kernel/drivers/staging/lustre/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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38 #include <linux/log2.h>
  39 #include <linux/ktime.h>
  40
  41 #include "../../include/linux/lnet/lib-lnet.h"
  42
  43 #define D_LNI D_CONSOLE
  44
  45 lnet_t the_lnet;                           /* THE state of the network */
  46 EXPORT_SYMBOL(the_lnet);
  47
  48 static char *ip2nets = "";
  49 module_param(ip2nets, charp, 0444);
  50 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  51
  52 static char *networks = "";
  53 module_param(networks, charp, 0444);
  54 MODULE_PARM_DESC(networks, "local networks");
  55
  56 static char *routes = "";
  57 module_param(routes, charp, 0444);
  58 MODULE_PARM_DESC(routes, "routes to non-local networks");
  59
  60 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  61 module_param(rnet_htable_size, int, 0444);
  62 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  63
  64 static char *
  65 lnet_get_routes(void)
  66 {
  67         return routes;
  68 }
  69
  70 static char *
  71 lnet_get_networks(void)
  72 {
  73         char *nets;
  74         int rc;
  75
  76         if (*networks != 0 && *ip2nets != 0) {
  77                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or 'ip2nets' but not both at once\n");
  78                 return NULL;
  79         }
  80
  81         if (*ip2nets != 0) {
  82                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  83                 return (rc == 0) ? nets : NULL;
  84         }
  85
  86         if (*networks != 0)
  87                 return networks;
  88
  89         return "tcp";
  90 }
  91
  92 static void
  93 lnet_init_locks(void)
  94 {
  95         spin_lock_init(&the_lnet.ln_eq_wait_lock);
  96         init_waitqueue_head(&the_lnet.ln_eq_waitq);
  97         mutex_init(&the_lnet.ln_lnd_mutex);
  98         mutex_init(&the_lnet.ln_api_mutex);
  99 }
 100
 101 static int
 102 lnet_create_remote_nets_table(void)
 103 {
 104         int i;
 105         struct list_head *hash;
 106
 107         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 108         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 109         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 110         if (hash == NULL) {
 111                 CERROR("Failed to create remote nets hash table\n");
 112                 return -ENOMEM;
 113         }
 114
 115         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 116                 INIT_LIST_HEAD(&hash[i]);
 117         the_lnet.ln_remote_nets_hash = hash;
 118         return 0;
 119 }
 120
 121 static void
 122 lnet_destroy_remote_nets_table(void)
 123 {
 124         int i;
 125
 126         if (the_lnet.ln_remote_nets_hash == NULL)
 127                 return;
 128
 129         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 130                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 131
 132         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 133                     LNET_REMOTE_NETS_HASH_SIZE *
 134                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 135         the_lnet.ln_remote_nets_hash = NULL;
 136 }
 137
 138 static void
 139 lnet_destroy_locks(void)
 140 {
 141         if (the_lnet.ln_res_lock != NULL) {
 142                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 143                 the_lnet.ln_res_lock = NULL;
 144         }
 145
 146         if (the_lnet.ln_net_lock != NULL) {
 147                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 148                 the_lnet.ln_net_lock = NULL;
 149         }
 150 }
 151
 152 static int
 153 lnet_create_locks(void)
 154 {
 155         lnet_init_locks();
 156
 157         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 158         if (the_lnet.ln_res_lock == NULL)
 159                 goto failed;
 160
 161         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 162         if (the_lnet.ln_net_lock == NULL)
 163                 goto failed;
 164
 165         return 0;
 166
 167  failed:
 168         lnet_destroy_locks();
 169         return -ENOMEM;
 170 }
 171
 172 static void lnet_assert_wire_constants(void)
 173 {
 174         /* Wire protocol assertions generated by 'wirecheck'
 175          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 176          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 177          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 178
 179         /* Constants... */
 180         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 181         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 182         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 183         CLASSERT(LNET_MSG_ACK == 0);
 184         CLASSERT(LNET_MSG_PUT == 1);
 185         CLASSERT(LNET_MSG_GET == 2);
 186         CLASSERT(LNET_MSG_REPLY == 3);
 187         CLASSERT(LNET_MSG_HELLO == 4);
 188
 189         /* Checks for struct ptl_handle_wire_t */
 190         CLASSERT((int)sizeof(lnet_handle_wire_t) == 16);
 191         CLASSERT((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 192         CLASSERT((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 193         CLASSERT((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 194         CLASSERT((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 195
 196         /* Checks for struct lnet_magicversion_t */
 197         CLASSERT((int)sizeof(lnet_magicversion_t) == 8);
 198         CLASSERT((int)offsetof(lnet_magicversion_t, magic) == 0);
 199         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 200         CLASSERT((int)offsetof(lnet_magicversion_t, version_major) == 4);
 201         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 202         CLASSERT((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 203         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 204
 205         /* Checks for struct lnet_hdr_t */
 206         CLASSERT((int)sizeof(lnet_hdr_t) == 72);
 207         CLASSERT((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 208         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 209         CLASSERT((int)offsetof(lnet_hdr_t, src_nid) == 8);
 210         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 211         CLASSERT((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 212         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 213         CLASSERT((int)offsetof(lnet_hdr_t, src_pid) == 20);
 214         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 215         CLASSERT((int)offsetof(lnet_hdr_t, type) == 24);
 216         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 217         CLASSERT((int)offsetof(lnet_hdr_t, payload_length) == 28);
 218         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 219         CLASSERT((int)offsetof(lnet_hdr_t, msg) == 32);
 220         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 221
 222         /* Ack */
 223         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 224         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 225         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 226         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 227         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 228         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 229
 230         /* Put */
 231         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 232         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 233         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 234         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 235         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 236         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 237         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 238         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 239         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 240         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 241
 242         /* Get */
 243         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 244         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 245         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 246         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 247         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 248         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 249         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 250         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 251         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 252         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 253
 254         /* Reply */
 255         CLASSERT((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 256         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 257
 258         /* Hello */
 259         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 260         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 261         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 262         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 263 }
 264
 265 static lnd_t *
 266 lnet_find_lnd_by_type(__u32 type)
 267 {
 268         lnd_t *lnd;
 269         struct list_head *tmp;
 270
 271         /* holding lnd mutex */
 272         list_for_each(tmp, &the_lnet.ln_lnds) {
 273                 lnd = list_entry(tmp, lnd_t, lnd_list);
 274
 275                 if (lnd->lnd_type == type)
 276                         return lnd;
 277         }
 278
 279         return NULL;
 280 }
 281
 282 void
 283 lnet_register_lnd(lnd_t *lnd)
 284 {
 285         mutex_lock(&the_lnet.ln_lnd_mutex);
 286
 287         LASSERT(the_lnet.ln_init);
 288         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 289         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 290
 291         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 292         lnd->lnd_refcount = 0;
 293
 294         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 295
 296         mutex_unlock(&the_lnet.ln_lnd_mutex);
 297 }
 298 EXPORT_SYMBOL(lnet_register_lnd);
 299
 300 void
 301 lnet_unregister_lnd(lnd_t *lnd)
 302 {
 303         mutex_lock(&the_lnet.ln_lnd_mutex);
 304
 305         LASSERT(the_lnet.ln_init);
 306         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 307         LASSERT(lnd->lnd_refcount == 0);
 308
 309         list_del(&lnd->lnd_list);
 310         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 311
 312         mutex_unlock(&the_lnet.ln_lnd_mutex);
 313 }
 314 EXPORT_SYMBOL(lnet_unregister_lnd);
 315
 316 void
 317 lnet_counters_get(lnet_counters_t *counters)
 318 {
 319         lnet_counters_t *ctr;
 320         int i;
 321
 322         memset(counters, 0, sizeof(*counters));
 323
 324         lnet_net_lock(LNET_LOCK_EX);
 325
 326         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 327                 counters->msgs_max     += ctr->msgs_max;
 328                 counters->msgs_alloc   += ctr->msgs_alloc;
 329                 counters->errors       += ctr->errors;
 330                 counters->send_count   += ctr->send_count;
 331                 counters->recv_count   += ctr->recv_count;
 332                 counters->route_count  += ctr->route_count;
 333                 counters->drop_count   += ctr->drop_count;
 334                 counters->send_length  += ctr->send_length;
 335                 counters->recv_length  += ctr->recv_length;
 336                 counters->route_length += ctr->route_length;
 337                 counters->drop_length  += ctr->drop_length;
 338
 339         }
 340         lnet_net_unlock(LNET_LOCK_EX);
 341 }
 342 EXPORT_SYMBOL(lnet_counters_get);
 343
 344 void
 345 lnet_counters_reset(void)
 346 {
 347         lnet_counters_t *counters;
 348         int i;
 349
 350         lnet_net_lock(LNET_LOCK_EX);
 351
 352         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 353                 memset(counters, 0, sizeof(lnet_counters_t));
 354
 355         lnet_net_unlock(LNET_LOCK_EX);
 356 }
 357 EXPORT_SYMBOL(lnet_counters_reset);
 358
 359 static __u64
 360 lnet_create_interface_cookie(void)
 361 {
 362         /* NB the interface cookie in wire handles guards against delayed
 363          * replies and ACKs appearing valid after reboot.
 364          */
 365         return ktime_get_ns();
 366 }
 367
 368 static char *
 369 lnet_res_type2str(int type)
 370 {
 371         switch (type) {
 372         default:
 373                 LBUG();
 374         case LNET_COOKIE_TYPE_MD:
 375                 return "MD";
 376         case LNET_COOKIE_TYPE_ME:
 377                 return "ME";
 378         case LNET_COOKIE_TYPE_EQ:
 379                 return "EQ";
 380         }
 381 }
 382
 383 static void
 384 lnet_res_container_cleanup(struct lnet_res_container *rec)
 385 {
 386         int count = 0;
 387
 388         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 389                 return;
 390
 391         while (!list_empty(&rec->rec_active)) {
 392                 struct list_head *e = rec->rec_active.next;
 393
 394                 list_del_init(e);
 395                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 396                         lnet_eq_free(list_entry(e, lnet_eq_t, eq_list));
 397
 398                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 399                         lnet_md_free(list_entry(e, lnet_libmd_t, md_list));
 400
 401                 } else { /* NB: Active MEs should be attached on portals */
 402                         LBUG();
 403                 }
 404                 count++;
 405         }
 406
 407         if (count > 0) {
 408                 /* Found alive MD/ME/EQ, user really should unlink/free
 409                  * all of them before finalize LNet, but if someone didn't,
 410                  * we have to recycle garbage for him */
 411                 CERROR("%d active elements on exit of %s container\n",
 412                        count, lnet_res_type2str(rec->rec_type));
 413         }
 414
 415         if (rec->rec_lh_hash != NULL) {
 416                 LIBCFS_FREE(rec->rec_lh_hash,
 417                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 418                 rec->rec_lh_hash = NULL;
 419         }
 420
 421         rec->rec_type = 0; /* mark it as finalized */
 422 }
 423
 424 static int
 425 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 426 {
 427         int rc = 0;
 428         int i;
 429
 430         LASSERT(rec->rec_type == 0);
 431
 432         rec->rec_type = type;
 433         INIT_LIST_HEAD(&rec->rec_active);
 434         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 435
 436         /* Arbitrary choice of hash table size */
 437         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 438                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 439         if (rec->rec_lh_hash == NULL) {
 440                 rc = -ENOMEM;
 441                 goto out;
 442         }
 443
 444         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 445                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 446
 447         return 0;
 448
 449 out:
 450         CERROR("Failed to setup %s resource container\n",
 451                lnet_res_type2str(type));
 452         lnet_res_container_cleanup(rec);
 453         return rc;
 454 }
 455
 456 static void
 457 lnet_res_containers_destroy(struct lnet_res_container **recs)
 458 {
 459         struct lnet_res_container *rec;
 460         int i;
 461
 462         cfs_percpt_for_each(rec, i, recs)
 463                 lnet_res_container_cleanup(rec);
 464
 465         cfs_percpt_free(recs);
 466 }
 467
 468 static struct lnet_res_container **
 469 lnet_res_containers_create(int type)
 470 {
 471         struct lnet_res_container **recs;
 472         struct lnet_res_container *rec;
 473         int rc;
 474         int i;
 475
 476         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 477         if (recs == NULL) {
 478                 CERROR("Failed to allocate %s resource containers\n",
 479                        lnet_res_type2str(type));
 480                 return NULL;
 481         }
 482
 483         cfs_percpt_for_each(rec, i, recs) {
 484                 rc = lnet_res_container_setup(rec, i, type);
 485                 if (rc != 0) {
 486                         lnet_res_containers_destroy(recs);
 487                         return NULL;
 488                 }
 489         }
 490
 491         return recs;
 492 }
 493
 494 lnet_libhandle_t *
 495 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 496 {
 497         /* ALWAYS called with lnet_res_lock held */
 498         struct list_head *head;
 499         lnet_libhandle_t *lh;
 500         unsigned int hash;
 501
 502         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 503                 return NULL;
 504
 505         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 506         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 507
 508         list_for_each_entry(lh, head, lh_hash_chain) {
 509                 if (lh->lh_cookie == cookie)
 510                         return lh;
 511         }
 512
 513         return NULL;
 514 }
 515
 516 void
 517 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 518 {
 519         /* ALWAYS called with lnet_res_lock held */
 520         unsigned int ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 521         unsigned int hash;
 522
 523         lh->lh_cookie = rec->rec_lh_cookie;
 524         rec->rec_lh_cookie += 1 << ibits;
 525
 526         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 527
 528         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 529 }
 530
 531 int lnet_unprepare(void);
 532
 533 static int
 534 lnet_prepare(lnet_pid_t requested_pid)
 535 {
 536         /* Prepare to bring up the network */
 537         struct lnet_res_container **recs;
 538         int rc = 0;
 539
 540         LASSERT(the_lnet.ln_refcount == 0);
 541
 542         the_lnet.ln_routing = 0;
 543
 544         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
 545         the_lnet.ln_pid = requested_pid;
 546
 547         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 548         INIT_LIST_HEAD(&the_lnet.ln_nis);
 549         INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 550         INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 551         INIT_LIST_HEAD(&the_lnet.ln_routers);
 552
 553         rc = lnet_create_remote_nets_table();
 554         if (rc != 0)
 555                 goto failed;
 556
 557         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 558
 559         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 560                                                 sizeof(lnet_counters_t));
 561         if (the_lnet.ln_counters == NULL) {
 562                 CERROR("Failed to allocate counters for LNet\n");
 563                 rc = -ENOMEM;
 564                 goto failed;
 565         }
 566
 567         rc = lnet_peer_tables_create();
 568         if (rc != 0)
 569                 goto failed;
 570
 571         rc = lnet_msg_containers_create();
 572         if (rc != 0)
 573                 goto failed;
 574
 575         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 576                                       LNET_COOKIE_TYPE_EQ);
 577         if (rc != 0)
 578                 goto failed;
 579
 580         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
 581         if (recs == NULL) {
 582                 rc = -ENOMEM;
 583                 goto failed;
 584         }
 585
 586         the_lnet.ln_me_containers = recs;
 587
 588         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
 589         if (recs == NULL) {
 590                 rc = -ENOMEM;
 591                 goto failed;
 592         }
 593
 594         the_lnet.ln_md_containers = recs;
 595
 596         rc = lnet_portals_create();
 597         if (rc != 0) {
 598                 CERROR("Failed to create portals for LNet: %d\n", rc);
 599                 goto failed;
 600         }
 601
 602         return 0;
 603
 604  failed:
 605         lnet_unprepare();
 606         return rc;
 607 }
 608
 609 int
 610 lnet_unprepare(void)
 611 {
 612         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 613          * have shut down already, so it is safe to unlink and free all
 614          * descriptors, even those that appear committed to a network op (eg MD
 615          * with non-zero pending count) */
 616
 617         lnet_fail_nid(LNET_NID_ANY, 0);
 618
 619         LASSERT(the_lnet.ln_refcount == 0);
 620         LASSERT(list_empty(&the_lnet.ln_test_peers));
 621         LASSERT(list_empty(&the_lnet.ln_nis));
 622         LASSERT(list_empty(&the_lnet.ln_nis_cpt));
 623         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
 624
 625         lnet_portals_destroy();
 626
 627         if (the_lnet.ln_md_containers != NULL) {
 628                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 629                 the_lnet.ln_md_containers = NULL;
 630         }
 631
 632         if (the_lnet.ln_me_containers != NULL) {
 633                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 634                 the_lnet.ln_me_containers = NULL;
 635         }
 636
 637         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 638
 639         lnet_msg_containers_destroy();
 640         lnet_peer_tables_destroy();
 641         lnet_rtrpools_free();
 642
 643         if (the_lnet.ln_counters != NULL) {
 644                 cfs_percpt_free(the_lnet.ln_counters);
 645                 the_lnet.ln_counters = NULL;
 646         }
 647         lnet_destroy_remote_nets_table();
 648
 649         return 0;
 650 }
 651
 652 lnet_ni_t  *
 653 lnet_net2ni_locked(__u32 net, int cpt)
 654 {
 655         struct list_head *tmp;
 656         lnet_ni_t *ni;
 657
 658         LASSERT(cpt != LNET_LOCK_EX);
 659
 660         list_for_each(tmp, &the_lnet.ln_nis) {
 661                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 662
 663                 if (LNET_NIDNET(ni->ni_nid) == net) {
 664                         lnet_ni_addref_locked(ni, cpt);
 665                         return ni;
 666                 }
 667         }
 668
 669         return NULL;
 670 }
 671
 672 lnet_ni_t *
 673 lnet_net2ni(__u32 net)
 674 {
 675         lnet_ni_t *ni;
 676
 677         lnet_net_lock(0);
 678         ni = lnet_net2ni_locked(net, 0);
 679         lnet_net_unlock(0);
 680
 681         return ni;
 682 }
 683 EXPORT_SYMBOL(lnet_net2ni);
 684
 685 static unsigned int
 686 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 687 {
 688         __u64 key = nid;
 689         unsigned int val;
 690
 691         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 692
 693         if (number == 1)
 694                 return 0;
 695
 696         val = hash_long(key, LNET_CPT_BITS);
 697         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 698         if (val < number)
 699                 return val;
 700
 701         return (unsigned int)(key + val + (val >> 1)) % number;
 702 }
 703
 704 int
 705 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 706 {
 707         struct lnet_ni *ni;
 708
 709         /* must called with hold of lnet_net_lock */
 710         if (LNET_CPT_NUMBER == 1)
 711                 return 0; /* the only one */
 712
 713         /* take lnet_net_lock(any) would be OK */
 714         if (!list_empty(&the_lnet.ln_nis_cpt)) {
 715                 list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 716                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 717                                 continue;
 718
 719                         LASSERT(ni->ni_cpts != NULL);
 720                         return ni->ni_cpts[lnet_nid_cpt_hash
 721                                            (nid, ni->ni_ncpts)];
 722                 }
 723         }
 724
 725         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 726 }
 727
 728 int
 729 lnet_cpt_of_nid(lnet_nid_t nid)
 730 {
 731         int cpt;
 732         int cpt2;
 733
 734         if (LNET_CPT_NUMBER == 1)
 735                 return 0; /* the only one */
 736
 737         if (list_empty(&the_lnet.ln_nis_cpt))
 738                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 739
 740         cpt = lnet_net_lock_current();
 741         cpt2 = lnet_cpt_of_nid_locked(nid);
 742         lnet_net_unlock(cpt);
 743
 744         return cpt2;
 745 }
 746 EXPORT_SYMBOL(lnet_cpt_of_nid);
 747
 748 int
 749 lnet_islocalnet(__u32 net)
 750 {
 751         struct lnet_ni *ni;
 752         int cpt;
 753
 754         cpt = lnet_net_lock_current();
 755
 756         ni = lnet_net2ni_locked(net, cpt);
 757         if (ni != NULL)
 758                 lnet_ni_decref_locked(ni, cpt);
 759
 760         lnet_net_unlock(cpt);
 761
 762         return ni != NULL;
 763 }
 764
 765 lnet_ni_t  *
 766 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 767 {
 768         struct lnet_ni *ni;
 769         struct list_head *tmp;
 770
 771         LASSERT(cpt != LNET_LOCK_EX);
 772
 773         list_for_each(tmp, &the_lnet.ln_nis) {
 774                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 775
 776                 if (ni->ni_nid == nid) {
 777                         lnet_ni_addref_locked(ni, cpt);
 778                         return ni;
 779                 }
 780         }
 781
 782         return NULL;
 783 }
 784
 785 int
 786 lnet_islocalnid(lnet_nid_t nid)
 787 {
 788         struct lnet_ni *ni;
 789         int cpt;
 790
 791         cpt = lnet_net_lock_current();
 792         ni = lnet_nid2ni_locked(nid, cpt);
 793         if (ni != NULL)
 794                 lnet_ni_decref_locked(ni, cpt);
 795         lnet_net_unlock(cpt);
 796
 797         return ni != NULL;
 798 }
 799
 800 int
 801 lnet_count_acceptor_nis(void)
 802 {
 803         /* Return the # of NIs that need the acceptor. */
 804         int count = 0;
 805         struct list_head *tmp;
 806         struct lnet_ni *ni;
 807         int cpt;
 808
 809         cpt = lnet_net_lock_current();
 810         list_for_each(tmp, &the_lnet.ln_nis) {
 811                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 812
 813                 if (ni->ni_lnd->lnd_accept != NULL)
 814                         count++;
 815         }
 816
 817         lnet_net_unlock(cpt);
 818
 819         return count;
 820 }
 821
 822 static int
 823 lnet_ni_tq_credits(lnet_ni_t *ni)
 824 {
 825         int credits;
 826
 827         LASSERT(ni->ni_ncpts >= 1);
 828
 829         if (ni->ni_ncpts == 1)
 830                 return ni->ni_maxtxcredits;
 831
 832         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
 833         credits = max(credits, 8 * ni->ni_peertxcredits);
 834         credits = min(credits, ni->ni_maxtxcredits);
 835
 836         return credits;
 837 }
 838
 839 static void
 840 lnet_shutdown_lndnis(void)
 841 {
 842         int i;
 843         int islo;
 844         lnet_ni_t *ni;
 845
 846         /* NB called holding the global mutex */
 847
 848         /* All quiet on the API front */
 849         LASSERT(!the_lnet.ln_shutdown);
 850         LASSERT(the_lnet.ln_refcount == 0);
 851         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
 852
 853         lnet_net_lock(LNET_LOCK_EX);
 854         the_lnet.ln_shutdown = 1;       /* flag shutdown */
 855
 856         /* Unlink NIs from the global table */
 857         while (!list_empty(&the_lnet.ln_nis)) {
 858                 ni = list_entry(the_lnet.ln_nis.next,
 859                                     lnet_ni_t, ni_list);
 860                 /* move it to zombie list and nobody can find it anymore */
 861                 list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
 862                 lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
 863
 864                 if (!list_empty(&ni->ni_cptlist)) {
 865                         list_del_init(&ni->ni_cptlist);
 866                         lnet_ni_decref_locked(ni, 0);
 867                 }
 868         }
 869
 870         /* Drop the cached eqwait NI. */
 871         if (the_lnet.ln_eq_waitni != NULL) {
 872                 lnet_ni_decref_locked(the_lnet.ln_eq_waitni, 0);
 873                 the_lnet.ln_eq_waitni = NULL;
 874         }
 875
 876         /* Drop the cached loopback NI. */
 877         if (the_lnet.ln_loni != NULL) {
 878                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
 879                 the_lnet.ln_loni = NULL;
 880         }
 881
 882         lnet_net_unlock(LNET_LOCK_EX);
 883
 884         /* Clear lazy portals and drop delayed messages which hold refs
 885          * on their lnet_msg_t::msg_rxpeer */
 886         for (i = 0; i < the_lnet.ln_nportals; i++)
 887                 LNetClearLazyPortal(i);
 888
 889         /* Clear the peer table and wait for all peers to go (they hold refs on
 890          * their NIs) */
 891         lnet_peer_tables_cleanup();
 892
 893         lnet_net_lock(LNET_LOCK_EX);
 894         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
 895          * and shut them down in guaranteed thread context */
 896         i = 2;
 897         while (!list_empty(&the_lnet.ln_nis_zombie)) {
 898                 int *ref;
 899                 int j;
 900
 901                 ni = list_entry(the_lnet.ln_nis_zombie.next,
 902                                     lnet_ni_t, ni_list);
 903                 list_del_init(&ni->ni_list);
 904                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
 905                         if (*ref == 0)
 906                                 continue;
 907                         /* still busy, add it back to zombie list */
 908                         list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
 909                         break;
 910                 }
 911
 912                 if (!list_empty(&ni->ni_list)) {
 913                         lnet_net_unlock(LNET_LOCK_EX);
 914                         ++i;
 915                         if ((i & (-i)) == i) {
 916                                 CDEBUG(D_WARNING, "Waiting for zombie LNI %s\n",
 917                                        libcfs_nid2str(ni->ni_nid));
 918                         }
 919                         set_current_state(TASK_UNINTERRUPTIBLE);
 920                         schedule_timeout(cfs_time_seconds(1));
 921                         lnet_net_lock(LNET_LOCK_EX);
 922                         continue;
 923                 }
 924
 925                 ni->ni_lnd->lnd_refcount--;
 926                 lnet_net_unlock(LNET_LOCK_EX);
 927
 928                 islo = ni->ni_lnd->lnd_type == LOLND;
 929
 930                 LASSERT(!in_interrupt());
 931                 (ni->ni_lnd->lnd_shutdown)(ni);
 932
 933                 /* can't deref lnd anymore now; it might have unregistered
 934                  * itself...  */
 935
 936                 if (!islo)
 937                         CDEBUG(D_LNI, "Removed LNI %s\n",
 938                                libcfs_nid2str(ni->ni_nid));
 939
 940                 lnet_ni_free(ni);
 941                 i = 2;
 942
 943                 lnet_net_lock(LNET_LOCK_EX);
 944         }
 945
 946         the_lnet.ln_shutdown = 0;
 947         lnet_net_unlock(LNET_LOCK_EX);
 948
 949         if (the_lnet.ln_network_tokens != NULL) {
 950                 LIBCFS_FREE(the_lnet.ln_network_tokens,
 951                             the_lnet.ln_network_tokens_nob);
 952                 the_lnet.ln_network_tokens = NULL;
 953         }
 954 }
 955
 956 static int
 957 lnet_startup_lndnis(void)
 958 {
 959         lnd_t *lnd;
 960         struct lnet_ni *ni;
 961         struct lnet_tx_queue *tq;
 962         struct list_head nilist;
 963         int i;
 964         int rc = 0;
 965         __u32 lnd_type;
 966         int nicount = 0;
 967         char *nets = lnet_get_networks();
 968
 969         INIT_LIST_HEAD(&nilist);
 970
 971         if (nets == NULL)
 972                 goto failed;
 973
 974         rc = lnet_parse_networks(&nilist, nets);
 975         if (rc != 0)
 976                 goto failed;
 977
 978         while (!list_empty(&nilist)) {
 979                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
 980                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 981
 982                 LASSERT(libcfs_isknown_lnd(lnd_type));
 983
 984                 if (lnd_type == CIBLND    ||
 985                     lnd_type == OPENIBLND ||
 986                     lnd_type == IIBLND    ||
 987                     lnd_type == VIBLND) {
 988                         CERROR("LND %s obsoleted\n",
 989                                libcfs_lnd2str(lnd_type));
 990                         goto failed;
 991                 }
 992
 993                 mutex_lock(&the_lnet.ln_lnd_mutex);
 994                 lnd = lnet_find_lnd_by_type(lnd_type);
 995
 996                 if (lnd == NULL) {
 997                         mutex_unlock(&the_lnet.ln_lnd_mutex);
 998                         rc = request_module("%s",
 999                                                 libcfs_lnd2modname(lnd_type));
1000                         mutex_lock(&the_lnet.ln_lnd_mutex);
1001
1002                         lnd = lnet_find_lnd_by_type(lnd_type);
1003                         if (lnd == NULL) {
1004                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1005                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1006                                        libcfs_lnd2str(lnd_type),
1007                                        libcfs_lnd2modname(lnd_type), rc);
1008                                 goto failed;
1009                         }
1010                 }
1011
1012                 lnet_net_lock(LNET_LOCK_EX);
1013                 lnd->lnd_refcount++;
1014                 lnet_net_unlock(LNET_LOCK_EX);
1015
1016                 ni->ni_lnd = lnd;
1017
1018                 rc = (lnd->lnd_startup)(ni);
1019
1020                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1021
1022                 if (rc != 0) {
1023                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1024                                            rc, libcfs_lnd2str(lnd->lnd_type));
1025                         lnet_net_lock(LNET_LOCK_EX);
1026                         lnd->lnd_refcount--;
1027                         lnet_net_unlock(LNET_LOCK_EX);
1028                         goto failed;
1029                 }
1030
1031                 LASSERT(ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1032
1033                 list_del(&ni->ni_list);
1034
1035                 lnet_net_lock(LNET_LOCK_EX);
1036                 /* refcount for ln_nis */
1037                 lnet_ni_addref_locked(ni, 0);
1038                 list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1039                 if (ni->ni_cpts != NULL) {
1040                         list_add_tail(&ni->ni_cptlist,
1041                                           &the_lnet.ln_nis_cpt);
1042                         lnet_ni_addref_locked(ni, 0);
1043                 }
1044
1045                 lnet_net_unlock(LNET_LOCK_EX);
1046
1047                 if (lnd->lnd_type == LOLND) {
1048                         lnet_ni_addref(ni);
1049                         LASSERT(the_lnet.ln_loni == NULL);
1050                         the_lnet.ln_loni = ni;
1051                         continue;
1052                 }
1053
1054                 if (ni->ni_peertxcredits == 0 ||
1055                     ni->ni_maxtxcredits == 0) {
1056                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1057                                            libcfs_lnd2str(lnd->lnd_type),
1058                                            ni->ni_peertxcredits == 0 ?
1059                                            "" : "per-peer ");
1060                         goto failed;
1061                 }
1062
1063                 cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1064                         tq->tq_credits_min =
1065                         tq->tq_credits_max =
1066                         tq->tq_credits = lnet_ni_tq_credits(ni);
1067                 }
1068
1069                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1070                        libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1071                        lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1072                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1073
1074                 nicount++;
1075         }
1076
1077         if (the_lnet.ln_eq_waitni != NULL && nicount > 1) {
1078                 lnd_type = the_lnet.ln_eq_waitni->ni_lnd->lnd_type;
1079                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network\n",
1080                                    libcfs_lnd2str(lnd_type));
1081                 goto failed;
1082         }
1083
1084         return 0;
1085
1086  failed:
1087         lnet_shutdown_lndnis();
1088
1089         while (!list_empty(&nilist)) {
1090                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1091                 list_del(&ni->ni_list);
1092                 lnet_ni_free(ni);
1093         }
1094
1095         return -ENETDOWN;
1096 }
1097
1098 /**
1099  * Initialize LNet library.
1100  *
1101  * Only userspace program needs to call this function - it's automatically
1102  * called in the kernel at module loading time. Caller has to call lnet_fini()
1103  * after a call to lnet_init(), if and only if the latter returned 0. It must
1104  * be called exactly once.
1105  *
1106  * \return 0 on success, and -ve on failures.
1107  */
1108 int
1109 lnet_init(void)
1110 {
1111         int rc;
1112
1113         lnet_assert_wire_constants();
1114         LASSERT(!the_lnet.ln_init);
1115
1116         memset(&the_lnet, 0, sizeof(the_lnet));
1117
1118         /* refer to global cfs_cpt_table for now */
1119         the_lnet.ln_cpt_table   = cfs_cpt_table;
1120         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1121
1122         LASSERT(the_lnet.ln_cpt_number > 0);
1123         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1124                 /* we are under risk of consuming all lh_cookie */
1125                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), please change setting of CPT-table and retry\n",
1126                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1127                 return -1;
1128         }
1129
1130         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1131                 the_lnet.ln_cpt_bits++;
1132
1133         rc = lnet_create_locks();
1134         if (rc != 0) {
1135                 CERROR("Can't create LNet global locks: %d\n", rc);
1136                 return -1;
1137         }
1138
1139         the_lnet.ln_refcount = 0;
1140         the_lnet.ln_init = 1;
1141         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1142         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1143         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1144         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1145
1146         /* The hash table size is the number of bits it takes to express the set
1147          * ln_num_routes, minus 1 (better to under estimate than over so we
1148          * don't waste memory). */
1149         if (rnet_htable_size <= 0)
1150                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1151         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1152                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1153         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1154                                            order_base_2(rnet_htable_size) - 1);
1155
1156         /* All LNDs apart from the LOLND are in separate modules.  They
1157          * register themselves when their module loads, and unregister
1158          * themselves when their module is unloaded. */
1159         lnet_register_lnd(&the_lolnd);
1160         return 0;
1161 }
1162 EXPORT_SYMBOL(lnet_init);
1163
1164 /**
1165  * Finalize LNet library.
1166  *
1167  * Only userspace program needs to call this function. It can be called
1168  * at most once.
1169  *
1170  * \pre lnet_init() called with success.
1171  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1172  */
1173 void
1174 lnet_fini(void)
1175 {
1176         LASSERT(the_lnet.ln_init);
1177         LASSERT(the_lnet.ln_refcount == 0);
1178
1179         while (!list_empty(&the_lnet.ln_lnds))
1180                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1181                                                    lnd_t, lnd_list));
1182         lnet_destroy_locks();
1183
1184         the_lnet.ln_init = 0;
1185 }
1186 EXPORT_SYMBOL(lnet_fini);
1187
1188 /**
1189  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1190  *
1191  * Userspace program should call this after a successful call to lnet_init().
1192  * Users must call this function at least once before any other functions.
1193  * For each successful call there must be a corresponding call to
1194  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1195  * ignored.
1196  *
1197  * The PID used by LNet may be different from the one requested.
1198  * See LNetGetId().
1199  *
1200  * \param requested_pid PID requested by the caller.
1201  *
1202  * \return >= 0 on success, and < 0 error code on failures.
1203  */
1204 int
1205 LNetNIInit(lnet_pid_t requested_pid)
1206 {
1207         int im_a_router = 0;
1208         int rc;
1209
1210         mutex_lock(&the_lnet.ln_api_mutex);
1211
1212         LASSERT(the_lnet.ln_init);
1213         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1214
1215         if (the_lnet.ln_refcount > 0) {
1216                 rc = the_lnet.ln_refcount++;
1217                 goto out;
1218         }
1219
1220         if (requested_pid == LNET_PID_ANY) {
1221                 /* Don't instantiate LNET just for me */
1222                 rc = -ENETDOWN;
1223                 goto failed0;
1224         }
1225
1226         rc = lnet_prepare(requested_pid);
1227         if (rc != 0)
1228                 goto failed0;
1229
1230         rc = lnet_startup_lndnis();
1231         if (rc != 0)
1232                 goto failed1;
1233
1234         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1235         if (rc != 0)
1236                 goto failed2;
1237
1238         rc = lnet_check_routes();
1239         if (rc != 0)
1240                 goto failed2;
1241
1242         rc = lnet_rtrpools_alloc(im_a_router);
1243         if (rc != 0)
1244                 goto failed2;
1245
1246         rc = lnet_acceptor_start();
1247         if (rc != 0)
1248                 goto failed2;
1249
1250         the_lnet.ln_refcount = 1;
1251         /* Now I may use my own API functions... */
1252
1253         /* NB router checker needs the_lnet.ln_ping_info in
1254          * lnet_router_checker -> lnet_update_ni_status_locked */
1255         rc = lnet_ping_target_init();
1256         if (rc != 0)
1257                 goto failed3;
1258
1259         rc = lnet_router_checker_start();
1260         if (rc != 0)
1261                 goto failed4;
1262
1263         lnet_router_debugfs_init();
1264         goto out;
1265
1266  failed4:
1267         lnet_ping_target_fini();
1268  failed3:
1269         the_lnet.ln_refcount = 0;
1270         lnet_acceptor_stop();
1271  failed2:
1272         lnet_destroy_routes();
1273         lnet_shutdown_lndnis();
1274  failed1:
1275         lnet_unprepare();
1276  failed0:
1277         LASSERT(rc < 0);
1278  out:
1279         mutex_unlock(&the_lnet.ln_api_mutex);
1280         return rc;
1281 }
1282 EXPORT_SYMBOL(LNetNIInit);
1283
1284 /**
1285  * Stop LNet interfaces, routing, and forwarding.
1286  *
1287  * Users must call this function once for each successful call to LNetNIInit().
1288  * Once the LNetNIFini() operation has been started, the results of pending
1289  * API operations are undefined.
1290  *
1291  * \return always 0 for current implementation.
1292  */
1293 int
1294 LNetNIFini(void)
1295 {
1296         mutex_lock(&the_lnet.ln_api_mutex);
1297
1298         LASSERT(the_lnet.ln_init);
1299         LASSERT(the_lnet.ln_refcount > 0);
1300
1301         if (the_lnet.ln_refcount != 1) {
1302                 the_lnet.ln_refcount--;
1303         } else {
1304                 LASSERT(!the_lnet.ln_niinit_self);
1305
1306                 lnet_router_debugfs_fini();
1307                 lnet_router_checker_stop();
1308                 lnet_ping_target_fini();
1309
1310                 /* Teardown fns that use my own API functions BEFORE here */
1311                 the_lnet.ln_refcount = 0;
1312
1313                 lnet_acceptor_stop();
1314                 lnet_destroy_routes();
1315                 lnet_shutdown_lndnis();
1316                 lnet_unprepare();
1317         }
1318
1319         mutex_unlock(&the_lnet.ln_api_mutex);
1320         return 0;
1321 }
1322 EXPORT_SYMBOL(LNetNIFini);
1323
1324 /**
1325  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
1326  * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
1327  * internal ioctl handler.
1328  *
1329  * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
1330  *
1331  * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
1332  * The data will be printed to system console. Don't use it excessively.
1333  * \param arg A pointer to lnet_process_id_t, process ID of the peer.
1334  *
1335  * \return Always return 0 when called by users directly (i.e., not via ioctl).
1336  */
1337 int
1338 LNetCtl(unsigned int cmd, void *arg)
1339 {
1340         struct libcfs_ioctl_data *data = arg;
1341         lnet_process_id_t id = {0};
1342         lnet_ni_t *ni;
1343         int rc;
1344         unsigned long secs_passed;
1345
1346         LASSERT(the_lnet.ln_init);
1347         LASSERT(the_lnet.ln_refcount > 0);
1348
1349         switch (cmd) {
1350         case IOC_LIBCFS_GET_NI:
1351                 rc = LNetGetId(data->ioc_count, &id);
1352                 data->ioc_nid = id.nid;
1353                 return rc;
1354
1355         case IOC_LIBCFS_FAIL_NID:
1356                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1357
1358         case IOC_LIBCFS_ADD_ROUTE:
1359                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1360                                     data->ioc_nid, data->ioc_priority);
1361                 return (rc != 0) ? rc : lnet_check_routes();
1362
1363         case IOC_LIBCFS_DEL_ROUTE:
1364                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1365
1366         case IOC_LIBCFS_GET_ROUTE:
1367                 return lnet_get_route(data->ioc_count,
1368                                       &data->ioc_net, &data->ioc_count,
1369                                       &data->ioc_nid, &data->ioc_flags,
1370                                       &data->ioc_priority);
1371         case IOC_LIBCFS_NOTIFY_ROUTER:
1372                 secs_passed = (ktime_get_real_seconds() - data->ioc_u64[0]);
1373                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1374                                    jiffies - secs_passed * HZ);
1375
1376         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1377                 /* This can be removed once lustre stops calling it */
1378                 return 0;
1379
1380         case IOC_LIBCFS_LNET_DIST:
1381                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1382                 if (rc < 0 && rc != -EHOSTUNREACH)
1383                         return rc;
1384
1385                 data->ioc_u32[0] = rc;
1386                 return 0;
1387
1388         case IOC_LIBCFS_TESTPROTOCOMPAT:
1389                 lnet_net_lock(LNET_LOCK_EX);
1390                 the_lnet.ln_testprotocompat = data->ioc_flags;
1391                 lnet_net_unlock(LNET_LOCK_EX);
1392                 return 0;
1393
1394         case IOC_LIBCFS_PING:
1395                 id.nid = data->ioc_nid;
1396                 id.pid = data->ioc_u32[0];
1397                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
1398                                (lnet_process_id_t *)data->ioc_pbuf1,
1399                                data->ioc_plen1/sizeof(lnet_process_id_t));
1400                 if (rc < 0)
1401                         return rc;
1402                 data->ioc_count = rc;
1403                 return 0;
1404
1405         case IOC_LIBCFS_DEBUG_PEER: {
1406                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1407                  * via an ioctl */
1408                 id = *((lnet_process_id_t *) arg);
1409
1410                 lnet_debug_peer(id.nid);
1411
1412                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
1413                 if (ni == NULL) {
1414                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
1415                 } else {
1416                         if (ni->ni_lnd->lnd_ctl == NULL) {
1417                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1418                                        libcfs_id2str(id));
1419                         } else {
1420                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1421                         }
1422
1423                         lnet_ni_decref(ni);
1424                 }
1425                 return 0;
1426         }
1427
1428         default:
1429                 ni = lnet_net2ni(data->ioc_net);
1430                 if (ni == NULL)
1431                         return -EINVAL;
1432
1433                 if (ni->ni_lnd->lnd_ctl == NULL)
1434                         rc = -EINVAL;
1435                 else
1436                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1437
1438                 lnet_ni_decref(ni);
1439                 return rc;
1440         }
1441         /* not reached */
1442 }
1443 EXPORT_SYMBOL(LNetCtl);
1444
1445 /**
1446  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
1447  * all interfaces share a same PID, as requested by LNetNIInit().
1448  *
1449  * \param index Index of the interface to look up.
1450  * \param id On successful return, this location will hold the
1451  * lnet_process_id_t ID of the interface.
1452  *
1453  * \retval 0 If an interface exists at \a index.
1454  * \retval -ENOENT If no interface has been found.
1455  */
1456 int
1457 LNetGetId(unsigned int index, lnet_process_id_t *id)
1458 {
1459         struct lnet_ni *ni;
1460         struct list_head *tmp;
1461         int cpt;
1462         int rc = -ENOENT;
1463
1464         LASSERT(the_lnet.ln_init);
1465
1466         /* LNetNI initilization failed? */
1467         if (the_lnet.ln_refcount == 0)
1468                 return rc;
1469
1470         cpt = lnet_net_lock_current();
1471
1472         list_for_each(tmp, &the_lnet.ln_nis) {
1473                 if (index-- != 0)
1474                         continue;
1475
1476                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1477
1478                 id->nid = ni->ni_nid;
1479                 id->pid = the_lnet.ln_pid;
1480                 rc = 0;
1481                 break;
1482         }
1483
1484         lnet_net_unlock(cpt);
1485         return rc;
1486 }
1487 EXPORT_SYMBOL(LNetGetId);
1488
1489 /**
1490  * Print a string representation of handle \a h into buffer \a str of
1491  * \a len bytes.
1492  */
1493 void
1494 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1495 {
1496         snprintf(str, len, "%#llx", h.cookie);
1497 }
1498 EXPORT_SYMBOL(LNetSnprintHandle);
1499
1500 static int
1501 lnet_create_ping_info(void)
1502 {
1503         int i;
1504         int n;
1505         int rc;
1506         unsigned int infosz;
1507         lnet_ni_t *ni;
1508         lnet_process_id_t id;
1509         lnet_ping_info_t *pinfo;
1510
1511         for (n = 0; ; n++) {
1512                 rc = LNetGetId(n, &id);
1513                 if (rc == -ENOENT)
1514                         break;
1515
1516                 LASSERT(rc == 0);
1517         }
1518
1519         infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
1520         LIBCFS_ALLOC(pinfo, infosz);
1521         if (pinfo == NULL) {
1522                 CERROR("Can't allocate ping info[%d]\n", n);
1523                 return -ENOMEM;
1524         }
1525
1526         pinfo->pi_nnis    = n;
1527         pinfo->pi_pid     = the_lnet.ln_pid;
1528         pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
1529         pinfo->pi_features = LNET_PING_FEAT_NI_STATUS;
1530
1531         for (i = 0; i < n; i++) {
1532                 lnet_ni_status_t *ns = &pinfo->pi_ni[i];
1533
1534                 rc = LNetGetId(i, &id);
1535                 LASSERT(rc == 0);
1536
1537                 ns->ns_nid    = id.nid;
1538                 ns->ns_status = LNET_NI_STATUS_UP;
1539
1540                 lnet_net_lock(0);
1541
1542                 ni = lnet_nid2ni_locked(id.nid, 0);
1543                 LASSERT(ni != NULL);
1544
1545                 lnet_ni_lock(ni);
1546                 LASSERT(ni->ni_status == NULL);
1547                 ni->ni_status = ns;
1548                 lnet_ni_unlock(ni);
1549
1550                 lnet_ni_decref_locked(ni, 0);
1551                 lnet_net_unlock(0);
1552         }
1553
1554         the_lnet.ln_ping_info = pinfo;
1555         return 0;
1556 }
1557
1558 static void
1559 lnet_destroy_ping_info(void)
1560 {
1561         struct lnet_ni *ni;
1562
1563         lnet_net_lock(0);
1564
1565         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1566                 lnet_ni_lock(ni);
1567                 ni->ni_status = NULL;
1568                 lnet_ni_unlock(ni);
1569         }
1570
1571         lnet_net_unlock(0);
1572
1573         LIBCFS_FREE(the_lnet.ln_ping_info,
1574                     offsetof(lnet_ping_info_t,
1575                              pi_ni[the_lnet.ln_ping_info->pi_nnis]));
1576         the_lnet.ln_ping_info = NULL;
1577 }
1578
1579 int
1580 lnet_ping_target_init(void)
1581 {
1582         lnet_md_t md = { NULL };
1583         lnet_handle_me_t meh;
1584         lnet_process_id_t id;
1585         int rc;
1586         int rc2;
1587         int infosz;
1588
1589         rc = lnet_create_ping_info();
1590         if (rc != 0)
1591                 return rc;
1592
1593         /* We can have a tiny EQ since we only need to see the unlink event on
1594          * teardown, which by definition is the last one! */
1595         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1596         if (rc != 0) {
1597                 CERROR("Can't allocate ping EQ: %d\n", rc);
1598                 goto failed_0;
1599         }
1600
1601         memset(&id, 0, sizeof(lnet_process_id_t));
1602         id.nid = LNET_NID_ANY;
1603         id.pid = LNET_PID_ANY;
1604
1605         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1606                           LNET_PROTO_PING_MATCHBITS, 0,
1607                           LNET_UNLINK, LNET_INS_AFTER,
1608                           &meh);
1609         if (rc != 0) {
1610                 CERROR("Can't create ping ME: %d\n", rc);
1611                 goto failed_1;
1612         }
1613
1614         /* initialize md content */
1615         infosz = offsetof(lnet_ping_info_t,
1616                           pi_ni[the_lnet.ln_ping_info->pi_nnis]);
1617         md.start     = the_lnet.ln_ping_info;
1618         md.length    = infosz;
1619         md.threshold = LNET_MD_THRESH_INF;
1620         md.max_size  = 0;
1621         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1622                        LNET_MD_MANAGE_REMOTE;
1623         md.user_ptr  = NULL;
1624         md.eq_handle = the_lnet.ln_ping_target_eq;
1625
1626         rc = LNetMDAttach(meh, md,
1627                           LNET_RETAIN,
1628                           &the_lnet.ln_ping_target_md);
1629         if (rc != 0) {
1630                 CERROR("Can't attach ping MD: %d\n", rc);
1631                 goto failed_2;
1632         }
1633
1634         return 0;
1635
1636  failed_2:
1637         rc2 = LNetMEUnlink(meh);
1638         LASSERT(rc2 == 0);
1639  failed_1:
1640         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1641         LASSERT(rc2 == 0);
1642  failed_0:
1643         lnet_destroy_ping_info();
1644         return rc;
1645 }
1646
1647 void
1648 lnet_ping_target_fini(void)
1649 {
1650         lnet_event_t event;
1651         int rc;
1652         int which;
1653         int timeout_ms = 1000;
1654         sigset_t blocked = cfs_block_allsigs();
1655
1656         LNetMDUnlink(the_lnet.ln_ping_target_md);
1657         /* NB md could be busy; this just starts the unlink */
1658
1659         for (;;) {
1660                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1661                                 timeout_ms, &event, &which);
1662
1663                 /* I expect overflow... */
1664                 LASSERT(rc >= 0 || rc == -EOVERFLOW);
1665
1666                 if (rc == 0) {
1667                         /* timed out: provide a diagnostic */
1668                         CWARN("Still waiting for ping MD to unlink\n");
1669                         timeout_ms *= 2;
1670                         continue;
1671                 }
1672
1673                 /* Got a valid event */
1674                 if (event.unlinked)
1675                         break;
1676         }
1677
1678         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1679         LASSERT(rc == 0);
1680         lnet_destroy_ping_info();
1681         cfs_restore_sigs(blocked);
1682 }
1683
1684 int
1685 lnet_ping(lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1686 {
1687         lnet_handle_eq_t eqh;
1688         lnet_handle_md_t mdh;
1689         lnet_event_t event;
1690         lnet_md_t md = { NULL };
1691         int which;
1692         int unlinked = 0;
1693         int replied = 0;
1694         const int a_long_time = 60000; /* mS */
1695         int infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
1696         lnet_ping_info_t *info;
1697         lnet_process_id_t tmpid;
1698         int i;
1699         int nob;
1700         int rc;
1701         int rc2;
1702         sigset_t blocked;
1703
1704         if (n_ids <= 0 ||
1705             id.nid == LNET_NID_ANY ||
1706             timeout_ms > 500000 ||            /* arbitrary limit! */
1707             n_ids > 20)                  /* arbitrary limit! */
1708                 return -EINVAL;
1709
1710         if (id.pid == LNET_PID_ANY)
1711                 id.pid = LUSTRE_SRV_LNET_PID;
1712
1713         LIBCFS_ALLOC(info, infosz);
1714         if (info == NULL)
1715                 return -ENOMEM;
1716
1717         /* NB 2 events max (including any unlink event) */
1718         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1719         if (rc != 0) {
1720                 CERROR("Can't allocate EQ: %d\n", rc);
1721                 goto out_0;
1722         }
1723
1724         /* initialize md content */
1725         md.start     = info;
1726         md.length    = infosz;
1727         md.threshold = 2; /*GET/REPLY*/
1728         md.max_size  = 0;
1729         md.options   = LNET_MD_TRUNCATE;
1730         md.user_ptr  = NULL;
1731         md.eq_handle = eqh;
1732
1733         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
1734         if (rc != 0) {
1735                 CERROR("Can't bind MD: %d\n", rc);
1736                 goto out_1;
1737         }
1738
1739         rc = LNetGet(LNET_NID_ANY, mdh, id,
1740                      LNET_RESERVED_PORTAL,
1741                      LNET_PROTO_PING_MATCHBITS, 0);
1742
1743         if (rc != 0) {
1744                 /* Don't CERROR; this could be deliberate! */
1745
1746                 rc2 = LNetMDUnlink(mdh);
1747                 LASSERT(rc2 == 0);
1748
1749                 /* NB must wait for the UNLINK event below... */
1750                 unlinked = 1;
1751                 timeout_ms = a_long_time;
1752         }
1753
1754         do {
1755                 /* MUST block for unlink to complete */
1756                 if (unlinked)
1757                         blocked = cfs_block_allsigs();
1758
1759                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1760
1761                 if (unlinked)
1762                         cfs_restore_sigs(blocked);
1763
1764                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1765                        (rc2 <= 0) ? -1 : event.type,
1766                        (rc2 <= 0) ? -1 : event.status,
1767                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1768
1769                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
1770
1771                 if (rc2 <= 0 || event.status != 0) {
1772                         /* timeout or error */
1773                         if (!replied && rc == 0)
1774                                 rc = (rc2 < 0) ? rc2 :
1775                                      (rc2 == 0) ? -ETIMEDOUT :
1776                                      event.status;
1777
1778                         if (!unlinked) {
1779                                 /* Ensure completion in finite time... */
1780                                 LNetMDUnlink(mdh);
1781                                 /* No assertion (racing with network) */
1782                                 unlinked = 1;
1783                                 timeout_ms = a_long_time;
1784                         } else if (rc2 == 0) {
1785                                 /* timed out waiting for unlink */
1786                                 CWARN("ping %s: late network completion\n",
1787                                       libcfs_id2str(id));
1788                         }
1789                 } else if (event.type == LNET_EVENT_REPLY) {
1790                         replied = 1;
1791                         rc = event.mlength;
1792                 }
1793
1794         } while (rc2 <= 0 || !event.unlinked);
1795
1796         if (!replied) {
1797                 if (rc >= 0)
1798                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1799                               libcfs_id2str(id));
1800                 rc = -EIO;
1801                 goto out_1;
1802         }
1803
1804         nob = rc;
1805         LASSERT(nob >= 0 && nob <= infosz);
1806
1807         rc = -EPROTO;                      /* if I can't parse... */
1808
1809         if (nob < 8) {
1810                 /* can't check magic/version */
1811                 CERROR("%s: ping info too short %d\n",
1812                        libcfs_id2str(id), nob);
1813                 goto out_1;
1814         }
1815
1816         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1817                 lnet_swap_pinginfo(info);
1818         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1819                 CERROR("%s: Unexpected magic %08x\n",
1820                        libcfs_id2str(id), info->pi_magic);
1821                 goto out_1;
1822         }
1823
1824         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
1825                 CERROR("%s: ping w/o NI status: 0x%x\n",
1826                        libcfs_id2str(id), info->pi_features);
1827                 goto out_1;
1828         }
1829
1830         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
1831                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1832                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
1833                 goto out_1;
1834         }
1835
1836         if (info->pi_nnis < n_ids)
1837                 n_ids = info->pi_nnis;
1838
1839         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
1840                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1841                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
1842                 goto out_1;
1843         }
1844
1845         rc = -EFAULT;                      /* If I SEGV... */
1846
1847         memset(&tmpid, 0, sizeof(tmpid));
1848         for (i = 0; i < n_ids; i++) {
1849                 tmpid.pid = info->pi_pid;
1850                 tmpid.nid = info->pi_ni[i].ns_nid;
1851                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1852                         goto out_1;
1853         }
1854         rc = info->pi_nnis;
1855
1856  out_1:
1857         rc2 = LNetEQFree(eqh);
1858         if (rc2 != 0)
1859                 CERROR("rc2 %d\n", rc2);
1860         LASSERT(rc2 == 0);
1861
1862  out_0:
1863         LIBCFS_FREE(info, infosz);
1864         return rc;
1865 }