lnet/lnet/api-ni.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LNET
  34 #include <linux/log2.h>
  35 #include <linux/ktime.h>
  36 #include <linux/moduleparam.h>
  37
  38 #include <lnet/lib-lnet.h>
  39
  40 #define D_LNI D_CONSOLE
  41
  42 struct lnet the_lnet;           /* THE state of the network */
  43 EXPORT_SYMBOL(the_lnet);
  44
  45 static char *ip2nets = "";
  46 module_param(ip2nets, charp, 0444);
  47 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  48
  49 static char *networks = "";
  50 module_param(networks, charp, 0444);
  51 MODULE_PARM_DESC(networks, "local networks");
  52
  53 static char *routes = "";
  54 module_param(routes, charp, 0444);
  55 MODULE_PARM_DESC(routes, "routes to non-local networks");
  56
  57 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  58 module_param(rnet_htable_size, int, 0444);
  59 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  60
  61 static int use_tcp_bonding = false;
  62 module_param(use_tcp_bonding, int, 0444);
  63 MODULE_PARM_DESC(use_tcp_bonding,
  64                  "Set to 1 to use socklnd bonding. 0 to use Multi-Rail");
  65
  66 unsigned int lnet_numa_range = 0;
  67 module_param(lnet_numa_range, uint, 0444);
  68 MODULE_PARM_DESC(lnet_numa_range,
  69                 "NUMA range to consider during Multi-Rail selection");
  70
  71 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
  72 static int intf_max_set(const char *val, struct kernel_param *kp);
  73 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
  74                   &lnet_interfaces_max, S_IRUGO|S_IWUSR);
  75 MODULE_PARM_DESC(lnet_interfaces_max,
  76                 "Maximum number of interfaces in a node.");
  77
  78 /*
  79  * This sequence number keeps track of how many times DLC was used to
  80  * update the local NIs. It is incremented when a NI is added or
  81  * removed and checked when sending a message to determine if there is
  82  * a need to re-run the selection algorithm. See lnet_select_pathway()
  83  * for more details on its usage.
  84  */
  85 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
  86
  87 static int lnet_ping(struct lnet_process_id id, signed long timeout,
  88                      struct lnet_process_id __user *ids, int n_ids);
  89
  90 static int
  91 intf_max_set(const char *val, struct kernel_param *kp)
  92 {
  93         int value, rc;
  94
  95         rc = kstrtoint(val, 0, &value);
  96         if (rc) {
  97                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
  98                 return rc;
  99         }
 100
 101         if (value < LNET_INTERFACES_MIN) {
 102                 CWARN("max interfaces provided are too small, setting to %d\n",
 103                       LNET_INTERFACES_MIN);
 104                 value = LNET_INTERFACES_MIN;
 105         }
 106
 107         *(int *)kp->arg = value;
 108
 109         return 0;
 110 }
 111
 112 static char *
 113 lnet_get_routes(void)
 114 {
 115         return routes;
 116 }
 117
 118 static char *
 119 lnet_get_networks(void)
 120 {
 121         char   *nets;
 122         int     rc;
 123
 124         if (*networks != 0 && *ip2nets != 0) {
 125                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 126                                    "'ip2nets' but not both at once\n");
 127                 return NULL;
 128         }
 129
 130         if (*ip2nets != 0) {
 131                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 132                 return (rc == 0) ? nets : NULL;
 133         }
 134
 135         if (*networks != 0)
 136                 return networks;
 137
 138         return "tcp";
 139 }
 140
 141 static void
 142 lnet_init_locks(void)
 143 {
 144         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 145         init_waitqueue_head(&the_lnet.ln_eq_waitq);
 146         init_waitqueue_head(&the_lnet.ln_rc_waitq);
 147         mutex_init(&the_lnet.ln_lnd_mutex);
 148         mutex_init(&the_lnet.ln_api_mutex);
 149 }
 150
 151 static void
 152 lnet_fini_locks(void)
 153 {
 154 }
 155
 156 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 157 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 158                                             *  MDs kmem_cache */
 159
 160 static int
 161 lnet_descriptor_setup(void)
 162 {
 163         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 164          * allocated in <size-xxx> kmem_cache).
 165          */
 166         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 167                                             0, 0, NULL);
 168         if (!lnet_mes_cachep)
 169                 return -ENOMEM;
 170
 171         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 172                                                   LNET_SMALL_MD_SIZE, 0, 0,
 173                                                   NULL);
 174         if (!lnet_small_mds_cachep)
 175                 return -ENOMEM;
 176
 177         return 0;
 178 }
 179
 180 static void
 181 lnet_descriptor_cleanup(void)
 182 {
 183
 184         if (lnet_small_mds_cachep) {
 185                 kmem_cache_destroy(lnet_small_mds_cachep);
 186                 lnet_small_mds_cachep = NULL;
 187         }
 188
 189         if (lnet_mes_cachep) {
 190                 kmem_cache_destroy(lnet_mes_cachep);
 191                 lnet_mes_cachep = NULL;
 192         }
 193 }
 194
 195 static int
 196 lnet_create_remote_nets_table(void)
 197 {
 198         int               i;
 199         struct list_head *hash;
 200
 201         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 202         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 203         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 204         if (hash == NULL) {
 205                 CERROR("Failed to create remote nets hash table\n");
 206                 return -ENOMEM;
 207         }
 208
 209         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 210                 INIT_LIST_HEAD(&hash[i]);
 211         the_lnet.ln_remote_nets_hash = hash;
 212         return 0;
 213 }
 214
 215 static void
 216 lnet_destroy_remote_nets_table(void)
 217 {
 218         int i;
 219
 220         if (the_lnet.ln_remote_nets_hash == NULL)
 221                 return;
 222
 223         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 224                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 225
 226         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 227                     LNET_REMOTE_NETS_HASH_SIZE *
 228                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 229         the_lnet.ln_remote_nets_hash = NULL;
 230 }
 231
 232 static void
 233 lnet_destroy_locks(void)
 234 {
 235         if (the_lnet.ln_res_lock != NULL) {
 236                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 237                 the_lnet.ln_res_lock = NULL;
 238         }
 239
 240         if (the_lnet.ln_net_lock != NULL) {
 241                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 242                 the_lnet.ln_net_lock = NULL;
 243         }
 244
 245         lnet_fini_locks();
 246 }
 247
 248 static int
 249 lnet_create_locks(void)
 250 {
 251         lnet_init_locks();
 252
 253         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 254         if (the_lnet.ln_res_lock == NULL)
 255                 goto failed;
 256
 257         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 258         if (the_lnet.ln_net_lock == NULL)
 259                 goto failed;
 260
 261         return 0;
 262
 263  failed:
 264         lnet_destroy_locks();
 265         return -ENOMEM;
 266 }
 267
 268 static void lnet_assert_wire_constants(void)
 269 {
 270         /* Wire protocol assertions generated by 'wirecheck'
 271          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 272          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 273          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 274
 275         /* Constants... */
 276         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 277         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 278         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 279         CLASSERT(LNET_MSG_ACK == 0);
 280         CLASSERT(LNET_MSG_PUT == 1);
 281         CLASSERT(LNET_MSG_GET == 2);
 282         CLASSERT(LNET_MSG_REPLY == 3);
 283         CLASSERT(LNET_MSG_HELLO == 4);
 284
 285         /* Checks for struct lnet_handle_wire */
 286         CLASSERT((int)sizeof(struct lnet_handle_wire) == 16);
 287         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_interface_cookie) == 0);
 288         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) == 8);
 289         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_object_cookie) == 8);
 290         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) == 8);
 291
 292         /* Checks for struct struct lnet_magicversion */
 293         CLASSERT((int)sizeof(struct lnet_magicversion) == 8);
 294         CLASSERT((int)offsetof(struct lnet_magicversion, magic) == 0);
 295         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->magic) == 4);
 296         CLASSERT((int)offsetof(struct lnet_magicversion, version_major) == 4);
 297         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_major) == 2);
 298         CLASSERT((int)offsetof(struct lnet_magicversion, version_minor) == 6);
 299         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_minor) == 2);
 300
 301         /* Checks for struct struct lnet_hdr */
 302         CLASSERT((int)sizeof(struct lnet_hdr) == 72);
 303         CLASSERT((int)offsetof(struct lnet_hdr, dest_nid) == 0);
 304         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_nid) == 8);
 305         CLASSERT((int)offsetof(struct lnet_hdr, src_nid) == 8);
 306         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_nid) == 8);
 307         CLASSERT((int)offsetof(struct lnet_hdr, dest_pid) == 16);
 308         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_pid) == 4);
 309         CLASSERT((int)offsetof(struct lnet_hdr, src_pid) == 20);
 310         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_pid) == 4);
 311         CLASSERT((int)offsetof(struct lnet_hdr, type) == 24);
 312         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->type) == 4);
 313         CLASSERT((int)offsetof(struct lnet_hdr, payload_length) == 28);
 314         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->payload_length) == 4);
 315         CLASSERT((int)offsetof(struct lnet_hdr, msg) == 32);
 316         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg) == 40);
 317
 318         /* Ack */
 319         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) == 32);
 320         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) == 16);
 321         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.match_bits) == 48);
 322         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) == 8);
 323         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.mlength) == 56);
 324         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) == 4);
 325
 326         /* Put */
 327         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) == 32);
 328         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) == 16);
 329         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.match_bits) == 48);
 330         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) == 8);
 331         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.hdr_data) == 56);
 332         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) == 8);
 333         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ptl_index) == 64);
 334         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) == 4);
 335         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.offset) == 68);
 336         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) == 4);
 337
 338         /* Get */
 339         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.return_wmd) == 32);
 340         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) == 16);
 341         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.match_bits) == 48);
 342         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) == 8);
 343         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.ptl_index) == 56);
 344         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) == 4);
 345         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.src_offset) == 60);
 346         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) == 4);
 347         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.sink_length) == 64);
 348         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) == 4);
 349
 350         /* Reply */
 351         CLASSERT((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) == 32);
 352         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) == 16);
 353
 354         /* Hello */
 355         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.incarnation) == 32);
 356         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) == 8);
 357         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.type) == 40);
 358         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) == 4);
 359 }
 360
 361 static struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 362 {
 363         struct lnet_lnd *lnd;
 364         struct list_head *tmp;
 365
 366         /* holding lnd mutex */
 367         list_for_each(tmp, &the_lnet.ln_lnds) {
 368                 lnd = list_entry(tmp, struct lnet_lnd, lnd_list);
 369
 370                 if (lnd->lnd_type == type)
 371                         return lnd;
 372         }
 373         return NULL;
 374 }
 375
 376 void
 377 lnet_register_lnd(struct lnet_lnd *lnd)
 378 {
 379         mutex_lock(&the_lnet.ln_lnd_mutex);
 380
 381         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 382         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 383
 384         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 385         lnd->lnd_refcount = 0;
 386
 387         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 388
 389         mutex_unlock(&the_lnet.ln_lnd_mutex);
 390 }
 391 EXPORT_SYMBOL(lnet_register_lnd);
 392
 393 void
 394 lnet_unregister_lnd(struct lnet_lnd *lnd)
 395 {
 396         mutex_lock(&the_lnet.ln_lnd_mutex);
 397
 398         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 399         LASSERT(lnd->lnd_refcount == 0);
 400
 401         list_del(&lnd->lnd_list);
 402         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 403
 404         mutex_unlock(&the_lnet.ln_lnd_mutex);
 405 }
 406 EXPORT_SYMBOL(lnet_unregister_lnd);
 407
 408 void
 409 lnet_counters_get(struct lnet_counters *counters)
 410 {
 411         struct lnet_counters *ctr;
 412         int             i;
 413
 414         memset(counters, 0, sizeof(*counters));
 415
 416         lnet_net_lock(LNET_LOCK_EX);
 417
 418         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 419                 counters->msgs_max     += ctr->msgs_max;
 420                 counters->msgs_alloc   += ctr->msgs_alloc;
 421                 counters->errors       += ctr->errors;
 422                 counters->send_count   += ctr->send_count;
 423                 counters->recv_count   += ctr->recv_count;
 424                 counters->route_count  += ctr->route_count;
 425                 counters->drop_count   += ctr->drop_count;
 426                 counters->send_length  += ctr->send_length;
 427                 counters->recv_length  += ctr->recv_length;
 428                 counters->route_length += ctr->route_length;
 429                 counters->drop_length  += ctr->drop_length;
 430
 431         }
 432         lnet_net_unlock(LNET_LOCK_EX);
 433 }
 434 EXPORT_SYMBOL(lnet_counters_get);
 435
 436 void
 437 lnet_counters_reset(void)
 438 {
 439         struct lnet_counters *counters;
 440         int             i;
 441
 442         lnet_net_lock(LNET_LOCK_EX);
 443
 444         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 445                 memset(counters, 0, sizeof(struct lnet_counters));
 446
 447         lnet_net_unlock(LNET_LOCK_EX);
 448 }
 449
 450 static char *
 451 lnet_res_type2str(int type)
 452 {
 453         switch (type) {
 454         default:
 455                 LBUG();
 456         case LNET_COOKIE_TYPE_MD:
 457                 return "MD";
 458         case LNET_COOKIE_TYPE_ME:
 459                 return "ME";
 460         case LNET_COOKIE_TYPE_EQ:
 461                 return "EQ";
 462         }
 463 }
 464
 465 static void
 466 lnet_res_container_cleanup(struct lnet_res_container *rec)
 467 {
 468         int     count = 0;
 469
 470         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 471                 return;
 472
 473         while (!list_empty(&rec->rec_active)) {
 474                 struct list_head *e = rec->rec_active.next;
 475
 476                 list_del_init(e);
 477                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 478                         lnet_eq_free(list_entry(e, struct lnet_eq, eq_list));
 479
 480                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 481                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
 482
 483                 } else { /* NB: Active MEs should be attached on portals */
 484                         LBUG();
 485                 }
 486                 count++;
 487         }
 488
 489         if (count > 0) {
 490                 /* Found alive MD/ME/EQ, user really should unlink/free
 491                  * all of them before finalize LNet, but if someone didn't,
 492                  * we have to recycle garbage for him */
 493                 CERROR("%d active elements on exit of %s container\n",
 494                        count, lnet_res_type2str(rec->rec_type));
 495         }
 496
 497         if (rec->rec_lh_hash != NULL) {
 498                 LIBCFS_FREE(rec->rec_lh_hash,
 499                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 500                 rec->rec_lh_hash = NULL;
 501         }
 502
 503         rec->rec_type = 0; /* mark it as finalized */
 504 }
 505
 506 static int
 507 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 508 {
 509         int     rc = 0;
 510         int     i;
 511
 512         LASSERT(rec->rec_type == 0);
 513
 514         rec->rec_type = type;
 515         INIT_LIST_HEAD(&rec->rec_active);
 516
 517         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 518
 519         /* Arbitrary choice of hash table size */
 520         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 521                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 522         if (rec->rec_lh_hash == NULL) {
 523                 rc = -ENOMEM;
 524                 goto out;
 525         }
 526
 527         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 528                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 529
 530         return 0;
 531
 532 out:
 533         CERROR("Failed to setup %s resource container\n",
 534                lnet_res_type2str(type));
 535         lnet_res_container_cleanup(rec);
 536         return rc;
 537 }
 538
 539 static void
 540 lnet_res_containers_destroy(struct lnet_res_container **recs)
 541 {
 542         struct lnet_res_container       *rec;
 543         int                             i;
 544
 545         cfs_percpt_for_each(rec, i, recs)
 546                 lnet_res_container_cleanup(rec);
 547
 548         cfs_percpt_free(recs);
 549 }
 550
 551 static struct lnet_res_container **
 552 lnet_res_containers_create(int type)
 553 {
 554         struct lnet_res_container       **recs;
 555         struct lnet_res_container       *rec;
 556         int                             rc;
 557         int                             i;
 558
 559         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 560         if (recs == NULL) {
 561                 CERROR("Failed to allocate %s resource containers\n",
 562                        lnet_res_type2str(type));
 563                 return NULL;
 564         }
 565
 566         cfs_percpt_for_each(rec, i, recs) {
 567                 rc = lnet_res_container_setup(rec, i, type);
 568                 if (rc != 0) {
 569                         lnet_res_containers_destroy(recs);
 570                         return NULL;
 571                 }
 572         }
 573
 574         return recs;
 575 }
 576
 577 struct lnet_libhandle *
 578 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 579 {
 580         /* ALWAYS called with lnet_res_lock held */
 581         struct list_head        *head;
 582         struct lnet_libhandle   *lh;
 583         unsigned int            hash;
 584
 585         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 586                 return NULL;
 587
 588         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 589         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 590
 591         list_for_each_entry(lh, head, lh_hash_chain) {
 592                 if (lh->lh_cookie == cookie)
 593                         return lh;
 594         }
 595
 596         return NULL;
 597 }
 598
 599 void
 600 lnet_res_lh_initialize(struct lnet_res_container *rec,
 601                        struct lnet_libhandle *lh)
 602 {
 603         /* ALWAYS called with lnet_res_lock held */
 604         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 605         unsigned int    hash;
 606
 607         lh->lh_cookie = rec->rec_lh_cookie;
 608         rec->rec_lh_cookie += 1 << ibits;
 609
 610         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 611
 612         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 613 }
 614
 615 static int lnet_unprepare(void);
 616
 617 static int
 618 lnet_prepare(lnet_pid_t requested_pid)
 619 {
 620         /* Prepare to bring up the network */
 621         struct lnet_res_container **recs;
 622         int                       rc = 0;
 623
 624         if (requested_pid == LNET_PID_ANY) {
 625                 /* Don't instantiate LNET just for me */
 626                 return -ENETDOWN;
 627         }
 628
 629         LASSERT(the_lnet.ln_refcount == 0);
 630
 631         the_lnet.ln_routing = 0;
 632
 633         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
 634         the_lnet.ln_pid = requested_pid;
 635
 636         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 637         INIT_LIST_HEAD(&the_lnet.ln_peers);
 638         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
 639         INIT_LIST_HEAD(&the_lnet.ln_nets);
 640         INIT_LIST_HEAD(&the_lnet.ln_routers);
 641         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
 642         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
 643
 644         rc = lnet_descriptor_setup();
 645         if (rc != 0)
 646                 goto failed;
 647
 648         rc = lnet_create_remote_nets_table();
 649         if (rc != 0)
 650                 goto failed;
 651
 652         /*
 653          * NB the interface cookie in wire handles guards against delayed
 654          * replies and ACKs appearing valid after reboot.
 655          */
 656         the_lnet.ln_interface_cookie = ktime_get_real_ns();
 657
 658         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 659                                                 sizeof(struct lnet_counters));
 660         if (the_lnet.ln_counters == NULL) {
 661                 CERROR("Failed to allocate counters for LNet\n");
 662                 rc = -ENOMEM;
 663                 goto failed;
 664         }
 665
 666         rc = lnet_peer_tables_create();
 667         if (rc != 0)
 668                 goto failed;
 669
 670         rc = lnet_msg_containers_create();
 671         if (rc != 0)
 672                 goto failed;
 673
 674         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 675                                       LNET_COOKIE_TYPE_EQ);
 676         if (rc != 0)
 677                 goto failed;
 678
 679         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
 680         if (recs == NULL) {
 681                 rc = -ENOMEM;
 682                 goto failed;
 683         }
 684
 685         the_lnet.ln_me_containers = recs;
 686
 687         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
 688         if (recs == NULL) {
 689                 rc = -ENOMEM;
 690                 goto failed;
 691         }
 692
 693         the_lnet.ln_md_containers = recs;
 694
 695         rc = lnet_portals_create();
 696         if (rc != 0) {
 697                 CERROR("Failed to create portals for LNet: %d\n", rc);
 698                 goto failed;
 699         }
 700
 701         return 0;
 702
 703  failed:
 704         lnet_unprepare();
 705         return rc;
 706 }
 707
 708 static int
 709 lnet_unprepare (void)
 710 {
 711         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 712          * have shut down already, so it is safe to unlink and free all
 713          * descriptors, even those that appear committed to a network op (eg MD
 714          * with non-zero pending count) */
 715
 716         lnet_fail_nid(LNET_NID_ANY, 0);
 717
 718         LASSERT(the_lnet.ln_refcount == 0);
 719         LASSERT(list_empty(&the_lnet.ln_test_peers));
 720         LASSERT(list_empty(&the_lnet.ln_nets));
 721
 722         lnet_portals_destroy();
 723
 724         if (the_lnet.ln_md_containers != NULL) {
 725                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 726                 the_lnet.ln_md_containers = NULL;
 727         }
 728
 729         if (the_lnet.ln_me_containers != NULL) {
 730                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 731                 the_lnet.ln_me_containers = NULL;
 732         }
 733
 734         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 735
 736         lnet_msg_containers_destroy();
 737         lnet_peer_uninit();
 738         lnet_rtrpools_free(0);
 739
 740         if (the_lnet.ln_counters != NULL) {
 741                 cfs_percpt_free(the_lnet.ln_counters);
 742                 the_lnet.ln_counters = NULL;
 743         }
 744         lnet_destroy_remote_nets_table();
 745         lnet_descriptor_cleanup();
 746
 747         return 0;
 748 }
 749
 750 struct lnet_ni  *
 751 lnet_net2ni_locked(__u32 net_id, int cpt)
 752 {
 753         struct lnet_ni   *ni;
 754         struct lnet_net  *net;
 755
 756         LASSERT(cpt != LNET_LOCK_EX);
 757
 758         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 759                 if (net->net_id == net_id) {
 760                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
 761                                         ni_netlist);
 762                         return ni;
 763                 }
 764         }
 765
 766         return NULL;
 767 }
 768
 769 struct lnet_ni *
 770 lnet_net2ni_addref(__u32 net)
 771 {
 772         struct lnet_ni *ni;
 773
 774         lnet_net_lock(0);
 775         ni = lnet_net2ni_locked(net, 0);
 776         if (ni)
 777                 lnet_ni_addref_locked(ni, 0);
 778         lnet_net_unlock(0);
 779
 780         return ni;
 781 }
 782 EXPORT_SYMBOL(lnet_net2ni_addref);
 783
 784 struct lnet_net *
 785 lnet_get_net_locked(__u32 net_id)
 786 {
 787         struct lnet_net  *net;
 788
 789         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 790                 if (net->net_id == net_id)
 791                         return net;
 792         }
 793
 794         return NULL;
 795 }
 796
 797 unsigned int
 798 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 799 {
 800         __u64           key = nid;
 801         unsigned int    val;
 802
 803         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 804
 805         if (number == 1)
 806                 return 0;
 807
 808         val = hash_long(key, LNET_CPT_BITS);
 809         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 810         if (val < number)
 811                 return val;
 812
 813         return (unsigned int)(key + val + (val >> 1)) % number;
 814 }
 815
 816 int
 817 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
 818 {
 819         struct lnet_net *net;
 820
 821         /* must called with hold of lnet_net_lock */
 822         if (LNET_CPT_NUMBER == 1)
 823                 return 0; /* the only one */
 824
 825         /*
 826          * If NI is provided then use the CPT identified in the NI cpt
 827          * list if one exists. If one doesn't exist, then that NI is
 828          * associated with all CPTs and it follows that the net it belongs
 829          * to is implicitly associated with all CPTs, so just hash the nid
 830          * and return that.
 831          */
 832         if (ni != NULL) {
 833                 if (ni->ni_cpts != NULL)
 834                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
 835                                                              ni->ni_ncpts)];
 836                 else
 837                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 838         }
 839
 840         /* no NI provided so look at the net */
 841         net = lnet_get_net_locked(LNET_NIDNET(nid));
 842
 843         if (net != NULL && net->net_cpts != NULL) {
 844                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
 845         }
 846
 847         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 848 }
 849
 850 int
 851 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
 852 {
 853         int     cpt;
 854         int     cpt2;
 855
 856         if (LNET_CPT_NUMBER == 1)
 857                 return 0; /* the only one */
 858
 859         cpt = lnet_net_lock_current();
 860
 861         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
 862
 863         lnet_net_unlock(cpt);
 864
 865         return cpt2;
 866 }
 867 EXPORT_SYMBOL(lnet_cpt_of_nid);
 868
 869 int
 870 lnet_islocalnet(__u32 net_id)
 871 {
 872         struct lnet_net *net;
 873         int             cpt;
 874         bool            local;
 875
 876         cpt = lnet_net_lock_current();
 877
 878         net = lnet_get_net_locked(net_id);
 879
 880         local = net != NULL;
 881
 882         lnet_net_unlock(cpt);
 883
 884         return local;
 885 }
 886
 887 bool
 888 lnet_is_ni_healthy_locked(struct lnet_ni *ni)
 889 {
 890         if (ni->ni_state == LNET_NI_STATE_ACTIVE ||
 891             ni->ni_state == LNET_NI_STATE_DEGRADED)
 892                 return true;
 893
 894         return false;
 895 }
 896
 897 struct lnet_ni  *
 898 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 899 {
 900         struct lnet_net  *net;
 901         struct lnet_ni   *ni;
 902
 903         LASSERT(cpt != LNET_LOCK_EX);
 904
 905         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 906                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
 907                         if (ni->ni_nid == nid)
 908                                 return ni;
 909                 }
 910         }
 911
 912         return NULL;
 913 }
 914
 915 struct lnet_ni *
 916 lnet_nid2ni_addref(lnet_nid_t nid)
 917 {
 918         struct lnet_ni *ni;
 919
 920         lnet_net_lock(0);
 921         ni = lnet_nid2ni_locked(nid, 0);
 922         if (ni)
 923                 lnet_ni_addref_locked(ni, 0);
 924         lnet_net_unlock(0);
 925
 926         return ni;
 927 }
 928 EXPORT_SYMBOL(lnet_nid2ni_addref);
 929
 930 int
 931 lnet_islocalnid(lnet_nid_t nid)
 932 {
 933         struct lnet_ni  *ni;
 934         int             cpt;
 935
 936         cpt = lnet_net_lock_current();
 937         ni = lnet_nid2ni_locked(nid, cpt);
 938         lnet_net_unlock(cpt);
 939
 940         return ni != NULL;
 941 }
 942
 943 int
 944 lnet_count_acceptor_nets(void)
 945 {
 946         /* Return the # of NIs that need the acceptor. */
 947         int              count = 0;
 948         struct lnet_net  *net;
 949         int              cpt;
 950
 951         cpt = lnet_net_lock_current();
 952         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 953                 /* all socklnd type networks should have the acceptor
 954                  * thread started */
 955                 if (net->net_lnd->lnd_accept != NULL)
 956                         count++;
 957         }
 958
 959         lnet_net_unlock(cpt);
 960
 961         return count;
 962 }
 963
 964 struct lnet_ping_buffer *
 965 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
 966 {
 967         struct lnet_ping_buffer *pbuf;
 968
 969         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
 970         if (pbuf) {
 971                 pbuf->pb_nnis = nnis;
 972                 atomic_set(&pbuf->pb_refcnt, 1);
 973         }
 974
 975         return pbuf;
 976 }
 977
 978 void
 979 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
 980 {
 981         LASSERT(lnet_ping_buffer_numref(pbuf) == 0);
 982         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
 983 }
 984
 985 static struct lnet_ping_buffer *
 986 lnet_ping_target_create(int nnis)
 987 {
 988         struct lnet_ping_buffer *pbuf;
 989
 990         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
 991         if (pbuf == NULL) {
 992                 CERROR("Can't allocate ping source [%d]\n", nnis);
 993                 return NULL;
 994         }
 995
 996         pbuf->pb_info.pi_nnis = nnis;
 997         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
 998         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
 999         pbuf->pb_info.pi_features = LNET_PING_FEAT_NI_STATUS;
1000
1001         return pbuf;
1002 }
1003
1004 static inline int
1005 lnet_get_net_ni_count_locked(struct lnet_net *net)
1006 {
1007         struct lnet_ni  *ni;
1008         int             count = 0;
1009
1010         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1011                 count++;
1012
1013         return count;
1014 }
1015
1016 static inline int
1017 lnet_get_net_ni_count_pre(struct lnet_net *net)
1018 {
1019         struct lnet_ni  *ni;
1020         int             count = 0;
1021
1022         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1023                 count++;
1024
1025         return count;
1026 }
1027
1028 static inline int
1029 lnet_get_ni_count(void)
1030 {
1031         struct lnet_ni  *ni;
1032         struct lnet_net *net;
1033         int             count = 0;
1034
1035         lnet_net_lock(0);
1036
1037         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1038                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1039                         count++;
1040         }
1041
1042         lnet_net_unlock(0);
1043
1044         return count;
1045 }
1046
1047 int
1048 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1049 {
1050         if (!pinfo)
1051                 return -EINVAL;
1052         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1053                 return -EPROTO;
1054         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1055                 return -EPROTO;
1056         /* Loopback is guaranteed to be present */
1057         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1058                 return -ERANGE;
1059         if (LNET_NETTYP(LNET_NIDNET(LNET_PING_INFO_LONI(pinfo))) != LOLND)
1060                 return -EPROTO;
1061         return 0;
1062 }
1063
1064 static void
1065 lnet_ping_target_destroy(void)
1066 {
1067         struct lnet_net *net;
1068         struct lnet_ni  *ni;
1069
1070         lnet_net_lock(LNET_LOCK_EX);
1071
1072         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1073                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1074                         lnet_ni_lock(ni);
1075                         ni->ni_status = NULL;
1076                         lnet_ni_unlock(ni);
1077                 }
1078         }
1079
1080         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1081         the_lnet.ln_ping_target = NULL;
1082
1083         lnet_net_unlock(LNET_LOCK_EX);
1084 }
1085
1086 static void
1087 lnet_ping_target_event_handler(struct lnet_event *event)
1088 {
1089         struct lnet_ping_buffer *pbuf = event->md.user_ptr;
1090
1091         if (event->unlinked)
1092                 lnet_ping_buffer_decref(pbuf);
1093 }
1094
1095 static int
1096 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1097                        struct lnet_handle_md *ping_mdh,
1098                        int ni_count, bool set_eq)
1099 {
1100         struct lnet_process_id id = {
1101                 .nid = LNET_NID_ANY,
1102                 .pid = LNET_PID_ANY
1103         };
1104         struct lnet_handle_me me_handle;
1105         struct lnet_md md = { NULL };
1106         int rc, rc2;
1107
1108         if (set_eq) {
1109                 rc = LNetEQAlloc(0, lnet_ping_target_event_handler,
1110                                  &the_lnet.ln_ping_target_eq);
1111                 if (rc != 0) {
1112                         CERROR("Can't allocate ping buffer EQ: %d\n", rc);
1113                         return rc;
1114                 }
1115         }
1116
1117         *ppbuf = lnet_ping_target_create(ni_count);
1118         if (*ppbuf == NULL) {
1119                 rc = -ENOMEM;
1120                 goto fail_free_eq;
1121         }
1122
1123         /* Ping target ME/MD */
1124         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1125                           LNET_PROTO_PING_MATCHBITS, 0,
1126                           LNET_UNLINK, LNET_INS_AFTER,
1127                           &me_handle);
1128         if (rc != 0) {
1129                 CERROR("Can't create ping target ME: %d\n", rc);
1130                 goto fail_decref_ping_buffer;
1131         }
1132
1133         /* initialize md content */
1134         md.start     = &(*ppbuf)->pb_info;
1135         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1136         md.threshold = LNET_MD_THRESH_INF;
1137         md.max_size  = 0;
1138         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1139                        LNET_MD_MANAGE_REMOTE;
1140         md.eq_handle = the_lnet.ln_ping_target_eq;
1141         md.user_ptr  = *ppbuf;
1142
1143         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, ping_mdh);
1144         if (rc != 0) {
1145                 CERROR("Can't attach ping target MD: %d\n", rc);
1146                 goto fail_unlink_ping_me;
1147         }
1148         lnet_ping_buffer_addref(*ppbuf);
1149
1150         return 0;
1151
1152 fail_unlink_ping_me:
1153         rc2 = LNetMEUnlink(me_handle);
1154         LASSERT(rc2 == 0);
1155 fail_decref_ping_buffer:
1156         LASSERT(lnet_ping_buffer_numref(*ppbuf) == 1);
1157         lnet_ping_buffer_decref(*ppbuf);
1158         *ppbuf = NULL;
1159 fail_free_eq:
1160         if (set_eq) {
1161                 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1162                 LASSERT(rc2 == 0);
1163         }
1164         return rc;
1165 }
1166
1167 static void
1168 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1169                     struct lnet_handle_md *ping_mdh)
1170 {
1171         sigset_t        blocked = cfs_block_allsigs();
1172
1173         LNetMDUnlink(*ping_mdh);
1174         LNetInvalidateMDHandle(ping_mdh);
1175
1176         /* NB the MD could be busy; this just starts the unlink */
1177         while (lnet_ping_buffer_numref(pbuf) > 1) {
1178                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1179                 set_current_state(TASK_UNINTERRUPTIBLE);
1180                 schedule_timeout(cfs_time_seconds(1));
1181         }
1182
1183         cfs_restore_sigs(blocked);
1184 }
1185
1186 static void
1187 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1188 {
1189         struct lnet_ni          *ni;
1190         struct lnet_net         *net;
1191         struct lnet_ni_status *ns;
1192         int                     i;
1193         int                     rc;
1194
1195         i = 0;
1196         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1197                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1198                         LASSERT(i < pbuf->pb_nnis);
1199
1200                         ns = &pbuf->pb_info.pi_ni[i];
1201
1202                         ns->ns_nid = ni->ni_nid;
1203
1204                         lnet_ni_lock(ni);
1205                         ns->ns_status = (ni->ni_status != NULL) ?
1206                                          ni->ni_status->ns_status :
1207                                                 LNET_NI_STATUS_UP;
1208                         ni->ni_status = ns;
1209                         lnet_ni_unlock(ni);
1210
1211                         i++;
1212                 }
1213         }
1214         /*
1215          * We (ab)use the ns_status of the loopback interface to
1216          * transmit the sequence number. The first interface listed
1217          * must be the loopback interface.
1218          */
1219         rc = lnet_ping_info_validate(&pbuf->pb_info);
1220         if (rc) {
1221                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1222                 LBUG();
1223         }
1224         LNET_PING_BUFFER_SEQNO(pbuf) =
1225                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1226 }
1227
1228 static void
1229 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1230                         struct lnet_handle_md ping_mdh)
1231 {
1232         struct lnet_ping_buffer *old_pbuf = NULL;
1233         struct lnet_handle_md old_ping_md;
1234
1235         /* switch the NIs to point to the new ping info created */
1236         lnet_net_lock(LNET_LOCK_EX);
1237
1238         if (!the_lnet.ln_routing)
1239                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1240
1241         /* Ensure only known feature bits have been set. */
1242         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1243         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1244
1245         lnet_ping_target_install_locked(pbuf);
1246
1247         if (the_lnet.ln_ping_target) {
1248                 old_pbuf = the_lnet.ln_ping_target;
1249                 old_ping_md = the_lnet.ln_ping_target_md;
1250         }
1251         the_lnet.ln_ping_target_md = ping_mdh;
1252         the_lnet.ln_ping_target = pbuf;
1253
1254         lnet_net_unlock(LNET_LOCK_EX);
1255
1256         if (old_pbuf) {
1257                 /* unlink and free the old ping info */
1258                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1259                 lnet_ping_buffer_decref(old_pbuf);
1260         }
1261 }
1262
1263 static void
1264 lnet_ping_target_fini(void)
1265 {
1266         int             rc;
1267
1268         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1269                             &the_lnet.ln_ping_target_md);
1270
1271         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1272         LASSERT(rc == 0);
1273
1274         lnet_ping_target_destroy();
1275 }
1276
1277 static int
1278 lnet_ni_tq_credits(struct lnet_ni *ni)
1279 {
1280         int     credits;
1281
1282         LASSERT(ni->ni_ncpts >= 1);
1283
1284         if (ni->ni_ncpts == 1)
1285                 return ni->ni_net->net_tunables.lct_max_tx_credits;
1286
1287         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
1288         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
1289         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
1290
1291         return credits;
1292 }
1293
1294 static void
1295 lnet_ni_unlink_locked(struct lnet_ni *ni)
1296 {
1297         if (!list_empty(&ni->ni_cptlist)) {
1298                 list_del_init(&ni->ni_cptlist);
1299                 lnet_ni_decref_locked(ni, 0);
1300         }
1301
1302         /* move it to zombie list and nobody can find it anymore */
1303         LASSERT(!list_empty(&ni->ni_netlist));
1304         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
1305         lnet_ni_decref_locked(ni, 0);
1306 }
1307
1308 static void
1309 lnet_clear_zombies_nis_locked(struct lnet_net *net)
1310 {
1311         int             i;
1312         int             islo;
1313         struct lnet_ni  *ni;
1314         struct list_head *zombie_list = &net->net_ni_zombie;
1315
1316         /*
1317          * Now wait for the NIs I just nuked to show up on the zombie
1318          * list and shut them down in guaranteed thread context
1319          */
1320         i = 2;
1321         while (!list_empty(zombie_list)) {
1322                 int     *ref;
1323                 int     j;
1324
1325                 ni = list_entry(zombie_list->next,
1326                                 struct lnet_ni, ni_netlist);
1327                 list_del_init(&ni->ni_netlist);
1328                 /* the ni should be in deleting state. If it's not it's
1329                  * a bug */
1330                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
1331                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1332                         if (*ref == 0)
1333                                 continue;
1334                         /* still busy, add it back to zombie list */
1335                         list_add(&ni->ni_netlist, zombie_list);
1336                         break;
1337                 }
1338
1339                 if (!list_empty(&ni->ni_netlist)) {
1340                         lnet_net_unlock(LNET_LOCK_EX);
1341                         ++i;
1342                         if ((i & (-i)) == i) {
1343                                 CDEBUG(D_WARNING,
1344                                        "Waiting for zombie LNI %s\n",
1345                                        libcfs_nid2str(ni->ni_nid));
1346                         }
1347                         set_current_state(TASK_UNINTERRUPTIBLE);
1348                         schedule_timeout(cfs_time_seconds(1));
1349                         lnet_net_lock(LNET_LOCK_EX);
1350                         continue;
1351                 }
1352
1353                 lnet_net_unlock(LNET_LOCK_EX);
1354
1355                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
1356
1357                 LASSERT(!in_interrupt());
1358                 (net->net_lnd->lnd_shutdown)(ni);
1359
1360                 if (!islo)
1361                         CDEBUG(D_LNI, "Removed LNI %s\n",
1362                               libcfs_nid2str(ni->ni_nid));
1363
1364                 lnet_ni_free(ni);
1365                 i = 2;
1366                 lnet_net_lock(LNET_LOCK_EX);
1367         }
1368 }
1369
1370 /* shutdown down the NI and release refcount */
1371 static void
1372 lnet_shutdown_lndni(struct lnet_ni *ni)
1373 {
1374         int i;
1375         struct lnet_net *net = ni->ni_net;
1376
1377         lnet_net_lock(LNET_LOCK_EX);
1378         ni->ni_state = LNET_NI_STATE_DELETING;
1379         lnet_ni_unlink_locked(ni);
1380         lnet_incr_dlc_seq();
1381         lnet_net_unlock(LNET_LOCK_EX);
1382
1383         /* clear messages for this NI on the lazy portal */
1384         for (i = 0; i < the_lnet.ln_nportals; i++)
1385                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
1386
1387         lnet_net_lock(LNET_LOCK_EX);
1388         lnet_clear_zombies_nis_locked(net);
1389         lnet_net_unlock(LNET_LOCK_EX);
1390 }
1391
1392 static void
1393 lnet_shutdown_lndnet(struct lnet_net *net)
1394 {
1395         struct lnet_ni *ni;
1396
1397         lnet_net_lock(LNET_LOCK_EX);
1398
1399         net->net_state = LNET_NET_STATE_DELETING;
1400
1401         list_del_init(&net->net_list);
1402
1403         while (!list_empty(&net->net_ni_list)) {
1404                 ni = list_entry(net->net_ni_list.next,
1405                                 struct lnet_ni, ni_netlist);
1406                 lnet_net_unlock(LNET_LOCK_EX);
1407                 lnet_shutdown_lndni(ni);
1408                 lnet_net_lock(LNET_LOCK_EX);
1409         }
1410
1411         lnet_net_unlock(LNET_LOCK_EX);
1412
1413         /* Do peer table cleanup for this net */
1414         lnet_peer_tables_cleanup(net);
1415
1416         lnet_net_lock(LNET_LOCK_EX);
1417         /*
1418          * decrement ref count on lnd only when the entire network goes
1419          * away
1420          */
1421         net->net_lnd->lnd_refcount--;
1422
1423         lnet_net_unlock(LNET_LOCK_EX);
1424
1425         lnet_net_free(net);
1426 }
1427
1428 static void
1429 lnet_shutdown_lndnets(void)
1430 {
1431         struct lnet_net *net;
1432
1433         /* NB called holding the global mutex */
1434
1435         /* All quiet on the API front */
1436         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
1437         LASSERT(the_lnet.ln_refcount == 0);
1438
1439         lnet_net_lock(LNET_LOCK_EX);
1440         the_lnet.ln_state = LNET_STATE_STOPPING;
1441
1442         while (!list_empty(&the_lnet.ln_nets)) {
1443                 /*
1444                  * move the nets to the zombie list to avoid them being
1445                  * picked up for new work. LONET is also included in the
1446                  * Nets that will be moved to the zombie list
1447                  */
1448                 net = list_entry(the_lnet.ln_nets.next,
1449                                  struct lnet_net, net_list);
1450                 list_move(&net->net_list, &the_lnet.ln_net_zombie);
1451         }
1452
1453         /* Drop the cached loopback Net. */
1454         if (the_lnet.ln_loni != NULL) {
1455                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1456                 the_lnet.ln_loni = NULL;
1457         }
1458         lnet_net_unlock(LNET_LOCK_EX);
1459
1460         /* iterate through the net zombie list and delete each net */
1461         while (!list_empty(&the_lnet.ln_net_zombie)) {
1462                 net = list_entry(the_lnet.ln_net_zombie.next,
1463                                  struct lnet_net, net_list);
1464                 lnet_shutdown_lndnet(net);
1465         }
1466
1467         lnet_net_lock(LNET_LOCK_EX);
1468         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
1469         lnet_net_unlock(LNET_LOCK_EX);
1470 }
1471
1472 static int
1473 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
1474 {
1475         int                     rc = -EINVAL;
1476         struct lnet_tx_queue    *tq;
1477         int                     i;
1478         struct lnet_net         *net = ni->ni_net;
1479
1480         mutex_lock(&the_lnet.ln_lnd_mutex);
1481
1482         if (tun) {
1483                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
1484                 ni->ni_lnd_tunables_set = true;
1485         }
1486
1487         rc = (net->net_lnd->lnd_startup)(ni);
1488
1489         mutex_unlock(&the_lnet.ln_lnd_mutex);
1490
1491         if (rc != 0) {
1492                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1493                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
1494                 lnet_net_lock(LNET_LOCK_EX);
1495                 net->net_lnd->lnd_refcount--;
1496                 lnet_net_unlock(LNET_LOCK_EX);
1497                 goto failed0;
1498         }
1499
1500         ni->ni_state = LNET_NI_STATE_ACTIVE;
1501
1502         /* We keep a reference on the loopback net through the loopback NI */
1503         if (net->net_lnd->lnd_type == LOLND) {
1504                 lnet_ni_addref(ni);
1505                 LASSERT(the_lnet.ln_loni == NULL);
1506                 the_lnet.ln_loni = ni;
1507                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
1508                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
1509                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
1510                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
1511                 return 0;
1512         }
1513
1514         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
1515             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
1516                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1517                                    libcfs_lnd2str(net->net_lnd->lnd_type),
1518                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
1519                                         "" : "per-peer ");
1520                 /* shutdown the NI since if we get here then it must've already
1521                  * been started
1522                  */
1523                 lnet_shutdown_lndni(ni);
1524                 return -EINVAL;
1525         }
1526
1527         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1528                 tq->tq_credits_min =
1529                 tq->tq_credits_max =
1530                 tq->tq_credits = lnet_ni_tq_credits(ni);
1531         }
1532
1533         atomic_set(&ni->ni_tx_credits,
1534                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
1535
1536         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1537                 libcfs_nid2str(ni->ni_nid),
1538                 ni->ni_net->net_tunables.lct_peer_tx_credits,
1539                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1540                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
1541                 ni->ni_net->net_tunables.lct_peer_timeout);
1542
1543         return 0;
1544 failed0:
1545         lnet_ni_free(ni);
1546         return rc;
1547 }
1548
1549 static int
1550 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
1551 {
1552         struct lnet_ni *ni;
1553         struct lnet_net *net_l = NULL;
1554         struct list_head        local_ni_list;
1555         int                     rc;
1556         int                     ni_count = 0;
1557         __u32                   lnd_type;
1558         struct lnet_lnd *lnd;
1559         int                     peer_timeout =
1560                 net->net_tunables.lct_peer_timeout;
1561         int                     maxtxcredits =
1562                 net->net_tunables.lct_max_tx_credits;
1563         int                     peerrtrcredits =
1564                 net->net_tunables.lct_peer_rtr_credits;
1565
1566         INIT_LIST_HEAD(&local_ni_list);
1567
1568         /*
1569          * make sure that this net is unique. If it isn't then
1570          * we are adding interfaces to an already existing network, and
1571          * 'net' is just a convenient way to pass in the list.
1572          * if it is unique we need to find the LND and load it if
1573          * necessary.
1574          */
1575         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
1576                 lnd_type = LNET_NETTYP(net->net_id);
1577
1578                 LASSERT(libcfs_isknown_lnd(lnd_type));
1579
1580                 mutex_lock(&the_lnet.ln_lnd_mutex);
1581                 lnd = lnet_find_lnd_by_type(lnd_type);
1582
1583                 if (lnd == NULL) {
1584                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1585                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
1586                         mutex_lock(&the_lnet.ln_lnd_mutex);
1587
1588                         lnd = lnet_find_lnd_by_type(lnd_type);
1589                         if (lnd == NULL) {
1590                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1591                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1592                                 libcfs_lnd2str(lnd_type),
1593                                 libcfs_lnd2modname(lnd_type), rc);
1594 #ifndef HAVE_MODULE_LOADING_SUPPORT
1595                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1596                                                 "compiled with kernel module "
1597                                                 "loading support.");
1598 #endif
1599                                 rc = -EINVAL;
1600                                 goto failed0;
1601                         }
1602                 }
1603
1604                 lnet_net_lock(LNET_LOCK_EX);
1605                 lnd->lnd_refcount++;
1606                 lnet_net_unlock(LNET_LOCK_EX);
1607
1608                 net->net_lnd = lnd;
1609
1610                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1611
1612                 net_l = net;
1613         }
1614
1615         /*
1616          * net_l: if the network being added is unique then net_l
1617          *        will point to that network
1618          *        if the network being added is not unique then
1619          *        net_l points to the existing network.
1620          *
1621          * When we enter the loop below, we'll pick NIs off he
1622          * network beign added and start them up, then add them to
1623          * a local ni list. Once we've successfully started all
1624          * the NIs then we join the local NI list (of started up
1625          * networks) with the net_l->net_ni_list, which should
1626          * point to the correct network to add the new ni list to
1627          *
1628          * If any of the new NIs fail to start up, then we want to
1629          * iterate through the local ni list, which should include
1630          * any NIs which were successfully started up, and shut
1631          * them down.
1632          *
1633          * After than we want to delete the network being added,
1634          * to avoid a memory leak.
1635          */
1636
1637         /*
1638          * When a network uses TCP bonding then all its interfaces
1639          * must be specified when the network is first defined: the
1640          * TCP bonding code doesn't allow for interfaces to be added
1641          * or removed.
1642          */
1643         if (net_l != net && net_l != NULL && use_tcp_bonding &&
1644             LNET_NETTYP(net_l->net_id) == SOCKLND) {
1645                 rc = -EINVAL;
1646                 goto failed0;
1647         }
1648
1649         while (!list_empty(&net->net_ni_added)) {
1650                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
1651                                 ni_netlist);
1652                 list_del_init(&ni->ni_netlist);
1653
1654                 /* make sure that the the NI we're about to start
1655                  * up is actually unique. if it's not fail. */
1656                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
1657                                         ni->ni_interfaces[0])) {
1658                         rc = -EINVAL;
1659                         goto failed1;
1660                 }
1661
1662                 /* adjust the pointer the parent network, just in case it
1663                  * the net is a duplicate */
1664                 ni->ni_net = net_l;
1665
1666                 rc = lnet_startup_lndni(ni, tun);
1667
1668                 LASSERT(ni->ni_net->net_tunables.lct_peer_timeout <= 0 ||
1669                         ni->ni_net->net_lnd->lnd_query != NULL);
1670
1671                 if (rc < 0)
1672                         goto failed1;
1673
1674                 lnet_ni_addref(ni);
1675                 list_add_tail(&ni->ni_netlist, &local_ni_list);
1676
1677                 ni_count++;
1678         }
1679
1680         lnet_net_lock(LNET_LOCK_EX);
1681         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
1682         lnet_incr_dlc_seq();
1683         lnet_net_unlock(LNET_LOCK_EX);
1684
1685         /* if the network is not unique then we don't want to keep
1686          * it around after we're done. Free it. Otherwise add that
1687          * net to the global the_lnet.ln_nets */
1688         if (net_l != net && net_l != NULL) {
1689                 /*
1690                  * TODO - note. currently the tunables can not be updated
1691                  * once added
1692                  */
1693                 lnet_net_free(net);
1694         } else {
1695                 net->net_state = LNET_NET_STATE_ACTIVE;
1696                 /*
1697                  * restore tunables after it has been overwitten by the
1698                  * lnd
1699                  */
1700                 if (peer_timeout != -1)
1701                         net->net_tunables.lct_peer_timeout = peer_timeout;
1702                 if (maxtxcredits != -1)
1703                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
1704                 if (peerrtrcredits != -1)
1705                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
1706
1707                 lnet_net_lock(LNET_LOCK_EX);
1708                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
1709                 lnet_net_unlock(LNET_LOCK_EX);
1710         }
1711
1712         return ni_count;
1713
1714 failed1:
1715         /*
1716          * shutdown the new NIs that are being started up
1717          * free the NET being started
1718          */
1719         while (!list_empty(&local_ni_list)) {
1720                 ni = list_entry(local_ni_list.next, struct lnet_ni,
1721                                 ni_netlist);
1722
1723                 lnet_shutdown_lndni(ni);
1724         }
1725
1726 failed0:
1727         lnet_net_free(net);
1728
1729         return rc;
1730 }
1731
1732 static int
1733 lnet_startup_lndnets(struct list_head *netlist)
1734 {
1735         struct lnet_net         *net;
1736         int                     rc;
1737         int                     ni_count = 0;
1738
1739         /*
1740          * Change to running state before bringing up the LNDs. This
1741          * allows lnet_shutdown_lndnets() to assert that we've passed
1742          * through here.
1743          */
1744         lnet_net_lock(LNET_LOCK_EX);
1745         the_lnet.ln_state = LNET_STATE_RUNNING;
1746         lnet_net_unlock(LNET_LOCK_EX);
1747
1748         while (!list_empty(netlist)) {
1749                 net = list_entry(netlist->next, struct lnet_net, net_list);
1750                 list_del_init(&net->net_list);
1751
1752                 rc = lnet_startup_lndnet(net, NULL);
1753
1754                 if (rc < 0)
1755                         goto failed;
1756
1757                 ni_count += rc;
1758         }
1759
1760         return ni_count;
1761 failed:
1762         lnet_shutdown_lndnets();
1763
1764         return rc;
1765 }
1766
1767 /**
1768  * Initialize LNet library.
1769  *
1770  * Automatically called at module loading time. Caller has to call
1771  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
1772  * latter returned 0. It must be called exactly once.
1773  *
1774  * \retval 0 on success
1775  * \retval -ve on failures.
1776  */
1777 int lnet_lib_init(void)
1778 {
1779         int rc;
1780
1781         lnet_assert_wire_constants();
1782
1783         memset(&the_lnet, 0, sizeof(the_lnet));
1784
1785         /* refer to global cfs_cpt_table for now */
1786         the_lnet.ln_cpt_table   = cfs_cpt_table;
1787         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1788
1789         LASSERT(the_lnet.ln_cpt_number > 0);
1790         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1791                 /* we are under risk of consuming all lh_cookie */
1792                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1793                        "please change setting of CPT-table and retry\n",
1794                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1795                 return -E2BIG;
1796         }
1797
1798         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1799                 the_lnet.ln_cpt_bits++;
1800
1801         rc = lnet_create_locks();
1802         if (rc != 0) {
1803                 CERROR("Can't create LNet global locks: %d\n", rc);
1804                 return rc;
1805         }
1806
1807         the_lnet.ln_refcount = 0;
1808         LNetInvalidateEQHandle(&the_lnet.ln_rc_eqh);
1809         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1810         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
1811         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1812         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1813
1814         /* The hash table size is the number of bits it takes to express the set
1815          * ln_num_routes, minus 1 (better to under estimate than over so we
1816          * don't waste memory). */
1817         if (rnet_htable_size <= 0)
1818                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1819         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1820                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1821         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1822                                            order_base_2(rnet_htable_size) - 1);
1823
1824         /* All LNDs apart from the LOLND are in separate modules.  They
1825          * register themselves when their module loads, and unregister
1826          * themselves when their module is unloaded. */
1827         lnet_register_lnd(&the_lolnd);
1828         return 0;
1829 }
1830
1831 /**
1832  * Finalize LNet library.
1833  *
1834  * \pre lnet_lib_init() called with success.
1835  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1836  */
1837 void lnet_lib_exit(void)
1838 {
1839         LASSERT(the_lnet.ln_refcount == 0);
1840
1841         while (!list_empty(&the_lnet.ln_lnds))
1842                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1843                                                struct lnet_lnd, lnd_list));
1844         lnet_destroy_locks();
1845 }
1846
1847 /**
1848  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1849  *
1850  * Users must call this function at least once before any other functions.
1851  * For each successful call there must be a corresponding call to
1852  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1853  * ignored.
1854  *
1855  * The PID used by LNet may be different from the one requested.
1856  * See LNetGetId().
1857  *
1858  * \param requested_pid PID requested by the caller.
1859  *
1860  * \return >= 0 on success, and < 0 error code on failures.
1861  */
1862 int
1863 LNetNIInit(lnet_pid_t requested_pid)
1864 {
1865         int                     im_a_router = 0;
1866         int                     rc;
1867         int                     ni_count;
1868         struct lnet_ping_buffer *pbuf;
1869         struct lnet_handle_md   ping_mdh;
1870         struct list_head        net_head;
1871         struct lnet_net         *net;
1872
1873         INIT_LIST_HEAD(&net_head);
1874
1875         mutex_lock(&the_lnet.ln_api_mutex);
1876
1877         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1878
1879         if (the_lnet.ln_refcount > 0) {
1880                 rc = the_lnet.ln_refcount++;
1881                 mutex_unlock(&the_lnet.ln_api_mutex);
1882                 return rc;
1883         }
1884
1885         rc = lnet_prepare(requested_pid);
1886         if (rc != 0) {
1887                 mutex_unlock(&the_lnet.ln_api_mutex);
1888                 return rc;
1889         }
1890
1891         /* create a network for Loopback network */
1892         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
1893         if (net == NULL) {
1894                 rc = -ENOMEM;
1895                 goto err_empty_list;
1896         }
1897
1898         /* Add in the loopback NI */
1899         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
1900                 rc = -ENOMEM;
1901                 goto err_empty_list;
1902         }
1903
1904         /* If LNet is being initialized via DLC it is possible
1905          * that the user requests not to load module parameters (ones which
1906          * are supported by DLC) on initialization.  Therefore, make sure not
1907          * to load networks, routes and forwarding from module parameters
1908          * in this case.  On cleanup in case of failure only clean up
1909          * routes if it has been loaded */
1910         if (!the_lnet.ln_nis_from_mod_params) {
1911                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
1912                                          use_tcp_bonding);
1913                 if (rc < 0)
1914                         goto err_empty_list;
1915         }
1916
1917         ni_count = lnet_startup_lndnets(&net_head);
1918         if (ni_count < 0) {
1919                 rc = ni_count;
1920                 goto err_empty_list;
1921         }
1922
1923         if (!the_lnet.ln_nis_from_mod_params) {
1924                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1925                 if (rc != 0)
1926                         goto err_shutdown_lndnis;
1927
1928                 rc = lnet_check_routes();
1929                 if (rc != 0)
1930                         goto err_destroy_routes;
1931
1932                 rc = lnet_rtrpools_alloc(im_a_router);
1933                 if (rc != 0)
1934                         goto err_destroy_routes;
1935         }
1936
1937         rc = lnet_acceptor_start();
1938         if (rc != 0)
1939                 goto err_destroy_routes;
1940
1941         the_lnet.ln_refcount = 1;
1942         /* Now I may use my own API functions... */
1943
1944         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
1945         if (rc != 0)
1946                 goto err_acceptor_stop;
1947
1948         lnet_ping_target_update(pbuf, ping_mdh);
1949
1950         rc = lnet_router_checker_start();
1951         if (rc != 0)
1952                 goto err_stop_ping;
1953
1954         lnet_fault_init();
1955         lnet_proc_init();
1956
1957         mutex_unlock(&the_lnet.ln_api_mutex);
1958
1959         return 0;
1960
1961 err_stop_ping:
1962         lnet_ping_target_fini();
1963 err_acceptor_stop:
1964         the_lnet.ln_refcount = 0;
1965         lnet_acceptor_stop();
1966 err_destroy_routes:
1967         if (!the_lnet.ln_nis_from_mod_params)
1968                 lnet_destroy_routes();
1969 err_shutdown_lndnis:
1970         lnet_shutdown_lndnets();
1971 err_empty_list:
1972         lnet_unprepare();
1973         LASSERT(rc < 0);
1974         mutex_unlock(&the_lnet.ln_api_mutex);
1975         while (!list_empty(&net_head)) {
1976                 struct lnet_net *net;
1977
1978                 net = list_entry(net_head.next, struct lnet_net, net_list);
1979                 list_del_init(&net->net_list);
1980                 lnet_net_free(net);
1981         }
1982         return rc;
1983 }
1984 EXPORT_SYMBOL(LNetNIInit);
1985
1986 /**
1987  * Stop LNet interfaces, routing, and forwarding.
1988  *
1989  * Users must call this function once for each successful call to LNetNIInit().
1990  * Once the LNetNIFini() operation has been started, the results of pending
1991  * API operations are undefined.
1992  *
1993  * \return always 0 for current implementation.
1994  */
1995 int
1996 LNetNIFini()
1997 {
1998         mutex_lock(&the_lnet.ln_api_mutex);
1999
2000         LASSERT(the_lnet.ln_refcount > 0);
2001
2002         if (the_lnet.ln_refcount != 1) {
2003                 the_lnet.ln_refcount--;
2004         } else {
2005                 LASSERT(!the_lnet.ln_niinit_self);
2006
2007                 lnet_fault_fini();
2008
2009                 lnet_proc_fini();
2010                 lnet_router_checker_stop();
2011                 lnet_ping_target_fini();
2012
2013                 /* Teardown fns that use my own API functions BEFORE here */
2014                 the_lnet.ln_refcount = 0;
2015
2016                 lnet_acceptor_stop();
2017                 lnet_destroy_routes();
2018                 lnet_shutdown_lndnets();
2019                 lnet_unprepare();
2020         }
2021
2022         mutex_unlock(&the_lnet.ln_api_mutex);
2023         return 0;
2024 }
2025 EXPORT_SYMBOL(LNetNIFini);
2026
2027 /**
2028  * Grabs the ni data from the ni structure and fills the out
2029  * parameters
2030  *
2031  * \param[in] ni network        interface structure
2032  * \param[out] cfg_ni           NI config information
2033  * \param[out] tun              network and LND tunables
2034  */
2035 static void
2036 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2037                    struct lnet_ioctl_config_lnd_tunables *tun,
2038                    struct lnet_ioctl_element_stats *stats,
2039                    __u32 tun_size)
2040 {
2041         size_t min_size = 0;
2042         int i;
2043
2044         if (!ni || !cfg_ni || !tun)
2045                 return;
2046
2047         if (ni->ni_interfaces[0] != NULL) {
2048                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2049                         if (ni->ni_interfaces[i] != NULL) {
2050                                 strncpy(cfg_ni->lic_ni_intf[i],
2051                                         ni->ni_interfaces[i],
2052                                         sizeof(cfg_ni->lic_ni_intf[i]));
2053                         }
2054                 }
2055         }
2056
2057         cfg_ni->lic_nid = ni->ni_nid;
2058         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2059                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2060         else
2061                 cfg_ni->lic_status = ni->ni_status->ns_status;
2062         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2063         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2064
2065         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2066
2067         if (stats) {
2068                 stats->iel_send_count = atomic_read(&ni->ni_stats.send_count);
2069                 stats->iel_recv_count = atomic_read(&ni->ni_stats.recv_count);
2070         }
2071
2072         /*
2073          * tun->lt_tun will always be present, but in order to be
2074          * backwards compatible, we need to deal with the cases when
2075          * tun->lt_tun is smaller than what the kernel has, because it
2076          * comes from an older version of a userspace program, then we'll
2077          * need to copy as much information as we have available space.
2078          */
2079         min_size = tun_size - sizeof(tun->lt_cmn);
2080         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2081
2082         /* copy over the cpts */
2083         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2084             ni->ni_cpts == NULL)  {
2085                 for (i = 0; i < ni->ni_ncpts; i++)
2086                         cfg_ni->lic_cpts[i] = i;
2087         } else {
2088                 for (i = 0;
2089                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2090                      i < LNET_MAX_SHOW_NUM_CPT;
2091                      i++)
2092                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2093         }
2094         cfg_ni->lic_ncpts = ni->ni_ncpts;
2095 }
2096
2097 /**
2098  * NOTE: This is a legacy function left in the code to be backwards
2099  * compatible with older userspace programs. It should eventually be
2100  * removed.
2101  *
2102  * Grabs the ni data from the ni structure and fills the out
2103  * parameters
2104  *
2105  * \param[in] ni network        interface structure
2106  * \param[out] config           config information
2107  */
2108 static void
2109 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2110                          struct lnet_ioctl_config_data *config)
2111 {
2112         struct lnet_ioctl_net_config *net_config;
2113         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2114         size_t min_size, tunable_size = 0;
2115         int i;
2116
2117         if (!ni || !config)
2118                 return;
2119
2120         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2121         if (!net_config)
2122                 return;
2123
2124         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2125                      ARRAY_SIZE(net_config->ni_interfaces));
2126
2127         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2128                 if (!ni->ni_interfaces[i])
2129                         break;
2130
2131                 strncpy(net_config->ni_interfaces[i],
2132                         ni->ni_interfaces[i],
2133                         sizeof(net_config->ni_interfaces[i]));
2134         }
2135
2136         config->cfg_nid = ni->ni_nid;
2137         config->cfg_config_u.cfg_net.net_peer_timeout =
2138                 ni->ni_net->net_tunables.lct_peer_timeout;
2139         config->cfg_config_u.cfg_net.net_max_tx_credits =
2140                 ni->ni_net->net_tunables.lct_max_tx_credits;
2141         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2142                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2143         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2144                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2145
2146         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2147                 net_config->ni_status = LNET_NI_STATUS_UP;
2148         else
2149                 net_config->ni_status = ni->ni_status->ns_status;
2150
2151         if (ni->ni_cpts) {
2152                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2153
2154                 for (i = 0; i < num_cpts; i++)
2155                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2156
2157                 config->cfg_ncpts = num_cpts;
2158         }
2159
2160         /*
2161          * See if user land tools sent in a newer and larger version
2162          * of struct lnet_tunables than what the kernel uses.
2163          */
2164         min_size = sizeof(*config) + sizeof(*net_config);
2165
2166         if (config->cfg_hdr.ioc_len > min_size)
2167                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2168
2169         /* Don't copy too much data to user space */
2170         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2171         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2172
2173         if (lnd_cfg && min_size) {
2174                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2175                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2176
2177                 /* Tell user land that kernel side has less data */
2178                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2179                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2180                         config->cfg_hdr.ioc_len -= min_size;
2181                 }
2182         }
2183 }
2184
2185 struct lnet_ni *
2186 lnet_get_ni_idx_locked(int idx)
2187 {
2188         struct lnet_ni          *ni;
2189         struct lnet_net         *net;
2190
2191         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2192                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2193                         if (idx-- == 0)
2194                                 return ni;
2195                 }
2196         }
2197
2198         return NULL;
2199 }
2200
2201 struct lnet_ni *
2202 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2203 {
2204         struct lnet_ni          *ni;
2205         struct lnet_net         *net = mynet;
2206
2207         if (prev == NULL) {
2208                 if (net == NULL)
2209                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2210                                         net_list);
2211                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2212                                 ni_netlist);
2213
2214                 return ni;
2215         }
2216
2217         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2218                 /* if you reached the end of the ni list and the net is
2219                  * specified, then there are no more nis in that net */
2220                 if (net != NULL)
2221                         return NULL;
2222
2223                 /* we reached the end of this net ni list. move to the
2224                  * next net */
2225                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2226                         /* no more nets and no more NIs. */
2227                         return NULL;
2228
2229                 /* get the next net */
2230                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2231                                  net_list);
2232                 /* get the ni on it */
2233                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2234                                 ni_netlist);
2235
2236                 return ni;
2237         }
2238
2239         /* there are more nis left */
2240         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
2241
2242         return ni;
2243 }
2244
2245 int
2246 lnet_get_net_config(struct lnet_ioctl_config_data *config)
2247 {
2248         struct lnet_ni *ni;
2249         int cpt;
2250         int rc = -ENOENT;
2251         int idx = config->cfg_count;
2252
2253         cpt = lnet_net_lock_current();
2254
2255         ni = lnet_get_ni_idx_locked(idx);
2256
2257         if (ni != NULL) {
2258                 rc = 0;
2259                 lnet_ni_lock(ni);
2260                 lnet_fill_ni_info_legacy(ni, config);
2261                 lnet_ni_unlock(ni);
2262         }
2263
2264         lnet_net_unlock(cpt);
2265         return rc;
2266 }
2267
2268 int
2269 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
2270                    struct lnet_ioctl_config_lnd_tunables *tun,
2271                    struct lnet_ioctl_element_stats *stats,
2272                    __u32 tun_size)
2273 {
2274         struct lnet_ni          *ni;
2275         int                     cpt;
2276         int                     rc = -ENOENT;
2277
2278         if (!cfg_ni || !tun || !stats)
2279                 return -EINVAL;
2280
2281         cpt = lnet_net_lock_current();
2282
2283         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
2284
2285         if (ni) {
2286                 rc = 0;
2287                 lnet_ni_lock(ni);
2288                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
2289                 lnet_ni_unlock(ni);
2290         }
2291
2292         lnet_net_unlock(cpt);
2293         return rc;
2294 }
2295
2296 static int lnet_add_net_common(struct lnet_net *net,
2297                                struct lnet_ioctl_config_lnd_tunables *tun)
2298 {
2299         __u32                   net_id;
2300         struct lnet_ping_buffer *pbuf;
2301         struct lnet_handle_md   ping_mdh;
2302         int                     rc;
2303         struct lnet_remotenet *rnet;
2304         int                     net_ni_count;
2305         int                     num_acceptor_nets;
2306
2307         lnet_net_lock(LNET_LOCK_EX);
2308         rnet = lnet_find_rnet_locked(net->net_id);
2309         lnet_net_unlock(LNET_LOCK_EX);
2310         /*
2311          * make sure that the net added doesn't invalidate the current
2312          * configuration LNet is keeping
2313          */
2314         if (rnet) {
2315                 CERROR("Adding net %s will invalidate routing configuration\n",
2316                        libcfs_net2str(net->net_id));
2317                 lnet_net_free(net);
2318                 return -EUSERS;
2319         }
2320
2321         /*
2322          * make sure you calculate the correct number of slots in the ping
2323          * buffer. Since the ping info is a flattened list of all the NIs,
2324          * we should allocate enough slots to accomodate the number of NIs
2325          * which will be added.
2326          *
2327          * since ni hasn't been configured yet, use
2328          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
2329          */
2330         net_ni_count = lnet_get_net_ni_count_pre(net);
2331
2332         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
2333                                     net_ni_count + lnet_get_ni_count(),
2334                                     false);
2335         if (rc < 0) {
2336                 lnet_net_free(net);
2337                 return rc;
2338         }
2339
2340         if (tun)
2341                 memcpy(&net->net_tunables,
2342                        &tun->lt_cmn, sizeof(net->net_tunables));
2343         else
2344                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
2345
2346         /*
2347          * before starting this network get a count of the current TCP
2348          * networks which require the acceptor thread running. If that
2349          * count is == 0 before we start up this network, then we'd want to
2350          * start up the acceptor thread after starting up this network
2351          */
2352         num_acceptor_nets = lnet_count_acceptor_nets();
2353
2354         net_id = net->net_id;
2355
2356         rc = lnet_startup_lndnet(net,
2357                                  (tun) ? &tun->lt_tun : NULL);
2358         if (rc < 0)
2359                 goto failed;
2360
2361         lnet_net_lock(LNET_LOCK_EX);
2362         net = lnet_get_net_locked(net_id);
2363         lnet_net_unlock(LNET_LOCK_EX);
2364
2365         LASSERT(net);
2366
2367         /*
2368          * Start the acceptor thread if this is the first network
2369          * being added that requires the thread.
2370          */
2371         if (net->net_lnd->lnd_accept && num_acceptor_nets == 0) {
2372                 rc = lnet_acceptor_start();
2373                 if (rc < 0) {
2374                         /* shutdown the net that we just started */
2375                         CERROR("Failed to start up acceptor thread\n");
2376                         lnet_shutdown_lndnet(net);
2377                         goto failed;
2378                 }
2379         }
2380
2381         lnet_net_lock(LNET_LOCK_EX);
2382         lnet_peer_net_added(net);
2383         lnet_net_unlock(LNET_LOCK_EX);
2384
2385         lnet_ping_target_update(pbuf, ping_mdh);
2386
2387         return 0;
2388
2389 failed:
2390         lnet_ping_md_unlink(pbuf, &ping_mdh);
2391         lnet_ping_buffer_decref(pbuf);
2392         return rc;
2393 }
2394
2395 static int lnet_handle_legacy_ip2nets(char *ip2nets,
2396                                       struct lnet_ioctl_config_lnd_tunables *tun)
2397 {
2398         struct lnet_net *net;
2399         char *nets;
2400         int rc;
2401         struct list_head net_head;
2402
2403         INIT_LIST_HEAD(&net_head);
2404
2405         rc = lnet_parse_ip2nets(&nets, ip2nets);
2406         if (rc < 0)
2407                 return rc;
2408
2409         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
2410         if (rc < 0)
2411                 return rc;
2412
2413         mutex_lock(&the_lnet.ln_api_mutex);
2414         while (!list_empty(&net_head)) {
2415                 net = list_entry(net_head.next, struct lnet_net, net_list);
2416                 list_del_init(&net->net_list);
2417                 rc = lnet_add_net_common(net, tun);
2418                 if (rc < 0)
2419                         goto out;
2420         }
2421
2422 out:
2423         mutex_unlock(&the_lnet.ln_api_mutex);
2424
2425         while (!list_empty(&net_head)) {
2426                 net = list_entry(net_head.next, struct lnet_net, net_list);
2427                 list_del_init(&net->net_list);
2428                 lnet_net_free(net);
2429         }
2430         return rc;
2431 }
2432
2433 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
2434 {
2435         struct lnet_net *net;
2436         struct lnet_ni *ni;
2437         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
2438         int rc, i;
2439         __u32 net_id;
2440
2441         /* get the tunables if they are available */
2442         if (conf->lic_cfg_hdr.ioc_len >=
2443             sizeof(*conf) + sizeof(*tun))
2444                 tun = (struct lnet_ioctl_config_lnd_tunables *)
2445                         conf->lic_bulk;
2446
2447         /* handle legacy ip2nets from DLC */
2448         if (conf->lic_legacy_ip2nets[0] != '\0')
2449                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
2450                                                   tun);
2451
2452         net_id = LNET_NIDNET(conf->lic_nid);
2453
2454         net = lnet_net_alloc(net_id, NULL);
2455         if (!net)
2456                 return -ENOMEM;
2457
2458         for (i = 0; i < conf->lic_ncpts; i++) {
2459                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
2460                         return -EINVAL;
2461         }
2462
2463         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
2464                                        conf->lic_ni_intf[0]);
2465         if (!ni)
2466                 return -ENOMEM;
2467
2468         mutex_lock(&the_lnet.ln_api_mutex);
2469
2470         rc = lnet_add_net_common(net, tun);
2471
2472         mutex_unlock(&the_lnet.ln_api_mutex);
2473
2474         return rc;
2475 }
2476
2477 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
2478 {
2479         struct lnet_net  *net;
2480         struct lnet_ni *ni;
2481         __u32 net_id = LNET_NIDNET(conf->lic_nid);
2482         struct lnet_ping_buffer *pbuf;
2483         struct lnet_handle_md  ping_mdh;
2484         int               rc;
2485         int               net_count;
2486         __u32             addr;
2487
2488         /* don't allow userspace to shutdown the LOLND */
2489         if (LNET_NETTYP(net_id) == LOLND)
2490                 return -EINVAL;
2491
2492         mutex_lock(&the_lnet.ln_api_mutex);
2493
2494         lnet_net_lock(0);
2495
2496         net = lnet_get_net_locked(net_id);
2497         if (!net) {
2498                 CERROR("net %s not found\n",
2499                        libcfs_net2str(net_id));
2500                 rc = -ENOENT;
2501                 goto unlock_net;
2502         }
2503
2504         addr = LNET_NIDADDR(conf->lic_nid);
2505         if (addr == 0) {
2506                 /* remove the entire net */
2507                 net_count = lnet_get_net_ni_count_locked(net);
2508
2509                 lnet_net_unlock(0);
2510
2511                 /* create and link a new ping info, before removing the old one */
2512                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
2513                                         lnet_get_ni_count() - net_count,
2514                                         false);
2515                 if (rc != 0)
2516                         goto unlock_api_mutex;
2517
2518                 lnet_shutdown_lndnet(net);
2519
2520                 if (lnet_count_acceptor_nets() == 0)
2521                         lnet_acceptor_stop();
2522
2523                 lnet_ping_target_update(pbuf, ping_mdh);
2524
2525                 goto unlock_api_mutex;
2526         }
2527
2528         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
2529         if (!ni) {
2530                 CERROR("nid %s not found\n",
2531                        libcfs_nid2str(conf->lic_nid));
2532                 rc = -ENOENT;
2533                 goto unlock_net;
2534         }
2535
2536         net_count = lnet_get_net_ni_count_locked(net);
2537
2538         lnet_net_unlock(0);
2539
2540         /* create and link a new ping info, before removing the old one */
2541         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
2542                                   lnet_get_ni_count() - 1, false);
2543         if (rc != 0)
2544                 goto unlock_api_mutex;
2545
2546         lnet_shutdown_lndni(ni);
2547
2548         if (lnet_count_acceptor_nets() == 0)
2549                 lnet_acceptor_stop();
2550
2551         lnet_ping_target_update(pbuf, ping_mdh);
2552
2553         /* check if the net is empty and remove it if it is */
2554         if (net_count == 1)
2555                 lnet_shutdown_lndnet(net);
2556
2557         goto unlock_api_mutex;
2558
2559 unlock_net:
2560         lnet_net_unlock(0);
2561 unlock_api_mutex:
2562         mutex_unlock(&the_lnet.ln_api_mutex);
2563
2564         return rc;
2565 }
2566
2567 /*
2568  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
2569  * They are only expected to be called for unique networks.
2570  * That can be as a result of older DLC library
2571  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
2572  */
2573 int
2574 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
2575 {
2576         struct lnet_net         *net;
2577         struct list_head        net_head;
2578         int                     rc;
2579         struct lnet_ioctl_config_lnd_tunables tun;
2580         char *nets = conf->cfg_config_u.cfg_net.net_intf;
2581
2582         INIT_LIST_HEAD(&net_head);
2583
2584         /* Create a net/ni structures for the network string */
2585         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
2586         if (rc <= 0)
2587                 return rc == 0 ? -EINVAL : rc;
2588
2589         mutex_lock(&the_lnet.ln_api_mutex);
2590
2591         if (rc > 1) {
2592                 rc = -EINVAL; /* only add one network per call */
2593                 goto out_unlock_clean;
2594         }
2595
2596         net = list_entry(net_head.next, struct lnet_net, net_list);
2597         list_del_init(&net->net_list);
2598
2599         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
2600
2601         memset(&tun, 0, sizeof(tun));
2602
2603         tun.lt_cmn.lct_peer_timeout =
2604           conf->cfg_config_u.cfg_net.net_peer_timeout;
2605         tun.lt_cmn.lct_peer_tx_credits =
2606           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
2607         tun.lt_cmn.lct_peer_rtr_credits =
2608           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
2609         tun.lt_cmn.lct_max_tx_credits =
2610           conf->cfg_config_u.cfg_net.net_max_tx_credits;
2611
2612         rc = lnet_add_net_common(net, &tun);
2613
2614 out_unlock_clean:
2615         mutex_unlock(&the_lnet.ln_api_mutex);
2616         while (!list_empty(&net_head)) {
2617                 /* net_head list is empty in success case */
2618                 net = list_entry(net_head.next, struct lnet_net, net_list);
2619                 list_del_init(&net->net_list);
2620                 lnet_net_free(net);
2621         }
2622         return rc;
2623 }
2624
2625 int
2626 lnet_dyn_del_net(__u32 net_id)
2627 {
2628         struct lnet_net  *net;
2629         struct lnet_ping_buffer *pbuf;
2630         struct lnet_handle_md ping_mdh;
2631         int               rc;
2632         int               net_ni_count;
2633
2634         /* don't allow userspace to shutdown the LOLND */
2635         if (LNET_NETTYP(net_id) == LOLND)
2636                 return -EINVAL;
2637
2638         mutex_lock(&the_lnet.ln_api_mutex);
2639
2640         lnet_net_lock(0);
2641
2642         net = lnet_get_net_locked(net_id);
2643         if (net == NULL) {
2644                 lnet_net_unlock(0);
2645                 rc = -EINVAL;
2646                 goto out;
2647         }
2648
2649         net_ni_count = lnet_get_net_ni_count_locked(net);
2650
2651         lnet_net_unlock(0);
2652
2653         /* create and link a new ping info, before removing the old one */
2654         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
2655                                     lnet_get_ni_count() - net_ni_count, false);
2656         if (rc != 0)
2657                 goto out;
2658
2659         lnet_shutdown_lndnet(net);
2660
2661         if (lnet_count_acceptor_nets() == 0)
2662                 lnet_acceptor_stop();
2663
2664         lnet_ping_target_update(pbuf, ping_mdh);
2665
2666 out:
2667         mutex_unlock(&the_lnet.ln_api_mutex);
2668
2669         return rc;
2670 }
2671
2672 void lnet_incr_dlc_seq(void)
2673 {
2674         atomic_inc(&lnet_dlc_seq_no);
2675 }
2676
2677 __u32 lnet_get_dlc_seq_locked(void)
2678 {
2679         return atomic_read(&lnet_dlc_seq_no);
2680 }
2681
2682 /**
2683  * LNet ioctl handler.
2684  *
2685  */
2686 int
2687 LNetCtl(unsigned int cmd, void *arg)
2688 {
2689         struct libcfs_ioctl_data *data = arg;
2690         struct lnet_ioctl_config_data *config;
2691         struct lnet_process_id    id = {0};
2692         struct lnet_ni           *ni;
2693         int                       rc;
2694
2695         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
2696                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
2697
2698         switch (cmd) {
2699         case IOC_LIBCFS_GET_NI:
2700                 rc = LNetGetId(data->ioc_count, &id);
2701                 data->ioc_nid = id.nid;
2702                 return rc;
2703
2704         case IOC_LIBCFS_FAIL_NID:
2705                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
2706
2707         case IOC_LIBCFS_ADD_ROUTE:
2708                 config = arg;
2709
2710                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2711                         return -EINVAL;
2712
2713                 mutex_lock(&the_lnet.ln_api_mutex);
2714                 rc = lnet_add_route(config->cfg_net,
2715                                     config->cfg_config_u.cfg_route.rtr_hop,
2716                                     config->cfg_nid,
2717                                     config->cfg_config_u.cfg_route.
2718                                         rtr_priority);
2719                 if (rc == 0) {
2720                         rc = lnet_check_routes();
2721                         if (rc != 0)
2722                                 lnet_del_route(config->cfg_net,
2723                                                config->cfg_nid);
2724                 }
2725                 mutex_unlock(&the_lnet.ln_api_mutex);
2726                 return rc;
2727
2728         case IOC_LIBCFS_DEL_ROUTE:
2729                 config = arg;
2730
2731                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2732                         return -EINVAL;
2733
2734                 mutex_lock(&the_lnet.ln_api_mutex);
2735                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
2736                 mutex_unlock(&the_lnet.ln_api_mutex);
2737                 return rc;
2738
2739         case IOC_LIBCFS_GET_ROUTE:
2740                 config = arg;
2741
2742                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2743                         return -EINVAL;
2744
2745                 mutex_lock(&the_lnet.ln_api_mutex);
2746                 rc = lnet_get_route(config->cfg_count,
2747                                     &config->cfg_net,
2748                                     &config->cfg_config_u.cfg_route.rtr_hop,
2749                                     &config->cfg_nid,
2750                                     &config->cfg_config_u.cfg_route.rtr_flags,
2751                                     &config->cfg_config_u.cfg_route.
2752                                         rtr_priority);
2753                 mutex_unlock(&the_lnet.ln_api_mutex);
2754                 return rc;
2755
2756         case IOC_LIBCFS_GET_LOCAL_NI: {
2757                 struct lnet_ioctl_config_ni *cfg_ni;
2758                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
2759                 struct lnet_ioctl_element_stats *stats;
2760                 __u32 tun_size;
2761
2762                 cfg_ni = arg;
2763                 /* get the tunables if they are available */
2764                 if (cfg_ni->lic_cfg_hdr.ioc_len <
2765                     sizeof(*cfg_ni) + sizeof(*stats)+ sizeof(*tun))
2766                         return -EINVAL;
2767
2768                 stats = (struct lnet_ioctl_element_stats *)
2769                         cfg_ni->lic_bulk;
2770                 tun = (struct lnet_ioctl_config_lnd_tunables *)
2771                                 (cfg_ni->lic_bulk + sizeof(*stats));
2772
2773                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
2774                         sizeof(*stats);
2775
2776                 mutex_lock(&the_lnet.ln_api_mutex);
2777                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
2778                 mutex_unlock(&the_lnet.ln_api_mutex);
2779                 return rc;
2780         }
2781
2782         case IOC_LIBCFS_GET_NET: {
2783                 size_t total = sizeof(*config) +
2784                                sizeof(struct lnet_ioctl_net_config);
2785                 config = arg;
2786
2787                 if (config->cfg_hdr.ioc_len < total)
2788                         return -EINVAL;
2789
2790                 mutex_lock(&the_lnet.ln_api_mutex);
2791                 rc = lnet_get_net_config(config);
2792                 mutex_unlock(&the_lnet.ln_api_mutex);
2793                 return rc;
2794         }
2795
2796         case IOC_LIBCFS_GET_LNET_STATS:
2797         {
2798                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
2799
2800                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
2801                         return -EINVAL;
2802
2803                 mutex_lock(&the_lnet.ln_api_mutex);
2804                 lnet_counters_get(&lnet_stats->st_cntrs);
2805                 mutex_unlock(&the_lnet.ln_api_mutex);
2806                 return 0;
2807         }
2808
2809         case IOC_LIBCFS_CONFIG_RTR:
2810                 config = arg;
2811
2812                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2813                         return -EINVAL;
2814
2815                 mutex_lock(&the_lnet.ln_api_mutex);
2816                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
2817                         rc = lnet_rtrpools_enable();
2818                         mutex_unlock(&the_lnet.ln_api_mutex);
2819                         return rc;
2820                 }
2821                 lnet_rtrpools_disable();
2822                 mutex_unlock(&the_lnet.ln_api_mutex);
2823                 return 0;
2824
2825         case IOC_LIBCFS_ADD_BUF:
2826                 config = arg;
2827
2828                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2829                         return -EINVAL;
2830
2831                 mutex_lock(&the_lnet.ln_api_mutex);
2832                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
2833                                                 buf_tiny,
2834                                           config->cfg_config_u.cfg_buffers.
2835                                                 buf_small,
2836                                           config->cfg_config_u.cfg_buffers.
2837                                                 buf_large);
2838                 mutex_unlock(&the_lnet.ln_api_mutex);
2839                 return rc;
2840
2841         case IOC_LIBCFS_SET_NUMA_RANGE: {
2842                 struct lnet_ioctl_set_value *numa;
2843                 numa = arg;
2844                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
2845                         return -EINVAL;
2846                 lnet_net_lock(LNET_LOCK_EX);
2847                 lnet_numa_range = numa->sv_value;
2848                 lnet_net_unlock(LNET_LOCK_EX);
2849                 return 0;
2850         }
2851
2852         case IOC_LIBCFS_GET_NUMA_RANGE: {
2853                 struct lnet_ioctl_set_value *numa;
2854                 numa = arg;
2855                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
2856                         return -EINVAL;
2857                 numa->sv_value = lnet_numa_range;
2858                 return 0;
2859         }
2860
2861         case IOC_LIBCFS_GET_BUF: {
2862                 struct lnet_ioctl_pool_cfg *pool_cfg;
2863                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
2864
2865                 config = arg;
2866
2867                 if (config->cfg_hdr.ioc_len < total)
2868                         return -EINVAL;
2869
2870                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
2871
2872                 mutex_lock(&the_lnet.ln_api_mutex);
2873                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
2874                 mutex_unlock(&the_lnet.ln_api_mutex);
2875                 return rc;
2876         }
2877
2878         case IOC_LIBCFS_ADD_PEER_NI: {
2879                 struct lnet_ioctl_peer_cfg *cfg = arg;
2880
2881                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
2882                         return -EINVAL;
2883
2884                 mutex_lock(&the_lnet.ln_api_mutex);
2885                 rc = lnet_add_peer_ni_to_peer(cfg->prcfg_prim_nid,
2886                                               cfg->prcfg_cfg_nid,
2887                                               cfg->prcfg_mr);
2888                 mutex_unlock(&the_lnet.ln_api_mutex);
2889                 return rc;
2890         }
2891
2892         case IOC_LIBCFS_DEL_PEER_NI: {
2893                 struct lnet_ioctl_peer_cfg *cfg = arg;
2894
2895                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
2896                         return -EINVAL;
2897
2898                 mutex_lock(&the_lnet.ln_api_mutex);
2899                 rc = lnet_del_peer_ni_from_peer(cfg->prcfg_prim_nid,
2900                                                 cfg->prcfg_cfg_nid);
2901                 mutex_unlock(&the_lnet.ln_api_mutex);
2902                 return rc;
2903         }
2904
2905         case IOC_LIBCFS_GET_PEER_INFO: {
2906                 struct lnet_ioctl_peer *peer_info = arg;
2907
2908                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
2909                         return -EINVAL;
2910
2911                 mutex_lock(&the_lnet.ln_api_mutex);
2912                 rc = lnet_get_peer_ni_info(
2913                    peer_info->pr_count,
2914                    &peer_info->pr_nid,
2915                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
2916                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
2917                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
2918                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
2919                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
2920                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
2921                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
2922                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
2923                 mutex_unlock(&the_lnet.ln_api_mutex);
2924                 return rc;
2925         }
2926
2927         case IOC_LIBCFS_GET_PEER_NI: {
2928                 struct lnet_ioctl_peer_cfg *cfg = arg;
2929                 struct lnet_peer_ni_credit_info __user *lpni_cri;
2930                 struct lnet_ioctl_element_stats __user *lpni_stats;
2931                 size_t usr_size = sizeof(*lpni_cri) + sizeof(*lpni_stats);
2932
2933                 if ((cfg->prcfg_hdr.ioc_len != sizeof(*cfg)) ||
2934                     (cfg->prcfg_size != usr_size))
2935                         return -EINVAL;
2936
2937                 lpni_cri = cfg->prcfg_bulk;
2938                 lpni_stats = cfg->prcfg_bulk + sizeof(*lpni_cri);
2939
2940                 mutex_lock(&the_lnet.ln_api_mutex);
2941                 rc = lnet_get_peer_info(cfg->prcfg_count, &cfg->prcfg_prim_nid,
2942                                         &cfg->prcfg_cfg_nid, &cfg->prcfg_mr,
2943                                         lpni_cri, lpni_stats);
2944                 mutex_unlock(&the_lnet.ln_api_mutex);
2945                 return rc;
2946         }
2947
2948         case IOC_LIBCFS_NOTIFY_ROUTER: {
2949                 unsigned long jiffies_passed;
2950
2951                 jiffies_passed = ktime_get_real_seconds() - data->ioc_u64[0];
2952                 jiffies_passed = cfs_time_seconds(jiffies_passed);
2953
2954                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
2955                                    jiffies - jiffies_passed);
2956         }
2957
2958         case IOC_LIBCFS_LNET_DIST:
2959                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
2960                 if (rc < 0 && rc != -EHOSTUNREACH)
2961                         return rc;
2962
2963                 data->ioc_u32[0] = rc;
2964                 return 0;
2965
2966         case IOC_LIBCFS_TESTPROTOCOMPAT:
2967                 lnet_net_lock(LNET_LOCK_EX);
2968                 the_lnet.ln_testprotocompat = data->ioc_flags;
2969                 lnet_net_unlock(LNET_LOCK_EX);
2970                 return 0;
2971
2972         case IOC_LIBCFS_LNET_FAULT:
2973                 return lnet_fault_ctl(data->ioc_flags, data);
2974
2975         case IOC_LIBCFS_PING: {
2976                 signed long timeout;
2977
2978                 id.nid = data->ioc_nid;
2979                 id.pid = data->ioc_u32[0];
2980
2981                 /* Don't block longer than 2 minutes */
2982                 if (data->ioc_u32[1] > 120 * MSEC_PER_SEC)
2983                         return -EINVAL;
2984
2985                 /* If timestamp is negative then disable timeout */
2986                 if ((s32)data->ioc_u32[1] < 0)
2987                         timeout = MAX_SCHEDULE_TIMEOUT;
2988                 else
2989                         timeout = msecs_to_jiffies(data->ioc_u32[1]);
2990
2991                 rc = lnet_ping(id, timeout, data->ioc_pbuf1,
2992                                data->ioc_plen1 / sizeof(struct lnet_process_id));
2993                 if (rc < 0)
2994                         return rc;
2995                 data->ioc_count = rc;
2996                 return 0;
2997         }
2998
2999         default:
3000                 ni = lnet_net2ni_addref(data->ioc_net);
3001                 if (ni == NULL)
3002                         return -EINVAL;
3003
3004                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
3005                         rc = -EINVAL;
3006                 else
3007                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
3008
3009                 lnet_ni_decref(ni);
3010                 return rc;
3011         }
3012         /* not reached */
3013 }
3014 EXPORT_SYMBOL(LNetCtl);
3015
3016 void LNetDebugPeer(struct lnet_process_id id)
3017 {
3018         lnet_debug_peer(id.nid);
3019 }
3020 EXPORT_SYMBOL(LNetDebugPeer);
3021
3022 /**
3023  * Determine if the specified peer \a nid is on the local node.
3024  *
3025  * \param nid   peer nid to check
3026  *
3027  * \retval true         If peer NID is on the local node.
3028  * \retval false        If peer NID is not on the local node.
3029  */
3030 bool LNetIsPeerLocal(lnet_nid_t nid)
3031 {
3032         struct lnet_net *net;
3033         struct lnet_ni *ni;
3034         int cpt;
3035
3036         cpt = lnet_net_lock_current();
3037         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3038                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3039                         if (ni->ni_nid == nid) {
3040                                 lnet_net_unlock(cpt);
3041                                 return true;
3042                         }
3043                 }
3044         }
3045         lnet_net_unlock(cpt);
3046
3047         return false;
3048 }
3049 EXPORT_SYMBOL(LNetIsPeerLocal);
3050
3051 /**
3052  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
3053  * Note that all interfaces share a same PID, as requested by LNetNIInit().
3054  *
3055  * \param index Index of the interface to look up.
3056  * \param id On successful return, this location will hold the
3057  * struct lnet_process_id ID of the interface.
3058  *
3059  * \retval 0 If an interface exists at \a index.
3060  * \retval -ENOENT If no interface has been found.
3061  */
3062 int
3063 LNetGetId(unsigned int index, struct lnet_process_id *id)
3064 {
3065         struct lnet_ni   *ni;
3066         struct lnet_net  *net;
3067         int               cpt;
3068         int               rc = -ENOENT;
3069
3070         LASSERT(the_lnet.ln_refcount > 0);
3071
3072         cpt = lnet_net_lock_current();
3073
3074         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3075                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3076                         if (index-- != 0)
3077                                 continue;
3078
3079                         id->nid = ni->ni_nid;
3080                         id->pid = the_lnet.ln_pid;
3081                         rc = 0;
3082                         break;
3083                 }
3084         }
3085
3086         lnet_net_unlock(cpt);
3087         return rc;
3088 }
3089 EXPORT_SYMBOL(LNetGetId);
3090
3091 static int lnet_ping(struct lnet_process_id id, signed long timeout,
3092                      struct lnet_process_id __user *ids, int n_ids)
3093 {
3094         struct lnet_handle_eq eqh;
3095         struct lnet_handle_md mdh;
3096         struct lnet_event event;
3097         struct lnet_md md = { NULL };
3098         int which;
3099         int unlinked = 0;
3100         int replied = 0;
3101         const signed long a_long_time = msecs_to_jiffies(60 * MSEC_PER_SEC);
3102         struct lnet_ping_buffer *pbuf;
3103         struct lnet_process_id tmpid;
3104         int i;
3105         int nob;
3106         int rc;
3107         int rc2;
3108         sigset_t blocked;
3109
3110         /* n_ids limit is arbitrary */
3111         if (n_ids <= 0 || n_ids > lnet_interfaces_max || id.nid == LNET_NID_ANY)
3112                 return -EINVAL;
3113
3114         if (id.pid == LNET_PID_ANY)
3115                 id.pid = LNET_PID_LUSTRE;
3116
3117         pbuf = lnet_ping_buffer_alloc(n_ids, GFP_NOFS);
3118         if (!pbuf)
3119                 return -ENOMEM;
3120
3121         /* NB 2 events max (including any unlink event) */
3122         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
3123         if (rc != 0) {
3124                 CERROR("Can't allocate EQ: %d\n", rc);
3125                 goto fail_ping_buffer_decref;
3126         }
3127
3128         /* initialize md content */
3129         md.start     = &pbuf->pb_info;
3130         md.length    = LNET_PING_INFO_SIZE(n_ids);
3131         md.threshold = 2; /*GET/REPLY*/
3132         md.max_size  = 0;
3133         md.options   = LNET_MD_TRUNCATE;
3134         md.user_ptr  = NULL;
3135         md.eq_handle = eqh;
3136
3137         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
3138         if (rc != 0) {
3139                 CERROR("Can't bind MD: %d\n", rc);
3140                 goto fail_free_eq;
3141         }
3142
3143         rc = LNetGet(LNET_NID_ANY, mdh, id,
3144                      LNET_RESERVED_PORTAL,
3145                      LNET_PROTO_PING_MATCHBITS, 0);
3146
3147         if (rc != 0) {
3148                 /* Don't CERROR; this could be deliberate! */
3149
3150                 rc2 = LNetMDUnlink(mdh);
3151                 LASSERT(rc2 == 0);
3152
3153                 /* NB must wait for the UNLINK event below... */
3154                 unlinked = 1;
3155                 timeout = a_long_time;
3156         }
3157
3158         do {
3159                 /* MUST block for unlink to complete */
3160                 if (unlinked)
3161                         blocked = cfs_block_allsigs();
3162
3163                 rc2 = LNetEQPoll(&eqh, 1, timeout, &event, &which);
3164
3165                 if (unlinked)
3166                         cfs_restore_sigs(blocked);
3167
3168                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
3169                        (rc2 <= 0) ? -1 : event.type,
3170                        (rc2 <= 0) ? -1 : event.status,
3171                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
3172
3173                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
3174
3175                 if (rc2 <= 0 || event.status != 0) {
3176                         /* timeout or error */
3177                         if (!replied && rc == 0)
3178                                 rc = (rc2 < 0) ? rc2 :
3179                                      (rc2 == 0) ? -ETIMEDOUT :
3180                                      event.status;
3181
3182                         if (!unlinked) {
3183                                 /* Ensure completion in finite time... */
3184                                 LNetMDUnlink(mdh);
3185                                 /* No assertion (racing with network) */
3186                                 unlinked = 1;
3187                                 timeout = a_long_time;
3188                         } else if (rc2 == 0) {
3189                                 /* timed out waiting for unlink */
3190                                 CWARN("ping %s: late network completion\n",
3191                                       libcfs_id2str(id));
3192                         }
3193                 } else if (event.type == LNET_EVENT_REPLY) {
3194                         replied = 1;
3195                         rc = event.mlength;
3196                 }
3197
3198         } while (rc2 <= 0 || !event.unlinked);
3199
3200         if (!replied) {
3201                 if (rc >= 0)
3202                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
3203                               libcfs_id2str(id));
3204                 rc = -EIO;
3205                 goto fail_free_eq;
3206         }
3207
3208         nob = rc;
3209         LASSERT(nob >= 0 && nob <= LNET_PING_INFO_SIZE(n_ids));
3210
3211         rc = -EPROTO;                           /* if I can't parse... */
3212
3213         if (nob < 8) {
3214                 /* can't check magic/version */
3215                 CERROR("%s: ping info too short %d\n",
3216                        libcfs_id2str(id), nob);
3217                 goto fail_free_eq;
3218         }
3219
3220         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
3221                 lnet_swap_pinginfo(pbuf);
3222         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
3223                 CERROR("%s: Unexpected magic %08x\n",
3224                        libcfs_id2str(id), pbuf->pb_info.pi_magic);
3225                 goto fail_free_eq;
3226         }
3227
3228         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
3229                 CERROR("%s: ping w/o NI status: 0x%x\n",
3230                        libcfs_id2str(id), pbuf->pb_info.pi_features);
3231                 goto fail_free_eq;
3232         }
3233
3234         if (nob < LNET_PING_INFO_SIZE(0)) {
3235                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
3236                        nob, (int)LNET_PING_INFO_SIZE(0));
3237                 goto fail_free_eq;
3238         }
3239
3240         if (pbuf->pb_info.pi_nnis < n_ids)
3241                 n_ids = pbuf->pb_info.pi_nnis;
3242
3243         if (nob < LNET_PING_INFO_SIZE(n_ids)) {
3244                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
3245                        nob, (int)LNET_PING_INFO_SIZE(n_ids));
3246                 goto fail_free_eq;
3247         }
3248
3249         rc = -EFAULT;                           /* If I SEGV... */
3250
3251         memset(&tmpid, 0, sizeof(tmpid));
3252         for (i = 0; i < n_ids; i++) {
3253                 tmpid.pid = pbuf->pb_info.pi_pid;
3254                 tmpid.nid = pbuf->pb_info.pi_ni[i].ns_nid;
3255                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
3256                         goto fail_free_eq;
3257         }
3258         rc = pbuf->pb_info.pi_nnis;
3259
3260  fail_free_eq:
3261         rc2 = LNetEQFree(eqh);
3262         if (rc2 != 0)
3263                 CERROR("rc2 %d\n", rc2);
3264         LASSERT(rc2 == 0);
3265
3266  fail_ping_buffer_decref:
3267         lnet_ping_buffer_decref(pbuf);
3268         return rc;
3269 }