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 <lnet/lib-lnet.h>
  36
  37 #define D_LNI D_CONSOLE
  38
  39 lnet_t      the_lnet;                           /* THE state of the network */
  40 EXPORT_SYMBOL(the_lnet);
  41
  42 static char *ip2nets = "";
  43 module_param(ip2nets, charp, 0444);
  44 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  45
  46 static char *networks = "";
  47 module_param(networks, charp, 0444);
  48 MODULE_PARM_DESC(networks, "local networks");
  49
  50 static char *routes = "";
  51 module_param(routes, charp, 0444);
  52 MODULE_PARM_DESC(routes, "routes to non-local networks");
  53
  54 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  55 module_param(rnet_htable_size, int, 0444);
  56 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  57
  58 static int lnet_ping(lnet_process_id_t id, int timeout_ms,
  59                      lnet_process_id_t __user *ids, int n_ids);
  60
  61 static char *
  62 lnet_get_routes(void)
  63 {
  64         return routes;
  65 }
  66
  67 static char *
  68 lnet_get_networks(void)
  69 {
  70         char   *nets;
  71         int     rc;
  72
  73         if (*networks != 0 && *ip2nets != 0) {
  74                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  75                                    "'ip2nets' but not both at once\n");
  76                 return NULL;
  77         }
  78
  79         if (*ip2nets != 0) {
  80                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  81                 return (rc == 0) ? nets : NULL;
  82         }
  83
  84         if (*networks != 0)
  85                 return networks;
  86
  87         return "tcp";
  88 }
  89
  90 static void
  91 lnet_init_locks(void)
  92 {
  93         spin_lock_init(&the_lnet.ln_eq_wait_lock);
  94         init_waitqueue_head(&the_lnet.ln_eq_waitq);
  95         init_waitqueue_head(&the_lnet.ln_rc_waitq);
  96         mutex_init(&the_lnet.ln_lnd_mutex);
  97         mutex_init(&the_lnet.ln_api_mutex);
  98 }
  99
 100 static void
 101 lnet_fini_locks(void)
 102 {
 103 }
 104
 105 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 106 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 107                                             *  MDs kmem_cache */
 108
 109 static int
 110 lnet_descriptor_setup(void)
 111 {
 112         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 113          * allocated in <size-xxx> kmem_cache).
 114          */
 115         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(lnet_me_t),
 116                                             0, 0, NULL);
 117         if (!lnet_mes_cachep)
 118                 return -ENOMEM;
 119
 120         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 121                                                   LNET_SMALL_MD_SIZE, 0, 0,
 122                                                   NULL);
 123         if (!lnet_small_mds_cachep)
 124                 return -ENOMEM;
 125
 126         return 0;
 127 }
 128
 129 static void
 130 lnet_descriptor_cleanup(void)
 131 {
 132
 133         if (lnet_small_mds_cachep) {
 134                 kmem_cache_destroy(lnet_small_mds_cachep);
 135                 lnet_small_mds_cachep = NULL;
 136         }
 137
 138         if (lnet_mes_cachep) {
 139                 kmem_cache_destroy(lnet_mes_cachep);
 140                 lnet_mes_cachep = NULL;
 141         }
 142 }
 143
 144 static int
 145 lnet_create_remote_nets_table(void)
 146 {
 147         int               i;
 148         struct list_head *hash;
 149
 150         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 151         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 152         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 153         if (hash == NULL) {
 154                 CERROR("Failed to create remote nets hash table\n");
 155                 return -ENOMEM;
 156         }
 157
 158         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 159                 INIT_LIST_HEAD(&hash[i]);
 160         the_lnet.ln_remote_nets_hash = hash;
 161         return 0;
 162 }
 163
 164 static void
 165 lnet_destroy_remote_nets_table(void)
 166 {
 167         int i;
 168
 169         if (the_lnet.ln_remote_nets_hash == NULL)
 170                 return;
 171
 172         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 173                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 174
 175         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 176                     LNET_REMOTE_NETS_HASH_SIZE *
 177                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 178         the_lnet.ln_remote_nets_hash = NULL;
 179 }
 180
 181 static void
 182 lnet_destroy_locks(void)
 183 {
 184         if (the_lnet.ln_res_lock != NULL) {
 185                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 186                 the_lnet.ln_res_lock = NULL;
 187         }
 188
 189         if (the_lnet.ln_net_lock != NULL) {
 190                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 191                 the_lnet.ln_net_lock = NULL;
 192         }
 193
 194         lnet_fini_locks();
 195 }
 196
 197 static int
 198 lnet_create_locks(void)
 199 {
 200         lnet_init_locks();
 201
 202         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 203         if (the_lnet.ln_res_lock == NULL)
 204                 goto failed;
 205
 206         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 207         if (the_lnet.ln_net_lock == NULL)
 208                 goto failed;
 209
 210         return 0;
 211
 212  failed:
 213         lnet_destroy_locks();
 214         return -ENOMEM;
 215 }
 216
 217 static void lnet_assert_wire_constants(void)
 218 {
 219         /* Wire protocol assertions generated by 'wirecheck'
 220          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 221          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 222          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 223
 224         /* Constants... */
 225         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 226         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 227         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 228         CLASSERT(LNET_MSG_ACK == 0);
 229         CLASSERT(LNET_MSG_PUT == 1);
 230         CLASSERT(LNET_MSG_GET == 2);
 231         CLASSERT(LNET_MSG_REPLY == 3);
 232         CLASSERT(LNET_MSG_HELLO == 4);
 233
 234         /* Checks for struct ptl_handle_wire_t */
 235         CLASSERT((int)sizeof(lnet_handle_wire_t) == 16);
 236         CLASSERT((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 237         CLASSERT((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 238         CLASSERT((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 239         CLASSERT((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 240
 241         /* Checks for struct lnet_magicversion_t */
 242         CLASSERT((int)sizeof(lnet_magicversion_t) == 8);
 243         CLASSERT((int)offsetof(lnet_magicversion_t, magic) == 0);
 244         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 245         CLASSERT((int)offsetof(lnet_magicversion_t, version_major) == 4);
 246         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 247         CLASSERT((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 248         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 249
 250         /* Checks for struct lnet_hdr_t */
 251         CLASSERT((int)sizeof(lnet_hdr_t) == 72);
 252         CLASSERT((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 253         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 254         CLASSERT((int)offsetof(lnet_hdr_t, src_nid) == 8);
 255         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 256         CLASSERT((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 257         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 258         CLASSERT((int)offsetof(lnet_hdr_t, src_pid) == 20);
 259         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 260         CLASSERT((int)offsetof(lnet_hdr_t, type) == 24);
 261         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 262         CLASSERT((int)offsetof(lnet_hdr_t, payload_length) == 28);
 263         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 264         CLASSERT((int)offsetof(lnet_hdr_t, msg) == 32);
 265         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 266
 267         /* Ack */
 268         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 269         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 270         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 271         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 272         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 273         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 274
 275         /* Put */
 276         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 277         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 278         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 279         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 280         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 281         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 282         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 283         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 284         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 285         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 286
 287         /* Get */
 288         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 289         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 290         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 291         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 292         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 293         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 294         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 295         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 296         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 297         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 298
 299         /* Reply */
 300         CLASSERT((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 301         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 302
 303         /* Hello */
 304         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 305         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 306         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 307         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 308 }
 309
 310 static lnd_t *lnet_find_lnd_by_type(__u32 type)
 311 {
 312         lnd_t            *lnd;
 313         struct list_head *tmp;
 314
 315         /* holding lnd mutex */
 316         list_for_each(tmp, &the_lnet.ln_lnds) {
 317                 lnd = list_entry(tmp, lnd_t, lnd_list);
 318
 319                 if (lnd->lnd_type == type)
 320                         return lnd;
 321         }
 322         return NULL;
 323 }
 324
 325 void
 326 lnet_register_lnd (lnd_t *lnd)
 327 {
 328         mutex_lock(&the_lnet.ln_lnd_mutex);
 329
 330         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 331         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 332
 333         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 334         lnd->lnd_refcount = 0;
 335
 336         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 337
 338         mutex_unlock(&the_lnet.ln_lnd_mutex);
 339 }
 340 EXPORT_SYMBOL(lnet_register_lnd);
 341
 342 void
 343 lnet_unregister_lnd (lnd_t *lnd)
 344 {
 345         mutex_lock(&the_lnet.ln_lnd_mutex);
 346
 347         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 348         LASSERT(lnd->lnd_refcount == 0);
 349
 350         list_del(&lnd->lnd_list);
 351         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 352
 353         mutex_unlock(&the_lnet.ln_lnd_mutex);
 354 }
 355 EXPORT_SYMBOL(lnet_unregister_lnd);
 356
 357 void
 358 lnet_counters_get(lnet_counters_t *counters)
 359 {
 360         lnet_counters_t *ctr;
 361         int             i;
 362
 363         memset(counters, 0, sizeof(*counters));
 364
 365         lnet_net_lock(LNET_LOCK_EX);
 366
 367         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 368                 counters->msgs_max     += ctr->msgs_max;
 369                 counters->msgs_alloc   += ctr->msgs_alloc;
 370                 counters->errors       += ctr->errors;
 371                 counters->send_count   += ctr->send_count;
 372                 counters->recv_count   += ctr->recv_count;
 373                 counters->route_count  += ctr->route_count;
 374                 counters->drop_count   += ctr->drop_count;
 375                 counters->send_length  += ctr->send_length;
 376                 counters->recv_length  += ctr->recv_length;
 377                 counters->route_length += ctr->route_length;
 378                 counters->drop_length  += ctr->drop_length;
 379
 380         }
 381         lnet_net_unlock(LNET_LOCK_EX);
 382 }
 383 EXPORT_SYMBOL(lnet_counters_get);
 384
 385 void
 386 lnet_counters_reset(void)
 387 {
 388         lnet_counters_t *counters;
 389         int             i;
 390
 391         lnet_net_lock(LNET_LOCK_EX);
 392
 393         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 394                 memset(counters, 0, sizeof(lnet_counters_t));
 395
 396         lnet_net_unlock(LNET_LOCK_EX);
 397 }
 398
 399 static __u64 lnet_create_interface_cookie(void)
 400 {
 401         /* NB the interface cookie in wire handles guards against delayed
 402          * replies and ACKs appearing valid after reboot. Initialisation time,
 403          * even if it's only implemented to millisecond resolution is probably
 404          * easily good enough. */
 405         struct timeval tv;
 406         __u64          cookie;
 407         do_gettimeofday(&tv);
 408         cookie = tv.tv_sec;
 409         cookie *= 1000000;
 410         cookie += tv.tv_usec;
 411         return cookie;
 412 }
 413
 414 static char *
 415 lnet_res_type2str(int type)
 416 {
 417         switch (type) {
 418         default:
 419                 LBUG();
 420         case LNET_COOKIE_TYPE_MD:
 421                 return "MD";
 422         case LNET_COOKIE_TYPE_ME:
 423                 return "ME";
 424         case LNET_COOKIE_TYPE_EQ:
 425                 return "EQ";
 426         }
 427 }
 428
 429 static void
 430 lnet_res_container_cleanup(struct lnet_res_container *rec)
 431 {
 432         int     count = 0;
 433
 434         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 435                 return;
 436
 437         while (!list_empty(&rec->rec_active)) {
 438                 struct list_head *e = rec->rec_active.next;
 439
 440                 list_del_init(e);
 441                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 442                         lnet_eq_free(list_entry(e, lnet_eq_t, eq_list));
 443
 444                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 445                         lnet_md_free(list_entry(e, lnet_libmd_t, md_list));
 446
 447                 } else { /* NB: Active MEs should be attached on portals */
 448                         LBUG();
 449                 }
 450                 count++;
 451         }
 452
 453         if (count > 0) {
 454                 /* Found alive MD/ME/EQ, user really should unlink/free
 455                  * all of them before finalize LNet, but if someone didn't,
 456                  * we have to recycle garbage for him */
 457                 CERROR("%d active elements on exit of %s container\n",
 458                        count, lnet_res_type2str(rec->rec_type));
 459         }
 460
 461         if (rec->rec_lh_hash != NULL) {
 462                 LIBCFS_FREE(rec->rec_lh_hash,
 463                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 464                 rec->rec_lh_hash = NULL;
 465         }
 466
 467         rec->rec_type = 0; /* mark it as finalized */
 468 }
 469
 470 static int
 471 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 472 {
 473         int     rc = 0;
 474         int     i;
 475
 476         LASSERT(rec->rec_type == 0);
 477
 478         rec->rec_type = type;
 479         INIT_LIST_HEAD(&rec->rec_active);
 480
 481         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 482
 483         /* Arbitrary choice of hash table size */
 484         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 485                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 486         if (rec->rec_lh_hash == NULL) {
 487                 rc = -ENOMEM;
 488                 goto out;
 489         }
 490
 491         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 492                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 493
 494         return 0;
 495
 496 out:
 497         CERROR("Failed to setup %s resource container\n",
 498                lnet_res_type2str(type));
 499         lnet_res_container_cleanup(rec);
 500         return rc;
 501 }
 502
 503 static void
 504 lnet_res_containers_destroy(struct lnet_res_container **recs)
 505 {
 506         struct lnet_res_container       *rec;
 507         int                             i;
 508
 509         cfs_percpt_for_each(rec, i, recs)
 510                 lnet_res_container_cleanup(rec);
 511
 512         cfs_percpt_free(recs);
 513 }
 514
 515 static struct lnet_res_container **
 516 lnet_res_containers_create(int type)
 517 {
 518         struct lnet_res_container       **recs;
 519         struct lnet_res_container       *rec;
 520         int                             rc;
 521         int                             i;
 522
 523         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 524         if (recs == NULL) {
 525                 CERROR("Failed to allocate %s resource containers\n",
 526                        lnet_res_type2str(type));
 527                 return NULL;
 528         }
 529
 530         cfs_percpt_for_each(rec, i, recs) {
 531                 rc = lnet_res_container_setup(rec, i, type);
 532                 if (rc != 0) {
 533                         lnet_res_containers_destroy(recs);
 534                         return NULL;
 535                 }
 536         }
 537
 538         return recs;
 539 }
 540
 541 lnet_libhandle_t *
 542 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 543 {
 544         /* ALWAYS called with lnet_res_lock held */
 545         struct list_head        *head;
 546         lnet_libhandle_t        *lh;
 547         unsigned int            hash;
 548
 549         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 550                 return NULL;
 551
 552         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 553         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 554
 555         list_for_each_entry(lh, head, lh_hash_chain) {
 556                 if (lh->lh_cookie == cookie)
 557                         return lh;
 558         }
 559
 560         return NULL;
 561 }
 562
 563 void
 564 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 565 {
 566         /* ALWAYS called with lnet_res_lock held */
 567         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 568         unsigned int    hash;
 569
 570         lh->lh_cookie = rec->rec_lh_cookie;
 571         rec->rec_lh_cookie += 1 << ibits;
 572
 573         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 574
 575         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 576 }
 577
 578 static int lnet_unprepare(void);
 579
 580 static int
 581 lnet_prepare(lnet_pid_t requested_pid)
 582 {
 583         /* Prepare to bring up the network */
 584         struct lnet_res_container **recs;
 585         int                       rc = 0;
 586
 587         if (requested_pid == LNET_PID_ANY) {
 588                 /* Don't instantiate LNET just for me */
 589                 return -ENETDOWN;
 590         }
 591
 592         LASSERT(the_lnet.ln_refcount == 0);
 593
 594         the_lnet.ln_routing = 0;
 595
 596         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
 597         the_lnet.ln_pid = requested_pid;
 598
 599         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 600         INIT_LIST_HEAD(&the_lnet.ln_nis);
 601         INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 602         INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 603         INIT_LIST_HEAD(&the_lnet.ln_routers);
 604         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
 605         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
 606
 607         rc = lnet_descriptor_setup();
 608         if (rc != 0)
 609                 goto failed;
 610
 611         rc = lnet_create_remote_nets_table();
 612         if (rc != 0)
 613                 goto failed;
 614
 615         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 616
 617         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 618                                                 sizeof(lnet_counters_t));
 619         if (the_lnet.ln_counters == NULL) {
 620                 CERROR("Failed to allocate counters for LNet\n");
 621                 rc = -ENOMEM;
 622                 goto failed;
 623         }
 624
 625         rc = lnet_peer_tables_create();
 626         if (rc != 0)
 627                 goto failed;
 628
 629         rc = lnet_msg_containers_create();
 630         if (rc != 0)
 631                 goto failed;
 632
 633         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 634                                       LNET_COOKIE_TYPE_EQ);
 635         if (rc != 0)
 636                 goto failed;
 637
 638         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
 639         if (recs == NULL) {
 640                 rc = -ENOMEM;
 641                 goto failed;
 642         }
 643
 644         the_lnet.ln_me_containers = recs;
 645
 646         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
 647         if (recs == NULL) {
 648                 rc = -ENOMEM;
 649                 goto failed;
 650         }
 651
 652         the_lnet.ln_md_containers = recs;
 653
 654         rc = lnet_portals_create();
 655         if (rc != 0) {
 656                 CERROR("Failed to create portals for LNet: %d\n", rc);
 657                 goto failed;
 658         }
 659
 660         return 0;
 661
 662  failed:
 663         lnet_unprepare();
 664         return rc;
 665 }
 666
 667 static int
 668 lnet_unprepare (void)
 669 {
 670         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 671          * have shut down already, so it is safe to unlink and free all
 672          * descriptors, even those that appear committed to a network op (eg MD
 673          * with non-zero pending count) */
 674
 675         lnet_fail_nid(LNET_NID_ANY, 0);
 676
 677         LASSERT(the_lnet.ln_refcount == 0);
 678         LASSERT(list_empty(&the_lnet.ln_test_peers));
 679         LASSERT(list_empty(&the_lnet.ln_nis));
 680         LASSERT(list_empty(&the_lnet.ln_nis_cpt));
 681         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
 682
 683         lnet_portals_destroy();
 684
 685         if (the_lnet.ln_md_containers != NULL) {
 686                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 687                 the_lnet.ln_md_containers = NULL;
 688         }
 689
 690         if (the_lnet.ln_me_containers != NULL) {
 691                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 692                 the_lnet.ln_me_containers = NULL;
 693         }
 694
 695         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 696
 697         lnet_msg_containers_destroy();
 698         lnet_peer_tables_destroy();
 699         lnet_rtrpools_free(0);
 700
 701         if (the_lnet.ln_counters != NULL) {
 702                 cfs_percpt_free(the_lnet.ln_counters);
 703                 the_lnet.ln_counters = NULL;
 704         }
 705         lnet_destroy_remote_nets_table();
 706         lnet_descriptor_cleanup();
 707
 708         return 0;
 709 }
 710
 711 lnet_ni_t  *
 712 lnet_net2ni_locked(__u32 net, int cpt)
 713 {
 714         struct list_head *tmp;
 715         lnet_ni_t        *ni;
 716
 717         LASSERT(cpt != LNET_LOCK_EX);
 718
 719         list_for_each(tmp, &the_lnet.ln_nis) {
 720                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 721
 722                 if (LNET_NIDNET(ni->ni_nid) == net) {
 723                         lnet_ni_addref_locked(ni, cpt);
 724                         return ni;
 725                 }
 726         }
 727
 728         return NULL;
 729 }
 730
 731 lnet_ni_t *
 732 lnet_net2ni(__u32 net)
 733 {
 734         lnet_ni_t *ni;
 735
 736         lnet_net_lock(0);
 737         ni = lnet_net2ni_locked(net, 0);
 738         lnet_net_unlock(0);
 739
 740         return ni;
 741 }
 742 EXPORT_SYMBOL(lnet_net2ni);
 743
 744 static unsigned int
 745 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 746 {
 747         __u64           key = nid;
 748         unsigned int    val;
 749
 750         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 751
 752         if (number == 1)
 753                 return 0;
 754
 755         val = hash_long(key, LNET_CPT_BITS);
 756         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 757         if (val < number)
 758                 return val;
 759
 760         return (unsigned int)(key + val + (val >> 1)) % number;
 761 }
 762
 763 int
 764 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 765 {
 766         struct lnet_ni *ni;
 767
 768         /* must called with hold of lnet_net_lock */
 769         if (LNET_CPT_NUMBER == 1)
 770                 return 0; /* the only one */
 771
 772         /* take lnet_net_lock(any) would be OK */
 773         if (!list_empty(&the_lnet.ln_nis_cpt)) {
 774                 list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 775                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 776                                 continue;
 777
 778                         LASSERT(ni->ni_cpts != NULL);
 779                         return ni->ni_cpts[lnet_nid_cpt_hash
 780                                            (nid, ni->ni_ncpts)];
 781                 }
 782         }
 783
 784         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 785 }
 786
 787 int
 788 lnet_cpt_of_nid(lnet_nid_t nid)
 789 {
 790         int     cpt;
 791         int     cpt2;
 792
 793         if (LNET_CPT_NUMBER == 1)
 794                 return 0; /* the only one */
 795
 796         if (list_empty(&the_lnet.ln_nis_cpt))
 797                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 798
 799         cpt = lnet_net_lock_current();
 800         cpt2 = lnet_cpt_of_nid_locked(nid);
 801         lnet_net_unlock(cpt);
 802
 803         return cpt2;
 804 }
 805 EXPORT_SYMBOL(lnet_cpt_of_nid);
 806
 807 int
 808 lnet_islocalnet(__u32 net)
 809 {
 810         struct lnet_ni  *ni;
 811         int             cpt;
 812
 813         cpt = lnet_net_lock_current();
 814
 815         ni = lnet_net2ni_locked(net, cpt);
 816         if (ni != NULL)
 817                 lnet_ni_decref_locked(ni, cpt);
 818
 819         lnet_net_unlock(cpt);
 820
 821         return ni != NULL;
 822 }
 823
 824 lnet_ni_t  *
 825 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 826 {
 827         struct lnet_ni   *ni;
 828         struct list_head *tmp;
 829
 830         LASSERT(cpt != LNET_LOCK_EX);
 831
 832         list_for_each(tmp, &the_lnet.ln_nis) {
 833                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 834
 835                 if (ni->ni_nid == nid) {
 836                         lnet_ni_addref_locked(ni, cpt);
 837                         return ni;
 838                 }
 839         }
 840
 841         return NULL;
 842 }
 843
 844 int
 845 lnet_islocalnid(lnet_nid_t nid)
 846 {
 847         struct lnet_ni  *ni;
 848         int             cpt;
 849
 850         cpt = lnet_net_lock_current();
 851         ni = lnet_nid2ni_locked(nid, cpt);
 852         if (ni != NULL)
 853                 lnet_ni_decref_locked(ni, cpt);
 854         lnet_net_unlock(cpt);
 855
 856         return ni != NULL;
 857 }
 858
 859 int
 860 lnet_count_acceptor_nis (void)
 861 {
 862         /* Return the # of NIs that need the acceptor. */
 863         int              count = 0;
 864         struct list_head *tmp;
 865         struct lnet_ni   *ni;
 866         int              cpt;
 867
 868         cpt = lnet_net_lock_current();
 869         list_for_each(tmp, &the_lnet.ln_nis) {
 870                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 871
 872                 if (ni->ni_lnd->lnd_accept != NULL)
 873                         count++;
 874         }
 875
 876         lnet_net_unlock(cpt);
 877
 878         return count;
 879 }
 880
 881 static lnet_ping_info_t *
 882 lnet_ping_info_create(int num_ni)
 883 {
 884         lnet_ping_info_t *ping_info;
 885         unsigned int     infosz;
 886
 887         infosz = offsetof(lnet_ping_info_t, pi_ni[num_ni]);
 888         LIBCFS_ALLOC(ping_info, infosz);
 889         if (ping_info == NULL) {
 890                 CERROR("Can't allocate ping info[%d]\n", num_ni);
 891                 return NULL;
 892         }
 893
 894         ping_info->pi_nnis = num_ni;
 895         ping_info->pi_pid = the_lnet.ln_pid;
 896         ping_info->pi_magic = LNET_PROTO_PING_MAGIC;
 897         ping_info->pi_features = LNET_PING_FEAT_NI_STATUS;
 898
 899         return ping_info;
 900 }
 901
 902 static inline int
 903 lnet_get_ni_count(void)
 904 {
 905         struct lnet_ni *ni;
 906         int            count = 0;
 907
 908         lnet_net_lock(0);
 909
 910         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list)
 911                 count++;
 912
 913         lnet_net_unlock(0);
 914
 915         return count;
 916 }
 917
 918 static inline void
 919 lnet_ping_info_free(lnet_ping_info_t *pinfo)
 920 {
 921         LIBCFS_FREE(pinfo,
 922                     offsetof(lnet_ping_info_t,
 923                              pi_ni[pinfo->pi_nnis]));
 924 }
 925
 926 static void
 927 lnet_ping_info_destroy(void)
 928 {
 929         struct lnet_ni  *ni;
 930
 931         lnet_net_lock(LNET_LOCK_EX);
 932
 933         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 934                 lnet_ni_lock(ni);
 935                 ni->ni_status = NULL;
 936                 lnet_ni_unlock(ni);
 937         }
 938
 939         lnet_ping_info_free(the_lnet.ln_ping_info);
 940         the_lnet.ln_ping_info = NULL;
 941
 942         lnet_net_unlock(LNET_LOCK_EX);
 943 }
 944
 945 static void
 946 lnet_ping_event_handler(lnet_event_t *event)
 947 {
 948         lnet_ping_info_t *pinfo = event->md.user_ptr;
 949
 950         if (event->unlinked)
 951                 pinfo->pi_features = LNET_PING_FEAT_INVAL;
 952 }
 953
 954 static int
 955 lnet_ping_info_setup(lnet_ping_info_t **ppinfo, lnet_handle_md_t *md_handle,
 956                      int ni_count, bool set_eq)
 957 {
 958         lnet_handle_me_t  me_handle;
 959         lnet_process_id_t id = {LNET_NID_ANY, LNET_PID_ANY};
 960         lnet_md_t         md = {NULL};
 961         int               rc, rc2;
 962
 963         if (set_eq) {
 964                 rc = LNetEQAlloc(0, lnet_ping_event_handler,
 965                                  &the_lnet.ln_ping_target_eq);
 966                 if (rc != 0) {
 967                         CERROR("Can't allocate ping EQ: %d\n", rc);
 968                         return rc;
 969                 }
 970         }
 971
 972         *ppinfo = lnet_ping_info_create(ni_count);
 973         if (*ppinfo == NULL) {
 974                 rc = -ENOMEM;
 975                 goto failed_0;
 976         }
 977
 978         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
 979                           LNET_PROTO_PING_MATCHBITS, 0,
 980                           LNET_UNLINK, LNET_INS_AFTER,
 981                           &me_handle);
 982         if (rc != 0) {
 983                 CERROR("Can't create ping ME: %d\n", rc);
 984                 goto failed_1;
 985         }
 986
 987         /* initialize md content */
 988         md.start     = *ppinfo;
 989         md.length    = offsetof(lnet_ping_info_t,
 990                                 pi_ni[(*ppinfo)->pi_nnis]);
 991         md.threshold = LNET_MD_THRESH_INF;
 992         md.max_size  = 0;
 993         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
 994                        LNET_MD_MANAGE_REMOTE;
 995         md.user_ptr  = NULL;
 996         md.eq_handle = the_lnet.ln_ping_target_eq;
 997         md.user_ptr = *ppinfo;
 998
 999         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, md_handle);
1000         if (rc != 0) {
1001                 CERROR("Can't attach ping MD: %d\n", rc);
1002                 goto failed_2;
1003         }
1004
1005         return 0;
1006
1007 failed_2:
1008         rc2 = LNetMEUnlink(me_handle);
1009         LASSERT(rc2 == 0);
1010 failed_1:
1011         lnet_ping_info_free(*ppinfo);
1012         *ppinfo = NULL;
1013 failed_0:
1014         if (set_eq)
1015                 LNetEQFree(the_lnet.ln_ping_target_eq);
1016         return rc;
1017 }
1018
1019 static void
1020 lnet_ping_md_unlink(lnet_ping_info_t *pinfo, lnet_handle_md_t *md_handle)
1021 {
1022         sigset_t        blocked = cfs_block_allsigs();
1023
1024         LNetMDUnlink(*md_handle);
1025         LNetInvalidateHandle(md_handle);
1026
1027         /* NB md could be busy; this just starts the unlink */
1028         while (pinfo->pi_features != LNET_PING_FEAT_INVAL) {
1029                 CDEBUG(D_NET, "Still waiting for ping MD to unlink\n");
1030                 set_current_state(TASK_UNINTERRUPTIBLE);
1031                 schedule_timeout(cfs_time_seconds(1));
1032         }
1033
1034         cfs_restore_sigs(blocked);
1035 }
1036
1037 static void
1038 lnet_ping_info_install_locked(lnet_ping_info_t *ping_info)
1039 {
1040         int                     i;
1041         lnet_ni_t               *ni;
1042         lnet_ni_status_t        *ns;
1043
1044         i = 0;
1045         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1046                 LASSERT(i < ping_info->pi_nnis);
1047
1048                 ns = &ping_info->pi_ni[i];
1049
1050                 ns->ns_nid = ni->ni_nid;
1051
1052                 lnet_ni_lock(ni);
1053                 ns->ns_status = (ni->ni_status != NULL) ?
1054                                 ni->ni_status->ns_status : LNET_NI_STATUS_UP;
1055                 ni->ni_status = ns;
1056                 lnet_ni_unlock(ni);
1057
1058                 i++;
1059         }
1060 }
1061
1062 static void
1063 lnet_ping_target_update(lnet_ping_info_t *pinfo, lnet_handle_md_t md_handle)
1064 {
1065         lnet_ping_info_t *old_pinfo = NULL;
1066         lnet_handle_md_t old_md;
1067
1068         /* switch the NIs to point to the new ping info created */
1069         lnet_net_lock(LNET_LOCK_EX);
1070
1071         if (!the_lnet.ln_routing)
1072                 pinfo->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1073         lnet_ping_info_install_locked(pinfo);
1074
1075         if (the_lnet.ln_ping_info != NULL) {
1076                 old_pinfo = the_lnet.ln_ping_info;
1077                 old_md = the_lnet.ln_ping_target_md;
1078         }
1079         the_lnet.ln_ping_target_md = md_handle;
1080         the_lnet.ln_ping_info = pinfo;
1081
1082         lnet_net_unlock(LNET_LOCK_EX);
1083
1084         if (old_pinfo != NULL) {
1085                 /* unlink the old ping info */
1086                 lnet_ping_md_unlink(old_pinfo, &old_md);
1087                 lnet_ping_info_free(old_pinfo);
1088         }
1089 }
1090
1091 static void
1092 lnet_ping_target_fini(void)
1093 {
1094         int             rc;
1095
1096         lnet_ping_md_unlink(the_lnet.ln_ping_info,
1097                             &the_lnet.ln_ping_target_md);
1098
1099         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1100         LASSERT(rc == 0);
1101
1102         lnet_ping_info_destroy();
1103 }
1104
1105 static int
1106 lnet_ni_tq_credits(lnet_ni_t *ni)
1107 {
1108         int     credits;
1109
1110         LASSERT(ni->ni_ncpts >= 1);
1111
1112         if (ni->ni_ncpts == 1)
1113                 return ni->ni_maxtxcredits;
1114
1115         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
1116         credits = max(credits, 8 * ni->ni_peertxcredits);
1117         credits = min(credits, ni->ni_maxtxcredits);
1118
1119         return credits;
1120 }
1121
1122 static void
1123 lnet_ni_unlink_locked(lnet_ni_t *ni)
1124 {
1125         if (!list_empty(&ni->ni_cptlist)) {
1126                 list_del_init(&ni->ni_cptlist);
1127                 lnet_ni_decref_locked(ni, 0);
1128         }
1129
1130         /* move it to zombie list and nobody can find it anymore */
1131         LASSERT(!list_empty(&ni->ni_list));
1132         list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1133         lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1134 }
1135
1136 static void
1137 lnet_clear_zombies_nis_locked(void)
1138 {
1139         int             i;
1140         int             islo;
1141         lnet_ni_t       *ni;
1142
1143         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1144          * and shut them down in guaranteed thread context */
1145         i = 2;
1146         while (!list_empty(&the_lnet.ln_nis_zombie)) {
1147                 int     *ref;
1148                 int     j;
1149
1150                 ni = list_entry(the_lnet.ln_nis_zombie.next,
1151                                 lnet_ni_t, ni_list);
1152                 list_del_init(&ni->ni_list);
1153                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1154                         if (*ref == 0)
1155                                 continue;
1156                         /* still busy, add it back to zombie list */
1157                         list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1158                         break;
1159                 }
1160
1161                 if (!list_empty(&ni->ni_list)) {
1162                         lnet_net_unlock(LNET_LOCK_EX);
1163                         ++i;
1164                         if ((i & (-i)) == i) {
1165                                 CDEBUG(D_WARNING,
1166                                        "Waiting for zombie LNI %s\n",
1167                                        libcfs_nid2str(ni->ni_nid));
1168                         }
1169                         set_current_state(TASK_UNINTERRUPTIBLE);
1170                         schedule_timeout(cfs_time_seconds(1));
1171                         lnet_net_lock(LNET_LOCK_EX);
1172                         continue;
1173                 }
1174
1175                 ni->ni_lnd->lnd_refcount--;
1176                 lnet_net_unlock(LNET_LOCK_EX);
1177
1178                 islo = ni->ni_lnd->lnd_type == LOLND;
1179
1180                 LASSERT(!in_interrupt());
1181                 (ni->ni_lnd->lnd_shutdown)(ni);
1182
1183                 /* can't deref lnd anymore now; it might have unregistered
1184                  * itself...  */
1185
1186                 if (!islo)
1187                         CDEBUG(D_LNI, "Removed LNI %s\n",
1188                               libcfs_nid2str(ni->ni_nid));
1189
1190                 lnet_ni_free(ni);
1191                 i = 2;
1192                 lnet_net_lock(LNET_LOCK_EX);
1193         }
1194 }
1195
1196 static void
1197 lnet_shutdown_lndnis(void)
1198 {
1199         int             i;
1200         lnet_ni_t       *ni;
1201
1202         /* NB called holding the global mutex */
1203
1204         /* All quiet on the API front */
1205         LASSERT(!the_lnet.ln_shutdown);
1206         LASSERT(the_lnet.ln_refcount == 0);
1207         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
1208
1209         lnet_net_lock(LNET_LOCK_EX);
1210         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1211
1212         /* Unlink NIs from the global table */
1213         while (!list_empty(&the_lnet.ln_nis)) {
1214                 ni = list_entry(the_lnet.ln_nis.next,
1215                                 lnet_ni_t, ni_list);
1216                 lnet_ni_unlink_locked(ni);
1217         }
1218
1219         /* Drop the cached loopback NI. */
1220         if (the_lnet.ln_loni != NULL) {
1221                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1222                 the_lnet.ln_loni = NULL;
1223         }
1224
1225         lnet_net_unlock(LNET_LOCK_EX);
1226
1227         /* Clear lazy portals and drop delayed messages which hold refs
1228          * on their lnet_msg_t::msg_rxpeer */
1229         for (i = 0; i < the_lnet.ln_nportals; i++)
1230                 LNetClearLazyPortal(i);
1231
1232         /* Clear the peer table and wait for all peers to go (they hold refs on
1233          * their NIs) */
1234         lnet_peer_tables_cleanup(NULL);
1235
1236         lnet_net_lock(LNET_LOCK_EX);
1237
1238         lnet_clear_zombies_nis_locked();
1239         the_lnet.ln_shutdown = 0;
1240         lnet_net_unlock(LNET_LOCK_EX);
1241 }
1242
1243 /* shutdown down the NI and release refcount */
1244 static void
1245 lnet_shutdown_lndni(struct lnet_ni *ni)
1246 {
1247         int i;
1248
1249         lnet_net_lock(LNET_LOCK_EX);
1250         lnet_ni_unlink_locked(ni);
1251         lnet_net_unlock(LNET_LOCK_EX);
1252
1253         /* clear messages for this NI on the lazy portal */
1254         for (i = 0; i < the_lnet.ln_nportals; i++)
1255                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
1256
1257         /* Do peer table cleanup for this ni */
1258         lnet_peer_tables_cleanup(ni);
1259
1260         lnet_net_lock(LNET_LOCK_EX);
1261         lnet_clear_zombies_nis_locked();
1262         lnet_net_unlock(LNET_LOCK_EX);
1263 }
1264
1265 static int
1266 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_ioctl_config_data *conf)
1267 {
1268         struct lnet_ioctl_config_lnd_tunables *lnd_tunables = NULL;
1269         int                     rc = -EINVAL;
1270         __u32                   lnd_type;
1271         lnd_t                   *lnd;
1272         struct lnet_tx_queue    *tq;
1273         int                     i;
1274
1275         lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1276
1277         LASSERT(libcfs_isknown_lnd(lnd_type));
1278
1279         if (lnd_type == CIBLND || lnd_type == OPENIBLND ||
1280             lnd_type == IIBLND || lnd_type == VIBLND) {
1281                 CERROR("LND %s obsoleted\n", libcfs_lnd2str(lnd_type));
1282                 goto failed0;
1283         }
1284
1285         /* Make sure this new NI is unique. */
1286         lnet_net_lock(LNET_LOCK_EX);
1287         rc = lnet_net_unique(LNET_NIDNET(ni->ni_nid), &the_lnet.ln_nis);
1288         lnet_net_unlock(LNET_LOCK_EX);
1289
1290         if (!rc) {
1291                 if (lnd_type == LOLND) {
1292                         lnet_ni_free(ni);
1293                         return 0;
1294                 }
1295
1296                 CERROR("Net %s is not unique\n",
1297                        libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
1298
1299                 rc = -EEXIST;
1300                 goto failed0;
1301         }
1302
1303         mutex_lock(&the_lnet.ln_lnd_mutex);
1304         lnd = lnet_find_lnd_by_type(lnd_type);
1305
1306         if (lnd == NULL) {
1307                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1308                 rc = request_module("%s", libcfs_lnd2modname(lnd_type));
1309                 mutex_lock(&the_lnet.ln_lnd_mutex);
1310
1311                 lnd = lnet_find_lnd_by_type(lnd_type);
1312                 if (lnd == NULL) {
1313                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1314                         CERROR("Can't load LND %s, module %s, rc=%d\n",
1315                                libcfs_lnd2str(lnd_type),
1316                                libcfs_lnd2modname(lnd_type), rc);
1317 #ifndef HAVE_MODULE_LOADING_SUPPORT
1318                         LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1319                                            "compiled with kernel module "
1320                                            "loading support.");
1321 #endif
1322                         rc = -EINVAL;
1323                         goto failed0;
1324                 }
1325         }
1326
1327         lnet_net_lock(LNET_LOCK_EX);
1328         lnd->lnd_refcount++;
1329         lnet_net_unlock(LNET_LOCK_EX);
1330
1331         ni->ni_lnd = lnd;
1332
1333         if (conf && conf->cfg_hdr.ioc_len > sizeof(*conf))
1334                 lnd_tunables = (struct lnet_ioctl_config_lnd_tunables *)conf->cfg_bulk;
1335
1336         if (lnd_tunables != NULL) {
1337                 LIBCFS_ALLOC(ni->ni_lnd_tunables,
1338                              sizeof(*ni->ni_lnd_tunables));
1339                 if (ni->ni_lnd_tunables == NULL) {
1340                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1341                         rc = -ENOMEM;
1342                         goto failed0;
1343                 }
1344                 memcpy(ni->ni_lnd_tunables, lnd_tunables,
1345                        sizeof(*ni->ni_lnd_tunables));
1346         }
1347
1348         /* If given some LND tunable parameters, parse those now to
1349          * override the values in the NI structure. */
1350         if (conf) {
1351                 if (conf->cfg_config_u.cfg_net.net_peer_rtr_credits >= 0)
1352                         ni->ni_peerrtrcredits =
1353                                 conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
1354                 if (conf->cfg_config_u.cfg_net.net_peer_timeout >= 0)
1355                         ni->ni_peertimeout =
1356                                 conf->cfg_config_u.cfg_net.net_peer_timeout;
1357                 if (conf->cfg_config_u.cfg_net.net_peer_tx_credits >= 0)
1358                         ni->ni_peertxcredits =
1359                                 conf->cfg_config_u.cfg_net.net_peer_tx_credits;
1360                 if (conf->cfg_config_u.cfg_net.net_max_tx_credits >= 0)
1361                         ni->ni_maxtxcredits =
1362                                 conf->cfg_config_u.cfg_net.net_max_tx_credits;
1363         }
1364
1365         rc = (lnd->lnd_startup)(ni);
1366
1367         mutex_unlock(&the_lnet.ln_lnd_mutex);
1368
1369         if (rc != 0) {
1370                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1371                                    rc, libcfs_lnd2str(lnd->lnd_type));
1372                 lnet_net_lock(LNET_LOCK_EX);
1373                 lnd->lnd_refcount--;
1374                 lnet_net_unlock(LNET_LOCK_EX);
1375                 goto failed0;
1376         }
1377
1378         LASSERT(ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1379
1380         lnet_net_lock(LNET_LOCK_EX);
1381         /* refcount for ln_nis */
1382         lnet_ni_addref_locked(ni, 0);
1383         list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1384         if (ni->ni_cpts != NULL) {
1385                 lnet_ni_addref_locked(ni, 0);
1386                 list_add_tail(&ni->ni_cptlist, &the_lnet.ln_nis_cpt);
1387         }
1388
1389         lnet_net_unlock(LNET_LOCK_EX);
1390
1391         if (lnd->lnd_type == LOLND) {
1392                 lnet_ni_addref(ni);
1393                 LASSERT(the_lnet.ln_loni == NULL);
1394                 the_lnet.ln_loni = ni;
1395                 return 0;
1396         }
1397
1398         if (ni->ni_peertxcredits == 0 || ni->ni_maxtxcredits == 0) {
1399                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1400                                    libcfs_lnd2str(lnd->lnd_type),
1401                                    ni->ni_peertxcredits == 0 ?
1402                                         "" : "per-peer ");
1403                 /* shutdown the NI since if we get here then it must've already
1404                  * been started
1405                  */
1406                 lnet_shutdown_lndni(ni);
1407                 return -EINVAL;
1408         }
1409
1410         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1411                 tq->tq_credits_min =
1412                 tq->tq_credits_max =
1413                 tq->tq_credits = lnet_ni_tq_credits(ni);
1414         }
1415
1416         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1417                 libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1418                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1419                 ni->ni_peerrtrcredits, ni->ni_peertimeout);
1420
1421         return 0;
1422 failed0:
1423         lnet_ni_free(ni);
1424         return rc;
1425 }
1426
1427 static int
1428 lnet_startup_lndnis(struct list_head *nilist)
1429 {
1430         struct lnet_ni          *ni;
1431         int                     rc;
1432         int                     ni_count = 0;
1433
1434         while (!list_empty(nilist)) {
1435                 ni = list_entry(nilist->next, lnet_ni_t, ni_list);
1436                 list_del(&ni->ni_list);
1437                 rc = lnet_startup_lndni(ni, NULL);
1438
1439                 if (rc < 0)
1440                         goto failed;
1441
1442                 ni_count++;
1443         }
1444
1445         return ni_count;
1446 failed:
1447         lnet_shutdown_lndnis();
1448
1449         return rc;
1450 }
1451
1452 /**
1453  * Initialize LNet library.
1454  *
1455  * Automatically called at module loading time. Caller has to call
1456  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
1457  * latter returned 0. It must be called exactly once.
1458  *
1459  * \retval 0 on success
1460  * \retval -ve on failures.
1461  */
1462 int lnet_lib_init(void)
1463 {
1464         int rc;
1465
1466         lnet_assert_wire_constants();
1467
1468         memset(&the_lnet, 0, sizeof(the_lnet));
1469
1470         /* refer to global cfs_cpt_table for now */
1471         the_lnet.ln_cpt_table   = cfs_cpt_table;
1472         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1473
1474         LASSERT(the_lnet.ln_cpt_number > 0);
1475         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1476                 /* we are under risk of consuming all lh_cookie */
1477                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1478                        "please change setting of CPT-table and retry\n",
1479                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1480                 return -E2BIG;
1481         }
1482
1483         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1484                 the_lnet.ln_cpt_bits++;
1485
1486         rc = lnet_create_locks();
1487         if (rc != 0) {
1488                 CERROR("Can't create LNet global locks: %d\n", rc);
1489                 return rc;
1490         }
1491
1492         the_lnet.ln_refcount = 0;
1493         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1494         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1495         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1496         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1497
1498         /* The hash table size is the number of bits it takes to express the set
1499          * ln_num_routes, minus 1 (better to under estimate than over so we
1500          * don't waste memory). */
1501         if (rnet_htable_size <= 0)
1502                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1503         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1504                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1505         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1506                                            order_base_2(rnet_htable_size) - 1);
1507
1508         /* All LNDs apart from the LOLND are in separate modules.  They
1509          * register themselves when their module loads, and unregister
1510          * themselves when their module is unloaded. */
1511         lnet_register_lnd(&the_lolnd);
1512         return 0;
1513 }
1514
1515 /**
1516  * Finalize LNet library.
1517  *
1518  * \pre lnet_lib_init() called with success.
1519  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1520  */
1521 void lnet_lib_exit(void)
1522 {
1523         LASSERT(the_lnet.ln_refcount == 0);
1524
1525         while (!list_empty(&the_lnet.ln_lnds))
1526                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1527                                                lnd_t, lnd_list));
1528         lnet_destroy_locks();
1529 }
1530
1531 /**
1532  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1533  *
1534  * Users must call this function at least once before any other functions.
1535  * For each successful call there must be a corresponding call to
1536  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1537  * ignored.
1538  *
1539  * The PID used by LNet may be different from the one requested.
1540  * See LNetGetId().
1541  *
1542  * \param requested_pid PID requested by the caller.
1543  *
1544  * \return >= 0 on success, and < 0 error code on failures.
1545  */
1546 int
1547 LNetNIInit(lnet_pid_t requested_pid)
1548 {
1549         int                     im_a_router = 0;
1550         int                     rc;
1551         int                     ni_count;
1552         lnet_ping_info_t        *pinfo;
1553         lnet_handle_md_t        md_handle;
1554         struct list_head        net_head;
1555
1556         INIT_LIST_HEAD(&net_head);
1557
1558         mutex_lock(&the_lnet.ln_api_mutex);
1559
1560         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1561
1562         if (the_lnet.ln_refcount > 0) {
1563                 rc = the_lnet.ln_refcount++;
1564                 mutex_unlock(&the_lnet.ln_api_mutex);
1565                 return rc;
1566         }
1567
1568         rc = lnet_prepare(requested_pid);
1569         if (rc != 0) {
1570                 mutex_unlock(&the_lnet.ln_api_mutex);
1571                 return rc;
1572         }
1573
1574         /* Add in the loopback network */
1575         if (lnet_ni_alloc(LNET_MKNET(LOLND, 0), NULL, &net_head) == NULL) {
1576                 rc = -ENOMEM;
1577                 goto failed0;
1578         }
1579
1580         /* If LNet is being initialized via DLC it is possible
1581          * that the user requests not to load module parameters (ones which
1582          * are supported by DLC) on initialization.  Therefore, make sure not
1583          * to load networks, routes and forwarding from module parameters
1584          * in this case.  On cleanup in case of failure only clean up
1585          * routes if it has been loaded */
1586         if (!the_lnet.ln_nis_from_mod_params) {
1587                 rc = lnet_parse_networks(&net_head,
1588                                          lnet_get_networks());
1589                 if (rc < 0)
1590                         goto failed0;
1591         }
1592
1593         ni_count = lnet_startup_lndnis(&net_head);
1594         if (ni_count < 0) {
1595                 rc = ni_count;
1596                 goto failed0;
1597         }
1598
1599         if (!the_lnet.ln_nis_from_mod_params) {
1600                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1601                 if (rc != 0)
1602                         goto failed1;
1603
1604                 rc = lnet_check_routes();
1605                 if (rc != 0)
1606                         goto failed2;
1607
1608                 rc = lnet_rtrpools_alloc(im_a_router);
1609                 if (rc != 0)
1610                         goto failed2;
1611         }
1612
1613         rc = lnet_acceptor_start();
1614         if (rc != 0)
1615                 goto failed2;
1616         the_lnet.ln_refcount = 1;
1617         /* Now I may use my own API functions... */
1618
1619         rc = lnet_ping_info_setup(&pinfo, &md_handle, ni_count, true);
1620         if (rc != 0)
1621                 goto failed3;
1622
1623         lnet_ping_target_update(pinfo, md_handle);
1624
1625         rc = lnet_router_checker_start();
1626         if (rc != 0)
1627                 goto failed4;
1628
1629         lnet_fault_init();
1630         lnet_proc_init();
1631
1632         mutex_unlock(&the_lnet.ln_api_mutex);
1633
1634         return 0;
1635
1636 failed4:
1637         lnet_ping_target_fini();
1638 failed3:
1639         the_lnet.ln_refcount = 0;
1640         lnet_acceptor_stop();
1641 failed2:
1642         if (!the_lnet.ln_nis_from_mod_params)
1643                 lnet_destroy_routes();
1644 failed1:
1645         lnet_shutdown_lndnis();
1646 failed0:
1647         lnet_unprepare();
1648         LASSERT(rc < 0);
1649         mutex_unlock(&the_lnet.ln_api_mutex);
1650         while (!list_empty(&net_head)) {
1651                 struct lnet_ni *ni;
1652                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1653                 list_del_init(&ni->ni_list);
1654                 lnet_ni_free(ni);
1655         }
1656         return rc;
1657 }
1658 EXPORT_SYMBOL(LNetNIInit);
1659
1660 /**
1661  * Stop LNet interfaces, routing, and forwarding.
1662  *
1663  * Users must call this function once for each successful call to LNetNIInit().
1664  * Once the LNetNIFini() operation has been started, the results of pending
1665  * API operations are undefined.
1666  *
1667  * \return always 0 for current implementation.
1668  */
1669 int
1670 LNetNIFini()
1671 {
1672         mutex_lock(&the_lnet.ln_api_mutex);
1673
1674         LASSERT(the_lnet.ln_refcount > 0);
1675
1676         if (the_lnet.ln_refcount != 1) {
1677                 the_lnet.ln_refcount--;
1678         } else {
1679                 LASSERT(!the_lnet.ln_niinit_self);
1680
1681                 lnet_fault_fini();
1682
1683                 lnet_proc_fini();
1684                 lnet_router_checker_stop();
1685                 lnet_ping_target_fini();
1686
1687                 /* Teardown fns that use my own API functions BEFORE here */
1688                 the_lnet.ln_refcount = 0;
1689
1690                 lnet_acceptor_stop();
1691                 lnet_destroy_routes();
1692                 lnet_shutdown_lndnis();
1693                 lnet_unprepare();
1694         }
1695
1696         mutex_unlock(&the_lnet.ln_api_mutex);
1697         return 0;
1698 }
1699 EXPORT_SYMBOL(LNetNIFini);
1700
1701 /**
1702  * Grabs the ni data from the ni structure and fills the out
1703  * parameters
1704  *
1705  * \param[in] ni network        interface structure
1706  * \param[out] cpt_count        the number of cpts the ni is on
1707  * \param[out] nid              Network Interface ID
1708  * \param[out] peer_timeout     NI peer timeout
1709  * \param[out] peer_tx_crdits   NI peer transmit credits
1710  * \param[out] peer_rtr_credits NI peer router credits
1711  * \param[out] max_tx_credits   NI max transmit credit
1712  * \param[out] net_config       Network configuration
1713  */
1714 static void
1715 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_data *config)
1716 {
1717         struct lnet_ioctl_net_config *net_config;
1718         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
1719         size_t min_size, tunable_size = 0;
1720         int i;
1721
1722         if (!ni || !config)
1723                 return;
1724
1725         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
1726         if (!net_config)
1727                 return;
1728
1729         CLASSERT(ARRAY_SIZE(ni->ni_interfaces) ==
1730                  ARRAY_SIZE(net_config->ni_interfaces));
1731
1732         if (ni->ni_interfaces[0] != NULL) {
1733                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
1734                         if (ni->ni_interfaces[i] != NULL) {
1735                                 strncpy(net_config->ni_interfaces[i],
1736                                         ni->ni_interfaces[i],
1737                                         sizeof(net_config->ni_interfaces[i]));
1738                         }
1739                 }
1740         }
1741
1742         config->cfg_nid = ni->ni_nid;
1743         config->cfg_config_u.cfg_net.net_peer_timeout = ni->ni_peertimeout;
1744         config->cfg_config_u.cfg_net.net_max_tx_credits = ni->ni_maxtxcredits;
1745         config->cfg_config_u.cfg_net.net_peer_tx_credits = ni->ni_peertxcredits;
1746         config->cfg_config_u.cfg_net.net_peer_rtr_credits = ni->ni_peerrtrcredits;
1747
1748         net_config->ni_status = ni->ni_status->ns_status;
1749
1750         for (i = 0;
1751              ni->ni_cpts != NULL && i < ni->ni_ncpts &&
1752              i < LNET_MAX_SHOW_NUM_CPT;
1753              i++)
1754                 net_config->ni_cpts[i] = ni->ni_cpts[i];
1755
1756         config->cfg_ncpts = ni->ni_ncpts;
1757
1758         /*
1759          * See if user land tools sent in a newer and larger version
1760          * of struct lnet_tunables than what the kernel uses.
1761          */
1762         min_size = sizeof(*config) + sizeof(*net_config);
1763
1764         if (config->cfg_hdr.ioc_len > min_size)
1765                 tunable_size = config->cfg_hdr.ioc_len - min_size;
1766
1767         /* Don't copy to much data to user space */
1768         min_size = min(tunable_size, sizeof(*ni->ni_lnd_tunables));
1769         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
1770
1771         if (ni->ni_lnd_tunables && lnd_cfg && min_size) {
1772                 memcpy(lnd_cfg, ni->ni_lnd_tunables, min_size);
1773                 config->cfg_config_u.cfg_net.net_interface_count = 1;
1774
1775                 /* Tell user land that kernel side has less data */
1776                 if (tunable_size > sizeof(*ni->ni_lnd_tunables)) {
1777                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
1778                         config->cfg_hdr.ioc_len -= min_size;
1779                 }
1780         }
1781 }
1782
1783 static int
1784 lnet_get_net_config(struct lnet_ioctl_config_data *config)
1785 {
1786         struct lnet_ni *ni;
1787         struct list_head *tmp;
1788         int idx = config->cfg_count;
1789         int rc = -ENOENT;
1790         int cpt;
1791
1792         if (unlikely(!config->cfg_bulk))
1793                 return -EINVAL;
1794
1795         cpt = lnet_net_lock_current();
1796
1797         list_for_each(tmp, &the_lnet.ln_nis) {
1798                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1799                 if (idx-- == 0) {
1800                         rc = 0;
1801                         lnet_ni_lock(ni);
1802                         lnet_fill_ni_info(ni, config);
1803                         lnet_ni_unlock(ni);
1804                         break;
1805                 }
1806         }
1807
1808         lnet_net_unlock(cpt);
1809         return rc;
1810 }
1811
1812 int
1813 lnet_dyn_add_ni(lnet_pid_t requested_pid, struct lnet_ioctl_config_data *conf)
1814 {
1815         char                    *nets = conf->cfg_config_u.cfg_net.net_intf;
1816         lnet_ping_info_t        *pinfo;
1817         lnet_handle_md_t        md_handle;
1818         struct lnet_ni          *ni;
1819         struct list_head        net_head;
1820         int                     rc;
1821         lnet_remotenet_t        *rnet;
1822
1823         INIT_LIST_HEAD(&net_head);
1824
1825         /* Create a ni structure for the network string */
1826         rc = lnet_parse_networks(&net_head, nets);
1827         if (rc <= 0)
1828                 return rc == 0 ? -EINVAL : rc;
1829
1830         mutex_lock(&the_lnet.ln_api_mutex);
1831
1832         if (rc > 1) {
1833                 rc = -EINVAL; /* only add one interface per call */
1834                 goto failed0;
1835         }
1836
1837         ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1838
1839         lnet_net_lock(LNET_LOCK_EX);
1840         rnet = lnet_find_net_locked(LNET_NIDNET(ni->ni_nid));
1841         lnet_net_unlock(LNET_LOCK_EX);
1842         /* make sure that the net added doesn't invalidate the current
1843          * configuration LNet is keeping */
1844         if (rnet != NULL) {
1845                 CERROR("Adding net %s will invalidate routing configuration\n",
1846                        nets);
1847                 rc = -EUSERS;
1848                 goto failed0;
1849         }
1850
1851         rc = lnet_ping_info_setup(&pinfo, &md_handle, 1 + lnet_get_ni_count(),
1852                                   false);
1853         if (rc != 0)
1854                 goto failed0;
1855
1856         list_del_init(&ni->ni_list);
1857
1858         rc = lnet_startup_lndni(ni, conf);
1859         if (rc != 0)
1860                 goto failed1;
1861
1862         if (ni->ni_lnd->lnd_accept != NULL) {
1863                 rc = lnet_acceptor_start();
1864                 if (rc < 0) {
1865                         /* shutdown the ni that we just started */
1866                         CERROR("Failed to start up acceptor thread\n");
1867                         lnet_shutdown_lndni(ni);
1868                         goto failed1;
1869                 }
1870         }
1871
1872         lnet_ping_target_update(pinfo, md_handle);
1873         mutex_unlock(&the_lnet.ln_api_mutex);
1874
1875         return 0;
1876
1877 failed1:
1878         lnet_ping_md_unlink(pinfo, &md_handle);
1879         lnet_ping_info_free(pinfo);
1880 failed0:
1881         mutex_unlock(&the_lnet.ln_api_mutex);
1882         while (!list_empty(&net_head)) {
1883                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1884                 list_del_init(&ni->ni_list);
1885                 lnet_ni_free(ni);
1886         }
1887         return rc;
1888 }
1889
1890 int
1891 lnet_dyn_del_ni(__u32 net)
1892 {
1893         lnet_ni_t        *ni;
1894         lnet_ping_info_t *pinfo;
1895         lnet_handle_md_t  md_handle;
1896         int               rc;
1897
1898         /* don't allow userspace to shutdown the LOLND */
1899         if (LNET_NETTYP(net) == LOLND)
1900                 return -EINVAL;
1901
1902         mutex_lock(&the_lnet.ln_api_mutex);
1903         /* create and link a new ping info, before removing the old one */
1904         rc = lnet_ping_info_setup(&pinfo, &md_handle,
1905                                   lnet_get_ni_count() - 1, false);
1906         if (rc != 0)
1907                 goto out;
1908
1909         ni = lnet_net2ni(net);
1910         if (ni == NULL) {
1911                 rc = -EINVAL;
1912                 goto failed;
1913         }
1914
1915         /* decrement the reference counter taken by lnet_net2ni() */
1916         lnet_ni_decref_locked(ni, 0);
1917
1918         lnet_shutdown_lndni(ni);
1919
1920         if (lnet_count_acceptor_nis() == 0)
1921                 lnet_acceptor_stop();
1922
1923         lnet_ping_target_update(pinfo, md_handle);
1924         goto out;
1925 failed:
1926         lnet_ping_md_unlink(pinfo, &md_handle);
1927         lnet_ping_info_free(pinfo);
1928 out:
1929         mutex_unlock(&the_lnet.ln_api_mutex);
1930
1931         return rc;
1932 }
1933
1934 /**
1935  * LNet ioctl handler.
1936  *
1937  */
1938 int
1939 LNetCtl(unsigned int cmd, void *arg)
1940 {
1941         struct libcfs_ioctl_data *data = arg;
1942         struct lnet_ioctl_config_data *config;
1943         lnet_process_id_t         id = {0};
1944         lnet_ni_t                *ni;
1945         int                       rc;
1946
1947         CLASSERT(LIBCFS_IOC_DATA_MAX >= sizeof(struct lnet_ioctl_net_config) +
1948                                         sizeof(struct lnet_ioctl_config_data));
1949
1950         switch (cmd) {
1951         case IOC_LIBCFS_GET_NI:
1952                 rc = LNetGetId(data->ioc_count, &id);
1953                 data->ioc_nid = id.nid;
1954                 return rc;
1955
1956         case IOC_LIBCFS_FAIL_NID:
1957                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1958
1959         case IOC_LIBCFS_ADD_ROUTE:
1960                 config = arg;
1961
1962                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1963                         return -EINVAL;
1964
1965                 mutex_lock(&the_lnet.ln_api_mutex);
1966                 rc = lnet_add_route(config->cfg_net,
1967                                     config->cfg_config_u.cfg_route.rtr_hop,
1968                                     config->cfg_nid,
1969                                     config->cfg_config_u.cfg_route.
1970                                         rtr_priority);
1971                 if (rc == 0) {
1972                         rc = lnet_check_routes();
1973                         if (rc != 0)
1974                                 lnet_del_route(config->cfg_net,
1975                                                config->cfg_nid);
1976                 }
1977                 mutex_unlock(&the_lnet.ln_api_mutex);
1978                 return rc;
1979
1980         case IOC_LIBCFS_DEL_ROUTE:
1981                 config = arg;
1982
1983                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1984                         return -EINVAL;
1985
1986                 mutex_lock(&the_lnet.ln_api_mutex);
1987                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
1988                 mutex_unlock(&the_lnet.ln_api_mutex);
1989                 return rc;
1990
1991         case IOC_LIBCFS_GET_ROUTE:
1992                 config = arg;
1993
1994                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1995                         return -EINVAL;
1996
1997                 return lnet_get_route(config->cfg_count,
1998                                       &config->cfg_net,
1999                                       &config->cfg_config_u.cfg_route.rtr_hop,
2000                                       &config->cfg_nid,
2001                                       &config->cfg_config_u.cfg_route.rtr_flags,
2002                                       &config->cfg_config_u.cfg_route.
2003                                         rtr_priority);
2004
2005         case IOC_LIBCFS_GET_NET: {
2006                 size_t total = sizeof(*config) +
2007                                sizeof(struct lnet_ioctl_net_config);
2008                 config = arg;
2009
2010                 if (config->cfg_hdr.ioc_len < total)
2011                         return -EINVAL;
2012
2013                 return lnet_get_net_config(config);
2014         }
2015
2016         case IOC_LIBCFS_GET_LNET_STATS:
2017         {
2018                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
2019
2020                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
2021                         return -EINVAL;
2022
2023                 lnet_counters_get(&lnet_stats->st_cntrs);
2024                 return 0;
2025         }
2026
2027         case IOC_LIBCFS_CONFIG_RTR:
2028                 config = arg;
2029
2030                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2031                         return -EINVAL;
2032
2033                 mutex_lock(&the_lnet.ln_api_mutex);
2034                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
2035                         rc = lnet_rtrpools_enable();
2036                         mutex_unlock(&the_lnet.ln_api_mutex);
2037                         return rc;
2038                 }
2039                 lnet_rtrpools_disable();
2040                 mutex_unlock(&the_lnet.ln_api_mutex);
2041                 return 0;
2042
2043         case IOC_LIBCFS_ADD_BUF:
2044                 config = arg;
2045
2046                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2047                         return -EINVAL;
2048
2049                 mutex_lock(&the_lnet.ln_api_mutex);
2050                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
2051                                                 buf_tiny,
2052                                           config->cfg_config_u.cfg_buffers.
2053                                                 buf_small,
2054                                           config->cfg_config_u.cfg_buffers.
2055                                                 buf_large);
2056                 mutex_unlock(&the_lnet.ln_api_mutex);
2057                 return rc;
2058
2059         case IOC_LIBCFS_GET_BUF: {
2060                 struct lnet_ioctl_pool_cfg *pool_cfg;
2061                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
2062
2063                 config = arg;
2064
2065                 if (config->cfg_hdr.ioc_len < total)
2066                         return -EINVAL;
2067
2068                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
2069                 return lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
2070         }
2071
2072         case IOC_LIBCFS_GET_PEER_INFO: {
2073                 struct lnet_ioctl_peer *peer_info = arg;
2074
2075                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
2076                         return -EINVAL;
2077
2078                 return lnet_get_peer_info(
2079                    peer_info->pr_count,
2080                    &peer_info->pr_nid,
2081                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
2082                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
2083                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
2084                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
2085                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
2086                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
2087                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_rtr_credits,
2088                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
2089         }
2090
2091         case IOC_LIBCFS_NOTIFY_ROUTER: {
2092                 unsigned long jiffies_passed;
2093
2094                 jiffies_passed = ktime_get_real_seconds() - data->ioc_u64[0];
2095                 jiffies_passed = cfs_time_seconds(jiffies_passed);
2096
2097                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
2098                                    jiffies - jiffies_passed);
2099         }
2100
2101         case IOC_LIBCFS_LNET_DIST:
2102                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
2103                 if (rc < 0 && rc != -EHOSTUNREACH)
2104                         return rc;
2105
2106                 data->ioc_u32[0] = rc;
2107                 return 0;
2108
2109         case IOC_LIBCFS_TESTPROTOCOMPAT:
2110                 lnet_net_lock(LNET_LOCK_EX);
2111                 the_lnet.ln_testprotocompat = data->ioc_flags;
2112                 lnet_net_unlock(LNET_LOCK_EX);
2113                 return 0;
2114
2115         case IOC_LIBCFS_LNET_FAULT:
2116                 return lnet_fault_ctl(data->ioc_flags, data);
2117
2118         case IOC_LIBCFS_PING:
2119                 id.nid = data->ioc_nid;
2120                 id.pid = data->ioc_u32[0];
2121                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
2122                                data->ioc_pbuf1,
2123                                data->ioc_plen1/sizeof(lnet_process_id_t));
2124                 if (rc < 0)
2125                         return rc;
2126                 data->ioc_count = rc;
2127                 return 0;
2128
2129         default:
2130                 ni = lnet_net2ni(data->ioc_net);
2131                 if (ni == NULL)
2132                         return -EINVAL;
2133
2134                 if (ni->ni_lnd->lnd_ctl == NULL)
2135                         rc = -EINVAL;
2136                 else
2137                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
2138
2139                 lnet_ni_decref(ni);
2140                 return rc;
2141         }
2142         /* not reached */
2143 }
2144 EXPORT_SYMBOL(LNetCtl);
2145
2146 void LNetDebugPeer(lnet_process_id_t id)
2147 {
2148         lnet_debug_peer(id.nid);
2149 }
2150 EXPORT_SYMBOL(LNetDebugPeer);
2151
2152 /**
2153  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
2154  * all interfaces share a same PID, as requested by LNetNIInit().
2155  *
2156  * \param index Index of the interface to look up.
2157  * \param id On successful return, this location will hold the
2158  * lnet_process_id_t ID of the interface.
2159  *
2160  * \retval 0 If an interface exists at \a index.
2161  * \retval -ENOENT If no interface has been found.
2162  */
2163 int
2164 LNetGetId(unsigned int index, lnet_process_id_t *id)
2165 {
2166         struct lnet_ni   *ni;
2167         struct list_head *tmp;
2168         int               cpt;
2169         int               rc = -ENOENT;
2170
2171         LASSERT(the_lnet.ln_refcount > 0);
2172
2173         cpt = lnet_net_lock_current();
2174
2175         list_for_each(tmp, &the_lnet.ln_nis) {
2176                 if (index-- != 0)
2177                         continue;
2178
2179                 ni = list_entry(tmp, lnet_ni_t, ni_list);
2180
2181                 id->nid = ni->ni_nid;
2182                 id->pid = the_lnet.ln_pid;
2183                 rc = 0;
2184                 break;
2185         }
2186
2187         lnet_net_unlock(cpt);
2188         return rc;
2189 }
2190 EXPORT_SYMBOL(LNetGetId);
2191
2192 /**
2193  * Print a string representation of handle \a h into buffer \a str of
2194  * \a len bytes.
2195  */
2196 void
2197 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
2198 {
2199         snprintf(str, len, "%#llx", h.cookie);
2200 }
2201 EXPORT_SYMBOL(LNetSnprintHandle);
2202
2203 static int
2204 lnet_ping(lnet_process_id_t id, int timeout_ms, lnet_process_id_t __user *ids,
2205           int n_ids)
2206 {
2207         lnet_handle_eq_t     eqh;
2208         lnet_handle_md_t     mdh;
2209         lnet_event_t         event;
2210         lnet_md_t            md = { NULL };
2211         int                  which;
2212         int                  unlinked = 0;
2213         int                  replied = 0;
2214         const int            a_long_time = 60000; /* mS */
2215         int                  infosz;
2216         lnet_ping_info_t    *info;
2217         lnet_process_id_t    tmpid;
2218         int                  i;
2219         int                  nob;
2220         int                  rc;
2221         int                  rc2;
2222         sigset_t         blocked;
2223
2224         infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
2225
2226         if (n_ids <= 0 ||
2227             id.nid == LNET_NID_ANY ||
2228             timeout_ms > 500000 ||              /* arbitrary limit! */
2229             n_ids > 20)                         /* arbitrary limit! */
2230                 return -EINVAL;
2231
2232         if (id.pid == LNET_PID_ANY)
2233                 id.pid = LNET_PID_LUSTRE;
2234
2235         LIBCFS_ALLOC(info, infosz);
2236         if (info == NULL)
2237                 return -ENOMEM;
2238
2239         /* NB 2 events max (including any unlink event) */
2240         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
2241         if (rc != 0) {
2242                 CERROR("Can't allocate EQ: %d\n", rc);
2243                 goto out_0;
2244         }
2245
2246         /* initialize md content */
2247         md.start     = info;
2248         md.length    = infosz;
2249         md.threshold = 2; /*GET/REPLY*/
2250         md.max_size  = 0;
2251         md.options   = LNET_MD_TRUNCATE;
2252         md.user_ptr  = NULL;
2253         md.eq_handle = eqh;
2254
2255         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
2256         if (rc != 0) {
2257                 CERROR("Can't bind MD: %d\n", rc);
2258                 goto out_1;
2259         }
2260
2261         rc = LNetGet(LNET_NID_ANY, mdh, id,
2262                      LNET_RESERVED_PORTAL,
2263                      LNET_PROTO_PING_MATCHBITS, 0);
2264
2265         if (rc != 0) {
2266                 /* Don't CERROR; this could be deliberate! */
2267
2268                 rc2 = LNetMDUnlink(mdh);
2269                 LASSERT(rc2 == 0);
2270
2271                 /* NB must wait for the UNLINK event below... */
2272                 unlinked = 1;
2273                 timeout_ms = a_long_time;
2274         }
2275
2276         do {
2277                 /* MUST block for unlink to complete */
2278                 if (unlinked)
2279                         blocked = cfs_block_allsigs();
2280
2281                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
2282
2283                 if (unlinked)
2284                         cfs_restore_sigs(blocked);
2285
2286                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
2287                        (rc2 <= 0) ? -1 : event.type,
2288                        (rc2 <= 0) ? -1 : event.status,
2289                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
2290
2291                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
2292
2293                 if (rc2 <= 0 || event.status != 0) {
2294                         /* timeout or error */
2295                         if (!replied && rc == 0)
2296                                 rc = (rc2 < 0) ? rc2 :
2297                                      (rc2 == 0) ? -ETIMEDOUT :
2298                                      event.status;
2299
2300                         if (!unlinked) {
2301                                 /* Ensure completion in finite time... */
2302                                 LNetMDUnlink(mdh);
2303                                 /* No assertion (racing with network) */
2304                                 unlinked = 1;
2305                                 timeout_ms = a_long_time;
2306                         } else if (rc2 == 0) {
2307                                 /* timed out waiting for unlink */
2308                                 CWARN("ping %s: late network completion\n",
2309                                       libcfs_id2str(id));
2310                         }
2311                 } else if (event.type == LNET_EVENT_REPLY) {
2312                         replied = 1;
2313                         rc = event.mlength;
2314                 }
2315
2316         } while (rc2 <= 0 || !event.unlinked);
2317
2318         if (!replied) {
2319                 if (rc >= 0)
2320                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
2321                               libcfs_id2str(id));
2322                 rc = -EIO;
2323                 goto out_1;
2324         }
2325
2326         nob = rc;
2327         LASSERT(nob >= 0 && nob <= infosz);
2328
2329         rc = -EPROTO;                           /* if I can't parse... */
2330
2331         if (nob < 8) {
2332                 /* can't check magic/version */
2333                 CERROR("%s: ping info too short %d\n",
2334                        libcfs_id2str(id), nob);
2335                 goto out_1;
2336         }
2337
2338         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2339                 lnet_swap_pinginfo(info);
2340         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2341                 CERROR("%s: Unexpected magic %08x\n",
2342                        libcfs_id2str(id), info->pi_magic);
2343                 goto out_1;
2344         }
2345
2346         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
2347                 CERROR("%s: ping w/o NI status: 0x%x\n",
2348                        libcfs_id2str(id), info->pi_features);
2349                 goto out_1;
2350         }
2351
2352         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2353                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2354                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2355                 goto out_1;
2356         }
2357
2358         if (info->pi_nnis < n_ids)
2359                 n_ids = info->pi_nnis;
2360
2361         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2362                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2363                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2364                 goto out_1;
2365         }
2366
2367         rc = -EFAULT;                           /* If I SEGV... */
2368
2369         memset(&tmpid, 0, sizeof(tmpid));
2370         for (i = 0; i < n_ids; i++) {
2371                 tmpid.pid = info->pi_pid;
2372                 tmpid.nid = info->pi_ni[i].ns_nid;
2373                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2374                         goto out_1;
2375         }
2376         rc = info->pi_nnis;
2377
2378  out_1:
2379         rc2 = LNetEQFree(eqh);
2380         if (rc2 != 0)
2381                 CERROR("rc2 %d\n", rc2);
2382         LASSERT(rc2 == 0);
2383
2384  out_0:
2385         LIBCFS_FREE(info, infosz);
2386         return rc;
2387 }