lnet_get_net_ni_bytes_locked(net);
3891
3892         lnet_net_unlock(0);
3893
3894         /* create and link a new ping info, before removing the old one */
3895         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3896                                     LNET_PING_INFO_HDR_SIZE +
3897                                     lnet_get_ni_bytes() - net_ni_bytes,
3898                                     false);
3899         if (rc != 0)
3900                 goto out;
3901
3902         lnet_shutdown_lndnet(net);
3903
3904         lnet_acceptor_stop();
3905
3906         lnet_ping_target_update(pbuf, ping_mdh);
3907
3908 out:
3909         mutex_unlock(&the_lnet.ln_api_mutex);
3910
3911         return rc;
3912 }
3913
3914 void lnet_mark_ping_buffer_for_update(void)
3915 {
3916         if (the_lnet.ln_routing)
3917                 return;
3918
3919         atomic_set(&the_lnet.ln_update_ping_buf, 1);
3920         complete(&the_lnet.ln_mt_wait_complete);
3921 }
3922 EXPORT_SYMBOL(lnet_mark_ping_buffer_for_update);
3923
3924 void lnet_update_ping_buffer(struct work_struct *work)
3925 {
3926         struct lnet_ping_buffer *pbuf;
3927         struct lnet_handle_md ping_mdh;
3928
3929         mutex_lock(&the_lnet.ln_api_mutex);
3930
3931         atomic_set(&the_lnet.ln_pb_update_ready, 1);
3932
3933         if ((the_lnet.ln_state == LNET_STATE_RUNNING) &&
3934             !lnet_ping_target_setup(&pbuf, &ping_mdh,
3935                                     LNET_PING_INFO_HDR_SIZE +
3936                                     lnet_get_ni_bytes(),
3937                                     false))
3938                 lnet_ping_target_update(pbuf, ping_mdh);
3939
3940
3941         mutex_unlock(&the_lnet.ln_api_mutex);
3942 }
3943
3944
3945 void lnet_queue_ping_buffer_update(void)
3946 {
3947         /* don't queue pb update if it is not needed */
3948         if (atomic_dec_if_positive(&the_lnet.ln_update_ping_buf) < 0)
3949                 return;
3950
3951         /* don't queue pb update if already queued and not processed */
3952         if (atomic_dec_if_positive(&the_lnet.ln_pb_update_ready) < 0)
3953                 return;
3954
3955         INIT_WORK(&the_lnet.ln_pb_update_work, lnet_update_ping_buffer);
3956         queue_work(the_lnet.ln_pb_update_wq, &the_lnet.ln_pb_update_work);
3957 }
3958
3959 void lnet_incr_dlc_seq(void)
3960 {
3961         atomic_inc(&lnet_dlc_seq_no);
3962 }
3963
3964 __u32 lnet_get_dlc_seq_locked(void)
3965 {
3966         return atomic_read(&lnet_dlc_seq_no);
3967 }
3968
3969 static void
3970 lnet_ni_set_healthv(struct lnet_nid *nid, int value)
3971 {
3972         bool all = nid_same(nid, &LNET_ANY_NID);
3973         struct lnet_net *net;
3974         struct lnet_ni *ni;
3975
3976         lnet_net_lock(LNET_LOCK_EX);
3977         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3978                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3979                         if (!all && !nid_same(&ni->ni_nid, nid))
3980                                 continue;
3981
3982                         atomic_set(&ni->ni_healthv, value);
3983                         if (list_empty(&ni->ni_recovery) &&
3984                             value < LNET_MAX_HEALTH_VALUE) {
3985                                 CERROR("manually adding local NI %s to recovery\n",
3986                                        libcfs_nidstr(&ni->ni_nid));
3987                                 list_add_tail(&ni->ni_recovery,
3988                                               &the_lnet.ln_mt_localNIRecovq);
3989                                 lnet_ni_addref_locked(ni, 0);
3990                         }
3991                         if (!all) {
3992                                 lnet_net_unlock(LNET_LOCK_EX);
3993                                 return;
3994                         }
3995                 }
3996         }
3997         lnet_net_unlock(LNET_LOCK_EX);
3998 }
3999
4000 static void
4001 lnet_ni_set_conns_per_peer(lnet_nid_t nid, int value, bool all)
4002 {
4003         struct lnet_net *net;
4004         struct lnet_ni *ni;
4005
4006         lnet_net_lock(LNET_LOCK_EX);
4007         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
4008                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
4009                         if (lnet_nid_to_nid4(&ni->ni_nid) != nid && !all)
4010                                 continue;
4011                         if (LNET_NETTYP(net->net_id) == SOCKLND)
4012                                 ni->ni_lnd_tunables.lnd_tun_u.lnd_sock.lnd_conns_per_peer = value;
4013                         else if (LNET_NETTYP(net->net_id) == O2IBLND)
4014                                 ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib.lnd_conns_per_peer = value;
4015                         if (!all) {
4016                                 lnet_net_unlock(LNET_LOCK_EX);
4017                                 return;
4018                         }
4019                 }
4020         }
4021         lnet_net_unlock(LNET_LOCK_EX);
4022 }
4023
4024 static int
4025 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
4026 {
4027         int cpt, rc = 0;
4028         struct lnet_ni *ni;
4029         struct lnet_nid nid;
4030
4031         lnet_nid4_to_nid(stats->hlni_nid, &nid);
4032         cpt = lnet_net_lock_current();
4033         ni = lnet_nid_to_ni_locked(&nid, cpt);
4034         if (!ni) {
4035                 rc = -ENOENT;
4036                 goto unlock;
4037         }
4038
4039         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
4040         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
4041         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
4042         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
4043         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
4044         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
4045         stats->hlni_fatal_error = atomic_read(&ni->ni_fatal_error_on);
4046         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
4047         stats->hlni_ping_count = ni->ni_ping_count;
4048         stats->hlni_next_ping = ni->ni_next_ping;
4049
4050 unlock:
4051         lnet_net_unlock(cpt);
4052
4053         return rc;
4054 }
4055
4056 static int
4057 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
4058 {
4059         struct lnet_ni *ni;
4060         int i = 0;
4061
4062         lnet_net_lock(LNET_LOCK_EX);
4063         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
4064                 if (!nid_is_nid4(&ni->ni_nid))
4065                         continue;
4066                 list->rlst_nid_array[i] = lnet_nid_to_nid4(&ni->ni_nid);
4067                 i++;
4068                 if (i >= LNET_MAX_SHOW_NUM_NID)
4069                         break;
4070         }
4071         lnet_net_unlock(LNET_LOCK_EX);
4072         list->rlst_num_nids = i;
4073
4074         return 0;
4075 }
4076
4077 static int
4078 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
4079 {
4080         struct lnet_peer_ni *lpni;
4081         int i = 0;
4082
4083         lnet_net_lock(LNET_LOCK_EX);
4084         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
4085                 list->rlst_nid_array[i] = lnet_nid_to_nid4(&lpni->lpni_nid);
4086                 i++;
4087                 if (i >= LNET_MAX_SHOW_NUM_NID)
4088                         break;
4089         }
4090         lnet_net_unlock(LNET_LOCK_EX);
4091         list->rlst_num_nids = i;
4092
4093         return 0;
4094 }
4095
4096 /**
4097  * LNet ioctl handler.
4098  *
4099  */
4100 int
4101 LNetCtl(unsigned int cmd, void *arg)
4102 {
4103         struct libcfs_ioctl_data *data = arg;
4104         struct lnet_ioctl_config_data *config;
4105         struct lnet_ni           *ni;
4106         struct lnet_nid           nid;
4107         int                       rc;
4108
4109         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
4110                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
4111
4112         switch (cmd) {
4113         case IOC_LIBCFS_GET_NI: {
4114                 struct lnet_processid id = {};
4115
4116                 rc = LNetGetId(data->ioc_count, &id, false);
4117                 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
4118                 return rc;
4119         }
4120         case IOC_LIBCFS_FAIL_NID:
4121                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
4122
4123         case IOC_LIBCFS_ADD_ROUTE: {
4124                 /* default router sensitivity to 1 */
4125                 unsigned int sensitivity = 1;
4126                 config = arg;
4127
4128                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4129                         return -EINVAL;
4130
4131                 if (config->cfg_config_u.cfg_route.rtr_sensitivity) {
4132                         sensitivity =
4133                           config->cfg_config_u.cfg_route.rtr_sensitivity;
4134                 }
4135
4136                 lnet_nid4_to_nid(config->cfg_nid, &nid);
4137                 mutex_lock(&the_lnet.ln_api_mutex);
4138                 rc = lnet_add_route(config->cfg_net,
4139                                     config->cfg_config_u.cfg_route.rtr_hop,
4140                                     &nid,
4141                                     config->cfg_config_u.cfg_route.
4142                                         rtr_priority, sensitivity);
4143                 mutex_unlock(&the_lnet.ln_api_mutex);
4144                 return rc;
4145         }
4146
4147         case IOC_LIBCFS_DEL_ROUTE:
4148                 config = arg;
4149
4150                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4151                         return -EINVAL;
4152
4153                 lnet_nid4_to_nid(config->cfg_nid, &nid);
4154                 mutex_lock(&the_lnet.ln_api_mutex);
4155                 rc = lnet_del_route(config->cfg_net, &nid);
4156                 mutex_unlock(&the_lnet.ln_api_mutex);
4157                 return rc;
4158
4159         case IOC_LIBCFS_GET_ROUTE:
4160                 config = arg;
4161
4162                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4163                         return -EINVAL;
4164
4165                 mutex_lock(&the_lnet.ln_api_mutex);
4166                 rc = lnet_get_route(config->cfg_count,
4167                                     &config->cfg_net,
4168                                     &config->cfg_config_u.cfg_route.rtr_hop,
4169                                     &config->cfg_nid,
4170                                     &config->cfg_config_u.cfg_route.rtr_flags,
4171                                     &config->cfg_config_u.cfg_route.
4172                                         rtr_priority,
4173                                     &config->cfg_config_u.cfg_route.
4174                                         rtr_sensitivity);
4175                 mutex_unlock(&the_lnet.ln_api_mutex);
4176                 return rc;
4177
4178         case IOC_LIBCFS_GET_LOCAL_NI: {
4179                 struct lnet_ioctl_config_ni *cfg_ni;
4180                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
4181                 struct lnet_ioctl_element_stats *stats;
4182                 __u32 tun_size;
4183
4184                 cfg_ni = arg;
4185
4186                 /* get the tunables if they are available */
4187                 if (cfg_ni->lic_cfg_hdr.ioc_len <
4188                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
4189                         return -EINVAL;
4190
4191                 stats = (struct lnet_ioctl_element_stats *)
4192                         cfg_ni->lic_bulk;
4193                 tun = (struct lnet_ioctl_config_lnd_tunables *)
4194                                 (cfg_ni->lic_bulk + sizeof(*stats));
4195
4196                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
4197                         sizeof(*stats);
4198
4199                 mutex_lock(&the_lnet.ln_api_mutex);
4200                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
4201                 mutex_unlock(&the_lnet.ln_api_mutex);
4202                 return rc;
4203         }
4204
4205         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
4206                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
4207                 int cpt;
4208
4209                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
4210                         return -EINVAL;
4211
4212                 mutex_lock(&the_lnet.ln_api_mutex);
4213
4214                 cpt = lnet_net_lock_current();
4215                 rc = lnet_get_ni_stats(msg_stats);
4216                 lnet_net_unlock(cpt);
4217
4218                 mutex_unlock(&the_lnet.ln_api_mutex);
4219
4220                 return rc;
4221         }
4222
4223         case IOC_LIBCFS_GET_NET: {
4224                 size_t total = sizeof(*config) +
4225                                sizeof(struct lnet_ioctl_net_config);
4226                 config = arg;
4227
4228                 if (config->cfg_hdr.ioc_len < total)
4229                         return -EINVAL;
4230
4231                 mutex_lock(&the_lnet.ln_api_mutex);
4232                 rc = lnet_get_net_config(config);
4233                 mutex_unlock(&the_lnet.ln_api_mutex);
4234                 return rc;
4235         }
4236
4237         case IOC_LIBCFS_GET_LNET_STATS:
4238         {
4239                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
4240
4241                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
4242                         return -EINVAL;
4243
4244                 mutex_lock(&the_lnet.ln_api_mutex);
4245                 rc = lnet_counters_get(&lnet_stats->st_cntrs);
4246                 mutex_unlock(&the_lnet.ln_api_mutex);
4247                 return rc;
4248         }
4249
4250         case IOC_LIBCFS_RESET_LNET_STATS:
4251         {
4252                 mutex_lock(&the_lnet.ln_api_mutex);
4253                 lnet_counters_reset();
4254                 mutex_unlock(&the_lnet.ln_api_mutex);
4255                 return 0;
4256         }
4257
4258         case IOC_LIBCFS_CONFIG_RTR:
4259                 config = arg;
4260
4261                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4262                         return -EINVAL;
4263
4264                 mutex_lock(&the_lnet.ln_api_mutex);
4265                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
4266                         rc = lnet_rtrpools_enable();
4267                         mutex_unlock(&the_lnet.ln_api_mutex);
4268                         return rc;
4269                 }
4270                 lnet_rtrpools_disable();
4271                 mutex_unlock(&the_lnet.ln_api_mutex);
4272                 return 0;
4273
4274         case IOC_LIBCFS_ADD_BUF:
4275                 config = arg;
4276
4277                 if (config->cfg_hdr.ioc_len < sizeof(*config))
4278                         return -EINVAL;
4279
4280                 mutex_lock(&the_lnet.ln_api_mutex);
4281                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
4282                                                 buf_tiny,
4283                                           config->cfg_config_u.cfg_buffers.
4284                                                 buf_small,
4285                                           config->cfg_config_u.cfg_buffers.
4286                                                 buf_large);
4287                 mutex_unlock(&the_lnet.ln_api_mutex);
4288                 return rc;
4289
4290         case IOC_LIBCFS_SET_NUMA_RANGE: {
4291                 struct lnet_ioctl_set_value *numa;
4292                 numa = arg;
4293                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
4294                         return -EINVAL;
4295                 lnet_net_lock(LNET_LOCK_EX);
4296                 lnet_numa_range = numa->sv_value;
4297                 lnet_net_unlock(LNET_LOCK_EX);
4298                 return 0;
4299         }
4300
4301         case IOC_LIBCFS_GET_NUMA_RANGE: {
4302                 struct lnet_ioctl_set_value *numa;
4303                 numa = arg;
4304                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
4305                         return -EINVAL;
4306                 numa->sv_value = lnet_numa_range;
4307                 return 0;
4308         }
4309
4310         case IOC_LIBCFS_GET_BUF: {
4311                 struct lnet_ioctl_pool_cfg *pool_cfg;
4312                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
4313
4314                 config = arg;
4315
4316                 if (config->cfg_hdr.ioc_len < total)
4317                         return -EINVAL;
4318
4319                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
4320
4321                 mutex_lock(&the_lnet.ln_api_mutex);
4322                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
4323                 mutex_unlock(&the_lnet.ln_api_mutex);
4324                 return rc;
4325         }
4326
4327         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
4328                 struct lnet_ioctl_local_ni_hstats *stats = arg;
4329
4330                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
4331                         return -EINVAL;
4332
4333                 mutex_lock(&the_lnet.ln_api_mutex);
4334                 rc = lnet_get_local_ni_hstats(stats);
4335                 mutex_unlock(&the_lnet.ln_api_mutex);
4336
4337                 return rc;
4338         }
4339
4340         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
4341                 struct lnet_ioctl_recovery_list *list = arg;
4342                 if (list->rlst_hdr.ioc_len < sizeof(*list))
4343                         return -EINVAL;
4344
4345                 mutex_lock(&the_lnet.ln_api_mutex);
4346                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
4347                         rc = lnet_get_local_ni_recovery_list(list);
4348                 else
4349                         rc = lnet_get_peer_ni_recovery_list(list);
4350                 mutex_unlock(&the_lnet.ln_api_mutex);
4351                 return rc;
4352         }
4353
4354         case IOC_LIBCFS_ADD_PEER_NI: {
4355                 struct lnet_ioctl_peer_cfg *cfg = arg;
4356                 struct lnet_nid prim_nid;
4357
4358                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4359                         return -EINVAL;
4360
4361                 mutex_lock(&the_lnet.ln_api_mutex);
4362                 lnet_nid4_to_nid(cfg->prcfg_prim_nid, &prim_nid);
4363                 lnet_nid4_to_nid(cfg->prcfg_cfg_nid, &nid);
4364                 rc = lnet_user_add_peer_ni(&prim_nid, &nid, cfg->prcfg_mr,
4365                                            cfg->prcfg_count == 1);
4366                 mutex_unlock(&the_lnet.ln_api_mutex);
4367                 return rc;
4368         }
4369
4370         case IOC_LIBCFS_DEL_PEER_NI: {
4371                 struct lnet_ioctl_peer_cfg *cfg = arg;
4372                 struct lnet_nid prim_nid;
4373
4374                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4375                         return -EINVAL;
4376
4377                 mutex_lock(&the_lnet.ln_api_mutex);
4378                 lnet_nid4_to_nid(cfg->prcfg_prim_nid, &prim_nid);
4379                 lnet_nid4_to_nid(cfg->prcfg_cfg_nid, &nid);
4380                 rc = lnet_del_peer_ni(&prim_nid,
4381                                       &nid,
4382                                       cfg->prcfg_count);
4383                 mutex_unlock(&the_lnet.ln_api_mutex);
4384                 return rc;
4385         }
4386
4387         case IOC_LIBCFS_GET_PEER_INFO: {
4388                 struct lnet_ioctl_peer *peer_info = arg;
4389
4390                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
4391                         return -EINVAL;
4392
4393                 mutex_lock(&the_lnet.ln_api_mutex);
4394                 rc = lnet_get_peer_ni_info(
4395                    peer_info->pr_count,
4396                    &peer_info->pr_nid,
4397                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
4398                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
4399                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
4400                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
4401                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
4402                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
4403                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
4404                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
4405                 mutex_unlock(&the_lnet.ln_api_mutex);
4406                 return rc;
4407         }
4408
4409         case IOC_LIBCFS_GET_PEER_NI: {
4410                 struct lnet_ioctl_peer_cfg *cfg = arg;
4411
4412                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4413                         return -EINVAL;
4414
4415                 mutex_lock(&the_lnet.ln_api_mutex);
4416                 rc = lnet_get_peer_info(cfg,
4417                                         (void __user *)cfg->prcfg_bulk);
4418                 mutex_unlock(&the_lnet.ln_api_mutex);
4419                 return rc;
4420         }
4421
4422         case IOC_LIBCFS_GET_PEER_LIST: {
4423                 struct lnet_ioctl_peer_cfg *cfg = arg;
4424
4425                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4426                         return -EINVAL;
4427
4428                 mutex_lock(&the_lnet.ln_api_mutex);
4429                 rc = lnet_get_peer_list(&cfg->prcfg_count, &cfg->prcfg_size,
4430                                 (struct lnet_process_id __user *)cfg->prcfg_bulk);
4431                 mutex_unlock(&the_lnet.ln_api_mutex);
4432                 return rc;
4433         }
4434
4435         case IOC_LIBCFS_SET_HEALHV: {
4436                 struct lnet_ioctl_reset_health_cfg *cfg = arg;
4437                 int value;
4438
4439                 if (cfg->rh_hdr.ioc_len < sizeof(*cfg))
4440                         return -EINVAL;
4441                 if (cfg->rh_value < 0 ||
4442                     cfg->rh_value > LNET_MAX_HEALTH_VALUE)
4443                         value = LNET_MAX_HEALTH_VALUE;
4444                 else
4445                         value = cfg->rh_value;
4446                 CDEBUG(D_NET, "Manually setting healthv to %d for %s:%s. all = %d\n",
4447                        value, (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) ?
4448                        "local" : "peer", libcfs_nid2str(cfg->rh_nid), cfg->rh_all);
4449                 lnet_nid4_to_nid(cfg->rh_nid, &nid);
4450                 mutex_lock(&the_lnet.ln_api_mutex);
4451                 if (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) {
4452                         if (cfg->rh_all)
4453                                 nid = LNET_ANY_NID;
4454                         lnet_ni_set_healthv(&nid, value);
4455                 } else {
4456                         lnet_peer_ni_set_healthv(&nid, value, cfg->rh_all);
4457                 }
4458                 mutex_unlock(&the_lnet.ln_api_mutex);
4459                 return 0;
4460         }
4461
4462         case IOC_LIBCFS_SET_PEER: {
4463                 struct lnet_ioctl_peer_cfg *cfg = arg;
4464                 struct lnet_peer *lp;
4465
4466                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
4467                         return -EINVAL;
4468
4469                 mutex_lock(&the_lnet.ln_api_mutex);
4470                 lnet_nid4_to_nid(cfg->prcfg_prim_nid, &nid);
4471                 lp = lnet_find_peer(&nid);
4472                 if (!lp) {
4473                         mutex_unlock(&the_lnet.ln_api_mutex);
4474                         return -ENOENT;
4475                 }
4476                 spin_lock(&lp->lp_lock);
4477                 lp->lp_state = cfg->prcfg_state;
4478                 spin_unlock(&lp->lp_lock);
4479                 lnet_peer_decref_locked(lp);
4480                 mutex_unlock(&the_lnet.ln_api_mutex);
4481                 CDEBUG(D_NET, "Set peer %s state to %u\n",
4482                        libcfs_nid2str(cfg->prcfg_prim_nid), cfg->prcfg_state);
4483                 return 0;
4484         }
4485
4486         case IOC_LIBCFS_SET_CONNS_PER_PEER: {
4487                 struct lnet_ioctl_reset_conns_per_peer_cfg *cfg = arg;
4488                 int value;
4489
4490                 if (cfg->rcpp_hdr.ioc_len < sizeof(*cfg))
4491                         return -EINVAL;
4492                 if (cfg->rcpp_value < 0)
4493                         value = 1;
4494                 else
4495                         value = cfg->rcpp_value;
4496                 CDEBUG(D_NET,
4497                        "Setting conns_per_peer to %d for %s. all = %d\n",
4498                        value, libcfs_nid2str(cfg->rcpp_nid), cfg->rcpp_all);
4499                 mutex_lock(&the_lnet.ln_api_mutex);
4500                 lnet_ni_set_conns_per_peer(cfg->rcpp_nid, value, cfg->rcpp_all);
4501                 mutex_unlock(&the_lnet.ln_api_mutex);
4502                 return 0;
4503         }
4504
4505         case IOC_LIBCFS_NOTIFY_ROUTER: {
4506                 /* Convert the user-supplied real time to monotonic.
4507                  * NB: "when" is always in the past
4508                  */
4509                 time64_t when = ktime_get_seconds() -
4510                                 (ktime_get_real_seconds() - data->ioc_u64[0]);
4511
4512                 lnet_nid4_to_nid(data->ioc_nid, &nid);
4513                 return lnet_notify(NULL, &nid, data->ioc_flags, false, when);
4514         }
4515
4516         case IOC_LIBCFS_LNET_DIST:
4517                 lnet_nid4_to_nid(data->ioc_nid, &nid);
4518                 rc = LNetDist(&nid, &nid, &data->ioc_u32[1]);
4519                 if (rc < 0 && rc != -EHOSTUNREACH)
4520                         return rc;
4521
4522                 data->ioc_nid = lnet_nid_to_nid4(&nid);
4523                 data->ioc_u32[0] = rc;
4524                 return 0;
4525
4526         case IOC_LIBCFS_TESTPROTOCOMPAT:
4527                 the_lnet.ln_testprotocompat = data->ioc_flags;
4528                 return 0;
4529
4530         case IOC_LIBCFS_LNET_FAULT:
4531                 return lnet_fault_ctl(data->ioc_flags, data);
4532
4533         case IOC_LIBCFS_PING_PEER: {
4534                 struct lnet_ioctl_ping_data *ping = arg;
4535                 struct lnet_process_id __user *ids = ping->ping_buf;
4536                 struct lnet_nid src_nid = LNET_ANY_NID;
4537                 struct lnet_genl_ping_list plist;
4538                 struct lnet_processid id;
4539                 struct lnet_peer *lp;
4540                 signed long timeout;
4541                 int count, i;
4542
4543                 /* Check if the supplied ping data supports source nid
4544                  * NB: This check is sufficient if lnet_ioctl_ping_data has
4545                  * additional fields added, but if they are re-ordered or
4546                  * fields removed then this will break. It is expected that
4547                  * these ioctls will be replaced with netlink implementation, so
4548                  * it is probably not worth coming up with a more robust version
4549                  * compatibility scheme.
4550                  */
4551                 if (ping->ping_hdr.ioc_len >= sizeof(struct lnet_ioctl_ping_data))
4552                         lnet_nid4_to_nid(ping->ping_src, &src_nid);
4553
4554                 /* If timeout is negative then set default of 3 minutes */
4555                 if (((s32)ping->op_param) <= 0 ||
4556                     ping->op_param > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
4557                         timeout = cfs_time_seconds(DEFAULT_PEER_TIMEOUT);
4558                 else
4559                         timeout = nsecs_to_jiffies(ping->op_param * NSEC_PER_MSEC);
4560
4561                 id.pid = ping->ping_id.pid;
4562                 lnet_nid4_to_nid(ping->ping_id.nid, &id.nid);
4563                 rc = lnet_ping(&id, &src_nid, timeout, &plist,
4564                                ping->ping_count);
4565                 if (rc < 0)
4566                         goto report_ping_err;
4567                 count = rc;
4568                 rc = 0;
4569
4570                 for (i = 0; i < count; i++) {
4571                         struct lnet_processid *result;
4572                         struct lnet_process_id tmpid;
4573
4574                         result = genradix_ptr(&plist.lgpl_list, i);
4575                         memset(&tmpid, 0, sizeof(tmpid));
4576                         tmpid.pid = result->pid;
4577                         tmpid.nid = lnet_nid_to_nid4(&result->nid);
4578                         if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid))) {
4579                                 rc = -EFAULT;
4580                                 goto report_ping_err;
4581                         }
4582                 }
4583
4584                 mutex_lock(&the_lnet.ln_api_mutex);
4585                 lp = lnet_find_peer(&id.nid);
4586                 if (lp) {
4587                         ping->ping_id.nid =
4588                                 lnet_nid_to_nid4(&lp->lp_primary_nid);
4589                         ping->mr_info = lnet_peer_is_multi_rail(lp);
4590                         lnet_peer_decref_locked(lp);
4591                 }
4592                 mutex_unlock(&the_lnet.ln_api_mutex);
4593
4594                 ping->ping_count = count;
4595 report_ping_err:
4596                 genradix_free(&plist.lgpl_list);
4597                 return rc;
4598         }
4599
4600         case IOC_LIBCFS_DISCOVER: {
4601                 struct lnet_ioctl_ping_data *discover = arg;
4602                 struct lnet_process_id __user *ids;
4603                 struct lnet_genl_ping_list dlists;
4604                 struct lnet_processid id;
4605                 struct lnet_peer *lp;
4606                 int count, i;
4607
4608                 if (discover->ping_count <= 0)
4609                         return -EINVAL;
4610
4611                 genradix_init(&dlists.lgpl_list);
4612                 /* If the user buffer has more space than the lnet_interfaces_max,
4613                  * then only fill it up to lnet_interfaces_max.
4614                  */
4615                 if (discover->ping_count > lnet_interfaces_max)
4616                         discover->ping_count = lnet_interfaces_max;
4617
4618                 id.pid = discover->ping_id.pid;
4619                 lnet_nid4_to_nid(discover->ping_id.nid, &id.nid);
4620                 rc = lnet_discover(&id, discover->op_param, &dlists);
4621                 if (rc < 0)
4622                         goto report_discover_err;
4623                 count = rc;
4624
4625                 ids = discover->ping_buf;
4626                 for (i = 0; i < count; i++) {
4627                         struct lnet_processid *result;
4628                         struct lnet_process_id tmpid;
4629
4630                         result = genradix_ptr(&dlists.lgpl_list, i);
4631                         memset(&tmpid, 0, sizeof(tmpid));
4632                         tmpid.pid = result->pid;
4633                         tmpid.nid = lnet_nid_to_nid4(&result->nid);
4634                         if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid))) {
4635                                 rc = -EFAULT;
4636                                 goto report_discover_err;
4637                         }
4638
4639                         if (i >= discover->ping_count)
4640                                 break;
4641                 }
4642                 rc = 0;
4643
4644                 mutex_lock(&the_lnet.ln_api_mutex);
4645                 lp = lnet_find_peer(&id.nid);
4646                 if (lp) {
4647                         discover->ping_id.nid =
4648                                 lnet_nid_to_nid4(&lp->lp_primary_nid);
4649                         discover->mr_info = lnet_peer_is_multi_rail(lp);
4650                         lnet_peer_decref_locked(lp);
4651                 }
4652                 mutex_unlock(&the_lnet.ln_api_mutex);
4653
4654                 discover->ping_count = count;
4655 report_discover_err:
4656                 genradix_free(&dlists.lgpl_list);
4657                 return rc;
4658         }
4659
4660         case IOC_LIBCFS_ADD_UDSP: {
4661                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4662                 __u32 bulk_size = ioc_udsp->iou_hdr.ioc_len;
4663
4664                 mutex_lock(&the_lnet.ln_api_mutex);
4665                 rc = lnet_udsp_demarshal_add(arg, bulk_size);
4666                 if (!rc) {
4667                         rc = lnet_udsp_apply_policies(NULL, false);
4668                         CDEBUG(D_NET, "policy application returned %d\n", rc);
4669                         rc = 0;
4670                 }
4671                 mutex_unlock(&the_lnet.ln_api_mutex);
4672
4673                 return rc;
4674         }
4675
4676         case IOC_LIBCFS_DEL_UDSP: {
4677                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4678                 int idx = ioc_udsp->iou_idx;
4679
4680                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4681                         return -EINVAL;
4682
4683                 mutex_lock(&the_lnet.ln_api_mutex);
4684                 rc = lnet_udsp_del_policy(idx);
4685                 mutex_unlock(&the_lnet.ln_api_mutex);
4686
4687                 return rc;
4688         }
4689
4690         case IOC_LIBCFS_GET_UDSP_SIZE: {
4691                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4692                 struct lnet_udsp *udsp;
4693
4694                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4695                         return -EINVAL;
4696
4697                 rc = 0;
4698
4699                 mutex_lock(&the_lnet.ln_api_mutex);
4700                 udsp = lnet_udsp_get_policy(ioc_udsp->iou_idx);
4701                 if (!udsp) {
4702                         rc = -ENOENT;
4703                 } else {
4704                         /* coming in iou_idx will hold the idx of the udsp
4705                          * to get the size of. going out the iou_idx will
4706                          * hold the size of the UDSP found at the passed
4707                          * in index.
4708                          */
4709                         ioc_udsp->iou_idx = lnet_get_udsp_size(udsp);
4710                         if (ioc_udsp->iou_idx < 0)
4711                                 rc = -EINVAL;
4712                 }
4713                 mutex_unlock(&the_lnet.ln_api_mutex);
4714
4715                 return rc;
4716         }
4717
4718         case IOC_LIBCFS_GET_UDSP: {
4719                 struct lnet_ioctl_udsp *ioc_udsp = arg;
4720                 struct lnet_udsp *udsp;
4721
4722                 if (ioc_udsp->iou_hdr.ioc_len < sizeof(*ioc_udsp))
4723                         return -EINVAL;
4724
4725                 rc = 0;
4726
4727                 mutex_lock(&the_lnet.ln_api_mutex);
4728                 udsp = lnet_udsp_get_policy(ioc_udsp->iou_idx);
4729                 if (!udsp)
4730                         rc = -ENOENT;
4731                 else
4732                         rc = lnet_udsp_marshal(udsp, ioc_udsp);
4733                 mutex_unlock(&the_lnet.ln_api_mutex);
4734
4735                 return rc;
4736         }
4737
4738         case IOC_LIBCFS_GET_CONST_UDSP_INFO: {
4739                 struct lnet_ioctl_construct_udsp_info *info = arg;
4740
4741                 if (info->cud_hdr.ioc_len < sizeof(*info))
4742                         return -EINVAL;
4743
4744                 CDEBUG(D_NET, "GET_UDSP_INFO for %s\n",
4745                        libcfs_nid2str(info->cud_nid));
4746
4747                 lnet_nid4_to_nid(info->cud_nid, &nid);
4748                 mutex_lock(&the_lnet.ln_api_mutex);
4749                 lnet_net_lock(0);
4750                 lnet_udsp_get_construct_info(info, &nid);
4751                 lnet_net_unlock(0);
4752                 mutex_unlock(&the_lnet.ln_api_mutex);
4753
4754                 return 0;
4755         }
4756
4757         default:
4758                 ni = lnet_net2ni_addref(data->ioc_net);
4759                 if (ni == NULL)
4760                         return -EINVAL;
4761
4762                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
4763                         rc = -EINVAL;
4764                 else
4765                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
4766
4767                 lnet_ni_decref(ni);
4768                 return rc <= 0 ? rc : 0;
4769         }
4770         /* not reached */
4771 }
4772 EXPORT_SYMBOL(LNetCtl);
4773
4774 static int lnet_net_conf_cmd(struct sk_buff *skb, struct genl_info *info)
4775 {
4776         int rc = 0;
4777
4778         if (info->nlhdr->nlmsg_flags & NLM_F_CREATE) {
4779                 /* NLM_F_EXCL means ignore module parameters */
4780                 if (info->nlhdr->nlmsg_flags & NLM_F_EXCL)
4781                         the_lnet.ln_nis_from_mod_params = true;
4782
4783                 if (info->nlhdr->nlmsg_flags & NLM_F_REPLACE)
4784                         the_lnet.ln_nis_use_large_nids = true;
4785
4786                 rc = lnet_configure(NULL);
4787                 switch (rc) {
4788                 case -ENETDOWN:
4789                         GENL_SET_ERR_MSG(info,
4790                                          "Network is down");
4791                         break;
4792                 case -ENODEV:
4793                         GENL_SET_ERR_MSG(info,
4794                                          "LNET is currently not loaded");
4795                         break;
4796                 case -EBUSY:
4797                         GENL_SET_ERR_MSG(info, "LNET busy");
4798                         break;
4799                 default:
4800                         break;
4801                 }
4802         } else {
4803                 rc = lnet_unconfigure();
4804         }
4805
4806         return rc;
4807 };
4808
4809 struct lnet_nid_cpt {
4810         struct lnet_nid lnc_nid;
4811         unsigned int lnc_cpt;
4812 };
4813
4814 struct lnet_genl_nid_cpt_list {
4815         unsigned int lgncl_index;
4816         unsigned int lgncl_list_count;
4817         GENRADIX(struct lnet_nid_cpt) lgncl_lnc_list;
4818 };
4819
4820 static inline struct lnet_genl_nid_cpt_list *
4821 lnet_cpt_of_nid_dump_ctx(struct netlink_callback *cb)
4822 {
4823         return (struct lnet_genl_nid_cpt_list *)cb->args[0];
4824 }
4825
4826 static int lnet_cpt_of_nid_show_done(struct netlink_callback *cb)
4827 {
4828         struct lnet_genl_nid_cpt_list *lgncl;
4829
4830         lgncl = lnet_cpt_of_nid_dump_ctx(cb);
4831
4832         if (lgncl) {
4833                 genradix_free(&lgncl->lgncl_lnc_list);
4834                 LIBCFS_FREE(lgncl, sizeof(*lgncl));
4835                 cb->args[0] = 0;
4836         }
4837
4838         return 0;
4839 }
4840
4841 static int lnet_cpt_of_nid_show_start(struct netlink_callback *cb)
4842 {
4843         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
4844 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
4845         struct netlink_ext_ack *extack = NULL;
4846 #endif
4847         struct lnet_genl_nid_cpt_list *lgncl;
4848         int msg_len = genlmsg_len(gnlh);
4849         struct nlattr *params, *top;
4850         int rem, rc = 0;
4851
4852 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
4853         extack = cb->extack;
4854 #endif
4855
4856         mutex_lock(&the_lnet.ln_api_mutex);
4857         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
4858                 NL_SET_ERR_MSG(extack, "Network is down");
4859                 mutex_unlock(&the_lnet.ln_api_mutex);
4860                 return -ENETDOWN;
4861         }
4862
4863         msg_len = genlmsg_len(gnlh);
4864         if (!msg_len) {
4865                 NL_SET_ERR_MSG(extack, "Missing NID argument(s)");
4866                 mutex_unlock(&the_lnet.ln_api_mutex);
4867                 return -ENOENT;
4868         }
4869
4870         LIBCFS_ALLOC(lgncl, sizeof(*lgncl));
4871         if (!lgncl) {
4872                 mutex_unlock(&the_lnet.ln_api_mutex);
4873                 return -ENOMEM;
4874         }
4875
4876         genradix_init(&lgncl->lgncl_lnc_list);
4877         lgncl->lgncl_list_count = 0;
4878         cb->args[0] = (long)lgncl;
4879
4880         params = genlmsg_data(gnlh);
4881         nla_for_each_attr(top, params, msg_len, rem) {
4882                 struct nlattr *nids;
4883                 int rem2;
4884
4885                 switch (nla_type(top)) {
4886                 case LN_SCALAR_ATTR_LIST:
4887                         nla_for_each_nested(nids, top, rem2) {
4888                                 char nidstr[LNET_NIDSTR_SIZE + 1];
4889                                 struct lnet_nid_cpt *lnc;
4890
4891                                 if (nla_type(nids) != LN_SCALAR_ATTR_VALUE)
4892                                         continue;
4893
4894                                 memset(nidstr, 0, sizeof(nidstr));
4895                                 rc = nla_strscpy(nidstr, nids, sizeof(nidstr));
4896                                 if (rc < 0) {
4897                                         NL_SET_ERR_MSG(extack,
4898                                                        "failed to get NID");
4899                                         GOTO(report_err, rc);
4900                                 }
4901
4902                                 lnc = genradix_ptr_alloc(&lgncl->lgncl_lnc_list,
4903                                                       lgncl->lgncl_list_count++,
4904                                                       GFP_KERNEL);
4905                                 if (!lnc) {
4906                                         NL_SET_ERR_MSG(extack,
4907                                                       "failed to allocate NID");
4908                                         GOTO(report_err, rc = -ENOMEM);
4909                                 }
4910
4911                                 rc = libcfs_strnid(&lnc->lnc_nid,
4912                                                    strim(nidstr));
4913                                 if (rc < 0) {
4914                                         NL_SET_ERR_MSG(extack, "invalid NID");
4915                                         GOTO(report_err, rc);
4916                                 }
4917                                 rc = 0;
4918                                 CDEBUG(D_NET, "nid: %s\n",
4919                                        libcfs_nidstr(&lnc->lnc_nid));
4920                         }
4921                         fallthrough;
4922                 default:
4923                         break;
4924                 }
4925         }
4926 report_err:
4927         mutex_unlock(&the_lnet.ln_api_mutex);
4928
4929         if (rc < 0)
4930                 lnet_cpt_of_nid_show_done(cb);
4931
4932         return rc;
4933 }
4934
4935 static const struct ln_key_list cpt_of_nid_props_list = {
4936         .lkl_maxattr                    = LNET_CPT_OF_NID_ATTR_MAX,
4937         .lkl_list                       = {
4938                 [LNET_CPT_OF_NID_ATTR_HDR]      = {
4939                         .lkp_value              = "cpt-of-nid",
4940                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
4941                         .lkp_data_type          = NLA_NUL_STRING,
4942                 },
4943                 [LNET_CPT_OF_NID_ATTR_NID]      = {
4944                         .lkp_value              = "nid",
4945                         .lkp_data_type          = NLA_STRING,
4946                 },
4947                 [LNET_CPT_OF_NID_ATTR_CPT]      = {
4948                         .lkp_value              = "cpt",
4949                         .lkp_data_type          = NLA_U32,
4950                 },
4951         },
4952 };
4953
4954 static int lnet_cpt_of_nid_show_dump(struct sk_buff *msg,
4955                                      struct netlink_callback *cb)
4956 {
4957         struct lnet_genl_nid_cpt_list *lgncl;
4958 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
4959         struct netlink_ext_ack *extack = NULL;
4960 #endif
4961         int portid = NETLINK_CB(cb->skb).portid;
4962         int seq = cb->nlh->nlmsg_seq;
4963         int idx;
4964         int rc = 0;
4965         bool need_hdr = true;
4966
4967 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
4968         extack = cb->extack;
4969 #endif
4970
4971         mutex_lock(&the_lnet.ln_api_mutex);
4972         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
4973                 NL_SET_ERR_MSG(extack, "Network is down");
4974                 GOTO(send_error, rc = -ENETDOWN);
4975         }
4976
4977         lgncl = lnet_cpt_of_nid_dump_ctx(cb);
4978         idx = lgncl->lgncl_index;
4979
4980         if (!lgncl->lgncl_index) {
4981                 const struct ln_key_list *all[] = {
4982                         &cpt_of_nid_props_list, NULL, NULL
4983                 };
4984
4985                 rc = lnet_genl_send_scalar_list(msg, portid, seq, &lnet_family,
4986                                                 NLM_F_CREATE | NLM_F_MULTI,
4987                                                 LNET_CMD_CPT_OF_NID, all);
4988                 if (rc < 0) {
4989                         NL_SET_ERR_MSG(extack, "failed to send key table");
4990                         GOTO(send_error, rc);
4991                 }
4992         }
4993
4994         while (idx < lgncl->lgncl_list_count) {
4995                 struct lnet_nid_cpt *lnc;
4996                 void *hdr;
4997                 int cpt;
4998
4999                 lnc = genradix_ptr(&lgncl->lgncl_lnc_list, idx++);
5000
5001                 cpt = lnet_nid_cpt_hash(&lnc->lnc_nid, LNET_CPT_NUMBER);
5002
5003                 CDEBUG(D_NET, "nid: %s cpt: %d\n", libcfs_nidstr(&lnc->lnc_nid), cpt);
5004                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
5005                                   NLM_F_MULTI, LNET_CMD_CPT_OF_NID);
5006                 if (!hdr) {
5007                         NL_SET_ERR_MSG(extack, "failed to send values");
5008                         genlmsg_cancel(msg, hdr);
5009                         GOTO(send_error, rc = -EMSGSIZE);
5010                 }
5011
5012                 if (need_hdr) {
5013                         nla_put_string(msg, LNET_CPT_OF_NID_ATTR_HDR, "");
5014                         need_hdr = false;
5015                 }
5016
5017                 nla_put_string(msg, LNET_CPT_OF_NID_ATTR_NID,
5018                                libcfs_nidstr(&lnc->lnc_nid));
5019                 nla_put_u32(msg, LNET_CPT_OF_NID_ATTR_CPT, cpt);
5020
5021                 genlmsg_end(msg, hdr);
5022         }
5023
5024         genradix_free(&lgncl->lgncl_lnc_list);
5025         rc = 0;
5026         lgncl->lgncl_index = idx;
5027
5028 send_error:
5029         mutex_unlock(&the_lnet.ln_api_mutex);
5030
5031         return lnet_nl_send_error(cb->skb, portid, seq, rc);
5032 }
5033
5034 #ifndef HAVE_NETLINK_CALLBACK_START
5035 static int lnet_old_cpt_of_nid_show_dump(struct sk_buff *msg,
5036                                          struct netlink_callback *cb)
5037 {
5038         if (!cb->args[0]) {
5039                 int rc = lnet_cpt_of_nid_show_start(cb);
5040
5041                 if (rc < 0)
5042                         return lnet_nl_send_error(cb->skb,
5043                                                   NETLINK_CB(cb->skb).portid,
5044                                                   cb->nlh->nlmsg_seq,
5045                                                   rc);
5046         }
5047
5048         return lnet_cpt_of_nid_show_dump(msg, cb);
5049 }
5050 #endif
5051
5052 /* This is the keys for the UDSP info which is used by many
5053  * Netlink commands.
5054  */
5055 static const struct ln_key_list udsp_info_list = {
5056         .lkl_maxattr                    = LNET_UDSP_INFO_ATTR_MAX,
5057         .lkl_list                       = {
5058                 [LNET_UDSP_INFO_ATTR_NET_PRIORITY]              = {
5059                         .lkp_value      = "net priority",
5060                         .lkp_data_type  = NLA_S32
5061                 },
5062                 [LNET_UDSP_INFO_ATTR_NID_PRIORITY]              = {
5063                         .lkp_value      = "nid priority",
5064                         .lkp_data_type  = NLA_S32
5065                 },
5066                 [LNET_UDSP_INFO_ATTR_PREF_RTR_NIDS_LIST]        = {
5067                         .lkp_value      = "Preferred gateway NIDs",
5068                         .lkp_key_format = LNKF_MAPPING,
5069                         .lkp_data_type  = NLA_NESTED,
5070                 },
5071                 [LNET_UDSP_INFO_ATTR_PREF_NIDS_LIST]            = {
5072                         .lkp_value      = "Preferred source NIDs",
5073                         .lkp_key_format = LNKF_MAPPING,
5074                         .lkp_data_type  = NLA_NESTED,
5075                 },
5076         },
5077 };
5078
5079 static const struct ln_key_list udsp_info_pref_nids_list = {
5080         .lkl_maxattr                    = LNET_UDSP_INFO_PREF_NIDS_ATTR_MAX,
5081         .lkl_list                       = {
5082                 [LNET_UDSP_INFO_PREF_NIDS_ATTR_INDEX]           = {
5083                         .lkp_value      = "NID-0",
5084                         .lkp_data_type  = NLA_NUL_STRING,
5085                 },
5086                 [LNET_UDSP_INFO_PREF_NIDS_ATTR_NID]             = {
5087                         .lkp_value      = "0@lo",
5088                         .lkp_data_type  = NLA_STRING,
5089                 },
5090         },
5091 };
5092
5093 static int lnet_udsp_info_send(struct sk_buff *msg, int attr,
5094                                struct lnet_nid *nid, bool remote)
5095 {
5096         struct lnet_ioctl_construct_udsp_info *udsp;
5097         struct nlattr *udsp_attr, *udsp_info;
5098         struct nlattr *udsp_list_attr;
5099         struct nlattr *udsp_list_info;
5100         int i;
5101
5102         CFS_ALLOC_PTR(udsp);
5103         if (!udsp)
5104                 return -ENOMEM;
5105
5106         udsp->cud_peer = remote;
5107         lnet_udsp_get_construct_info(udsp, nid);
5108
5109         udsp_info = nla_nest_start(msg, attr);
5110         udsp_attr = nla_nest_start(msg, 0);
5111         nla_put_s32(msg, LNET_UDSP_INFO_ATTR_NET_PRIORITY,
5112                     udsp->cud_net_priority);
5113         nla_put_s32(msg, LNET_UDSP_INFO_ATTR_NID_PRIORITY,
5114                     udsp->cud_nid_priority);
5115
5116         if (udsp->cud_pref_rtr_nid[0] == 0)
5117                 goto skip_list;
5118
5119         udsp_list_info = nla_nest_start(msg,
5120                                         LNET_UDSP_INFO_ATTR_PREF_RTR_NIDS_LIST);
5121         for (i = 0; i < LNET_MAX_SHOW_NUM_NID; i++) {
5122                 char tmp[8]; /* NID-"3 number"\0 */
5123
5124                 if (udsp->cud_pref_rtr_nid[i] == 0)
5125                         break;
5126
5127                 udsp_list_attr = nla_nest_start(msg, i);
5128                 snprintf(tmp, sizeof(tmp), "NID-%d", i);
5129                 nla_put_string(msg, LNET_UDSP_INFO_PREF_NIDS_ATTR_INDEX,
5130                                tmp);
5131                 nla_put_string(msg, LNET_UDSP_INFO_PREF_NIDS_ATTR_NID,
5132                                libcfs_nid2str(udsp->cud_pref_rtr_nid[i]));
5133                 nla_nest_end(msg, udsp_list_attr);
5134         }
5135         nla_nest_end(msg, udsp_list_info);
5136 skip_list:
5137         nla_nest_end(msg, udsp_attr);
5138         nla_nest_end(msg, udsp_info);
5139         LIBCFS_FREE(udsp, sizeof(*udsp));
5140
5141         return 0;
5142 }
5143
5144 /* LNet NI handling */
5145 static const struct ln_key_list net_props_list = {
5146         .lkl_maxattr                    = LNET_NET_ATTR_MAX,
5147         .lkl_list                       = {
5148                 [LNET_NET_ATTR_HDR]             = {
5149                         .lkp_value              = "net",
5150                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
5151                         .lkp_data_type          = NLA_NUL_STRING,
5152                 },
5153                 [LNET_NET_ATTR_TYPE]            = {
5154                         .lkp_value              = "net type",
5155                         .lkp_data_type          = NLA_STRING
5156                 },
5157                 [LNET_NET_ATTR_LOCAL]           = {
5158                         .lkp_value              = "local NI(s)",
5159                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
5160                         .lkp_data_type          = NLA_NESTED
5161                 },
5162         },
5163 };
5164
5165 static struct ln_key_list local_ni_list = {
5166         .lkl_maxattr                    = LNET_NET_LOCAL_NI_ATTR_MAX,
5167         .lkl_list                       = {
5168                 [LNET_NET_LOCAL_NI_ATTR_NID]            = {
5169                         .lkp_value              = "nid",
5170                         .lkp_data_type          = NLA_STRING
5171                 },
5172                 [LNET_NET_LOCAL_NI_ATTR_STATUS]         = {
5173                         .lkp_value              = "status",
5174                         .lkp_data_type          = NLA_STRING
5175                 },
5176                 [LNET_NET_LOCAL_NI_ATTR_INTERFACE]      = {
5177                         .lkp_value              = "interfaces",
5178                         .lkp_key_format         = LNKF_MAPPING,
5179                         .lkp_data_type          = NLA_NESTED
5180                 },
5181                 [LNET_NET_LOCAL_NI_ATTR_STATS]          = {
5182                         .lkp_value              = "statistics",
5183                         .lkp_key_format         = LNKF_MAPPING,
5184                         .lkp_data_type          = NLA_NESTED
5185                 },
5186                 [LNET_NET_LOCAL_NI_ATTR_UDSP_INFO]      = {
5187                         .lkp_value              = "udsp info",
5188                         .lkp_key_format         = LNKF_MAPPING,
5189                         .lkp_data_type          = NLA_NESTED
5190                 },
5191                 [LNET_NET_LOCAL_NI_ATTR_SEND_STATS]     = {
5192                         .lkp_value              = "sent_stats",
5193                         .lkp_key_format         = LNKF_MAPPING,
5194                         .lkp_data_type          = NLA_NESTED
5195                 },
5196                 [LNET_NET_LOCAL_NI_ATTR_RECV_STATS]     = {
5197                         .lkp_value              = "received_stats",
5198                         .lkp_key_format         = LNKF_MAPPING,
5199                         .lkp_data_type          = NLA_NESTED
5200                 },
5201                 [LNET_NET_LOCAL_NI_ATTR_DROPPED_STATS]  = {
5202                         .lkp_value              = "dropped_stats",
5203                         .lkp_key_format         = LNKF_MAPPING,
5204                         .lkp_data_type          = NLA_NESTED
5205
5206                 },
5207                 [LNET_NET_LOCAL_NI_ATTR_HEALTH_STATS]   = {
5208                         .lkp_value              = "health stats",
5209                         .lkp_key_format         = LNKF_MAPPING,
5210                         .lkp_data_type          = NLA_NESTED
5211                 },
5212                 [LNET_NET_LOCAL_NI_ATTR_TUNABLES]       = {
5213                         .lkp_value              = "tunables",
5214                         .lkp_key_format         = LNKF_MAPPING,
5215                         .lkp_data_type          = NLA_NESTED
5216                 },
5217                 [LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES]   = {
5218                         .lkp_value              = "lnd tunables",
5219                         .lkp_key_format         = LNKF_MAPPING,
5220                         .lkp_data_type          = NLA_NESTED
5221                 },
5222                 [LNET_NET_LOCAL_NI_DEV_CPT]             = {
5223                         .lkp_value              = "dev cpt",
5224                         .lkp_data_type          = NLA_S32,
5225                 },
5226                 [LNET_NET_LOCAL_NI_CPTS]                = {
5227                         .lkp_value              = "CPT",
5228                         .lkp_data_type          = NLA_STRING,
5229                 },
5230         },
5231 };
5232
5233 static const struct ln_key_list local_ni_interfaces_list = {
5234         .lkl_maxattr                    = LNET_NET_LOCAL_NI_INTF_ATTR_MAX,
5235         .lkl_list                       = {
5236                 [LNET_NET_LOCAL_NI_INTF_ATTR_TYPE] = {
5237                         .lkp_value      = "0",
5238                         .lkp_data_type  = NLA_STRING
5239                 },
5240         },
5241 };
5242
5243 static const struct ln_key_list local_ni_stats_list = {
5244         .lkl_maxattr                    = LNET_NET_LOCAL_NI_STATS_ATTR_MAX,
5245         .lkl_list                       = {
5246                 [LNET_NET_LOCAL_NI_STATS_ATTR_SEND_COUNT]       = {
5247                         .lkp_value      = "send_count",
5248                         .lkp_data_type  = NLA_U32
5249                 },
5250                 [LNET_NET_LOCAL_NI_STATS_ATTR_RECV_COUNT]       = {
5251                         .lkp_value      = "recv_count",
5252                         .lkp_data_type  = NLA_U32
5253                 },
5254                 [LNET_NET_LOCAL_NI_STATS_ATTR_DROP_COUNT]       = {
5255                         .lkp_value      = "drop_count",
5256                         .lkp_data_type  = NLA_U32
5257                 },
5258         },
5259 };
5260
5261 static const struct ln_key_list local_ni_msg_stats_list = {
5262         .lkl_maxattr                    = LNET_NET_LOCAL_NI_MSG_STATS_ATTR_MAX,
5263         .lkl_list                       = {
5264                 [LNET_NET_LOCAL_NI_MSG_STATS_ATTR_PUT_COUNT]    = {
5265                         .lkp_value      = "put",
5266                         .lkp_data_type  = NLA_U32
5267                 },
5268                 [LNET_NET_LOCAL_NI_MSG_STATS_ATTR_GET_COUNT]    = {
5269                         .lkp_value      = "get",
5270                         .lkp_data_type  = NLA_U32
5271                 },
5272                 [LNET_NET_LOCAL_NI_MSG_STATS_ATTR_REPLY_COUNT]  = {
5273                         .lkp_value      = "reply",
5274                         .lkp_data_type  = NLA_U32
5275                 },
5276                 [LNET_NET_LOCAL_NI_MSG_STATS_ATTR_ACK_COUNT]    = {
5277                         .lkp_value      = "ack",
5278                         .lkp_data_type  = NLA_U32
5279                 },
5280                 [LNET_NET_LOCAL_NI_MSG_STATS_ATTR_HELLO_COUNT]  = {
5281                         .lkp_value      = "hello",
5282                         .lkp_data_type  = NLA_U32
5283                 },
5284         },
5285 };
5286
5287 static const struct ln_key_list local_ni_health_stats_list = {
5288         .lkl_maxattr                    = LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_MAX,
5289         .lkl_list                       = {
5290                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_FATAL_ERRORS] = {
5291                         .lkp_value      = "fatal_error",
5292                         .lkp_data_type  = NLA_S32
5293                 },
5294                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_LEVEL] = {
5295                         .lkp_value      = "health value",
5296                         .lkp_data_type  = NLA_S32
5297                 },
5298                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_INTERRUPTS] = {
5299                         .lkp_value      = "interrupts",
5300                         .lkp_data_type  = NLA_U32
5301                 },
5302                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_DROPPED] = {
5303                         .lkp_value      = "dropped",
5304                         .lkp_data_type  = NLA_U32
5305                 },
5306                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_ABORTED] = {
5307                         .lkp_value      = "aborted",
5308                         .lkp_data_type  = NLA_U32
5309                 },
5310                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_NO_ROUTE] = {
5311                         .lkp_value      = "no route",
5312                         .lkp_data_type  = NLA_U32
5313                 },
5314                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_TIMEOUTS] = {
5315                         .lkp_value      = "timeouts",
5316                         .lkp_data_type  = NLA_U32
5317                 },
5318                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_ERROR] = {
5319                         .lkp_value      = "error",
5320                         .lkp_data_type  = NLA_U32
5321                 },
5322                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_PING_COUNT] = {
5323                         .lkp_value      = "ping_count",
5324                         .lkp_data_type  = NLA_U32,
5325                 },
5326                 [LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_NEXT_PING] = {
5327                         .lkp_value      = "next_ping",
5328                         .lkp_data_type  = NLA_U64
5329                 },
5330         },
5331 };
5332
5333 static const struct ln_key_list local_ni_tunables_list = {
5334         .lkl_maxattr                    = LNET_NET_LOCAL_NI_TUNABLES_ATTR_MAX,
5335         .lkl_list                       = {
5336                 [LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_TIMEOUT]  = {
5337                         .lkp_value      = "peer_timeout",
5338                         .lkp_data_type  = NLA_S32
5339                 },
5340                 [LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_CREDITS]  = {
5341                         .lkp_value      = "peer_credits",
5342                         .lkp_data_type  = NLA_S32
5343                 },
5344                 [LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_BUFFER_CREDITS] = {
5345                         .lkp_value      = "peer_buffer_credits",
5346                         .lkp_data_type  = NLA_S32
5347                 },
5348                 [LNET_NET_LOCAL_NI_TUNABLES_ATTR_CREDITS] = {
5349                         .lkp_value      = "credits",
5350                         .lkp_data_type  = NLA_S32
5351                 },
5352         },
5353 };
5354
5355 /* Use an index since the traversal is across LNet nets and ni collections */
5356 struct lnet_genl_net_list {
5357         unsigned int    lngl_net_id;
5358         unsigned int    lngl_idx;
5359 };
5360
5361 static inline struct lnet_genl_net_list *
5362 lnet_net_dump_ctx(struct netlink_callback *cb)
5363 {
5364         return (struct lnet_genl_net_list *)cb->args[0];
5365 }
5366
5367 static int lnet_net_show_done(struct netlink_callback *cb)
5368 {
5369         struct lnet_genl_net_list *nlist = lnet_net_dump_ctx(cb);
5370
5371         if (nlist) {
5372                 LIBCFS_FREE(nlist, sizeof(*nlist));
5373                 cb->args[0] = 0;
5374         }
5375
5376         return 0;
5377 }
5378
5379 /* LNet net ->start() handler for GET requests */
5380 static int lnet_net_show_start(struct netlink_callback *cb)
5381 {
5382         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
5383 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
5384         struct netlink_ext_ack *extack = NULL;
5385 #endif
5386         struct lnet_genl_net_list *nlist;
5387         int msg_len = genlmsg_len(gnlh);
5388         struct nlattr *params, *top;
5389         int rem, rc = 0;
5390
5391 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
5392         extack = cb->extack;
5393 #endif
5394         if (the_lnet.ln_refcount == 0) {
5395                 NL_SET_ERR_MSG(extack, "LNet stack down");
5396                 return -ENETDOWN;
5397         }
5398
5399         LIBCFS_ALLOC(nlist, sizeof(*nlist));
5400         if (!nlist)
5401                 return -ENOMEM;
5402
5403         nlist->lngl_net_id = LNET_NET_ANY;
5404         nlist->lngl_idx = 0;
5405         cb->args[0] = (long)nlist;
5406
5407         cb->min_dump_alloc = U16_MAX;
5408         if (!msg_len)
5409                 return 0;
5410
5411         params = genlmsg_data(gnlh);
5412         if (!(nla_type(params) & LN_SCALAR_ATTR_LIST)) {
5413                 NL_SET_ERR_MSG(extack, "invalid configuration");
5414                 return -EINVAL;
5415         }
5416
5417         nla_for_each_nested(top, params, rem) {
5418                 struct nlattr *net;
5419                 int rem2;
5420
5421                 nla_for_each_nested(net, top, rem2) {
5422                         char filter[LNET_NIDSTR_SIZE];
5423
5424                         if (nla_type(net) != LN_SCALAR_ATTR_VALUE ||
5425                             nla_strcmp(net, "net type") != 0)
5426                                 continue;
5427
5428                         net = nla_next(net, &rem2);
5429                         if (nla_type(net) != LN_SCALAR_ATTR_VALUE) {
5430                                 NL_SET_ERR_MSG(extack, "invalid config param");
5431                                 GOTO(report_err, rc = -EINVAL);
5432                         }
5433
5434                         rc = nla_strscpy(filter, net, sizeof(filter));
5435                         if (rc < 0) {
5436                                 NL_SET_ERR_MSG(extack, "failed to get param");
5437                                 GOTO(report_err, rc);
5438                         }
5439                         rc = 0;
5440
5441                         nlist->lngl_net_id = libcfs_str2net(filter);
5442                         if (nlist->lngl_net_id == LNET_NET_ANY) {
5443                                 NL_SET_ERR_MSG(extack, "cannot parse net");
5444                                 GOTO(report_err, rc = -ENOENT);
5445                         }
5446                 }
5447         }
5448 report_err:
5449         if (rc < 0)
5450                 lnet_net_show_done(cb);
5451
5452         return rc;
5453 }
5454
5455 static const struct ln_key_list net_update_props_list = {
5456         .lkl_maxattr                    = LNET_NET_ATTR_MAX,
5457         .lkl_list                       = {
5458                 [LNET_NET_ATTR_HDR]             = {
5459                         .lkp_value              = "",
5460                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
5461                         .lkp_data_type          = NLA_NUL_STRING,
5462                 },
5463                 [LNET_NET_ATTR_TYPE]            = {
5464                         .lkp_value              = "net type",
5465                         .lkp_data_type          = NLA_STRING
5466                 },
5467                 [LNET_NET_ATTR_LOCAL]           = {
5468                         .lkp_value              = "local NI(s)",
5469                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
5470                         .lkp_data_type          = NLA_NESTED
5471                 },
5472         },
5473 };
5474
5475 static int lnet_net_show_dump(struct sk_buff *msg,
5476                               struct netlink_callback *cb)
5477 {
5478         struct lnet_genl_net_list *nlist = lnet_net_dump_ctx(cb);
5479 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
5480         struct netlink_ext_ack *extack = NULL;
5481 #endif
5482         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
5483         int portid = NETLINK_CB(cb->skb).portid;
5484         bool found = false, started = false;
5485         const struct lnet_lnd *lnd = NULL;
5486         int idx = nlist->lngl_idx, rc = 0;
5487         int seq = cb->nlh->nlmsg_seq;
5488         struct lnet_net *net;
5489         void *hdr = NULL;
5490
5491 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
5492         extack = cb->extack;
5493 #endif
5494         lnet_net_lock(LNET_LOCK_EX);
5495
5496         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
5497                 struct nlattr *local_ni, *ni_attr;
5498                 struct lnet_ni *ni;
5499                 int dev = 0;
5500
5501                 if (nlist->lngl_net_id != LNET_NET_ANY &&
5502                     nlist->lngl_net_id != net->net_id)
5503                         continue;
5504
5505                 if (cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED &&
5506                     LNET_NETTYP(net->net_id) == LOLND)
5507                         continue;
5508
5509                 if (gnlh->version && LNET_NETTYP(net->net_id) != LOLND) {
5510                         if (!net->net_lnd) {
5511                                 NL_SET_ERR_MSG(extack,
5512                                                "LND not setup for NI");
5513                                 GOTO(net_unlock, rc = -ENODEV);
5514                         }
5515                         if (net->net_lnd != lnd)
5516                                 lnd = net->net_lnd;
5517                         else
5518                                 lnd = NULL;
5519                 }
5520
5521                 /* We need to resend the key table every time the base LND
5522                  * changed.
5523                  */
5524                 if (!idx || lnd) {
5525                         const struct ln_key_list *all[] = {
5526                                 &net_props_list, &local_ni_list,
5527                                 &local_ni_interfaces_list,
5528                                 &local_ni_stats_list,
5529                                 &udsp_info_list,
5530                                 &udsp_info_pref_nids_list,
5531                                 &udsp_info_pref_nids_list,
5532                                 &local_ni_msg_stats_list,
5533                                 &local_ni_msg_stats_list,
5534                                 &local_ni_msg_stats_list,
5535                                 &local_ni_health_stats_list,
5536                                 &local_ni_tunables_list,
5537                                 NULL, /* lnd tunables */
5538                                 NULL
5539                         };
5540                         int flags = NLM_F_CREATE | NLM_F_MULTI;
5541
5542                         if (lnd) {
5543                                 all[ARRAY_SIZE(all) - 2] = lnd->lnd_keys;
5544                                 if (idx) {
5545                                         all[0] = &net_update_props_list;
5546                                         flags |= NLM_F_REPLACE;
5547                                 }
5548                         }
5549
5550                         rc = lnet_genl_send_scalar_list(msg, portid, seq,
5551                                                         &lnet_family, flags,
5552                                                         LNET_CMD_NETS, all);
5553                         if (rc < 0) {
5554                                 NL_SET_ERR_MSG(extack, "failed to send key table");
5555                                 GOTO(net_unlock, rc);
5556                         }
5557                         started = true;
5558                 }
5559
5560                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
5561                                   NLM_F_MULTI, LNET_CMD_NETS);
5562                 if (!hdr) {
5563                         NL_SET_ERR_MSG(extack, "failed to send values");
5564                         GOTO(net_unlock, rc = -EMSGSIZE);
5565                 }
5566
5567                 if (started) {
5568                         nla_put_string(msg, LNET_NET_ATTR_HDR, "");
5569                         started = false;
5570                 }
5571
5572                 nla_put_string(msg, LNET_NET_ATTR_TYPE,
5573                                libcfs_net2str(net->net_id));
5574
5575                 local_ni = nla_nest_start(msg, LNET_NET_ATTR_LOCAL);
5576                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
5577                         char *status = "up";
5578
5579                         if (idx++ < nlist->lngl_idx)
5580                                 continue;
5581
5582                         ni_attr = nla_nest_start(msg, dev++);
5583                         found = true;
5584                         lnet_ni_lock(ni);
5585                         if (!(cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED)) {
5586                                 nla_put_string(msg, LNET_NET_LOCAL_NI_ATTR_NID,
5587                                                libcfs_nidstr(&ni->ni_nid));
5588                                 if (!nid_is_lo0(&ni->ni_nid) &&
5589                                     lnet_ni_get_status_locked(ni) != LNET_NI_STATUS_UP)
5590                                         status = "down";
5591                                 nla_put_string(msg, LNET_NET_LOCAL_NI_ATTR_STATUS,
5592                                                status);
5593                         }
5594
5595                         if (!nid_is_lo0(&ni->ni_nid) && ni->ni_interface) {
5596                                 struct nlattr *intf_nest, *intf_attr;
5597
5598                                 intf_nest = nla_nest_start(msg,
5599                                                            LNET_NET_LOCAL_NI_ATTR_INTERFACE);
5600                                 intf_attr = nla_nest_start(msg, 0);
5601                                 nla_put_string(msg,
5602                                                LNET_NET_LOCAL_NI_INTF_ATTR_TYPE,
5603                                                ni->ni_interface);
5604                                 nla_nest_end(msg, intf_attr);
5605                                 nla_nest_end(msg, intf_nest);
5606                         }
5607
5608                         if (gnlh->version) {
5609                                 char cpts[LNET_MAX_SHOW_NUM_CPT * 4 + 4], *cpt;
5610                                 struct lnet_ioctl_element_msg_stats msg_stats;
5611                                 struct lnet_ioctl_element_stats stats;
5612                                 size_t buf_len = sizeof(cpts), len;
5613                                 struct nlattr *health_attr, *health_stats;
5614                                 struct nlattr *send_attr, *send_stats;
5615                                 struct nlattr *recv_attr, *recv_stats;
5616                                 struct nlattr *drop_attr, *drop_stats;
5617                                 struct nlattr *stats_attr, *ni_stats;
5618                                 struct nlattr *tun_attr, *ni_tun;
5619                                 int j;
5620
5621                                 if (cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED) {
5622                                         lnet_ni_unlock(ni);
5623                                         goto skip_msg_stats;
5624                                 }
5625
5626                                 stats.iel_send_count = lnet_sum_stats(&ni->ni_stats,
5627                                                                       LNET_STATS_TYPE_SEND);
5628                                 stats.iel_recv_count = lnet_sum_stats(&ni->ni_stats,
5629                                                                       LNET_STATS_TYPE_RECV);
5630                                 stats.iel_drop_count = lnet_sum_stats(&ni->ni_stats,
5631                                                                       LNET_STATS_TYPE_DROP);
5632                                 lnet_ni_unlock(ni);
5633
5634                                 stats_attr = nla_nest_start(msg, LNET_NET_LOCAL_NI_ATTR_STATS);
5635                                 ni_stats = nla_nest_start(msg, 0);
5636                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_STATS_ATTR_SEND_COUNT,
5637                                             stats.iel_send_count);
5638                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_STATS_ATTR_RECV_COUNT,
5639                                             stats.iel_recv_count);
5640                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_STATS_ATTR_DROP_COUNT,
5641                                             stats.iel_drop_count);
5642                                 nla_nest_end(msg, ni_stats);
5643                                 nla_nest_end(msg, stats_attr);
5644
5645                                 if (gnlh->version < 4)
5646                                         goto skip_udsp;
5647
5648                                 /* UDSP info */
5649                                 rc = lnet_udsp_info_send(msg, LNET_NET_LOCAL_NI_ATTR_UDSP_INFO,
5650                                                          &ni->ni_nid, false);
5651                                 if (rc < 0) {
5652                                         NL_SET_ERR_MSG(extack,
5653                                                        "Failed to get udsp info");
5654                                         genlmsg_cancel(msg, hdr);
5655                                         GOTO(net_unlock, rc = -ENOMEM);
5656                                 }
5657 skip_udsp:
5658                                 if (gnlh->version < 2)
5659                                         goto skip_msg_stats;
5660
5661                                 msg_stats.im_idx = idx - 1;
5662                                 rc = lnet_get_ni_stats(&msg_stats);
5663                                 if (rc < 0) {
5664                                         NL_SET_ERR_MSG(extack,
5665                                                        "failed to get msg stats");
5666                                         genlmsg_cancel(msg, hdr);
5667                                         GOTO(net_unlock, rc = -ENOMEM);
5668                                 }
5669
5670                                 send_stats = nla_nest_start(msg, LNET_NET_LOCAL_NI_ATTR_SEND_STATS);
5671                                 send_attr = nla_nest_start(msg, 0);
5672                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_PUT_COUNT,
5673                                             msg_stats.im_send_stats.ico_get_count);
5674                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_GET_COUNT,
5675                                             msg_stats.im_send_stats.ico_put_count);
5676                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_REPLY_COUNT,
5677                                             msg_stats.im_send_stats.ico_reply_count);
5678                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_ACK_COUNT,
5679                                             msg_stats.im_send_stats.ico_ack_count);
5680                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_HELLO_COUNT,
5681                                             msg_stats.im_send_stats.ico_hello_count);
5682                                 nla_nest_end(msg, send_attr);
5683                                 nla_nest_end(msg, send_stats);
5684
5685                                 recv_stats = nla_nest_start(msg, LNET_NET_LOCAL_NI_ATTR_RECV_STATS);
5686                                 recv_attr = nla_nest_start(msg, 0);
5687                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_PUT_COUNT,
5688                                             msg_stats.im_recv_stats.ico_get_count);
5689                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_GET_COUNT,
5690                                             msg_stats.im_recv_stats.ico_put_count);
5691                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_REPLY_COUNT,
5692                                             msg_stats.im_recv_stats.ico_reply_count);
5693                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_ACK_COUNT,
5694                                             msg_stats.im_recv_stats.ico_ack_count);
5695                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_HELLO_COUNT,
5696                                             msg_stats.im_recv_stats.ico_hello_count);
5697                                 nla_nest_end(msg, recv_attr);
5698                                 nla_nest_end(msg, recv_stats);
5699
5700                                 drop_stats = nla_nest_start(msg,
5701                                                             LNET_NET_LOCAL_NI_ATTR_DROPPED_STATS);
5702                                 drop_attr = nla_nest_start(msg, 0);
5703                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_PUT_COUNT,
5704                                             msg_stats.im_drop_stats.ico_get_count);
5705                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_GET_COUNT,
5706                                             msg_stats.im_drop_stats.ico_put_count);
5707                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_REPLY_COUNT,
5708                                             msg_stats.im_drop_stats.ico_reply_count);
5709                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_ACK_COUNT,
5710                                             msg_stats.im_drop_stats.ico_ack_count);
5711                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_MSG_STATS_ATTR_HELLO_COUNT,
5712                                             msg_stats.im_drop_stats.ico_hello_count);
5713                                 nla_nest_end(msg, drop_attr);
5714                                 nla_nest_end(msg, drop_stats);
5715
5716                                 /* health stats */
5717                                 health_stats = nla_nest_start(msg,
5718                                                               LNET_NET_LOCAL_NI_ATTR_HEALTH_STATS);
5719                                 health_attr = nla_nest_start(msg, 0);
5720                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_FATAL_ERRORS,
5721                                             atomic_read(&ni->ni_fatal_error_on));
5722                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_LEVEL,
5723                                             atomic_read(&ni->ni_healthv));
5724                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_INTERRUPTS,
5725                                             atomic_read(&ni->ni_hstats.hlt_local_interrupt));
5726                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_DROPPED,
5727                                             atomic_read(&ni->ni_hstats.hlt_local_dropped));
5728                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_ABORTED,
5729                                             atomic_read(&ni->ni_hstats.hlt_local_aborted));
5730                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_NO_ROUTE,
5731                                             atomic_read(&ni->ni_hstats.hlt_local_no_route));
5732                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_TIMEOUTS,
5733                                             atomic_read(&ni->ni_hstats.hlt_local_timeout));
5734                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_ERROR,
5735                                             atomic_read(&ni->ni_hstats.hlt_local_error));
5736                                 nla_put_u32(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_PING_COUNT,
5737                                             ni->ni_ping_count);
5738                                 nla_put_u64_64bit(msg, LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_NEXT_PING,
5739                                                   ni->ni_next_ping,
5740                                                   LNET_NET_LOCAL_NI_HEALTH_STATS_ATTR_PAD);
5741                                 nla_nest_end(msg, health_attr);
5742                                 nla_nest_end(msg, health_stats);
5743 skip_msg_stats:
5744                                 /* Report net tunables */
5745                                 tun_attr = nla_nest_start(msg, LNET_NET_LOCAL_NI_ATTR_TUNABLES);
5746                                 ni_tun = nla_nest_start(msg, 0);
5747                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_TIMEOUT,
5748                                             ni->ni_net->net_tunables.lct_peer_timeout);
5749                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_CREDITS,
5750                                             ni->ni_net->net_tunables.lct_peer_tx_credits);
5751                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_BUFFER_CREDITS,
5752                                             ni->ni_net->net_tunables.lct_peer_rtr_credits);
5753                                 nla_put_s32(msg, LNET_NET_LOCAL_NI_TUNABLES_ATTR_CREDITS,
5754                                             ni->ni_net->net_tunables.lct_max_tx_credits);
5755                                 nla_nest_end(msg, ni_tun);
5756
5757                                 nla_nest_end(msg, tun_attr);
5758
5759                                 if (lnd && lnd->lnd_nl_get && lnd->lnd_keys) {
5760                                         struct nlattr *lnd_tun_attr, *lnd_ni_tun;
5761
5762                                         lnd_tun_attr = nla_nest_start(msg,
5763                                                                       LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES);
5764                                         lnd_ni_tun = nla_nest_start(msg, 0);
5765                                         rc = lnd->lnd_nl_get(LNET_CMD_NETS, msg,
5766                                                              LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES,
5767                                                              ni);
5768                                         if (rc < 0) {
5769                                                 NL_SET_ERR_MSG(extack,
5770                                                                "failed to get lnd tunables");
5771                                                 genlmsg_cancel(msg, hdr);
5772                                                 GOTO(net_unlock, rc);
5773                                         }
5774                                         nla_nest_end(msg, lnd_ni_tun);
5775                                         nla_nest_end(msg, lnd_tun_attr);
5776                                 }
5777
5778                                 if (!(cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED))
5779                                         nla_put_s32(msg, LNET_NET_LOCAL_NI_DEV_CPT,
5780                                                     ni->ni_dev_cpt);
5781
5782                                 /* Report cpts. We could send this as a nested list
5783                                  * of integers but older versions of the tools
5784                                  * except a string. The new versions can handle
5785                                  * both formats so in the future we can change
5786                                  * this to a nested list.
5787                                  */
5788                                 len = snprintf(cpts, buf_len, "\"[");
5789                                 cpt = cpts + len;
5790                                 buf_len -= len;
5791
5792                                 if (ni->ni_ncpts == LNET_CPT_NUMBER && !ni->ni_cpts)  {
5793                                         for (j = 0; j < ni->ni_ncpts; j++) {
5794                                                 len = snprintf(cpt, buf_len, "%d,", j);
5795                                                 buf_len -= len;
5796                                                 cpt += len;
5797                                         }
5798                                 } else {
5799                                         for (j = 0;
5800                                              ni->ni_cpts && j < ni->ni_ncpts &&
5801                                              j < LNET_MAX_SHOW_NUM_CPT; j++) {
5802                                                 len = snprintf(cpt, buf_len, "%d,",
5803                                                                ni->ni_cpts[j]);
5804                                                 buf_len -= len;
5805                                                 cpt += len;
5806                                         }
5807                                 }
5808                                 snprintf(cpt - 1, sizeof(cpts), "]\"");
5809
5810                                 nla_put_string(msg, LNET_NET_LOCAL_NI_CPTS, cpts);
5811                         } else {
5812                                 lnet_ni_unlock(ni);
5813                         }
5814                         nla_nest_end(msg, ni_attr);
5815                 }
5816                 nla_nest_end(msg, local_ni);
5817
5818                 genlmsg_end(msg, hdr);
5819         }
5820
5821         if (!found) {
5822                 struct nlmsghdr *nlh = nlmsg_hdr(msg);
5823
5824                 nlmsg_cancel(msg, nlh);
5825                 NL_SET_ERR_MSG(extack, "Network is down");
5826                 rc = -ESRCH;
5827         }
5828         nlist->lngl_idx = idx;
5829 net_unlock:
5830         lnet_net_unlock(LNET_LOCK_EX);
5831
5832         return lnet_nl_send_error(cb->skb, portid, seq, rc);
5833 }
5834
5835 #ifndef HAVE_NETLINK_CALLBACK_START
5836 static int lnet_old_net_show_dump(struct sk_buff *msg,
5837                                    struct netlink_callback *cb)
5838 {
5839         if (!cb->args[0]) {
5840                 int rc = lnet_net_show_start(cb);
5841
5842                 if (rc < 0)
5843                         return lnet_nl_send_error(cb->skb,
5844                                                   NETLINK_CB(cb->skb).portid,
5845                                                   cb->nlh->nlmsg_seq,
5846                                                   rc);
5847         }
5848
5849         return lnet_net_show_dump(msg, cb);
5850 }
5851 #endif
5852
5853 static int lnet_genl_parse_tunables(struct nlattr *settings,
5854                                     struct lnet_ioctl_config_lnd_tunables *tun)
5855 {
5856         struct nlattr *param;
5857         int rem, rc = 0;
5858
5859         nla_for_each_nested(param, settings, rem) {
5860                 int type = LNET_NET_LOCAL_NI_TUNABLES_ATTR_UNSPEC;
5861                 s64 num;
5862
5863                 if (nla_type(param) != LN_SCALAR_ATTR_VALUE)
5864                         continue;
5865
5866                 if (nla_strcmp(param, "peer_timeout") == 0)
5867                         type = LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_TIMEOUT;
5868                 else if (nla_strcmp(param, "peer_credits") == 0)
5869                         type = LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_CREDITS;
5870                 else if (nla_strcmp(param, "peer_buffer_credits") == 0)
5871                         type = LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_BUFFER_CREDITS;
5872                 else if (nla_strcmp(param, "credits") == 0)
5873                         type = LNET_NET_LOCAL_NI_TUNABLES_ATTR_CREDITS;
5874
5875                 param = nla_next(param, &rem);
5876                 if (nla_type(param) != LN_SCALAR_ATTR_INT_VALUE)
5877                         return -EINVAL;
5878
5879                 num = nla_get_s64(param);
5880                 switch (type) {
5881                 case LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_TIMEOUT:
5882                         if (num >= 0)
5883                                 tun->lt_cmn.lct_peer_timeout = num;
5884                         break;
5885                 case LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_CREDITS:
5886                         if (num > 0)
5887                                 tun->lt_cmn.lct_peer_tx_credits = num;
5888                         break;
5889                 case LNET_NET_LOCAL_NI_TUNABLES_ATTR_PEER_BUFFER_CREDITS:
5890                         if (num > 0)
5891                                 tun->lt_cmn.lct_peer_rtr_credits = num;
5892                         break;
5893                 case LNET_NET_LOCAL_NI_TUNABLES_ATTR_CREDITS:
5894                         if (num > 0)
5895                                 tun->lt_cmn.lct_max_tx_credits = num;
5896                         break;
5897                 default:
5898                         rc = -EINVAL;
5899                         break;
5900                 }
5901         }
5902         return rc;
5903 }
5904
5905 static int lnet_genl_parse_lnd_tunables(struct nlattr *settings,
5906                                         struct lnet_lnd_tunables *tun,
5907                                         const struct lnet_lnd *lnd)
5908 {
5909         const struct ln_key_list *list = lnd->lnd_keys;
5910         struct nlattr *param;
5911         int rem, rc = 0;
5912         int i = 1;
5913
5914         /* silently ignore these setting if the LND driver doesn't
5915          * support any LND tunables
5916          */
5917         if (!list || !lnd->lnd_nl_set || !list->lkl_maxattr)
5918                 return 0;
5919
5920         nla_for_each_nested(param, settings, rem) {
5921                 if (nla_type(param) != LN_SCALAR_ATTR_VALUE)
5922                         continue;
5923
5924                 for (i = 1; i <= list->lkl_maxattr; i++) {
5925                         if (!list->lkl_list[i].lkp_value ||
5926                             nla_strcmp(param, list->lkl_list[i].lkp_value) != 0)
5927                                 continue;
5928
5929                         param = nla_next(param, &rem);
5930                         rc = lnd->lnd_nl_set(LNET_CMD_NETS, param, i, tun);
5931                         if (rc < 0)
5932                                 return rc;
5933                 }
5934         }
5935
5936         return rc;
5937 }
5938
5939 static inline void
5940 lnet_genl_init_tunables(const struct lnet_lnd *lnd,
5941                         struct lnet_ioctl_config_lnd_tunables *tun)
5942 {
5943         const struct ln_key_list *list = lnd ? lnd->lnd_keys : NULL;
5944         int i;
5945
5946         tun->lt_cmn.lct_peer_timeout = -1;
5947         tun->lt_cmn.lct_peer_tx_credits = -1;
5948         tun->lt_cmn.lct_peer_rtr_credits = -1;
5949         tun->lt_cmn.lct_max_tx_credits = -1;
5950
5951         if (!list || !lnd->lnd_nl_set || !list->lkl_maxattr)
5952                 return;
5953
5954         /* init lnd tunables with default values */
5955         for (i = 1; i <= list->lkl_maxattr; i++)
5956                 lnd->lnd_nl_set(LNET_CMD_NETS, NULL, i, &tun->lt_tun);
5957 }
5958
5959 static int
5960 lnet_genl_parse_local_ni(struct nlattr *entry, struct genl_info *info,
5961                          int net_id, struct lnet_ioctl_config_ni *conf,
5962                          bool *ni_list)
5963 {
5964         struct lnet_ioctl_config_lnd_tunables *tun;
5965         struct lnet_nid nid = LNET_ANY_NID;
5966         const struct lnet_lnd *lnd = NULL;
5967         struct nlattr *settings;
5968         int healthv = -1;
5969         int rem3, rc = 0;
5970
5971         if (net_id != LNET_NET_ANY) {
5972                 lnd = lnet_load_lnd(LNET_NETTYP(net_id));
5973                 if (IS_ERR(lnd)) {
5974                         GENL_SET_ERR_MSG(info, "LND type not supported");
5975                         RETURN(PTR_ERR(lnd));
5976                 }
5977         }
5978
5979         LIBCFS_ALLOC(tun, sizeof(struct lnet_ioctl_config_lnd_tunables));
5980         if (!tun) {
5981                 GENL_SET_ERR_MSG(info, "cannot allocate memory for tunables");
5982                 GOTO(out, rc = -ENOMEM);
5983         }
5984
5985         /* Use LND defaults */
5986         lnet_genl_init_tunables(lnd, tun);
5987         conf->lic_ncpts = 0;
5988
5989         nla_for_each_nested(settings, entry, rem3) {
5990                 if (nla_type(settings) != LN_SCALAR_ATTR_VALUE)
5991                         continue;
5992
5993                 if (nla_strcmp(settings, "interfaces") == 0) {
5994                         struct nlattr *intf;
5995                         int rem4;
5996
5997                         settings = nla_next(settings, &rem3);
5998                         if (nla_type(settings) !=
5999                             LN_SCALAR_ATTR_LIST) {
6000                                 GENL_SET_ERR_MSG(info,
6001                                                  "invalid interfaces");
6002                                 GOTO(out, rc = -EINVAL);
6003                         }
6004
6005                         nla_for_each_nested(intf, settings, rem4) {
6006                                 intf = nla_next(intf, &rem4);
6007                                 if (nla_type(intf) !=
6008                                     LN_SCALAR_ATTR_VALUE) {
6009                                         GENL_SET_ERR_MSG(info,
6010                                                          "cannot parse interface");
6011                                         GOTO(out, rc = -EINVAL);
6012                                 }
6013
6014                                 rc = nla_strscpy(conf->lic_ni_intf, intf,
6015                                                  sizeof(conf->lic_ni_intf));
6016                                 if (rc < 0) {
6017                                         GENL_SET_ERR_MSG(info,
6018                                                          "failed to parse interfaces");
6019                                         GOTO(out, rc);
6020                                 }
6021                         }
6022                         *ni_list = true;
6023                 } else if (nla_strcmp(settings, "nid") == 0 &&
6024                            net_id != LNET_NET_ANY) {
6025                         char nidstr[LNET_NIDSTR_SIZE];
6026
6027                         settings = nla_next(settings, &rem3);
6028                         if (nla_type(settings) != LN_SCALAR_ATTR_VALUE) {
6029                                 GENL_SET_ERR_MSG(info, "cannot parse NID");
6030                                 GOTO(out, rc = -EINVAL);
6031                         }
6032
6033                         rc = nla_strscpy(nidstr, settings, sizeof(nidstr));
6034                         if (rc < 0) {
6035                                 GENL_SET_ERR_MSG(info,
6036                                                  "failed to parse NID");
6037                                 GOTO(out, rc);
6038                         }
6039
6040                         CDEBUG(D_NET, "Requested NID %s\n", nidstr);
6041                         rc = libcfs_strnid(&nid, strim(nidstr));
6042                         if (rc < 0) {
6043                                 GENL_SET_ERR_MSG(info, "unsupported NID");
6044                                 GOTO(out, rc);
6045                         }
6046
6047                         if (!(info->nlhdr->nlmsg_flags & NLM_F_REPLACE) &&
6048                              nid_same(&nid, &LNET_ANY_NID)) {
6049                                 GENL_SET_ERR_MSG(info, "any NID not supported");
6050                                 GOTO(out, rc = -EINVAL);
6051                         }
6052                         *ni_list = true;
6053                 } else if (nla_strcmp(settings, "health stats") == 0 &&
6054                            info->nlhdr->nlmsg_flags & NLM_F_REPLACE) {
6055                         struct nlattr *health;
6056                         int rem4;
6057
6058                         settings = nla_next(settings, &rem3);
6059                         if (nla_type(settings) != LN_SCALAR_ATTR_LIST) {
6060                                 GENL_SET_ERR_MSG(info,
6061                                                  "cannot parse health stats");
6062                                 GOTO(out, rc = -EINVAL);
6063                         }
6064
6065                         nla_for_each_nested(health, settings, rem4) {
6066                                 if (nla_type(health) != LN_SCALAR_ATTR_VALUE ||
6067                                     nla_strcmp(health, "health value") != 0) {
6068                                         GENL_SET_ERR_MSG(info,
6069                                                          "wrong health config format");
6070                                         GOTO(out, rc = -EINVAL);
6071                                 }
6072
6073                                 health = nla_next(health, &rem4);
6074                                 if (nla_type(health) !=
6075                                     LN_SCALAR_ATTR_INT_VALUE) {
6076                                         GENL_SET_ERR_MSG(info,
6077                                                          "invalid health config format");
6078                                         GOTO(out, rc = -EINVAL);
6079                                 }
6080
6081                                 healthv = nla_get_s64(health);
6082                                 clamp_t(s64, healthv, 0, LNET_MAX_HEALTH_VALUE);
6083                         }
6084                 } else if (nla_strcmp(settings, "tunables") == 0) {
6085                         settings = nla_next(settings, &rem3);
6086                         if (nla_type(settings) !=
6087                             LN_SCALAR_ATTR_LIST) {
6088                                 GENL_SET_ERR_MSG(info,
6089                                                  "invalid tunables");
6090                                 GOTO(out, rc = -EINVAL);
6091                         }
6092
6093                         rc = lnet_genl_parse_tunables(settings, tun);
6094                         if (rc < 0) {
6095                                 GENL_SET_ERR_MSG(info,
6096                                                  "failed to parse tunables");
6097                                 GOTO(out, rc);
6098                         }
6099                 } else if ((nla_strcmp(settings, "lnd tunables") == 0)) {
6100                         settings = nla_next(settings, &rem3);
6101                         if (nla_type(settings) !=
6102                             LN_SCALAR_ATTR_LIST) {
6103                                 GENL_SET_ERR_MSG(info,
6104                                                  "lnd tunables should be list\n");
6105                                 GOTO(out, rc = -EINVAL);
6106                         }
6107
6108                         if (info->nlhdr->nlmsg_flags & NLM_F_REPLACE &&
6109                             net_id == LNET_NET_ANY) {
6110                                 struct lnet_net *net;
6111
6112                                 lnet_net_lock(LNET_LOCK_EX);
6113                                 list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
6114                                         struct lnet_ni *ni;
6115
6116                                         if (!net->net_lnd)
6117                                                 continue;
6118
6119                                         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
6120                                                 if (!nid_same(&nid, &LNET_ANY_NID) &&
6121                                                     !nid_same(&nid, &ni->ni_nid))
6122                                                         continue;
6123
6124                                                 rc = lnet_genl_parse_lnd_tunables(settings,
6125                                                                                   &ni->ni_lnd_tunables,
6126                                                                                   net->net_lnd);
6127                                                 if (rc < 0) {
6128                                                         GENL_SET_ERR_MSG(info,
6129                                                                          "failed to parse lnd tunables");
6130                                                         lnet_net_unlock(LNET_LOCK_EX);
6131                                                         GOTO(out, rc);
6132                                                 }
6133                                         }
6134                                 }
6135                                 lnet_net_unlock(LNET_LOCK_EX);
6136                         } else {
6137                                 lnd = lnet_load_lnd(LNET_NETTYP(net_id));
6138                                 if (IS_ERR(lnd)) {
6139                                         GENL_SET_ERR_MSG(info,
6140                                                          "LND type not supported");
6141                                         GOTO(out, rc = PTR_ERR(lnd));
6142                                 }
6143
6144                                 rc = lnet_genl_parse_lnd_tunables(settings,
6145                                                                   &tun->lt_tun, lnd);
6146                                 if (rc < 0) {
6147                                         GENL_SET_ERR_MSG(info,
6148                                                          "failed to parse lnd tunables");
6149                                         GOTO(out, rc);
6150                                 }
6151                         }
6152                 } else if (nla_strcmp(settings, "CPT") == 0) {
6153                         struct nlattr *cpt;
6154                         int rem4;
6155
6156                         settings = nla_next(settings, &rem3);
6157                         if (nla_type(settings) != LN_SCALAR_ATTR_LIST) {
6158                                 GENL_SET_ERR_MSG(info,
6159                                                  "CPT should be list");
6160                                 GOTO(out, rc = -EINVAL);
6161                         }
6162
6163                         nla_for_each_nested(cpt, settings, rem4) {
6164                                 s64 core;
6165
6166                                 if (nla_type(cpt) !=
6167                                     LN_SCALAR_ATTR_INT_VALUE) {
6168                                         GENL_SET_ERR_MSG(info,
6169                                                          "invalid CPT config");
6170                                         GOTO(out, rc = -EINVAL);
6171                                 }
6172
6173                                 core = nla_get_s64(cpt);
6174                                 if (core >= LNET_CPT_NUMBER) {
6175                                         GENL_SET_ERR_MSG(info,
6176                                                          "invalid CPT value");
6177                                         GOTO(out, rc = -ERANGE);
6178                                 }
6179
6180                                 conf->lic_cpts[conf->lic_ncpts] = core;
6181                                 conf->lic_ncpts++;
6182                         }
6183                 }
6184         }
6185
6186         if (info->nlhdr->nlmsg_flags & NLM_F_CREATE) {
6187                 if (nid_same(&nid, &LNET_ANY_NID) &&
6188                     !strlen(conf->lic_ni_intf)) {
6189                         GENL_SET_ERR_MSG(info,
6190                                          "interface / NID is missing");
6191                         GOTO(out, rc);
6192                 }
6193
6194                 rc = lnet_dyn_add_ni(conf, net_id, &nid, tun);
6195                 switch (rc) {
6196                 case -ENOENT:
6197                         GENL_SET_ERR_MSG(info,
6198                                          "cannot parse net");
6199                         break;
6200                 case -ERANGE:
6201                         GENL_SET_ERR_MSG(info,
6202                                          "invalid CPT set");
6203                         break;
6204                 default:
6205                         GENL_SET_ERR_MSG(info,
6206                                          "cannot add LNet NI");
6207                 case 0:
6208                         break;
6209                 }
6210         } else if (info->nlhdr->nlmsg_flags & NLM_F_REPLACE && healthv != -1) {
6211                 lnet_ni_set_healthv(&nid, healthv);
6212         } else if (!(info->nlhdr->nlmsg_flags & (NLM_F_CREATE | NLM_F_REPLACE))) {
6213                 struct lnet_ni *ni;
6214
6215                 /* delete case */
6216                 rc = -ENODEV;
6217                 if (!strlen(conf->lic_ni_intf) &&
6218                     nid_same(&nid, &LNET_ANY_NID)) {
6219                         GENL_SET_ERR_MSG(info,
6220                                          "interface / NID is missing");
6221                         GOTO(out, rc);
6222                 }
6223
6224                 if (nid_same(&nid, &LNET_ANY_NID)) {
6225                         struct lnet_net *net;
6226                         bool found = false;
6227
6228                         lnet_net_lock(LNET_LOCK_EX);
6229                         net = lnet_get_net_locked(net_id);
6230                         if (!net) {
6231                                 GENL_SET_ERR_MSG(info,
6232                                                  "LNet net doesn't exist");
6233                                 lnet_net_unlock(LNET_LOCK_EX);
6234                                 GOTO(out, rc);
6235                         }
6236
6237                         list_for_each_entry(ni, &net->net_ni_list,
6238                                             ni_netlist) {
6239                                 if (!ni->ni_interface ||
6240                                     strcmp(ni->ni_interface,
6241                                           conf->lic_ni_intf) != 0)
6242                                         continue;
6243
6244                                 found = true;
6245                                 lnet_net_unlock(LNET_LOCK_EX);
6246                                 rc = lnet_dyn_del_ni(&ni->ni_nid);
6247                                 break;
6248                         }
6249
6250                         if (rc < 0 && !found) { /* will be -ENODEV */
6251                                 GENL_SET_ERR_MSG(info,
6252                                                  "interface invalid for deleting LNet NI");
6253                                 lnet_net_unlock(LNET_LOCK_EX);
6254                         }
6255                 } else {
6256                         rc = lnet_dyn_del_ni(&nid);
6257                 }
6258
6259                 if (rc < 0) {
6260                         GENL_SET_ERR_MSG(info,
6261                                          "cannot del LNet NI");
6262                         GOTO(out, rc);
6263                 }
6264         }
6265 out:
6266         if (tun)
6267                 LIBCFS_FREE(tun, sizeof(struct lnet_ioctl_config_lnd_tunables));
6268
6269         return rc;
6270 }
6271
6272 static int lnet_net_cmd(struct sk_buff *skb, struct genl_info *info)
6273 {
6274         struct nlmsghdr *nlh = nlmsg_hdr(skb);
6275         struct genlmsghdr *gnlh = nlmsg_data(nlh);
6276         struct nlattr *params = genlmsg_data(gnlh);
6277         int msg_len, rem, rc = 0;
6278         struct nlattr *attr;
6279
6280         msg_len = genlmsg_len(gnlh);
6281         if (!msg_len) {
6282                 GENL_SET_ERR_MSG(info, "no configuration");
6283                 return -ENOMSG;
6284         }
6285
6286         if (!(nla_type(params) & LN_SCALAR_ATTR_LIST)) {
6287                 GENL_SET_ERR_MSG(info, "invalid configuration");
6288                 return -EINVAL;
6289         }
6290
6291         nla_for_each_nested(attr, params, rem) {
6292                 bool ni_list = false, ipnets = false;
6293                 struct lnet_ioctl_config_ni conf;
6294                 u32 net_id = LNET_NET_ANY;
6295                 struct nlattr *entry;
6296                 int rem2;
6297
6298                 if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
6299                         continue;
6300
6301                 nla_for_each_nested(entry, attr, rem2) {
6302                         switch (nla_type(entry)) {
6303                         case LN_SCALAR_ATTR_VALUE: {
6304                                 ssize_t len;
6305
6306                                 memset(&conf, 0, sizeof(conf));
6307                                 if (nla_strcmp(entry, "ip2net") == 0) {
6308                                         entry = nla_next(entry, &rem2);
6309                                         if (nla_type(entry) !=
6310                                             LN_SCALAR_ATTR_VALUE) {
6311                                                 GENL_SET_ERR_MSG(info,
6312                                                                  "ip2net has invalid key");
6313                                                 GOTO(out, rc = -EINVAL);
6314                                         }
6315
6316                                         len = nla_strscpy(conf.lic_legacy_ip2nets,
6317                                                           entry,
6318                                                           sizeof(conf.lic_legacy_ip2nets));
6319                                         if (len < 0) {
6320                                                 GENL_SET_ERR_MSG(info,
6321                                                                  "ip2net key string is invalid");
6322                                                 GOTO(out, rc = len);
6323                                         }
6324                                         ni_list = true;
6325                                         ipnets = true;
6326                                 } else if (nla_strcmp(entry, "net type") == 0) {
6327                                         char tmp[LNET_NIDSTR_SIZE];
6328
6329                                         entry = nla_next(entry, &rem2);
6330                                         if (nla_type(entry) !=
6331                                             LN_SCALAR_ATTR_VALUE) {
6332                                                 GENL_SET_ERR_MSG(info,
6333                                                                  "net type has invalid key");
6334                                                 GOTO(out, rc = -EINVAL);
6335                                         }
6336
6337                                         len = nla_strscpy(tmp, entry,
6338                                                           sizeof(tmp));
6339                                         if (len < 0) {
6340                                                 GENL_SET_ERR_MSG(info,
6341                                                                  "net type key string is invalid");
6342                                                 GOTO(out, rc = len);
6343                                         }
6344
6345                                         net_id = libcfs_str2net(tmp);
6346                                         if (!net_id) {
6347                                                 GENL_SET_ERR_MSG(info,
6348                                                                  "cannot parse net");
6349                                                 GOTO(out, rc = -ENODEV);
6350                                         }
6351                                         if (LNET_NETTYP(net_id) == LOLND) {
6352                                                 GENL_SET_ERR_MSG(info,
6353                                                                  "setting @lo not allowed");
6354                                                 GOTO(out, rc = -ENODEV);
6355                                         }
6356                                         conf.lic_legacy_ip2nets[0] = '\0';
6357                                         conf.lic_ni_intf[0] = '\0';
6358                                         ni_list = false;
6359                                 }
6360                                 if (rc < 0)
6361                                         GOTO(out, rc);
6362                                 break;
6363                         }
6364                         case LN_SCALAR_ATTR_LIST: {
6365                                 struct nlattr *interface;
6366                                 int rem3;
6367
6368                                 ipnets = false;
6369                                 nla_for_each_nested(interface, entry, rem3) {
6370                                         rc = lnet_genl_parse_local_ni(interface, info,
6371                                                                       net_id, &conf,
6372                                                                       &ni_list);
6373                                         if (rc < 0)
6374                                                 GOTO(out, rc);
6375                                 }
6376                                 break;
6377                         }
6378                         /* it is possible a newer version of the user land send
6379                          * values older kernels doesn't handle. So silently
6380                          * ignore these values
6381                          */
6382                         default:
6383                                 break;
6384                         }
6385                 }
6386
6387                 /* Handle case of just sent NET with no list of NIDs */
6388                 if (!(info->nlhdr->nlmsg_flags & (NLM_F_CREATE | NLM_F_REPLACE)) &&
6389                     !ni_list) {
6390                         rc = lnet_dyn_del_net(net_id);
6391                         if (rc < 0) {
6392                                 GENL_SET_ERR_MSG(info,
6393                                                  "cannot del network");
6394                         }
6395                 } else if ((info->nlhdr->nlmsg_flags & NLM_F_CREATE) &&
6396                            ipnets && ni_list) {
6397                         rc = lnet_handle_legacy_ip2nets(conf.lic_legacy_ip2nets,
6398                                                         NULL);
6399                         if (rc < 0)
6400                                 GENL_SET_ERR_MSG(info,
6401                                                  "cannot setup ip2nets");
6402                 }
6403         }
6404 out:
6405         return rc;
6406 }
6407
6408 /* Called with ln_api_mutex */
6409 static int lnet_parse_peer_nis(struct nlattr *rlist, struct genl_info *info,
6410                                struct lnet_nid *pnid, bool mr,
6411                                bool *create_some)
6412 {
6413         struct lnet_nid snid = LNET_ANY_NID;
6414         struct nlattr *props;
6415         int rem, rc = 0;
6416         s64 num = -1;
6417
6418         nla_for_each_nested(props, rlist, rem) {
6419                 if (nla_type(props) != LN_SCALAR_ATTR_VALUE)
6420                         continue;
6421
6422                 if (nla_strcmp(props, "nid") == 0) {
6423                         char nidstr[LNET_NIDSTR_SIZE];
6424
6425                         props = nla_next(props, &rem);
6426                         if (nla_type(props) != LN_SCALAR_ATTR_VALUE) {
6427                                 GENL_SET_ERR_MSG(info,
6428                                                  "invalid secondary NID");
6429                                 GOTO(report_err, rc = -EINVAL);
6430                         }
6431
6432                         rc = nla_strscpy(nidstr, props, sizeof(nidstr));
6433                         if (rc < 0) {
6434                                 GENL_SET_ERR_MSG(info,
6435                                                  "failed to get secondary NID");
6436                                 GOTO(report_err, rc);
6437                         }
6438
6439                         rc = libcfs_strnid(&snid, strim(nidstr));
6440                         if (rc < 0) {
6441                                 GENL_SET_ERR_MSG(info, "unsupported secondary NID");
6442                                 GOTO(report_err, rc);
6443                         }
6444
6445                         if (LNET_NID_IS_ANY(&snid) || nid_same(&snid, pnid))
6446                                 *create_some = false;
6447                 } else if (nla_strcmp(props, "health stats") == 0) {
6448                         struct nlattr *health;
6449                         int rem2;
6450
6451                         props = nla_next(props, &rem);
6452                         if (nla_type(props) !=
6453                               LN_SCALAR_ATTR_LIST) {
6454                                 GENL_SET_ERR_MSG(info,
6455                                                  "invalid health configuration");
6456                                 GOTO(report_err, rc = -EINVAL);
6457                         }
6458
6459                         nla_for_each_nested(health, props, rem2) {
6460                                 if (nla_type(health) != LN_SCALAR_ATTR_VALUE ||
6461                                     nla_strcmp(health, "health value") != 0) {
6462                                         GENL_SET_ERR_MSG(info,
6463                                                          "wrong health config format");
6464                                         GOTO(report_err, rc = -EINVAL);
6465                                 }
6466
6467                                 health = nla_next(health, &rem2);
6468                                 if (nla_type(health) !=
6469                                     LN_SCALAR_ATTR_INT_VALUE) {
6470                                         GENL_SET_ERR_MSG(info,
6471                                                          "invalid health config format");
6472                                         GOTO(report_err, rc = -EINVAL);
6473                                 }
6474
6475                                 num = nla_get_s64(health);
6476                                 clamp_t(s64, num, 0, LNET_MAX_HEALTH_VALUE);
6477                         }
6478                 }
6479         }
6480
6481         if (info->nlhdr->nlmsg_flags & NLM_F_REPLACE && num != -1) {
6482                 lnet_peer_ni_set_healthv(pnid, num, !*create_some);
6483         } else if (info->nlhdr->nlmsg_flags & NLM_F_CREATE) {
6484                 bool lock_prim = info->nlhdr->nlmsg_flags & NLM_F_EXCL;
6485
6486                 rc = lnet_user_add_peer_ni(pnid, &snid, mr, lock_prim);
6487                 if (rc < 0)
6488                         GENL_SET_ERR_MSG(info,
6489                                          "failed to add peer");
6490         } else if (!(info->nlhdr->nlmsg_flags & NLM_F_CREATE) && *create_some) {
6491                 bool force = info->nlhdr->nlmsg_flags & NLM_F_EXCL;
6492
6493                 rc = lnet_del_peer_ni(pnid, &snid, force);
6494                 if (rc < 0)
6495                         GENL_SET_ERR_MSG(info,
6496                                          "failed to del peer");
6497         }
6498 report_err:
6499         return rc;
6500 }
6501
6502 static int lnet_peer_ni_cmd(struct sk_buff *skb, struct genl_info *info)
6503 {
6504         struct nlmsghdr *nlh = nlmsg_hdr(skb);
6505         struct genlmsghdr *gnlh = nlmsg_data(nlh);
6506         struct nlattr *params = genlmsg_data(gnlh);
6507         int msg_len, rem, rc = 0;
6508         struct lnet_nid pnid;
6509         struct nlattr *attr;
6510
6511         mutex_lock(&the_lnet.ln_api_mutex);
6512         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
6513                 GENL_SET_ERR_MSG(info, "Network is down");
6514                 mutex_unlock(&the_lnet.ln_api_mutex);
6515                 return -ENETDOWN;
6516         }
6517
6518         msg_len = genlmsg_len(gnlh);
6519         if (!msg_len) {
6520                 GENL_SET_ERR_MSG(info, "no configuration");
6521                 mutex_unlock(&the_lnet.ln_api_mutex);
6522                 return -ENOMSG;
6523         }
6524
6525         if (!(nla_type(params) & LN_SCALAR_ATTR_LIST)) {
6526                 GENL_SET_ERR_MSG(info, "invalid configuration");
6527                 mutex_unlock(&the_lnet.ln_api_mutex);
6528                 return -EINVAL;
6529         }
6530
6531         nla_for_each_nested(attr, params, rem) {
6532                 bool parse_peer_nis = false;
6533                 struct nlattr *pnid_prop;
6534                 int rem2;
6535
6536                 if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
6537                         continue;
6538
6539                 pnid = LNET_ANY_NID;
6540                 nla_for_each_nested(pnid_prop, attr, rem2) {
6541                         bool mr = true;
6542
6543                         if (nla_type(pnid_prop) != LN_SCALAR_ATTR_VALUE)
6544                                 continue;
6545
6546                         if (nla_strcmp(pnid_prop, "primary nid") == 0) {
6547                                 char nidstr[LNET_NIDSTR_SIZE];
6548
6549                                 pnid_prop = nla_next(pnid_prop, &rem2);
6550                                 if (nla_type(pnid_prop) !=
6551                                     LN_SCALAR_ATTR_VALUE) {
6552                                         GENL_SET_ERR_MSG(info,
6553                                                           "invalid primary NID type");
6554                                         GOTO(report_err, rc = -EINVAL);
6555                                 }
6556
6557                                 rc = nla_strscpy(nidstr, pnid_prop,
6558                                                  sizeof(nidstr));
6559                                 if (rc < 0) {
6560                                         GENL_SET_ERR_MSG(info,
6561                                                          "failed to get primary NID");
6562                                         GOTO(report_err, rc);
6563                                 }
6564
6565                                 rc = libcfs_strnid(&pnid, strim(nidstr));
6566                                 if (rc < 0) {
6567                                         GENL_SET_ERR_MSG(info,
6568                                                          "unsupported primary NID");
6569                                         GOTO(report_err, rc);
6570                                 }
6571
6572                                 /* we must create primary NID for peer ni
6573                                  * creation
6574                                  */
6575                                 if (info->nlhdr->nlmsg_flags & NLM_F_CREATE) {
6576                                         bool lock_prim;
6577
6578                                         lock_prim = info->nlhdr->nlmsg_flags & NLM_F_EXCL;
6579                                         rc = lnet_user_add_peer_ni(&pnid,
6580                                                                    &LNET_ANY_NID,
6581                                                                    true, lock_prim);
6582                                         if (rc < 0) {
6583                                                 GENL_SET_ERR_MSG(info,
6584                                                                  "failed to add primary peer");
6585                                                 GOTO(report_err, rc);
6586                                         }
6587                                 }
6588                         } else if (nla_strcmp(pnid_prop, "Multi-Rail") == 0) {
6589                                 pnid_prop = nla_next(pnid_prop, &rem2);
6590                                 if (nla_type(pnid_prop) !=
6591                                     LN_SCALAR_ATTR_INT_VALUE) {
6592                                         GENL_SET_ERR_MSG(info,
6593                                                           "invalid MR flag param");
6594                                         GOTO(report_err, rc = -EINVAL);
6595                                 }
6596
6597                                 if (nla_get_s64(pnid_prop) == 0)
6598                                         mr = false;
6599                         } else if (nla_strcmp(pnid_prop, "peer state") == 0) {
6600                                 struct lnet_peer_ni *lpni;
6601                                 struct lnet_peer *lp;
6602
6603                                 pnid_prop = nla_next(pnid_prop, &rem2);
6604                                 if (nla_type(pnid_prop) !=
6605                                     LN_SCALAR_ATTR_INT_VALUE) {
6606                                         GENL_SET_ERR_MSG(info,
6607                                                           "invalid peer state param");
6608                                         GOTO(report_err, rc = -EINVAL);
6609                                 }
6610
6611                                 lpni = lnet_peer_ni_find_locked(&pnid);
6612                                 if (!lpni) {
6613                                         GENL_SET_ERR_MSG(info,
6614                                                           "invalid peer state param");
6615                                         GOTO(report_err, rc = -ENOENT);
6616                                 }
6617                                 lnet_peer_ni_decref_locked(lpni);
6618                                 lp = lpni->lpni_peer_net->lpn_peer;
6619                                 lp->lp_state = nla_get_s64(pnid_prop);
6620                         } else if (nla_strcmp(pnid_prop, "peer ni") == 0) {
6621                                 struct nlattr *rlist;
6622                                 int rem3;
6623
6624                                 if (!(info->nlhdr->nlmsg_flags & NLM_F_REPLACE) &&
6625                                     LNET_NID_IS_ANY(&pnid)) {
6626                                         GENL_SET_ERR_MSG(info,
6627                                                          "missing required primary NID");
6628                                         GOTO(report_err, rc);
6629                                 }
6630
6631                                 pnid_prop = nla_next(pnid_prop, &rem2);
6632                                 if (nla_type(pnid_prop) !=
6633                                     LN_SCALAR_ATTR_LIST) {
6634                                         GENL_SET_ERR_MSG(info,
6635                                                           "invalid NIDs list");
6636                                         GOTO(report_err, rc = -EINVAL);
6637                                 }
6638
6639                                 parse_peer_nis = true;
6640                                 nla_for_each_nested(rlist, pnid_prop, rem3) {
6641                                         rc = lnet_parse_peer_nis(rlist, info,
6642                                                                  &pnid, mr,
6643                                                                  &parse_peer_nis);
6644                                         if (rc < 0)
6645                                                 GOTO(report_err, rc);
6646                                 }
6647                         }
6648                 }
6649
6650                 /* If we have remote peer ni's we already add /del peers */
6651                 if (parse_peer_nis)
6652                         continue;
6653
6654                 if (LNET_NID_IS_ANY(&pnid)) {
6655                         GENL_SET_ERR_MSG(info, "missing primary NID");
6656                         GOTO(report_err, rc);
6657                 }
6658
6659                 if (!(info->nlhdr->nlmsg_flags & NLM_F_CREATE)) {
6660                         bool force = info->nlhdr->nlmsg_flags & NLM_F_EXCL;
6661
6662                         rc = lnet_del_peer_ni(&pnid, &LNET_ANY_NID,
6663                                               force);
6664                         if (rc < 0) {
6665                                 GENL_SET_ERR_MSG(info,
6666                                                  "failed to del primary peer");
6667                                 GOTO(report_err, rc);
6668                         }
6669                 }
6670         }
6671 report_err:
6672         /* If we failed on creation and encounter a latter error then
6673          * delete the primary nid.
6674          */
6675         if (rc < 0 && info->nlhdr->nlmsg_flags & NLM_F_CREATE &&
6676             !LNET_NID_IS_ANY(&pnid))
6677                 lnet_del_peer_ni(&pnid, &LNET_ANY_NID,
6678                                  info->nlhdr->nlmsg_flags & NLM_F_EXCL);
6679         mutex_unlock(&the_lnet.ln_api_mutex);
6680
6681         return rc;
6682 }
6683
6684 /** LNet route handling */
6685
6686 /* We can't use struct lnet_ioctl_config_data since it lacks
6687  * support for large NIDS
6688  */
6689 struct lnet_route_properties {
6690         struct lnet_nid         lrp_gateway;
6691         u32                     lrp_net;
6692         s32                     lrp_hop;
6693         u32                     lrp_flags;
6694         u32                     lrp_priority;
6695         u32                     lrp_sensitivity;
6696 };
6697
6698 struct lnet_genl_route_list {
6699         unsigned int                            lgrl_index;
6700         unsigned int                            lgrl_count;
6701         GENRADIX(struct lnet_route_properties)  lgrl_list;
6702 };
6703
6704 static inline struct lnet_genl_route_list *
6705 lnet_route_dump_ctx(struct netlink_callback *cb)
6706 {
6707         return (struct lnet_genl_route_list *)cb->args[0];
6708 }
6709
6710 static int lnet_route_show_done(struct netlink_callback *cb)
6711 {
6712         struct lnet_genl_route_list *rlist = lnet_route_dump_ctx(cb);
6713
6714         if (rlist) {
6715                 genradix_free(&rlist->lgrl_list);
6716                 CFS_FREE_PTR(rlist);
6717         }
6718         cb->args[0] = 0;
6719
6720         return 0;
6721 }
6722
6723 static int lnet_scan_route(struct lnet_genl_route_list *rlist,
6724                     struct lnet_route_properties *settings)
6725 {
6726         struct lnet_remotenet *rnet;
6727         struct list_head *rn_list;
6728         struct lnet_route *route;
6729         int cpt, i, rc = 0;
6730
6731         cpt = lnet_net_lock_current();
6732
6733         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
6734                 rn_list = &the_lnet.ln_remote_nets_hash[i];
6735                 list_for_each_entry(rnet, rn_list, lrn_list) {
6736                         if (settings->lrp_net != LNET_NET_ANY &&
6737                             settings->lrp_net != rnet->lrn_net)
6738                                 continue;
6739
6740                         list_for_each_entry(route, &rnet->lrn_routes,
6741                                             lr_list) {
6742                                 struct lnet_route_properties *prop;
6743
6744                                 if (!LNET_NID_IS_ANY(&settings->lrp_gateway) &&
6745                                     !nid_same(&settings->lrp_gateway,
6746                                               &route->lr_nid)) {
6747                                         continue;
6748                                 }
6749
6750                                 if (settings->lrp_hop != -1 &&
6751                                     settings->lrp_hop != route->lr_hops)
6752                                         continue;
6753
6754                                 if (settings->lrp_priority != -1 &&
6755                                     settings->lrp_priority != route->lr_priority)
6756                                         continue;
6757
6758                                 if (settings->lrp_sensitivity != -1 &&
6759                                     settings->lrp_sensitivity !=
6760                                     route->lr_gateway->lp_health_sensitivity)
6761                                         continue;
6762
6763                                 prop = genradix_ptr_alloc(&rlist->lgrl_list,
6764                                                           rlist->lgrl_count++,
6765                                                           GFP_ATOMIC);
6766                                 if (!prop)
6767                                         GOTO(failed_alloc, rc = -ENOMEM);
6768
6769                                 prop->lrp_net = rnet->lrn_net;
6770                                 prop->lrp_gateway = route->lr_nid;
6771                                 prop->lrp_hop = route->lr_hops;
6772                                 prop->lrp_priority = route->lr_priority;
6773                                 prop->lrp_sensitivity =
6774                                         route->lr_gateway->lp_health_sensitivity;
6775                                 if (lnet_is_route_alive(route))
6776                                         prop->lrp_flags |= LNET_RT_ALIVE;
6777                                 else
6778                                         prop->lrp_flags &= ~LNET_RT_ALIVE;
6779                                 if (route->lr_single_hop)
6780                                         prop->lrp_flags &= ~LNET_RT_MULTI_HOP;
6781                                 else
6782                                         prop->lrp_flags |= LNET_RT_MULTI_HOP;
6783                         }
6784                 }
6785         }
6786
6787 failed_alloc:
6788         lnet_net_unlock(cpt);
6789         return rc;
6790 }
6791
6792 /* LNet route ->start() handler for GET requests */
6793 static int lnet_route_show_start(struct netlink_callback *cb)
6794 {
6795         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
6796 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
6797         struct netlink_ext_ack *extack = NULL;
6798 #endif
6799         struct lnet_genl_route_list *rlist;
6800         int msg_len = genlmsg_len(gnlh);
6801         int rc = 0;
6802
6803 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
6804         extack = cb->extack;
6805 #endif
6806         if (the_lnet.ln_refcount == 0 ||
6807             the_lnet.ln_state != LNET_STATE_RUNNING) {
6808                 NL_SET_ERR_MSG(extack, "Network is down");
6809                 return -ENETDOWN;
6810         }
6811
6812         CFS_ALLOC_PTR(rlist);
6813         if (!rlist) {
6814                 NL_SET_ERR_MSG(extack, "No memory for route list");
6815                 return -ENOMEM;
6816         }
6817
6818         genradix_init(&rlist->lgrl_list);
6819         rlist->lgrl_count = 0;
6820         rlist->lgrl_index = 0;
6821         cb->args[0] = (long)rlist;
6822
6823         mutex_lock(&the_lnet.ln_api_mutex);
6824         if (!msg_len) {
6825                 struct lnet_route_properties tmp = {
6826                         .lrp_gateway            = LNET_ANY_NID,
6827                         .lrp_net                = LNET_NET_ANY,
6828                         .lrp_hop                = -1,
6829                         .lrp_priority           = -1,
6830                         .lrp_sensitivity        = -1,
6831                 };
6832
6833                 rc = lnet_scan_route(rlist, &tmp);
6834                 if (rc < 0) {
6835                         NL_SET_ERR_MSG(extack,
6836                                        "failed to allocate router data");
6837                         GOTO(report_err, rc);
6838                 }
6839         } else {
6840                 struct nlattr *params = genlmsg_data(gnlh);
6841                 struct nlattr *attr;
6842                 int rem;
6843
6844                 nla_for_each_nested(attr, params, rem) {
6845                         struct lnet_route_properties tmp = {
6846                                 .lrp_gateway            = LNET_ANY_NID,
6847                                 .lrp_net                = LNET_NET_ANY,
6848                                 .lrp_hop                = -1,
6849                                 .lrp_priority           = -1,
6850                                 .lrp_sensitivity        = -1,
6851                         };
6852                         struct nlattr *route;
6853                         int rem2;
6854
6855                         if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
6856                                 continue;
6857
6858                         nla_for_each_nested(route, attr, rem2) {
6859                                 if (nla_type(route) != LN_SCALAR_ATTR_VALUE)
6860                                         continue;
6861
6862                                 if (nla_strcmp(route, "net") == 0) {
6863                                         char nw[LNET_NIDSTR_SIZE];
6864
6865                                         route = nla_next(route, &rem2);
6866                                         if (nla_type(route) !=
6867                                             LN_SCALAR_ATTR_VALUE) {
6868                                                 NL_SET_ERR_MSG(extack,
6869                                                                "invalid net param");
6870                                                 GOTO(report_err, rc = -EINVAL);
6871                                         }
6872
6873                                         rc = nla_strscpy(nw, route, sizeof(nw));
6874                                         if (rc < 0) {
6875                                                 NL_SET_ERR_MSG(extack,
6876                                                                "failed to get route param");
6877                                                 GOTO(report_err, rc);
6878                                         }
6879                                         rc = 0;
6880                                         tmp.lrp_net = libcfs_str2net(strim(nw));
6881                                 } else if (nla_strcmp(route, "gateway") == 0) {
6882                                         char gw[LNET_NIDSTR_SIZE];
6883
6884                                         route = nla_next(route, &rem2);
6885                                         if (nla_type(route) !=
6886                                             LN_SCALAR_ATTR_VALUE) {
6887                                                 NL_SET_ERR_MSG(extack,
6888                                                                "invalid gateway param");
6889                                                 GOTO(report_err, rc = -EINVAL);
6890                                         }
6891
6892                                         rc = nla_strscpy(gw, route, sizeof(gw));
6893                                         if (rc < 0) {
6894                                                 NL_SET_ERR_MSG(extack,
6895                                                                "failed to get route param");
6896                                                 GOTO(report_err, rc);
6897                                         }
6898
6899                                         rc = libcfs_strnid(&tmp.lrp_gateway, strim(gw));
6900                                         if (rc < 0) {
6901                                                 NL_SET_ERR_MSG(extack,
6902                                                                "cannot parse gateway");
6903                                                 GOTO(report_err, rc = -ENODEV);
6904                                         }
6905                                         rc = 0;
6906                                 } else if (nla_strcmp(route, "hop") == 0) {
6907                                         route = nla_next(route, &rem2);
6908                                         if (nla_type(route) !=
6909                                             LN_SCALAR_ATTR_INT_VALUE) {
6910                                                 NL_SET_ERR_MSG(extack,
6911                                                                "invalid hop param");
6912                                                 GOTO(report_err, rc = -EINVAL);
6913                                         }
6914
6915                                         tmp.lrp_hop = nla_get_s64(route);
6916                                         if (tmp.lrp_hop != -1)
6917                                                 clamp_t(s32, tmp.lrp_hop, 1, 127);
6918                                 } else if (nla_strcmp(route, "priority") == 0) {
6919                                         route = nla_next(route, &rem2);
6920                                         if (nla_type(route) !=
6921                                             LN_SCALAR_ATTR_INT_VALUE) {
6922                                                 NL_SET_ERR_MSG(extack,
6923                                                                "invalid priority param");
6924                                                 GOTO(report_err, rc = -EINVAL);
6925                                         }
6926
6927                                         tmp.lrp_priority = nla_get_s64(route);
6928                                 }
6929                         }
6930
6931                         rc = lnet_scan_route(rlist, &tmp);
6932                         if (rc < 0) {
6933                                 NL_SET_ERR_MSG(extack,
6934                                                "failed to allocate router data");
6935                                 GOTO(report_err, rc);
6936                         }
6937                 }
6938         }
6939 report_err:
6940         mutex_unlock(&the_lnet.ln_api_mutex);
6941
6942         if (rc < 0)
6943                 lnet_route_show_done(cb);
6944
6945         return rc;
6946 }
6947
6948 static const struct ln_key_list route_props_list = {
6949         .lkl_maxattr                    = LNET_ROUTE_ATTR_MAX,
6950         .lkl_list                       = {
6951                 [LNET_ROUTE_ATTR_HDR]                   = {
6952                         .lkp_value                      = "route",
6953                         .lkp_key_format                 = LNKF_SEQUENCE | LNKF_MAPPING,
6954                         .lkp_data_type                  = NLA_NUL_STRING,
6955                 },
6956                 [LNET_ROUTE_ATTR_NET]                   = {
6957                         .lkp_value                      = "net",
6958                         .lkp_data_type                  = NLA_STRING
6959                 },
6960                 [LNET_ROUTE_ATTR_GATEWAY]               = {
6961                         .lkp_value                      = "gateway",
6962                         .lkp_data_type                  = NLA_STRING
6963                 },
6964                 [LNET_ROUTE_ATTR_HOP]                   = {
6965                         .lkp_value                      = "hop",
6966                         .lkp_data_type                  = NLA_S32
6967                 },
6968                 [LNET_ROUTE_ATTR_PRIORITY]              = {
6969                         .lkp_value                      = "priority",
6970                         .lkp_data_type                  = NLA_U32
6971                 },
6972                 [LNET_ROUTE_ATTR_HEALTH_SENSITIVITY]    = {
6973                         .lkp_value                      = "health_sensitivity",
6974                         .lkp_data_type                  = NLA_U32
6975                 },
6976                 [LNET_ROUTE_ATTR_STATE] = {
6977                         .lkp_value                      = "state",
6978                         .lkp_data_type                  = NLA_STRING,
6979                 },
6980                 [LNET_ROUTE_ATTR_TYPE]  = {
6981                         .lkp_value                      = "type",
6982                         .lkp_data_type                  = NLA_STRING,
6983                 },
6984         },
6985 };
6986
6987
6988 static int lnet_route_show_dump(struct sk_buff *msg,
6989                                 struct netlink_callback *cb)
6990 {
6991         struct lnet_genl_route_list *rlist = lnet_route_dump_ctx(cb);
6992         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
6993 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
6994         struct netlink_ext_ack *extack = NULL;
6995 #endif
6996         int portid = NETLINK_CB(cb->skb).portid;
6997         int seq = cb->nlh->nlmsg_seq;
6998         int idx = rlist->lgrl_index;
6999         int msg_len = genlmsg_len(gnlh);
7000         int rc = 0;
7001
7002 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
7003         extack = cb->extack;
7004 #endif
7005         if (!rlist->lgrl_count) {
7006                 NL_SET_ERR_MSG(extack, "No routes found");
7007                 GOTO(send_error, rc = msg_len ? -ENOENT : 0);
7008         }
7009
7010         if (!idx) {
7011                 const struct ln_key_list *all[] = {
7012                         &route_props_list, NULL
7013                 };
7014
7015                 rc = lnet_genl_send_scalar_list(msg, portid, seq,
7016                                                 &lnet_family,
7017                                                 NLM_F_CREATE | NLM_F_MULTI,
7018                                                 LNET_CMD_ROUTES, all);
7019                 if (rc < 0) {
7020                         NL_SET_ERR_MSG(extack, "failed to send key table");
7021                         GOTO(send_error, rc);
7022                 }
7023         }
7024
7025         while (idx < rlist->lgrl_count) {
7026                 struct lnet_route_properties *prop;
7027                 void *hdr;
7028
7029                 prop = genradix_ptr(&rlist->lgrl_list, idx++);
7030
7031                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
7032                                   NLM_F_MULTI, LNET_CMD_ROUTES);
7033                 if (!hdr) {
7034                         NL_SET_ERR_MSG(extack, "failed to send values");
7035                         genlmsg_cancel(msg, hdr);
7036                         GOTO(send_error, rc = -EMSGSIZE);
7037                 }
7038
7039                 if (idx == 1)
7040                         nla_put_string(msg, LNET_ROUTE_ATTR_HDR, "");
7041
7042                 nla_put_string(msg, LNET_ROUTE_ATTR_NET,
7043                                libcfs_net2str(prop->lrp_net));
7044                 nla_put_string(msg, LNET_ROUTE_ATTR_GATEWAY,
7045                                libcfs_nidstr(&prop->lrp_gateway));
7046                 if (gnlh->version) {
7047                         nla_put_s32(msg, LNET_ROUTE_ATTR_HOP, prop->lrp_hop);
7048                         nla_put_u32(msg, LNET_ROUTE_ATTR_PRIORITY, prop->lrp_priority);
7049                         nla_put_u32(msg, LNET_ROUTE_ATTR_HEALTH_SENSITIVITY,
7050                                     prop->lrp_sensitivity);
7051
7052                         if (!(cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED)) {
7053                                 nla_put_string(msg, LNET_ROUTE_ATTR_STATE,
7054                                                prop->lrp_flags & LNET_RT_ALIVE ?
7055                                                "up" : "down");
7056                                 nla_put_string(msg, LNET_ROUTE_ATTR_TYPE,
7057                                                prop->lrp_flags & LNET_RT_MULTI_HOP ?
7058                                                "multi-hop" : "single-hop");
7059                         }
7060                 }
7061                 genlmsg_end(msg, hdr);
7062         }
7063         rlist->lgrl_index = idx;
7064 send_error:
7065         return lnet_nl_send_error(cb->skb, portid, seq, rc);
7066 };
7067
7068 #ifndef HAVE_NETLINK_CALLBACK_START
7069 static int lnet_old_route_show_dump(struct sk_buff *msg,
7070                                     struct netlink_callback *cb)
7071 {
7072         if (!cb->args[0]) {
7073                 int rc = lnet_route_show_start(cb);
7074
7075                 if (rc < 0)
7076                         return lnet_nl_send_error(cb->skb,
7077                                                   NETLINK_CB(cb->skb).portid,
7078                                                   cb->nlh->nlmsg_seq,
7079                                                   rc);
7080         }
7081
7082         return lnet_route_show_dump(msg, cb);
7083 }
7084 #endif /* !HAVE_NETLINK_CALLBACK_START */
7085
7086 /** LNet peer handling */
7087 struct lnet_genl_processid_list {
7088         unsigned int                    lgpl_index;
7089         unsigned int                    lgpl_count;
7090         GENRADIX(struct lnet_processid) lgpl_list;
7091 };
7092
7093 static inline struct lnet_genl_processid_list *
7094 lnet_peer_dump_ctx(struct netlink_callback *cb)
7095 {
7096         return (struct lnet_genl_processid_list *)cb->args[0];
7097 }
7098
7099 static int lnet_peer_ni_show_done(struct netlink_callback *cb)
7100 {
7101         struct lnet_genl_processid_list *plist = lnet_peer_dump_ctx(cb);
7102
7103         if (plist) {
7104                 genradix_free(&plist->lgpl_list);
7105                 CFS_FREE_PTR(plist);
7106         }
7107         cb->args[0] = 0;
7108
7109         return 0;
7110 }
7111
7112 /* LNet peer ->start() handler for GET requests */
7113 static int lnet_peer_ni_show_start(struct netlink_callback *cb)
7114 {
7115         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
7116 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
7117         struct netlink_ext_ack *extack = NULL;
7118 #endif
7119         struct lnet_genl_processid_list *plist;
7120         int msg_len = genlmsg_len(gnlh);
7121         int rc = 0;
7122
7123 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
7124         extack = cb->extack;
7125 #endif
7126         mutex_lock(&the_lnet.ln_api_mutex);
7127         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
7128                 NL_SET_ERR_MSG(extack, "Network is down");
7129                 mutex_unlock(&the_lnet.ln_api_mutex);
7130                 return -ENETDOWN;
7131         }
7132
7133         CFS_ALLOC_PTR(plist);
7134         if (!plist) {
7135                 NL_SET_ERR_MSG(extack, "No memory for peer list");
7136                 mutex_unlock(&the_lnet.ln_api_mutex);
7137                 return -ENOMEM;
7138         }
7139
7140         genradix_init(&plist->lgpl_list);
7141         plist->lgpl_count = 0;
7142         plist->lgpl_index = 0;
7143         cb->args[0] = (long)plist;
7144
7145         if (!msg_len) {
7146                 struct lnet_peer_table *ptable;
7147                 int cpt;
7148
7149                 cfs_percpt_for_each(ptable, cpt, the_lnet.ln_peer_tables) {
7150                         struct lnet_peer *lp;
7151
7152                         list_for_each_entry(lp, &ptable->pt_peer_list,
7153                                             lp_peer_list) {
7154                                 struct lnet_processid *lpi;
7155
7156                                 lpi = genradix_ptr_alloc(&plist->lgpl_list,
7157                                                          plist->lgpl_count++,
7158                                                          GFP_KERNEL);
7159                                 if (!lpi) {
7160                                         NL_SET_ERR_MSG(extack,
7161                                                       "failed to allocate NID");
7162                                         GOTO(report_err, rc = -ENOMEM);
7163                                 }
7164
7165                                 lpi->pid = LNET_PID_LUSTRE;
7166                                 lpi->nid = lp->lp_primary_nid;
7167                         }
7168                 }
7169         } else {
7170                 struct nlattr *params = genlmsg_data(gnlh);
7171                 struct nlattr *attr;
7172                 int rem;
7173
7174                 nla_for_each_nested(attr, params, rem) {
7175                         struct nlattr *nid;
7176                         int rem2;
7177
7178                         if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
7179                                 continue;
7180
7181                         nla_for_each_nested(nid, attr, rem2) {
7182                                 char addr[LNET_NIDSTR_SIZE];
7183                                 struct lnet_processid *id;
7184
7185                                 if (nla_type(nid) != LN_SCALAR_ATTR_VALUE ||
7186                                     nla_strcmp(nid, "primary nid") != 0)
7187                                         continue;
7188
7189                                 nid = nla_next(nid, &rem2);
7190                                 if (nla_type(nid) != LN_SCALAR_ATTR_VALUE) {
7191                                         NL_SET_ERR_MSG(extack,
7192                                                        "invalid primary nid param");
7193                                         GOTO(report_err, rc = -EINVAL);
7194                                 }
7195
7196                                 rc = nla_strscpy(addr, nid, sizeof(addr));
7197                                 if (rc < 0) {
7198                                         NL_SET_ERR_MSG(extack,
7199                                                        "failed to get primary nid param");
7200                                         GOTO(report_err, rc);
7201                                 }
7202
7203                                 id = genradix_ptr_alloc(&plist->lgpl_list,
7204                                                         plist->lgpl_count++,
7205                                                         GFP_KERNEL);
7206                                 if (!id) {
7207                                         NL_SET_ERR_MSG(extack, "failed to allocate NID");
7208                                         GOTO(report_err, rc = -ENOMEM);
7209                                 }
7210
7211                                 rc = libcfs_strid(id, strim(addr));
7212                                 if (rc < 0) {
7213                                         NL_SET_ERR_MSG(extack, "invalid NID");
7214                                         GOTO(report_err, rc);
7215                                 }
7216                                 rc = 0;
7217                         }
7218                 }
7219         }
7220 report_err:
7221         mutex_unlock(&the_lnet.ln_api_mutex);
7222
7223         if (rc < 0)
7224                 lnet_peer_ni_show_done(cb);
7225
7226         return rc;
7227 }
7228
7229 static const struct ln_key_list lnet_peer_ni_keys = {
7230         .lkl_maxattr                    = LNET_PEER_NI_ATTR_MAX,
7231         .lkl_list                       = {
7232                 [LNET_PEER_NI_ATTR_HDR]  = {
7233                         .lkp_value              = "peer",
7234                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
7235                         .lkp_data_type          = NLA_NUL_STRING,
7236                 },
7237                 [LNET_PEER_NI_ATTR_PRIMARY_NID] = {
7238                         .lkp_value              = "primary nid",
7239                         .lkp_data_type          = NLA_STRING,
7240                 },
7241                 [LNET_PEER_NI_ATTR_MULTIRAIL]   = {
7242                         .lkp_value              = "Multi-Rail",
7243                         .lkp_data_type          = NLA_FLAG
7244                 },
7245                 [LNET_PEER_NI_ATTR_STATE]       = {
7246                         .lkp_value              = "peer state",
7247                         .lkp_data_type          = NLA_U32
7248                 },
7249                 [LNET_PEER_NI_ATTR_PEER_NI_LIST] = {
7250                         .lkp_value              = "peer ni",
7251                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
7252                         .lkp_data_type          = NLA_NESTED,
7253                 },
7254         },
7255 };
7256
7257 static const struct ln_key_list lnet_peer_ni_list = {
7258         .lkl_maxattr                    = LNET_PEER_NI_LIST_ATTR_MAX,
7259         .lkl_list                       = {
7260                 [LNET_PEER_NI_LIST_ATTR_NID]            = {
7261                         .lkp_value                      = "nid",
7262                         .lkp_data_type                  = NLA_STRING,
7263                 },
7264                 [LNET_PEER_NI_LIST_ATTR_UDSP_INFO]      = {
7265                         .lkp_value                      = "udsp info",
7266                         .lkp_key_format                 = LNKF_MAPPING,
7267                         .lkp_data_type                  = NLA_NESTED,
7268                 },
7269                 [LNET_PEER_NI_LIST_ATTR_STATE]          = {
7270                         .lkp_value                      = "state",
7271                         .lkp_data_type                  = NLA_STRING,
7272                 },
7273                 [LNET_PEER_NI_LIST_ATTR_MAX_TX_CREDITS] = {
7274                         .lkp_value                      = "max_ni_tx_credits",
7275                         .lkp_data_type                  = NLA_U32,
7276                 },
7277                 [LNET_PEER_NI_LIST_ATTR_CUR_TX_CREDITS] = {
7278                         .lkp_value                      = "available_tx_credits",
7279                         .lkp_data_type                  = NLA_U32,
7280                 },
7281                 [LNET_PEER_NI_LIST_ATTR_MIN_TX_CREDITS] = {
7282                         .lkp_value                      = "min_tx_credits",
7283                         .lkp_data_type                  = NLA_U32,
7284                 },
7285                 [LNET_PEER_NI_LIST_ATTR_QUEUE_BUF_COUNT] = {
7286                         .lkp_value                      = "tx_q_num_of_buf",
7287                         .lkp_data_type                  = NLA_U32,
7288                 },
7289                 [LNET_PEER_NI_LIST_ATTR_CUR_RTR_CREDITS] = {
7290                         .lkp_value                      = "available_rtr_credits",
7291                         .lkp_data_type                  = NLA_U32,
7292                 },
7293                 [LNET_PEER_NI_LIST_ATTR_MIN_RTR_CREDITS] = {
7294                         .lkp_value                      = "min_rtr_credits",
7295                         .lkp_data_type                  = NLA_U32,
7296                 },
7297                 [LNET_PEER_NI_LIST_ATTR_REFCOUNT]       = {
7298                         .lkp_value                      = "refcount",
7299                         .lkp_data_type                  = NLA_U32,
7300                 },
7301                 [LNET_PEER_NI_LIST_ATTR_STATS_COUNT]    = {
7302                         .lkp_value                      = "statistics",
7303                         .lkp_key_format                 = LNKF_MAPPING,
7304                         .lkp_data_type                  = NLA_NESTED
7305                 },
7306                 [LNET_PEER_NI_LIST_ATTR_SENT_STATS]     = {
7307                         .lkp_value                      = "sent_stats",
7308                         .lkp_key_format                 = LNKF_MAPPING,
7309                         .lkp_data_type                  = NLA_NESTED
7310                 },
7311                 [LNET_PEER_NI_LIST_ATTR_RECV_STATS]     = {
7312                         .lkp_value                      = "received_stats",
7313                         .lkp_key_format                 = LNKF_MAPPING,
7314                         .lkp_data_type                  = NLA_NESTED
7315                 },
7316                 [LNET_PEER_NI_LIST_ATTR_DROP_STATS]     = {
7317                         .lkp_value                      = "dropped_stats",
7318                         .lkp_key_format                 = LNKF_MAPPING,
7319                         .lkp_data_type                  = NLA_NESTED
7320                 },
7321                 [LNET_PEER_NI_LIST_ATTR_HEALTH_STATS]   = {
7322                         .lkp_value                      = "health stats",
7323                         .lkp_key_format                 = LNKF_MAPPING,
7324                         .lkp_data_type                  = NLA_NESTED
7325                 },
7326         },
7327 };
7328
7329 static const struct ln_key_list lnet_peer_ni_list_stats_count = {
7330         .lkl_maxattr                    = LNET_PEER_NI_LIST_STATS_COUNT_ATTR_MAX,
7331         .lkl_list                       = {
7332                 [LNET_PEER_NI_LIST_STATS_COUNT_ATTR_SEND_COUNT] = {
7333                         .lkp_value                              = "send_count",
7334                         .lkp_data_type                          = NLA_U32,
7335                 },
7336                 [LNET_PEER_NI_LIST_STATS_COUNT_ATTR_RECV_COUNT] = {
7337                         .lkp_value                              = "recv_count",
7338                         .lkp_data_type                          = NLA_U32,
7339                 },
7340                 [LNET_PEER_NI_LIST_STATS_COUNT_ATTR_DROP_COUNT] = {
7341                         .lkp_value                              = "drop_count",
7342                         .lkp_data_type                          = NLA_U32,
7343                 },
7344         },
7345 };
7346
7347 static const struct ln_key_list lnet_peer_ni_list_stats = {
7348         .lkl_maxattr                    = LNET_PEER_NI_LIST_STATS_ATTR_MAX,
7349         .lkl_list                       = {
7350                 [LNET_PEER_NI_LIST_STATS_ATTR_PUT]      = {
7351                         .lkp_value                      = "put",
7352                         .lkp_data_type                  = NLA_U32,
7353                 },
7354                 [LNET_PEER_NI_LIST_STATS_ATTR_GET]      = {
7355                         .lkp_value                      = "get",
7356                         .lkp_data_type                  = NLA_U32,
7357                 },
7358                 [LNET_PEER_NI_LIST_STATS_ATTR_REPLY]    = {
7359                         .lkp_value                      = "reply",
7360                         .lkp_data_type                  = NLA_U32,
7361                 },
7362                 [LNET_PEER_NI_LIST_STATS_ATTR_ACK]      = {
7363                         .lkp_value                      = "ack",
7364                         .lkp_data_type                  = NLA_U32,
7365                 },
7366                 [LNET_PEER_NI_LIST_STATS_ATTR_HELLO]    = {
7367                         .lkp_value                      = "hello",
7368                         .lkp_data_type                  = NLA_U32,
7369                 },
7370         },
7371 };
7372
7373 static const struct ln_key_list lnet_peer_ni_list_health = {
7374         .lkl_maxattr                    = LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_MAX,
7375         .lkl_list                       = {
7376                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_VALUE]     = {
7377                         .lkp_value                      = "health value",
7378                         .lkp_data_type                  = NLA_S32,
7379                 },
7380                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_DROPPED]   = {
7381                         .lkp_value                      = "dropped",
7382                         .lkp_data_type                  = NLA_U32,
7383                 },
7384                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_TIMEOUT]   = {
7385                         .lkp_value                      = "timeout",
7386                         .lkp_data_type                  = NLA_U32,
7387                 },
7388                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_ERROR]     = {
7389                         .lkp_value                      = "error",
7390                         .lkp_data_type                  = NLA_U32,
7391                 },
7392                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_NETWORK_TIMEOUT] = {
7393                         .lkp_value                      = "network timeout",
7394                         .lkp_data_type                  = NLA_U32,
7395                 },
7396                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_PING_COUNT] = {
7397                         .lkp_value                      = "ping_count",
7398                         .lkp_data_type                  = NLA_U32,
7399                 },
7400                 [LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_NEXT_PING] = {
7401                         .lkp_value                      = "next_ping",
7402                         .lkp_data_type                  = NLA_S64,
7403                 },
7404         },
7405 };
7406
7407 static int lnet_peer_ni_show_dump(struct sk_buff *msg,
7408                                   struct netlink_callback *cb)
7409 {
7410         struct lnet_genl_processid_list *plist = lnet_peer_dump_ctx(cb);
7411         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
7412 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
7413         struct netlink_ext_ack *extack = NULL;
7414 #endif
7415         int portid = NETLINK_CB(cb->skb).portid;
7416         int seq = cb->nlh->nlmsg_seq;
7417         int idx = plist->lgpl_index;
7418         int msg_len = genlmsg_len(gnlh);
7419         int rc = 0;
7420
7421 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
7422         extack = cb->extack;
7423 #endif
7424         if (!plist->lgpl_count) {
7425                 NL_SET_ERR_MSG(extack, "No peers found");
7426                 GOTO(send_error, rc = msg_len ? -ENOENT : 0);
7427         }
7428
7429         if (!idx) {
7430                 const struct ln_key_list *all[] = {
7431                         &lnet_peer_ni_keys, &lnet_peer_ni_list,
7432                         &udsp_info_list, &udsp_info_pref_nids_list,
7433                         &udsp_info_pref_nids_list,
7434                         &lnet_peer_ni_list_stats_count,
7435                         &lnet_peer_ni_list_stats, /* send_stats */
7436                         &lnet_peer_ni_list_stats, /* recv_stats */
7437                         &lnet_peer_ni_list_stats, /* drop stats */
7438                         &lnet_peer_ni_list_health,
7439                         NULL
7440                 };
7441
7442                 rc = lnet_genl_send_scalar_list(msg, portid, seq,
7443                                                 &lnet_family,
7444                                                 NLM_F_CREATE | NLM_F_MULTI,
7445                                                 LNET_CMD_PEERS, all);
7446                 if (rc < 0) {
7447                         NL_SET_ERR_MSG(extack, "failed to send key table");
7448                         GOTO(send_error, rc);
7449                 }
7450         }
7451
7452         mutex_lock(&the_lnet.ln_api_mutex);
7453         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
7454                 NL_SET_ERR_MSG(extack, "Network is down");
7455                 GOTO(unlock_api_mutex, rc = -ENETDOWN);
7456         }
7457
7458         while (idx < plist->lgpl_count) {
7459                 struct lnet_processid *id;
7460                 struct lnet_peer_ni *lpni = NULL;
7461                 struct nlattr *nid_list;
7462                 struct lnet_peer *lp;
7463                 int count = 1;
7464                 void *hdr;
7465
7466                 id = genradix_ptr(&plist->lgpl_list, idx++);
7467                 if (nid_is_lo0(&id->nid))
7468                         continue;
7469
7470                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
7471                                   NLM_F_MULTI, LNET_CMD_PEERS);
7472                 if (!hdr) {
7473                         NL_SET_ERR_MSG(extack, "failed to send values");
7474                         genlmsg_cancel(msg, hdr);
7475                         GOTO(unlock_api_mutex, rc = -EMSGSIZE);
7476                 }
7477
7478                 lp = lnet_find_peer(&id->nid);
7479                 if (!lp) {
7480                         NL_SET_ERR_MSG(extack, "cannot find peer");
7481                         GOTO(unlock_api_mutex, rc = -ENOENT);
7482                 }
7483
7484                 if (idx == 1)
7485                         nla_put_string(msg, LNET_PEER_NI_ATTR_HDR, "");
7486
7487                 nla_put_string(msg, LNET_PEER_NI_ATTR_PRIMARY_NID,
7488                                libcfs_nidstr(&lp->lp_primary_nid));
7489                 if (lnet_peer_is_multi_rail(lp))
7490                         nla_put_flag(msg, LNET_PEER_NI_ATTR_MULTIRAIL);
7491
7492                 if (gnlh->version >= 3)
7493                         nla_put_u32(msg, LNET_PEER_NI_ATTR_STATE, lp->lp_state);
7494
7495                 nid_list = nla_nest_start(msg, LNET_PEER_NI_ATTR_PEER_NI_LIST);
7496                 while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL) {
7497                         struct nlattr *peer_nid = nla_nest_start(msg, count++);
7498
7499                         nla_put_string(msg, LNET_PEER_NI_LIST_ATTR_NID,
7500                                        libcfs_nidstr(&lpni->lpni_nid));
7501
7502                         if (gnlh->version >= 4) {
7503                                 rc = lnet_udsp_info_send(msg,
7504                                                          LNET_PEER_NI_LIST_ATTR_UDSP_INFO,
7505                                                          &lpni->lpni_nid, true);
7506                                 if (rc < 0) {
7507                                         lnet_peer_decref_locked(lp);
7508                                         NL_SET_ERR_MSG(extack,
7509                                                        "failed to get UDSP info");
7510                                         GOTO(unlock_api_mutex, rc);
7511                                 }
7512                         }
7513
7514                         if (cb->nlh->nlmsg_flags & NLM_F_DUMP_FILTERED)
7515                                 goto skip_state;
7516
7517                         if (lnet_isrouter(lpni) ||
7518                             lnet_peer_aliveness_enabled(lpni)) {
7519                                 nla_put_string(msg, LNET_PEER_NI_LIST_ATTR_STATE,
7520                                                lnet_is_peer_ni_alive(lpni) ?
7521                                                "up" : "down");
7522                         } else {
7523                                 nla_put_string(msg, LNET_PEER_NI_LIST_ATTR_STATE,
7524                                                "NA");
7525                         }
7526 skip_state:
7527                         if (gnlh->version) {
7528                                 struct lnet_ioctl_element_msg_stats lpni_msg_stats;
7529                                 struct nlattr *send_stats_list, *send_stats;
7530                                 struct nlattr *recv_stats_list, *recv_stats;
7531                                 struct nlattr *drop_stats_list, *drop_stats;
7532                                 struct nlattr *health_list, *health_stats;
7533                                 struct lnet_ioctl_element_stats stats;
7534                                 struct nlattr *stats_attr, *ni_stats;
7535
7536                                 nla_put_u32(msg,
7537                                             LNET_PEER_NI_LIST_ATTR_MAX_TX_CREDITS,
7538                                             lpni->lpni_net ?
7539                                                 lpni->lpni_net->net_tunables.lct_peer_tx_credits : 0);
7540                                 nla_put_u32(msg,
7541                                             LNET_PEER_NI_LIST_ATTR_CUR_TX_CREDITS,
7542                                             lpni->lpni_txcredits);
7543                                 nla_put_u32(msg,
7544                                             LNET_PEER_NI_LIST_ATTR_MIN_TX_CREDITS,
7545                                             lpni->lpni_mintxcredits);
7546                                 nla_put_u32(msg,
7547                                             LNET_PEER_NI_LIST_ATTR_QUEUE_BUF_COUNT,
7548                                             lpni->lpni_txqnob);
7549                                 nla_put_u32(msg,
7550                                             LNET_PEER_NI_LIST_ATTR_CUR_RTR_CREDITS,
7551                                             lpni->lpni_rtrcredits);
7552                                 nla_put_u32(msg,
7553                                             LNET_PEER_NI_LIST_ATTR_MIN_RTR_CREDITS,
7554                                             lpni->lpni_minrtrcredits);
7555                                 nla_put_u32(msg,
7556                                             LNET_PEER_NI_LIST_ATTR_REFCOUNT,
7557                                             kref_read(&lpni->lpni_kref));
7558
7559                                 memset(&stats, 0, sizeof(stats));
7560                                 stats.iel_send_count = lnet_sum_stats(&lpni->lpni_stats,
7561                                                                       LNET_STATS_TYPE_SEND);
7562                                 stats.iel_recv_count = lnet_sum_stats(&lpni->lpni_stats,
7563                                                                       LNET_STATS_TYPE_RECV);
7564                                 stats.iel_drop_count = lnet_sum_stats(&lpni->lpni_stats,
7565                                                                       LNET_STATS_TYPE_DROP);
7566
7567                                 stats_attr = nla_nest_start(msg,
7568                                                             LNET_PEER_NI_LIST_ATTR_STATS_COUNT);
7569                                 ni_stats = nla_nest_start(msg, 0);
7570                                 nla_put_u32(msg,
7571                                             LNET_PEER_NI_LIST_STATS_COUNT_ATTR_SEND_COUNT,
7572                                             stats.iel_send_count);
7573                                 nla_put_u32(msg,
7574                                             LNET_PEER_NI_LIST_STATS_COUNT_ATTR_RECV_COUNT,
7575                                             stats.iel_recv_count);
7576                                 nla_put_u32(msg,
7577                                             LNET_PEER_NI_LIST_STATS_COUNT_ATTR_DROP_COUNT,
7578                                             stats.iel_drop_count);
7579                                 nla_nest_end(msg, ni_stats);
7580                                 nla_nest_end(msg, stats_attr);
7581
7582                                 if (gnlh->version < 2)
7583                                         goto skip_msg_stats;
7584
7585                                 lnet_usr_translate_stats(&lpni_msg_stats, &lpni->lpni_stats);
7586
7587                                 send_stats_list = nla_nest_start(msg,
7588                                                                  LNET_PEER_NI_LIST_ATTR_SENT_STATS);
7589                                 send_stats = nla_nest_start(msg, 0);
7590                                 nla_put_u32(msg,
7591                                             LNET_PEER_NI_LIST_STATS_ATTR_PUT,
7592                                             lpni_msg_stats.im_send_stats.ico_put_count);
7593                                 nla_put_u32(msg,
7594                                             LNET_PEER_NI_LIST_STATS_ATTR_GET,
7595                                             lpni_msg_stats.im_send_stats.ico_get_count);
7596                                 nla_put_u32(msg,
7597                                             LNET_PEER_NI_LIST_STATS_ATTR_REPLY,
7598                                             lpni_msg_stats.im_send_stats.ico_reply_count);
7599                                 nla_put_u32(msg,
7600                                             LNET_PEER_NI_LIST_STATS_ATTR_ACK,
7601                                             lpni_msg_stats.im_send_stats.ico_ack_count);
7602                                 nla_put_u32(msg,
7603                                             LNET_PEER_NI_LIST_STATS_ATTR_HELLO,
7604                                             lpni_msg_stats.im_send_stats.ico_hello_count);
7605                                 nla_nest_end(msg, send_stats);
7606                                 nla_nest_end(msg, send_stats_list);
7607
7608                                 recv_stats_list = nla_nest_start(msg,
7609                                                                  LNET_PEER_NI_LIST_ATTR_RECV_STATS);
7610                                 recv_stats = nla_nest_start(msg, 0);
7611                                 nla_put_u32(msg,
7612                                             LNET_PEER_NI_LIST_STATS_ATTR_PUT,
7613                                             lpni_msg_stats.im_recv_stats.ico_put_count);
7614                                 nla_put_u32(msg,
7615                                             LNET_PEER_NI_LIST_STATS_ATTR_GET,
7616                                             lpni_msg_stats.im_recv_stats.ico_get_count);
7617                                 nla_put_u32(msg,
7618                                             LNET_PEER_NI_LIST_STATS_ATTR_REPLY,
7619                                             lpni_msg_stats.im_recv_stats.ico_reply_count);
7620                                 nla_put_u32(msg,
7621                                             LNET_PEER_NI_LIST_STATS_ATTR_ACK,
7622                                             lpni_msg_stats.im_recv_stats.ico_ack_count);
7623                                 nla_put_u32(msg,
7624                                             LNET_PEER_NI_LIST_STATS_ATTR_HELLO,
7625                                             lpni_msg_stats.im_recv_stats.ico_hello_count);
7626                                 nla_nest_end(msg, recv_stats);
7627                                 nla_nest_end(msg, recv_stats_list);
7628
7629                                 drop_stats_list = nla_nest_start(msg,
7630                                                                  LNET_PEER_NI_LIST_ATTR_DROP_STATS);
7631                                 drop_stats = nla_nest_start(msg, 0);
7632                                 nla_put_u32(msg,
7633                                             LNET_PEER_NI_LIST_STATS_ATTR_PUT,
7634                                             lpni_msg_stats.im_drop_stats.ico_put_count);
7635                                 nla_put_u32(msg,
7636                                             LNET_PEER_NI_LIST_STATS_ATTR_GET,
7637                                             lpni_msg_stats.im_drop_stats.ico_get_count);
7638                                 nla_put_u32(msg,
7639                                             LNET_PEER_NI_LIST_STATS_ATTR_REPLY,
7640                                             lpni_msg_stats.im_drop_stats.ico_reply_count);
7641                                 nla_put_u32(msg,
7642                                             LNET_PEER_NI_LIST_STATS_ATTR_ACK,
7643                                             lpni_msg_stats.im_drop_stats.ico_ack_count);
7644                                 nla_put_u32(msg,
7645                                             LNET_PEER_NI_LIST_STATS_ATTR_HELLO,
7646                                             lpni_msg_stats.im_drop_stats.ico_hello_count);
7647                                 nla_nest_end(msg, drop_stats);
7648                                 nla_nest_end(msg, drop_stats_list);
7649
7650                                 health_list = nla_nest_start(msg,
7651                                                              LNET_PEER_NI_LIST_ATTR_HEALTH_STATS);
7652                                 health_stats = nla_nest_start(msg, 0);
7653                                 nla_put_s32(msg,
7654                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_VALUE,
7655                                             atomic_read(&lpni->lpni_healthv));
7656                                 nla_put_u32(msg,
7657                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_DROPPED,
7658                                             atomic_read(&lpni->lpni_hstats.hlt_remote_dropped));
7659                                 nla_put_u32(msg,
7660                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_TIMEOUT,
7661                                             atomic_read(&lpni->lpni_hstats.hlt_remote_timeout));
7662                                 nla_put_u32(msg,
7663                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_ERROR,
7664                                             atomic_read(&lpni->lpni_hstats.hlt_remote_error));
7665                                 nla_put_u32(msg,
7666                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_NETWORK_TIMEOUT,
7667                                             atomic_read(&lpni->lpni_hstats.hlt_network_timeout));
7668                                 nla_put_u32(msg,
7669                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_PING_COUNT,
7670                                             lpni->lpni_ping_count);
7671                                 nla_put_s64(msg,
7672                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_NEXT_PING,
7673                                             lpni->lpni_next_ping,
7674                                             LNET_PEER_NI_LIST_HEALTH_STATS_ATTR_PAD);
7675                                 nla_nest_end(msg, health_stats);
7676                                 nla_nest_end(msg, health_list);
7677                         }
7678 skip_msg_stats:
7679                         nla_nest_end(msg, peer_nid);
7680                 }
7681                 nla_nest_end(msg, nid_list);
7682
7683                 genlmsg_end(msg, hdr);
7684                 lnet_peer_decref_locked(lp);
7685         }
7686         plist->lgpl_index = idx;
7687 unlock_api_mutex:
7688         mutex_unlock(&the_lnet.ln_api_mutex);
7689 send_error:
7690         return lnet_nl_send_error(cb->skb, portid, seq, rc);
7691 };
7692
7693 #ifndef HAVE_NETLINK_CALLBACK_START
7694 static int lnet_old_peer_ni_show_dump(struct sk_buff *msg,
7695                                       struct netlink_callback *cb)
7696 {
7697         if (!cb->args[0]) {
7698                 int rc = lnet_peer_ni_show_start(cb);
7699
7700                 if (rc < 0)
7701                         return lnet_nl_send_error(cb->skb,
7702                                                   NETLINK_CB(cb->skb).portid,
7703                                                   cb->nlh->nlmsg_seq,
7704                                                   rc);
7705         }
7706
7707         return lnet_peer_ni_show_dump(msg, cb);
7708 }
7709 #endif
7710
7711 static int lnet_route_cmd(struct sk_buff *skb, struct genl_info *info)
7712 {
7713         struct nlmsghdr *nlh = nlmsg_hdr(skb);
7714         struct genlmsghdr *gnlh = nlmsg_data(nlh);
7715         struct nlattr *params = genlmsg_data(gnlh);
7716         int msg_len, rem, rc = 0;
7717         struct nlattr *attr;
7718
7719         mutex_lock(&the_lnet.ln_api_mutex);
7720         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
7721                 GENL_SET_ERR_MSG(info, "Network is down");
7722                 mutex_unlock(&the_lnet.ln_api_mutex);
7723                 return -ENETDOWN;
7724         }
7725
7726         msg_len = genlmsg_len(gnlh);
7727         if (!msg_len) {
7728                 GENL_SET_ERR_MSG(info, "no configuration");
7729                 mutex_unlock(&the_lnet.ln_api_mutex);
7730                 return -ENOMSG;
7731         }
7732
7733         if (!(nla_type(params) & LN_SCALAR_ATTR_LIST)) {
7734                 GENL_SET_ERR_MSG(info, "invalid configuration");
7735                 mutex_unlock(&the_lnet.ln_api_mutex);
7736                 return -EINVAL;
7737         }
7738
7739         nla_for_each_nested(attr, params, rem) {
7740                 u32 net_id = LNET_NET_ANY, hops = LNET_UNDEFINED_HOPS;
7741                 u32 priority = 0, sensitivity = 1;
7742                 struct lnet_nid gw_nid = LNET_ANY_NID;
7743                 struct nlattr *route_prop;
7744                 bool alive = true;
7745                 s64 when = 0;
7746                 int rem2;
7747
7748                 if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
7749                         continue;
7750
7751                 nla_for_each_nested(route_prop, attr, rem2) {
7752                         char tmp[LNET_NIDSTR_SIZE];
7753                         ssize_t len;
7754                         s64 num;
7755
7756                         if (nla_type(route_prop) != LN_SCALAR_ATTR_VALUE)
7757                                 continue;
7758
7759                         if (nla_strcmp(route_prop, "net") == 0) {
7760                                 route_prop = nla_next(route_prop, &rem2);
7761                                 if (nla_type(route_prop) !=
7762                                     LN_SCALAR_ATTR_VALUE) {
7763                                         GENL_SET_ERR_MSG(info,
7764                                                          "net is invalid key");
7765                                         GOTO(report_err, rc = -EINVAL);
7766                                 }
7767
7768                                 len = nla_strscpy(tmp, route_prop, sizeof(tmp));
7769                                 if (len < 0) {
7770                                         GENL_SET_ERR_MSG(info,
7771                                                          "net key string is invalid");
7772                                         GOTO(report_err, rc = len);
7773                                 }
7774
7775                                 net_id = libcfs_str2net(tmp);
7776                                 if (!net_id) {
7777                                         GENL_SET_ERR_MSG(info,
7778                                                          "cannot parse remote net");
7779                                         GOTO(report_err, rc = -ENODEV);
7780                                 }
7781
7782                                 if (LNET_NETTYP(net_id) == LOLND) {
7783                                         GENL_SET_ERR_MSG(info,
7784                                                          "setting @lo not allowed");
7785                                         GOTO(report_err, rc = -EACCES);
7786                                 }
7787
7788                                 if (net_id == LNET_NET_ANY) {
7789                                         GENL_SET_ERR_MSG(info,
7790                                                          "setting LNET_NET_ANY not allowed");
7791                                         GOTO(report_err, rc = -ENXIO);
7792                                 }
7793                         } else if (nla_strcmp(route_prop, "gateway") == 0) {
7794                                 route_prop = nla_next(route_prop, &rem2);
7795                                 if (nla_type(route_prop) !=
7796                                     LN_SCALAR_ATTR_VALUE) {
7797                                         GENL_SET_ERR_MSG(info,
7798                                                          "gateway is invalid key");
7799                                         GOTO(report_err, rc = -EINVAL);
7800                                 }
7801
7802                                 len = nla_strscpy(tmp, route_prop, sizeof(tmp));
7803                                 if (len < 0) {
7804                                         GENL_SET_ERR_MSG(info,
7805                                                          "gateway string is invalid");
7806                                         GOTO(report_err, rc = len);
7807                                 }
7808
7809                                 rc = libcfs_strnid(&gw_nid, strim(tmp));
7810                                 if (rc < 0) {
7811                                         GENL_SET_ERR_MSG(info,
7812                                                          "cannot parse gateway");
7813                                         GOTO(report_err, rc = -ENODEV);
7814                                 }
7815                         } else if (nla_strcmp(route_prop, "state") == 0) {
7816                                 route_prop = nla_next(route_prop, &rem2);
7817                                 if (nla_type(route_prop) !=
7818                                     LN_SCALAR_ATTR_VALUE) {
7819                                         GENL_SET_ERR_MSG(info,
7820                                                          "state is invalid key");
7821                                         GOTO(report_err, rc = -EINVAL);
7822                                 }
7823
7824                                 if (nla_strcmp(route_prop, "down") == 0) {
7825                                         alive = false;
7826                                 } else if (nla_strcmp(route_prop, "up") == 0) {
7827                                         alive = true;
7828                                 } else {
7829                                         GENL_SET_ERR_MSG(info,
7830                                                          "status string bad value");
7831                                         GOTO(report_err, rc = -EINVAL);
7832                                 }
7833                         } else if (nla_strcmp(route_prop, "notify_time") == 0) {
7834                                 route_prop = nla_next(route_prop, &rem2);
7835                                 if (nla_type(route_prop) !=
7836                                     LN_SCALAR_ATTR_INT_VALUE) {
7837                                         GENL_SET_ERR_MSG(info,
7838                                                          "notify_time is invalid key");
7839                                         GOTO(report_err, rc = -EINVAL);
7840                                 }
7841
7842                                 when = nla_get_s64(route_prop);
7843                                 if (ktime_get_real_seconds() < when) {
7844                                         GENL_SET_ERR_MSG(info,
7845                                                          "notify_time is in the future");
7846                                         GOTO(report_err, rc = -EINVAL);
7847                                 }
7848                         } else if (nla_strcmp(route_prop, "hop") == 0) {
7849                                 route_prop = nla_next(route_prop, &rem2);
7850                                 if (nla_type(route_prop) !=
7851                                     LN_SCALAR_ATTR_INT_VALUE) {
7852                                         GENL_SET_ERR_MSG(info,
7853                                                          "hop has invalid key");
7854                                         GOTO(report_err, rc = -EINVAL);
7855                                 }
7856
7857                                 hops = nla_get_s64(route_prop);
7858                                 if ((hops < 1 || hops > 255) && hops != -1) {
7859                                         GENL_SET_ERR_MSG(info,
7860                                                          "invalid hop count must be between 1 and 255");
7861                                         GOTO(report_err, rc = -EINVAL);
7862                                 }
7863                         } else if (nla_strcmp(route_prop, "priority") == 0) {
7864                                 route_prop = nla_next(route_prop, &rem2);
7865                                 if (nla_type(route_prop) !=
7866                                     LN_SCALAR_ATTR_INT_VALUE) {
7867                                         GENL_SET_ERR_MSG(info,
7868                                                          "priority has invalid key");
7869                                         GOTO(report_err, rc = -EINVAL);
7870                                 }
7871
7872                                 num = nla_get_s64(route_prop);
7873                                 if (num < 0) {
7874                                         GENL_SET_ERR_MSG(info,
7875                                                          "invalid priority, must not be negative");
7876                                         GOTO(report_err, rc = -EINVAL);
7877                                 }
7878                                 priority = num;
7879                         } else if (nla_strcmp(route_prop,
7880                                               "health_sensitivity") == 0) {
7881                                 route_prop = nla_next(route_prop, &rem2);
7882                                 if (nla_type(route_prop) !=
7883                                     LN_SCALAR_ATTR_INT_VALUE) {
7884                                         GENL_SET_ERR_MSG(info,
7885                                                          "sensitivity has invalid key");
7886                                         GOTO(report_err, rc = -EINVAL);
7887                                 }
7888
7889                                 num = nla_get_s64(route_prop);
7890                                 if (num < 1) {
7891                                         GENL_SET_ERR_MSG(info,
7892                                                          "invalid health sensitivity, must be 1 or greater");
7893                                         GOTO(report_err, rc = -EINVAL);
7894                                 }
7895                                 sensitivity = num;
7896                         }
7897                 }
7898
7899                 if (net_id == LNET_NET_ANY) {
7900                         GENL_SET_ERR_MSG(info,
7901                                          "missing mandatory parameter: network");
7902                         GOTO(report_err, rc = -ENODEV);
7903                 }
7904
7905                 if (LNET_NID_IS_ANY(&gw_nid)) {
7906                         GENL_SET_ERR_MSG(info,
7907                                          "missing mandatory parameter: gateway");
7908                         GOTO(report_err, rc = -ENODEV);
7909                 }
7910
7911                 if (info->nlhdr->nlmsg_flags & NLM_F_REPLACE) {
7912                         /* Convert the user-supplied real time to monotonic.
7913                          * NB: "when" is always in the past
7914                          */
7915                         when = ktime_get_seconds() -
7916                                 (ktime_get_real_seconds() - when);
7917
7918                         mutex_unlock(&the_lnet.ln_api_mutex);
7919                         rc = lnet_notify(NULL, &gw_nid, alive, false, when);
7920                         mutex_lock(&the_lnet.ln_api_mutex);
7921                         if (rc < 0)
7922                                 GOTO(report_err, rc);
7923                         else if (the_lnet.ln_state != LNET_STATE_RUNNING)
7924                                 GOTO(report_err, rc = -ENETDOWN);
7925                 } else if (info->nlhdr->nlmsg_flags & NLM_F_CREATE) {
7926                         rc = lnet_add_route(net_id, hops, &gw_nid, priority,
7927                                             sensitivity);
7928                         if (rc < 0) {
7929                                 switch (rc) {
7930                                 case -EINVAL:
7931                                         GENL_SET_ERR_MSG(info,
7932                                                          "invalid settings for route creation");
7933                                         break;
7934                                 case -EHOSTUNREACH:
7935                                         GENL_SET_ERR_MSG(info,
7936                                                          "No interface configured on the same net as gateway");
7937                                         break;
7938                                 case -ESHUTDOWN:
7939                                         GENL_SET_ERR_MSG(info,
7940                                                          "Network is down");
7941                                         break;
7942                                 case -EEXIST:
7943                                         GENL_SET_ERR_MSG(info,
7944                                                          "Route already exists or the specified network is local");
7945                                         break;
7946                                 default:
7947                                         GENL_SET_ERR_MSG(info,
7948                                                          "failed to create route");
7949                                         break;
7950                                 }
7951                                 GOTO(report_err, rc);
7952                         }
7953                 } else if (!(info->nlhdr->nlmsg_flags & NLM_F_CREATE)) {
7954                         rc = lnet_del_route(net_id, &gw_nid);
7955                         if (rc < 0) {
7956                                 GENL_SET_ERR_MSG(info,
7957                                                  "failed to delete route");
7958                                 GOTO(report_err, rc);
7959                         }
7960                 }
7961         }
7962 report_err:
7963         mutex_unlock(&the_lnet.ln_api_mutex);
7964
7965         return rc;
7966 }
7967
7968 static inline struct lnet_genl_ping_list *
7969 lnet_ping_dump_ctx(struct netlink_callback *cb)
7970 {
7971         return (struct lnet_genl_ping_list *)cb->args[0];
7972 }
7973
7974 static int lnet_ping_show_done(struct netlink_callback *cb)
7975 {
7976         struct lnet_genl_ping_list *plist = lnet_ping_dump_ctx(cb);
7977
7978         if (plist) {
7979                 genradix_free(&plist->lgpl_failed);
7980                 genradix_free(&plist->lgpl_list);
7981                 LIBCFS_FREE(plist, sizeof(*plist));
7982                 cb->args[0] = 0;
7983         }
7984
7985         return 0;
7986 }
7987
7988 /* LNet ping ->start() handler for GET requests */
7989 static int lnet_ping_show_start(struct netlink_callback *cb)
7990 {
7991         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
7992 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
7993         struct netlink_ext_ack *extack = NULL;
7994 #endif
7995         struct lnet_genl_ping_list *plist;
7996         int msg_len = genlmsg_len(gnlh);
7997         struct nlattr *params, *top;
7998         int rem, rc = 0;
7999
8000 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
8001         extack = cb->extack;
8002 #endif
8003         if (the_lnet.ln_refcount == 0) {
8004                 NL_SET_ERR_MSG(extack, "Network is down");
8005                 return -ENETDOWN;
8006         }
8007
8008         if (!msg_len) {
8009                 NL_SET_ERR_MSG(extack, "Ping needs NID targets");
8010                 return -ENOENT;
8011         }
8012
8013         LIBCFS_ALLOC(plist, sizeof(*plist));
8014         if (!plist) {
8015                 NL_SET_ERR_MSG(extack, "failed to setup ping list");
8016                 return -ENOMEM;
8017         }
8018         genradix_init(&plist->lgpl_list);
8019         plist->lgpl_timeout = cfs_time_seconds(DEFAULT_PEER_TIMEOUT);
8020         plist->lgpl_src_nid = LNET_ANY_NID;
8021         plist->lgpl_index = 0;
8022         plist->lgpl_list_count = 0;
8023         cb->args[0] = (long)plist;
8024
8025         params = genlmsg_data(gnlh);
8026         nla_for_each_attr(top, params, msg_len, rem) {
8027                 struct nlattr *nids;
8028                 int rem2;
8029
8030                 switch (nla_type(top)) {
8031                 case LN_SCALAR_ATTR_VALUE:
8032                         if (nla_strcmp(top, "timeout") == 0) {
8033                                 s64 timeout;
8034
8035                                 top = nla_next(top, &rem);
8036                                 if (nla_type(top) != LN_SCALAR_ATTR_INT_VALUE) {
8037                                         NL_SET_ERR_MSG(extack,
8038                                                        "invalid timeout param");
8039                                         GOTO(report_err, rc = -EINVAL);
8040                                 }
8041
8042                                 /* If timeout is negative then set default of
8043                                  * 3 minutes
8044                                  */
8045                                 timeout = nla_get_s64(top);
8046                                 if (timeout > 0 &&
8047                                     timeout < (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
8048                                         plist->lgpl_timeout =
8049                                                 nsecs_to_jiffies(timeout * NSEC_PER_MSEC);
8050                         } else if (nla_strcmp(top, "source") == 0) {
8051                                 char nidstr[LNET_NIDSTR_SIZE + 1];
8052
8053                                 top = nla_next(top, &rem);
8054                                 if (nla_type(top) != LN_SCALAR_ATTR_VALUE) {
8055                                         NL_SET_ERR_MSG(extack,
8056                                                        "invalid source param");
8057                                         GOTO(report_err, rc = -EINVAL);
8058                                 }
8059
8060                                 rc = nla_strscpy(nidstr, top, sizeof(nidstr));
8061                                 if (rc < 0) {
8062                                         NL_SET_ERR_MSG(extack,
8063                                                        "failed to parse source nid");
8064                                         GOTO(report_err, rc);
8065                                 }
8066
8067                                 rc = libcfs_strnid(&plist->lgpl_src_nid,
8068                                                    strim(nidstr));
8069                                 if (rc < 0) {
8070                                         NL_SET_ERR_MSG(extack,
8071                                                        "invalid source nid");
8072                                         GOTO(report_err, rc);
8073                                 }
8074                                 rc = 0;
8075                         }
8076                         break;
8077                 case LN_SCALAR_ATTR_LIST:
8078                         nla_for_each_nested(nids, top, rem2) {
8079                                 char nid[LNET_NIDSTR_SIZE + 1];
8080                                 struct lnet_processid *id;
8081
8082                                 if (nla_type(nids) != LN_SCALAR_ATTR_VALUE)
8083                                         continue;
8084
8085                                 memset(nid, 0, sizeof(nid));
8086                                 rc = nla_strscpy(nid, nids, sizeof(nid));
8087                                 if (rc < 0) {
8088                                         NL_SET_ERR_MSG(extack,
8089                                                        "failed to get NID");
8090                                         GOTO(report_err, rc);
8091                                 }
8092
8093                                 id = genradix_ptr_alloc(&plist->lgpl_list,
8094                                                         plist->lgpl_list_count++,
8095                                                         GFP_KERNEL);
8096                                 if (!id) {
8097                                         NL_SET_ERR_MSG(extack,
8098                                                        "failed to allocate NID");
8099                                         GOTO(report_err, rc = -ENOMEM);
8100                                 }
8101
8102                                 rc = libcfs_strid(id, strim(nid));
8103                                 if (rc < 0) {
8104                                         NL_SET_ERR_MSG(extack, "cannot parse NID");
8105                                         GOTO(report_err, rc);
8106                                 }
8107                                 rc = 0;
8108                         }
8109                         fallthrough;
8110                 default:
8111                         break;
8112                 }
8113         }
8114 report_err:
8115         if (rc < 0)
8116                 lnet_ping_show_done(cb);
8117
8118         return rc;
8119 }
8120
8121 static const struct ln_key_list ping_err_props_list = {
8122         .lkl_maxattr                    = LNET_ERR_ATTR_MAX,
8123         .lkl_list                       = {
8124                 [LNET_ERR_ATTR_HDR]             = {
8125                         .lkp_value              = "manage",
8126                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8127                         .lkp_data_type          = NLA_NUL_STRING,
8128                 },
8129                 [LNET_ERR_ATTR_TYPE]            = {
8130                         .lkp_value              = "ping",
8131                         .lkp_data_type          = NLA_STRING,
8132                 },
8133                 [LNET_ERR_ATTR_ERRNO]           = {
8134                         .lkp_value              = "errno",
8135                         .lkp_data_type          = NLA_S16,
8136                 },
8137                 [LNET_ERR_ATTR_DESCR]           = {
8138                         .lkp_value              = "descr",
8139                         .lkp_data_type          = NLA_STRING,
8140                 },
8141         },
8142 };
8143
8144 static const struct ln_key_list ping_props_list = {
8145         .lkl_maxattr                    = LNET_PING_ATTR_MAX,
8146         .lkl_list                       = {
8147                 [LNET_PING_ATTR_HDR]            = {
8148                         .lkp_value              = "ping",
8149                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8150                         .lkp_data_type          = NLA_NUL_STRING,
8151                 },
8152                 [LNET_PING_ATTR_PRIMARY_NID]    = {
8153                         .lkp_value              = "primary nid",
8154                         .lkp_data_type          = NLA_STRING
8155                 },
8156                 [LNET_PING_ATTR_ERRNO]          = {
8157                         .lkp_value              = "errno",
8158                         .lkp_data_type          = NLA_S16
8159                 },
8160                 [LNET_PING_ATTR_MULTIRAIL]      = {
8161                         .lkp_value              = "Multi-Rail",
8162                         .lkp_data_type          = NLA_FLAG
8163                 },
8164                 [LNET_PING_ATTR_PEER_NI_LIST]   = {
8165                         .lkp_value              = "peer_ni",
8166                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8167                         .lkp_data_type          = NLA_NESTED
8168                 },
8169         },
8170 };
8171
8172 static const struct ln_key_list ping_peer_ni_list = {
8173         .lkl_maxattr                    = LNET_PING_PEER_NI_ATTR_MAX,
8174         .lkl_list                       = {
8175                 [LNET_PING_PEER_NI_ATTR_NID]    = {
8176                         .lkp_value              = "nid",
8177                         .lkp_data_type          = NLA_STRING
8178                 },
8179         },
8180 };
8181
8182 static int lnet_ping_show_dump(struct sk_buff *msg,
8183                                struct netlink_callback *cb)
8184 {
8185         struct lnet_genl_ping_list *plist = lnet_ping_dump_ctx(cb);
8186 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
8187         struct netlink_ext_ack *extack = NULL;
8188 #endif
8189         int portid = NETLINK_CB(cb->skb).portid;
8190         int seq = cb->nlh->nlmsg_seq;
8191         int idx = plist->lgpl_index;
8192         int rc = 0, i = 0;
8193
8194 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
8195         extack = cb->extack;
8196 #endif
8197         if (!plist->lgpl_index) {
8198                 const struct ln_key_list *all[] = {
8199                         &ping_props_list, &ping_peer_ni_list, NULL
8200                 };
8201
8202                 rc = lnet_genl_send_scalar_list(msg, portid, seq,
8203                                                 &lnet_family,
8204                                                 NLM_F_CREATE | NLM_F_MULTI,
8205                                                 LNET_CMD_PING, all);
8206                 if (rc < 0) {
8207                         NL_SET_ERR_MSG(extack, "failed to send key table");
8208                         GOTO(send_error, rc);
8209                 }
8210
8211                 genradix_init(&plist->lgpl_failed);
8212         }
8213
8214         while (idx < plist->lgpl_list_count) {
8215                 struct lnet_nid primary_nid = LNET_ANY_NID;
8216                 struct lnet_genl_ping_list peers;
8217                 struct lnet_processid *id;
8218                 struct nlattr *nid_list;
8219                 struct lnet_peer *lp;
8220                 bool mr_flag = false;
8221                 unsigned int count;
8222                 void *hdr = NULL;
8223
8224                 id = genradix_ptr(&plist->lgpl_list, idx++);
8225
8226                 rc = lnet_ping(id, &plist->lgpl_src_nid, plist->lgpl_timeout,
8227                                &peers, lnet_interfaces_max);
8228                 if (rc < 0) {
8229                         struct lnet_fail_ping *fail;
8230
8231                         fail = genradix_ptr_alloc(&plist->lgpl_failed,
8232                                                   plist->lgpl_failed_count++,
8233                                                   GFP_KERNEL);
8234                         if (!fail) {
8235                                 NL_SET_ERR_MSG(extack,
8236                                                "failed to allocate failed NID");
8237                                 GOTO(send_error, rc);
8238                         }
8239                         memset(fail->lfp_msg, '\0', sizeof(fail->lfp_msg));
8240                         snprintf(fail->lfp_msg, sizeof(fail->lfp_msg),
8241                                  "failed to ping %s",
8242                                  libcfs_nidstr(&id->nid));
8243                         fail->lfp_id = *id;
8244                         fail->lfp_errno = rc;
8245                         goto cant_reach;
8246                 }
8247
8248                 mutex_lock(&the_lnet.ln_api_mutex);
8249                 lp = lnet_find_peer(&id->nid);
8250                 if (lp) {
8251                         primary_nid = lp->lp_primary_nid;
8252                         mr_flag = lnet_peer_is_multi_rail(lp);
8253                         lnet_peer_decref_locked(lp);
8254                 }
8255                 mutex_unlock(&the_lnet.ln_api_mutex);
8256
8257                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
8258                                   NLM_F_MULTI, LNET_CMD_PING);
8259                 if (!hdr) {
8260                         NL_SET_ERR_MSG(extack, "failed to send values");
8261                         genlmsg_cancel(msg, hdr);
8262                         GOTO(send_error, rc = -EMSGSIZE);
8263                 }
8264
8265                 if (i++ == 0)
8266                         nla_put_string(msg, LNET_PING_ATTR_HDR, "");
8267
8268                 nla_put_string(msg, LNET_PING_ATTR_PRIMARY_NID,
8269                                libcfs_nidstr(&primary_nid));
8270                 if (mr_flag)
8271                         nla_put_flag(msg, LNET_PING_ATTR_MULTIRAIL);
8272
8273                 nid_list = nla_nest_start(msg, LNET_PING_ATTR_PEER_NI_LIST);
8274                 for (count = 0; count < rc; count++) {
8275                         struct lnet_processid *result;
8276                         struct nlattr *nid_attr;
8277                         char *idstr;
8278
8279                         result = genradix_ptr(&peers.lgpl_list, count);
8280                         if (nid_is_lo0(&result->nid))
8281                                 continue;
8282
8283                         nid_attr = nla_nest_start(msg, count + 1);
8284                         if (id->pid == LNET_PID_LUSTRE)
8285                                 idstr = libcfs_nidstr(&result->nid);
8286                         else
8287                                 idstr = libcfs_idstr(result);
8288                         nla_put_string(msg, LNET_PING_PEER_NI_ATTR_NID, idstr);
8289                         nla_nest_end(msg, nid_attr);
8290                 }
8291                 nla_nest_end(msg, nid_list);
8292                 genlmsg_end(msg, hdr);
8293 cant_reach:
8294                 genradix_free(&peers.lgpl_list);
8295         }
8296
8297         if (plist->lgpl_failed_count) {
8298                 int flags = NLM_F_CREATE | NLM_F_REPLACE | NLM_F_MULTI;
8299                 const struct ln_key_list *fail[] = {
8300                         &ping_err_props_list, NULL
8301                 };
8302
8303                 rc = lnet_genl_send_scalar_list(msg, portid, seq, &lnet_family,
8304                                                 flags, LNET_CMD_PING, fail);
8305                 if (rc < 0) {
8306                         NL_SET_ERR_MSG(extack,
8307                                        "failed to send new key table");
8308                         GOTO(send_error, rc);
8309                 }
8310
8311                 for (i = 0; i < plist->lgpl_failed_count; i++) {
8312                         struct lnet_fail_ping *fail;
8313                         void *hdr;
8314
8315                         fail = genradix_ptr(&plist->lgpl_failed, i);
8316
8317                         hdr = genlmsg_put(msg, portid, seq, &lnet_family,
8318                                           NLM_F_MULTI, LNET_CMD_PING);
8319                         if (!hdr) {
8320                                 NL_SET_ERR_MSG(extack,
8321                                                "failed to send failed values");
8322                                 genlmsg_cancel(msg, hdr);
8323                                 GOTO(send_error, rc = -EMSGSIZE);
8324                         }
8325
8326                         if (i == 0)
8327                                 nla_put_string(msg, LNET_ERR_ATTR_HDR, "");
8328
8329                         nla_put_string(msg, LNET_ERR_ATTR_TYPE, "\n");
8330                         nla_put_s16(msg, LNET_ERR_ATTR_ERRNO,
8331                                     fail->lfp_errno);
8332                         nla_put_string(msg, LNET_ERR_ATTR_DESCR,
8333                                        fail->lfp_msg);
8334                         genlmsg_end(msg, hdr);
8335                 }
8336         }
8337         genradix_free(&plist->lgpl_list);
8338         rc = 0; /* don't treat it as an error */
8339
8340         plist->lgpl_index = idx;
8341 send_error:
8342         return lnet_nl_send_error(cb->skb, portid, seq, rc);
8343 }
8344
8345 #ifndef HAVE_NETLINK_CALLBACK_START
8346 static int lnet_old_ping_show_dump(struct sk_buff *msg,
8347                                    struct netlink_callback *cb)
8348 {
8349         if (!cb->args[0]) {
8350                 int rc = lnet_ping_show_start(cb);
8351
8352                 if (rc < 0)
8353                         return lnet_nl_send_error(cb->skb,
8354                                                   NETLINK_CB(cb->skb).portid,
8355                                                   cb->nlh->nlmsg_seq,
8356                                                   rc);
8357         }
8358
8359         return lnet_ping_show_dump(msg, cb);
8360 }
8361 #endif
8362
8363 static const struct ln_key_list discover_err_props_list = {
8364         .lkl_maxattr                    = LNET_ERR_ATTR_MAX,
8365         .lkl_list                       = {
8366                 [LNET_ERR_ATTR_HDR]             = {
8367                         .lkp_value              = "manage",
8368                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8369                         .lkp_data_type          = NLA_NUL_STRING,
8370                 },
8371                 [LNET_ERR_ATTR_TYPE]            = {
8372                         .lkp_value              = "discover",
8373                         .lkp_data_type          = NLA_STRING,
8374                 },
8375                 [LNET_ERR_ATTR_ERRNO]           = {
8376                         .lkp_value              = "errno",
8377                         .lkp_data_type          = NLA_S16,
8378                 },
8379                 [LNET_ERR_ATTR_DESCR]           = {
8380                         .lkp_value              = "descr",
8381                         .lkp_data_type          = NLA_STRING,
8382                 },
8383         },
8384 };
8385
8386 static const struct ln_key_list discover_props_list = {
8387         .lkl_maxattr                    = LNET_PING_ATTR_MAX,
8388         .lkl_list                       = {
8389                 [LNET_PING_ATTR_HDR]            = {
8390                         .lkp_value              = "discover",
8391                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8392                         .lkp_data_type          = NLA_NUL_STRING,
8393                 },
8394                 [LNET_PING_ATTR_PRIMARY_NID]    = {
8395                         .lkp_value              = "primary nid",
8396                         .lkp_data_type          = NLA_STRING
8397                 },
8398                 [LNET_PING_ATTR_ERRNO]          = {
8399                         .lkp_value              = "errno",
8400                         .lkp_data_type          = NLA_S16
8401                 },
8402                 [LNET_PING_ATTR_MULTIRAIL]      = {
8403                         .lkp_value              = "Multi-Rail",
8404                         .lkp_data_type          = NLA_FLAG
8405                 },
8406                 [LNET_PING_ATTR_PEER_NI_LIST]   = {
8407                         .lkp_value              = "peer_ni",
8408                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8409                         .lkp_data_type          = NLA_NESTED
8410                 },
8411         },
8412 };
8413
8414 static int lnet_ping_cmd(struct sk_buff *skb, struct genl_info *info)
8415 {
8416         const struct ln_key_list *all[] = {
8417                 &discover_props_list, &ping_peer_ni_list, NULL
8418         };
8419         struct nlmsghdr *nlh = nlmsg_hdr(skb);
8420         struct genlmsghdr *gnlh = nlmsg_data(nlh);
8421         struct nlattr *params = genlmsg_data(gnlh);
8422         struct lnet_genl_ping_list dlists;
8423         int msg_len, rem, rc = 0, i;
8424         bool clear_hdr = false;
8425         struct sk_buff *reply;
8426         struct nlattr *attr;
8427         void *hdr = NULL;
8428
8429         msg_len = genlmsg_len(gnlh);
8430         if (!msg_len) {
8431                 GENL_SET_ERR_MSG(info, "no configuration");
8432                 return -ENOMSG;
8433         }
8434
8435         if (!(info->nlhdr->nlmsg_flags & NLM_F_CREATE)) {
8436                 GENL_SET_ERR_MSG(info, "only NLM_F_CREATE setting is allowed");
8437                 return -EINVAL;
8438         }
8439
8440         reply = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8441         if (!reply) {
8442                 GENL_SET_ERR_MSG(info,
8443                                  "fail to allocate reply");
8444                 return -ENOMEM;
8445         }
8446
8447         genradix_init(&dlists.lgpl_failed);
8448         dlists.lgpl_failed_count = 0;
8449         genradix_init(&dlists.lgpl_list);
8450         dlists.lgpl_list_count = 0;
8451
8452         rc = lnet_genl_send_scalar_list(reply, info->snd_portid,
8453                                         info->snd_seq, &lnet_family,
8454                                         NLM_F_CREATE | NLM_F_MULTI,
8455                                         LNET_CMD_PING, all);
8456         if (rc < 0) {
8457                 GENL_SET_ERR_MSG(info,
8458                                  "failed to send key table");
8459                 GOTO(report_err, rc);
8460         }
8461
8462         nla_for_each_attr(attr, params, msg_len, rem) {
8463                 struct nlattr *nids;
8464                 int rem2;
8465
8466                 /* We only care about the NID list to discover with */
8467                 if (nla_type(attr) != LN_SCALAR_ATTR_LIST)
8468                         continue;
8469
8470                 nla_for_each_nested(nids, attr, rem2) {
8471                         char nid[LNET_NIDSTR_SIZE + 1];
8472                         struct lnet_processid id;
8473                         struct nlattr *nid_list;
8474                         struct lnet_peer *lp;
8475                         ssize_t len;
8476
8477                         if (nla_type(nids) != LN_SCALAR_ATTR_VALUE)
8478                                 continue;
8479
8480                         memset(nid, 0, sizeof(nid));
8481                         rc = nla_strscpy(nid, nids, sizeof(nid));
8482                         if (rc < 0) {
8483                                 GENL_SET_ERR_MSG(info,
8484                                                  "failed to get NID");
8485                                 GOTO(report_err, rc);
8486                         }
8487
8488                         len = libcfs_strid(&id, strim(nid));
8489                         if (len < 0) {
8490                                 struct lnet_fail_ping *fail;
8491
8492                                 fail = genradix_ptr_alloc(&dlists.lgpl_failed,
8493                                                           dlists.lgpl_failed_count++,
8494                                                           GFP_KERNEL);
8495                                 if (!fail) {
8496                                         GENL_SET_ERR_MSG(info,
8497                                                          "failed to allocate improper NID");
8498                                         GOTO(report_err, rc = -ENOMEM);
8499                                 }
8500                                 memset(fail->lfp_msg, '\0', sizeof(fail->lfp_msg));
8501                                 snprintf(fail->lfp_msg, sizeof(fail->lfp_msg),
8502                                          "cannot parse NID '%s'", strim(nid));
8503                                 fail->lfp_id = id;
8504                                 fail->lfp_errno = len;
8505                                 continue;
8506                         }
8507
8508                         if (LNET_NID_IS_ANY(&id.nid))
8509                                 continue;
8510
8511                         rc = lnet_discover(&id,
8512                                            info->nlhdr->nlmsg_flags & NLM_F_EXCL,
8513                                            &dlists);
8514                         if (rc < 0) {
8515                                 struct lnet_fail_ping *fail;
8516
8517                                 fail = genradix_ptr_alloc(&dlists.lgpl_failed,
8518                                                           dlists.lgpl_failed_count++,
8519                                                           GFP_KERNEL);
8520                                 if (!fail) {
8521                                         GENL_SET_ERR_MSG(info,
8522                                                          "failed to allocate failed NID");
8523                                         GOTO(report_err, rc = -ENOMEM);
8524                                 }
8525                                 memset(fail->lfp_msg, '\0', sizeof(fail->lfp_msg));
8526                                 snprintf(fail->lfp_msg, sizeof(fail->lfp_msg),
8527                                          "failed to discover %s",
8528                                          libcfs_nidstr(&id.nid));
8529                                 fail->lfp_id = id;
8530                                 fail->lfp_errno = rc;
8531                                 continue;
8532                         }
8533
8534                         /* create the genetlink message header */
8535                         hdr = genlmsg_put(reply, info->snd_portid, info->snd_seq,
8536                                           &lnet_family, NLM_F_MULTI, LNET_CMD_PING);
8537                         if (!hdr) {
8538                                 GENL_SET_ERR_MSG(info,
8539                                                  "failed to allocate hdr");
8540                                 GOTO(report_err, rc = -ENOMEM);
8541                         }
8542
8543                         if (!clear_hdr) {
8544                                 nla_put_string(reply, LNET_PING_ATTR_HDR, "");
8545                                 clear_hdr = true;
8546                         }
8547
8548                         lp = lnet_find_peer(&id.nid);
8549                         if (lp) {
8550                                 nla_put_string(reply, LNET_PING_ATTR_PRIMARY_NID,
8551                                                libcfs_nidstr(&lp->lp_primary_nid));
8552                                 if (lnet_peer_is_multi_rail(lp))
8553                                         nla_put_flag(reply, LNET_PING_ATTR_MULTIRAIL);
8554                                 lnet_peer_decref_locked(lp);
8555                         }
8556
8557                         nid_list = nla_nest_start(reply, LNET_PING_ATTR_PEER_NI_LIST);
8558                         for (i = 0; i < dlists.lgpl_list_count; i++) {
8559                                 struct lnet_processid *found;
8560                                 struct nlattr *nid_attr;
8561                                 char *idstr;
8562
8563                                 found = genradix_ptr(&dlists.lgpl_list, i);
8564                                 if (nid_is_lo0(&found->nid))
8565                                         continue;
8566
8567                                 nid_attr = nla_nest_start(reply, i + 1);
8568                                 if (id.pid == LNET_PID_LUSTRE)
8569                                         idstr = libcfs_nidstr(&found->nid);
8570                                 else
8571                                         idstr = libcfs_idstr(found);
8572                                 nla_put_string(reply, LNET_PING_PEER_NI_ATTR_NID, idstr);
8573                                 nla_nest_end(reply, nid_attr);
8574                         }
8575                         nla_nest_end(reply, nid_list);
8576
8577                         genlmsg_end(reply, hdr);
8578                 }
8579         }
8580
8581         if (dlists.lgpl_failed_count) {
8582                 int flags = NLM_F_CREATE | NLM_F_REPLACE | NLM_F_MULTI;
8583                 const struct ln_key_list *fail[] = {
8584                         &discover_err_props_list, NULL
8585                 };
8586
8587                 rc = lnet_genl_send_scalar_list(reply, info->snd_portid,
8588                                                 info->snd_seq, &lnet_family,
8589                                                 flags, LNET_CMD_PING, fail);
8590                 if (rc < 0) {
8591                         GENL_SET_ERR_MSG(info,
8592                                          "failed to send new key table");
8593                         GOTO(report_err, rc);
8594                 }
8595
8596                 for (i = 0; i < dlists.lgpl_failed_count; i++) {
8597                         struct lnet_fail_ping *fail;
8598
8599                         hdr = genlmsg_put(reply, info->snd_portid, info->snd_seq,
8600                                           &lnet_family, NLM_F_MULTI, LNET_CMD_PING);
8601                         if (!hdr) {
8602                                 GENL_SET_ERR_MSG(info,
8603                                                  "failed to send failed values");
8604                                 GOTO(report_err, rc = -ENOMSG);
8605                         }
8606
8607                         fail = genradix_ptr(&dlists.lgpl_failed, i);
8608                         if (i == 0)
8609                                 nla_put_string(reply, LNET_ERR_ATTR_HDR, "");
8610
8611                         nla_put_string(reply, LNET_ERR_ATTR_TYPE, "\n");
8612                         nla_put_s16(reply, LNET_ERR_ATTR_ERRNO,
8613                                     fail->lfp_errno);
8614                         nla_put_string(reply, LNET_ERR_ATTR_DESCR,
8615                                        fail->lfp_msg);
8616                         genlmsg_end(reply, hdr);
8617                 }
8618         }
8619
8620         nlh = nlmsg_put(reply, info->snd_portid, info->snd_seq, NLMSG_DONE, 0,
8621                         NLM_F_MULTI);
8622         if (!nlh) {
8623                 genlmsg_cancel(reply, hdr);
8624                 GENL_SET_ERR_MSG(info,
8625                                  "failed to finish message");
8626                 GOTO(report_err, rc = -EMSGSIZE);
8627         }
8628
8629 report_err:
8630         genradix_free(&dlists.lgpl_failed);
8631         genradix_free(&dlists.lgpl_list);
8632
8633         if (rc < 0) {
8634                 genlmsg_cancel(reply, hdr);
8635                 nlmsg_free(reply);
8636         } else {
8637                 rc = genlmsg_reply(reply, info);
8638         }
8639
8640         return rc;
8641 }
8642
8643 #define lnet_peer_dist_show_done        lnet_peer_ni_show_done
8644
8645 static int lnet_peer_dist_show_start(struct netlink_callback *cb)
8646 {
8647         struct genlmsghdr *gnlh = nlmsg_data(cb->nlh);
8648 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
8649         struct netlink_ext_ack *extack = NULL;
8650 #endif
8651         struct lnet_genl_processid_list *plist;
8652         int msg_len = genlmsg_len(gnlh);
8653         struct nlattr *params, *top;
8654         int rem, rc = 0;
8655
8656 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
8657         extack = cb->extack;
8658 #endif
8659         mutex_lock(&the_lnet.ln_api_mutex);
8660         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
8661                 NL_SET_ERR_MSG(extack, "Network is down");
8662                 mutex_unlock(&the_lnet.ln_api_mutex);
8663                 return -ENETDOWN;
8664         }
8665
8666         msg_len = genlmsg_len(gnlh);
8667         if (!msg_len) {
8668                 NL_SET_ERR_MSG(extack, "Missing NID argument(s)");
8669                 mutex_unlock(&the_lnet.ln_api_mutex);
8670                 return -ENOENT;
8671         }
8672
8673         CFS_ALLOC_PTR(plist);
8674         if (!plist) {
8675                 NL_SET_ERR_MSG(extack, "No memory for peer NID list");
8676                 mutex_unlock(&the_lnet.ln_api_mutex);
8677                 return -ENOMEM;
8678         }
8679
8680         genradix_init(&plist->lgpl_list);
8681         plist->lgpl_count = 0;
8682         plist->lgpl_index = 0;
8683         cb->args[0] = (long)plist;
8684
8685         params = genlmsg_data(gnlh);
8686         nla_for_each_attr(top, params, msg_len, rem) {
8687                 struct nlattr *nids;
8688                 int rem2;
8689
8690                 if (nla_type(top) != LN_SCALAR_ATTR_LIST)
8691                         continue;
8692
8693                 nla_for_each_nested(nids, top, rem2) {
8694                         char nidstr[LNET_NIDSTR_SIZE + 1];
8695                         struct lnet_processid *id;
8696
8697                         if (nla_type(nids) != LN_SCALAR_ATTR_VALUE)
8698                                 continue;
8699
8700                         memset(nidstr, 0, sizeof(nidstr));
8701                         rc = nla_strscpy(nidstr, nids, sizeof(nidstr));
8702                         if (rc < 0) {
8703                                 NL_SET_ERR_MSG(extack,
8704                                                "failed to get NID");
8705                                 GOTO(report_err, rc);
8706                         }
8707
8708                         id = genradix_ptr_alloc(&plist->lgpl_list,
8709                                                 plist->lgpl_count++,
8710                                                 GFP_KERNEL);
8711                         if (!id) {
8712                                 NL_SET_ERR_MSG(extack, "failed to allocate NID");
8713                                 GOTO(report_err, rc = -ENOMEM);
8714                         }
8715
8716                         rc = libcfs_strid(id, strim(nidstr));
8717                         if (rc < 0) {
8718                                 NL_SET_ERR_MSG(extack, "invalid NID");
8719                                 GOTO(report_err, rc);
8720                         }
8721                         rc = 0;
8722                 }
8723         }
8724 report_err:
8725         mutex_unlock(&the_lnet.ln_api_mutex);
8726
8727         if (rc < 0)
8728                 lnet_peer_dist_show_done(cb);
8729
8730         return rc;
8731 }
8732
8733 static const struct ln_key_list peer_dist_props_list = {
8734         .lkl_maxattr                    = LNET_PEER_DIST_ATTR_MAX,
8735         .lkl_list                       = {
8736                 [LNET_PEER_DIST_ATTR_HDR]       = {
8737                         .lkp_value              = "peer",
8738                         .lkp_key_format         = LNKF_SEQUENCE | LNKF_MAPPING,
8739                         .lkp_data_type          = NLA_NUL_STRING,
8740                 },
8741                 [LNET_PEER_DIST_ATTR_NID]       = {
8742                         .lkp_value              = "nid",
8743                         .lkp_data_type          = NLA_STRING
8744                 },
8745                 [LNET_PEER_DIST_ATTR_DIST]      = {
8746                         .lkp_value              = "distance",
8747                         .lkp_data_type          = NLA_U32
8748                 },
8749                 [LNET_PEER_DIST_ATTR_ORDER]     = {
8750                         .lkp_value              = "order",
8751                         .lkp_data_type          = NLA_U32
8752                 },
8753         },
8754 };
8755
8756 static int lnet_peer_dist_show_dump(struct sk_buff *msg,
8757                                     struct netlink_callback *cb)
8758 {
8759         struct lnet_genl_processid_list *plist = lnet_peer_dump_ctx(cb);
8760 #ifdef HAVE_NL_PARSE_WITH_EXT_ACK
8761         struct netlink_ext_ack *extack = NULL;
8762 #endif
8763         int portid = NETLINK_CB(cb->skb).portid;
8764         int seq = cb->nlh->nlmsg_seq;
8765         int idx = plist->lgpl_index;
8766         int rc = 0;
8767
8768 #ifdef HAVE_NL_DUMP_WITH_EXT_ACK
8769         extack = cb->extack;
8770 #endif
8771         if (!idx) {
8772                 const struct ln_key_list *all[] = {
8773                         &peer_dist_props_list, NULL
8774                 };
8775
8776                 rc = lnet_genl_send_scalar_list(msg, portid, seq,
8777                                                 &lnet_family,
8778                                                 NLM_F_CREATE | NLM_F_MULTI,
8779                                                 LNET_CMD_PEER_DIST, all);
8780                 if (rc < 0) {
8781                         NL_SET_ERR_MSG(extack, "failed to send key table");
8782                         GOTO(send_error, rc);
8783                 }
8784         }
8785
8786         while (idx < plist->lgpl_count) {
8787                 struct lnet_processid *id;
8788                 void *hdr;
8789                 u32 order;
8790                 int dist;
8791
8792                 id = genradix_ptr(&plist->lgpl_list, idx++);
8793                 if (nid_is_lo0(&id->nid))
8794                         continue;
8795
8796                 dist = LNetDist(&id->nid, &id->nid, &order);
8797                 if (dist < 0) {
8798                         if (dist == -EHOSTUNREACH)
8799                                 continue;
8800
8801                         rc = dist;
8802                         return rc;
8803                 }
8804
8805                 hdr = genlmsg_put(msg, portid, seq, &lnet_family,
8806                                   NLM_F_MULTI, LNET_CMD_PEER_DIST);
8807                 if (!hdr) {
8808                         NL_SET_ERR_MSG(extack, "failed to send values");
8809                         genlmsg_cancel(msg, hdr);
8810                         GOTO(send_error, rc = -EMSGSIZE);
8811                 }
8812
8813                 if (idx == 1)
8814                         nla_put_string(msg, LNET_PEER_DIST_ATTR_HDR, "");
8815
8816                 nla_put_string(msg, LNET_PEER_DIST_ATTR_NID,
8817                                libcfs_nidstr(&id->nid));
8818                 nla_put_u32(msg, LNET_PEER_DIST_ATTR_DIST, dist);
8819                 nla_put_u32(msg, LNET_PEER_DIST_ATTR_ORDER, order);
8820
8821                 genlmsg_end(msg, hdr);
8822         }
8823
8824         plist->lgpl_index = idx;
8825 send_error:
8826         return lnet_nl_send_error(cb->skb, portid, seq, rc);
8827 }
8828
8829 #ifndef HAVE_NETLINK_CALLBACK_START
8830 static int lnet_old_peer_dist_show_dump(struct sk_buff *msg,
8831                                         struct netlink_callback *cb)
8832 {
8833         if (!cb->args[0]) {
8834                 int rc = lnet_peer_dist_show_start(cb);
8835
8836                 if (rc < 0)
8837                         return lnet_nl_send_error(cb->skb,
8838                                                   NETLINK_CB(cb->skb).portid,
8839                                                   cb->nlh->nlmsg_seq,
8840                                                   rc);
8841         }
8842
8843         return lnet_peer_dist_show_dump(msg, cb);
8844 }
8845 #endif
8846
8847 static const struct genl_multicast_group lnet_mcast_grps[] = {
8848         { .name =       "ip2net",       },
8849         { .name =       "net",          },
8850         { .name =       "peer",         },
8851         { .name =       "route",        },
8852         { .name =       "ping",         },
8853         { .name =       "discover",     },
8854         { .name =       "cpt-of-nid",   },
8855 };
8856
8857 static const struct genl_ops lnet_genl_ops[] = {
8858         {
8859                 .cmd            = LNET_CMD_CONFIGURE,
8860                 .flags          = GENL_ADMIN_PERM,
8861                 .doit           = lnet_net_conf_cmd,
8862         },
8863         {
8864                 .cmd            = LNET_CMD_NETS,
8865                 .flags          = GENL_ADMIN_PERM,
8866 #ifdef HAVE_NETLINK_CALLBACK_START
8867                 .start          = lnet_net_show_start,
8868                 .dumpit         = lnet_net_show_dump,
8869 #else
8870                 .dumpit         = lnet_old_net_show_dump,
8871 #endif
8872                 .done           = lnet_net_show_done,
8873                 .doit           = lnet_net_cmd,
8874         },
8875         {
8876                 .cmd            = LNET_CMD_PEERS,
8877                 .flags          = GENL_ADMIN_PERM,
8878 #ifdef HAVE_NETLINK_CALLBACK_START
8879                 .start          = lnet_peer_ni_show_start,
8880                 .dumpit         = lnet_peer_ni_show_dump,
8881 #else
8882                 .dumpit         = lnet_old_peer_ni_show_dump,
8883 #endif
8884                 .done           = lnet_peer_ni_show_done,
8885                 .doit           = lnet_peer_ni_cmd,
8886         },
8887         {
8888                 .cmd            = LNET_CMD_ROUTES,
8889                 .flags          = GENL_ADMIN_PERM,
8890 #ifdef HAVE_NETLINK_CALLBACK_START
8891                 .start          = lnet_route_show_start,
8892                 .dumpit         = lnet_route_show_dump,
8893 #else
8894                 .dumpit         = lnet_old_route_show_dump,
8895 #endif
8896                 .done           = lnet_route_show_done,
8897                 .doit           = lnet_route_cmd,
8898         },
8899         {
8900                 .cmd            = LNET_CMD_PING,
8901                 .flags          = GENL_ADMIN_PERM,
8902 #ifdef HAVE_NETLINK_CALLBACK_START
8903                 .start          = lnet_ping_show_start,
8904                 .dumpit         = lnet_ping_show_dump,
8905 #else
8906                 .dumpit         = lnet_old_ping_show_dump,
8907 #endif
8908                 .done           = lnet_ping_show_done,
8909                 .doit           = lnet_ping_cmd,
8910         },
8911         {
8912                 .cmd            = LNET_CMD_CPT_OF_NID,
8913 #ifdef HAVE_NETLINK_CALLBACK_START
8914                 .start          = lnet_cpt_of_nid_show_start,
8915                 .dumpit         = lnet_cpt_of_nid_show_dump,
8916 #else
8917                 .dumpit         = lnet_old_cpt_of_nid_show_dump,
8918 #endif
8919                 .done           = lnet_cpt_of_nid_show_done,
8920         },
8921         {
8922                 .cmd            = LNET_CMD_PEER_DIST,
8923 #ifdef HAVE_NETLINK_CALLBACK_START
8924                 .start          = lnet_peer_dist_show_start,
8925                 .dumpit         = lnet_peer_dist_show_dump,
8926 #else
8927                 .dumpit         = lnet_old_peer_dist_show_dump,
8928 #endif
8929                 .done           = lnet_peer_dist_show_done,
8930         },
8931 };
8932
8933 static struct genl_family lnet_family = {
8934         .name           = LNET_GENL_NAME,
8935         .version        = LNET_GENL_VERSION,
8936         .module         = THIS_MODULE,
8937         .parallel_ops   = true,
8938         .netnsok        = true,
8939         .ops            = lnet_genl_ops,
8940         .n_ops          = ARRAY_SIZE(lnet_genl_ops),
8941         .mcgrps         = lnet_mcast_grps,
8942         .n_mcgrps       = ARRAY_SIZE(lnet_mcast_grps),
8943 #ifdef GENL_FAMILY_HAS_RESV_START_OP
8944         .resv_start_op  = __LNET_CMD_MAX_PLUS_ONE,
8945 #endif
8946 };
8947
8948 void LNetDebugPeer(struct lnet_processid *id)
8949 {
8950         lnet_debug_peer(&id->nid);
8951 }
8952 EXPORT_SYMBOL(LNetDebugPeer);
8953
8954 /**
8955  * Determine if the specified peer \a nid is on the local node.
8956  *
8957  * \param nid   peer nid to check
8958  *
8959  * \retval true         If peer NID is on the local node.
8960  * \retval false        If peer NID is not on the local node.
8961  */
8962 bool LNetIsPeerLocal(struct lnet_nid *nid)
8963 {
8964         struct lnet_net *net;
8965         struct lnet_ni *ni;
8966         int cpt;
8967
8968         cpt = lnet_net_lock_current();
8969         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
8970                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
8971                         if (nid_same(&ni->ni_nid, nid)) {
8972                                 lnet_net_unlock(cpt);
8973                                 return true;
8974                         }
8975                 }
8976         }
8977         lnet_net_unlock(cpt);
8978
8979         return false;
8980 }
8981 EXPORT_SYMBOL(LNetIsPeerLocal);
8982
8983 /**
8984  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
8985  * Note that all interfaces share a same PID, as requested by LNetNIInit().
8986  *
8987  * @index       Index of the interface to look up.
8988  * @id          On successful return, this location will hold the
8989  *              struct lnet_process_id ID of the interface.
8990  * @large_nids  Report large NIDs if this is true.
8991  *
8992  * RETURN       0 If an interface exists at \a index.
8993  *              -ENOENT If no interface has been found.
8994  */
8995 int
8996 LNetGetId(unsigned int index, struct lnet_processid *id, bool large_nids)
8997 {
8998         struct lnet_ni   *ni;
8999         struct lnet_net  *net;
9000         int               cpt;
9001         int               rc = -ENOENT;
9002
9003         LASSERT(the_lnet.ln_refcount > 0);
9004
9005         cpt = lnet_net_lock_current();
9006
9007         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
9008                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
9009                         if (!large_nids && !nid_is_nid4(&ni->ni_nid))
9010                                 continue;
9011
9012                         if (index-- != 0)
9013                                 continue;
9014
9015                         id->nid = ni->ni_nid;
9016                         id->pid = the_lnet.ln_pid;
9017                         rc = 0;
9018                         break;
9019                 }
9020         }
9021
9022         lnet_net_unlock(cpt);
9023         return rc;
9024 }
9025 EXPORT_SYMBOL(LNetGetId);
9026
9027 struct ping_data {
9028         int rc;
9029         int replied;
9030         int pd_unlinked;
9031         struct lnet_handle_md mdh;
9032         struct completion completion;
9033 };
9034
9035 static void
9036 lnet_ping_event_handler(struct lnet_event *event)
9037 {
9038         struct ping_data *pd = event->md_user_ptr;
9039
9040         CDEBUG(D_NET, "ping event (%d %d)%s\n",
9041                event->type, event->status,
9042                event->unlinked ? " unlinked" : "");
9043
9044         if (event->status) {
9045                 if (!pd->rc)
9046                         pd->rc = event->status;
9047         } else if (event->type == LNET_EVENT_REPLY) {
9048                 pd->replied = 1;
9049                 pd->rc = event->mlength;
9050         }
9051
9052         if (event->unlinked)
9053                 pd->pd_unlinked = 1;
9054
9055         if (event->unlinked ||
9056             (event->type == LNET_EVENT_SEND && event->status))
9057                 complete(&pd->completion);
9058 }
9059
9060 /* Max buffer we allow to be sent. Larger values will cause IB failures */
9061 #define LNET_PING_BUFFER_MAX    3960
9062
9063 static int lnet_ping(struct lnet_processid *id, struct lnet_nid *src_nid,
9064                      signed long timeout, struct lnet_genl_ping_list *plist,
9065                      int n_ids)
9066 {
9067         int id_bytes = sizeof(struct lnet_ni_status); /* For 0@lo */
9068         struct lnet_md md = { NULL };
9069         struct ping_data pd = { 0 };
9070         struct lnet_ping_buffer *pbuf;
9071         struct lnet_processid pid;
9072         struct lnet_ping_iter pi;
9073         int i = 0;
9074         u32 *st;
9075         int nob;
9076         int rc;
9077         int rc2;
9078
9079         genradix_init(&plist->lgpl_list);
9080
9081         /* n_ids limit is arbitrary */
9082         if (n_ids <= 0 || LNET_NID_IS_ANY(&id->nid))
9083                 return -EINVAL;
9084
9085         /* if the user buffer has more space than the lnet_interfaces_max
9086          * then only fill it up to lnet_interfaces_max
9087          */
9088         if (n_ids > lnet_interfaces_max)
9089                 n_ids = lnet_interfaces_max;
9090
9091         if (id->pid == LNET_PID_ANY)
9092                 id->pid = LNET_PID_LUSTRE;
9093
9094         /* Allocate maximum possible NID size */
9095         id_bytes += lnet_ping_sts_size(&LNET_ANY_NID) * n_ids;
9096         if (id_bytes > LNET_PING_BUFFER_MAX)
9097                 id_bytes = LNET_PING_BUFFER_MAX;
9098
9099         pbuf = lnet_ping_buffer_alloc(id_bytes, GFP_NOFS);
9100         if (!pbuf)
9101                 return -ENOMEM;
9102
9103         /* initialize md content */
9104         md.start     = &pbuf->pb_info;
9105         md.length    = id_bytes;
9106         md.threshold = 2; /* GET/REPLY */
9107         md.max_size  = 0;
9108         md.options   = LNET_MD_TRUNCATE;
9109         md.user_ptr  = &pd;
9110         md.handler   = lnet_ping_event_handler;
9111
9112         init_completion(&pd.completion);
9113
9114         rc = LNetMDBind(&md, LNET_UNLINK, &pd.mdh);
9115         if (rc != 0) {
9116                 CERROR("Can't bind MD: %d\n", rc);
9117                 goto fail_ping_buffer_decref;
9118         }
9119
9120         rc = LNetGet(src_nid, pd.mdh, id, LNET_RESERVED_PORTAL,
9121                      LNET_PROTO_PING_MATCHBITS, 0, false);
9122         if (rc != 0) {
9123                 /* Don't CERROR; this could be deliberate! */
9124                 rc2 = LNetMDUnlink(pd.mdh);
9125                 LASSERT(rc2 == 0);
9126
9127                 /* NB must wait for the UNLINK event below... */
9128         }
9129
9130         /* Ensure completion in finite time... */
9131         wait_for_completion_timeout(&pd.completion, timeout);
9132         if (!pd.pd_unlinked) {
9133                 LNetMDUnlink(pd.mdh);
9134                 wait_for_completion(&pd.completion);
9135         }
9136
9137         if (!pd.replied) {
9138                 rc = pd.rc ?: -EIO;
9139                 goto fail_ping_buffer_decref;
9140         }
9141
9142         nob = pd.rc;
9143         LASSERT(nob >= 0 && nob <= id_bytes);
9144
9145         rc = -EPROTO;           /* if I can't parse... */
9146
9147         if (nob < LNET_PING_INFO_HDR_SIZE) {
9148                 CERROR("%s: ping info too short %d\n",
9149                        libcfs_idstr(id), nob);
9150                 goto fail_ping_buffer_decref;
9151         }
9152
9153         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
9154                 lnet_swap_pinginfo(pbuf);
9155         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
9156                 CERROR("%s: Unexpected magic %08x\n",
9157                        libcfs_idstr(id), pbuf->pb_info.pi_magic);
9158                 goto fail_ping_buffer_decref;
9159         }
9160
9161         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
9162                 CERROR("%s: ping w/o NI status: 0x%x\n",
9163                        libcfs_idstr(id), pbuf->pb_info.pi_features);
9164                 goto fail_ping_buffer_decref;
9165         }
9166
9167         /* Test if smaller than lnet_pinginfo with just one pi_ni status info.
9168          * That one might contain size when large nids are used.
9169          */
9170         if (nob < offsetof(struct lnet_ping_info, pi_ni[1])) {
9171                 CERROR("%s: Short reply %d(%lu min)\n",
9172                        libcfs_idstr(id), nob,
9173                        offsetof(struct lnet_ping_info, pi_ni[1]));
9174                 goto fail_ping_buffer_decref;
9175         }
9176
9177         if (ping_info_count_entries(pbuf) < n_ids) {
9178                 n_ids = ping_info_count_entries(pbuf);
9179                 id_bytes = lnet_ping_info_size(&pbuf->pb_info);
9180         }
9181
9182         if (nob < id_bytes) {
9183                 CERROR("%s: Short reply %d(%d expected)\n",
9184                        libcfs_idstr(id), nob, id_bytes);
9185                 goto fail_ping_buffer_decref;
9186         }
9187
9188         for (st = ping_iter_first(&pi, pbuf, &pid.nid);
9189              st;
9190              st = ping_iter_next(&pi, &pid.nid)) {
9191                 id = genradix_ptr_alloc(&plist->lgpl_list, i++, GFP_KERNEL);
9192                 if (!id) {
9193                         rc = -ENOMEM;
9194                         goto fail_ping_buffer_decref;
9195                 }
9196
9197                 id->pid = pbuf->pb_info.pi_pid;
9198                 id->nid = pid.nid;
9199         }
9200         rc = i;
9201 fail_ping_buffer_decref:
9202         lnet_ping_buffer_decref(pbuf);
9203         return rc;
9204 }
9205
9206 static int
9207 lnet_discover(struct lnet_processid *pid, u32 force,
9208               struct lnet_genl_ping_list *dlist)
9209 {
9210         struct lnet_peer_ni *lpni;
9211         struct lnet_peer_ni *p;
9212         struct lnet_peer *lp;
9213         int cpt;
9214         int rc;
9215
9216         if (LNET_NID_IS_ANY(&pid->nid))
9217                 return -EINVAL;
9218
9219         if (pid->pid == LNET_PID_ANY)
9220                 pid->pid = LNET_PID_LUSTRE;
9221
9222         cpt = lnet_net_lock_current();
9223         lpni = lnet_peerni_by_nid_locked(&pid->nid, NULL, cpt);
9224         if (IS_ERR(lpni)) {
9225                 rc = PTR_ERR(lpni);
9226                 goto out;
9227         }
9228
9229         /*
9230          * Clearing the NIDS_UPTODATE flag ensures the peer will
9231          * be discovered, provided discovery has not been disabled.
9232          */
9233         lp = lpni->lpni_peer_net->lpn_peer;
9234         spin_lock(&lp->lp_lock);
9235         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
9236         /* If the force flag is set, force a PING and PUSH as well. */
9237         if (force)
9238                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
9239         spin_unlock(&lp->lp_lock);
9240         rc = lnet_discover_peer_locked(lpni, cpt, true);
9241         if (rc)
9242                 goto out_decref;
9243
9244         /* The lpni (or lp) for this NID may have changed and our ref is
9245          * the only thing keeping the old one around. Release the ref
9246          * and lookup the lpni again
9247          */
9248         lnet_peer_ni_decref_locked(lpni);
9249         lpni = lnet_peer_ni_find_locked(&pid->nid);
9250         if (!lpni) {
9251                 rc = -ENOENT;
9252                 goto out;
9253         }
9254         lp = lpni->lpni_peer_net->lpn_peer;
9255
9256         dlist->lgpl_list_count = 0;
9257         p = NULL;
9258         while ((p = lnet_get_next_peer_ni_locked(lp, NULL, p)) != NULL) {
9259                 struct lnet_processid *id;
9260
9261                 id = genradix_ptr_alloc(&dlist->lgpl_list,
9262                                         dlist->lgpl_list_count++, GFP_ATOMIC);
9263                 if (!id) {
9264                         rc = -ENOMEM;
9265                         goto out_decref;
9266                 }
9267                 id->pid = pid->pid;
9268                 id->nid = p->lpni_nid;
9269         }
9270         rc = dlist->lgpl_list_count;
9271
9272 out_decref:
9273         lnet_peer_ni_decref_locked(lpni);
9274 out:
9275         lnet_net_unlock(cpt);
9276
9277         return rc;
9278 }
9279
9280 /**
9281  * Retrieve peer discovery status.
9282  *
9283  * \retval 1 if lnet_peer_discovery_disabled is 0
9284  * \retval 0 if lnet_peer_discovery_disabled is 1
9285  */
9286 int
9287 LNetGetPeerDiscoveryStatus(void)
9288 {
9289         return !lnet_peer_discovery_disabled;
9290 }
9291 EXPORT_SYMBOL(LNetGetPeerDiscoveryStatus);