lnet/lnet/api-ni.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.sf.net/projects/lustre/
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_LNET
  23 #include <lnet/lib-lnet.h>
  24
  25 #ifdef __KERNEL__
  26 #define D_LNI D_CONSOLE
  27 #else
  28 #define D_LNI D_CONFIG
  29 #endif
  30
  31 lnet_t      the_lnet;                           /* THE state of the network */
  32
  33 #ifdef __KERNEL__
  34
  35 static char *ip2nets = "";
  36 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  37                 "LNET network <- IP table");
  38
  39 static char *networks = "";
  40 CFS_MODULE_PARM(networks, "s", charp, 0444,
  41                 "local networks");
  42
  43 static char *routes = "";
  44 CFS_MODULE_PARM(routes, "s", charp, 0444,
  45                 "routes to non-local networks");
  46
  47 static char *portals_compatibility = "none";
  48 CFS_MODULE_PARM(portals_compatibility, "s", charp, 0444,
  49                 "wire protocol compatibility: 'strong'|'weak'|'none'");
  50
  51 char *
  52 lnet_get_routes(void)
  53 {
  54         return routes;
  55 }
  56
  57 char *
  58 lnet_get_networks(void)
  59 {
  60         char   *nets;
  61         int     rc;
  62
  63         if (*networks != 0 && *ip2nets != 0) {
  64                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  65                                    "'ip2nets' but not both at once\n");
  66                 return NULL;
  67         }
  68
  69         if (*ip2nets != 0) {
  70                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  71                 return (rc == 0) ? nets : NULL;
  72         }
  73
  74         if (*networks != 0)
  75                 return networks;
  76
  77         return "tcp";
  78 }
  79
  80 int
  81 lnet_get_portals_compatibility(void)
  82 {
  83         if (!strcmp(portals_compatibility, "none")) {
  84                 return 0;
  85         }
  86
  87         if (!strcmp(portals_compatibility, "weak")) {
  88                 return 1;
  89                 LCONSOLE_WARN("Starting in weak portals-compatible mode\n");
  90         }
  91
  92         if (!strcmp(portals_compatibility, "strong")) {
  93                 return 2;
  94                 LCONSOLE_WARN("Starting in strong portals-compatible mode\n");
  95         }
  96
  97         LCONSOLE_ERROR_MSG(0x102, "portals_compatibility=\"%s\" not supported\n",
  98                            portals_compatibility);
  99         return -EINVAL;
 100 }
 101
 102 void
 103 lnet_init_locks(void)
 104 {
 105         spin_lock_init (&the_lnet.ln_lock);
 106         cfs_waitq_init (&the_lnet.ln_waitq);
 107         init_mutex(&the_lnet.ln_lnd_mutex);
 108         init_mutex(&the_lnet.ln_api_mutex);
 109 }
 110
 111 void
 112 lnet_fini_locks(void)
 113 {
 114 }
 115
 116 #else
 117
 118 char *
 119 lnet_get_routes(void)
 120 {
 121         char *str = getenv("LNET_ROUTES");
 122
 123         return (str == NULL) ? "" : str;
 124 }
 125
 126 char *
 127 lnet_get_networks (void)
 128 {
 129         static char       default_networks[256];
 130         char             *networks = getenv ("LNET_NETWORKS");
 131         char             *ip2nets  = getenv ("LNET_IP2NETS");
 132         char             *str;
 133         char             *sep;
 134         int               len;
 135         int               nob;
 136         int               rc;
 137         struct list_head *tmp;
 138
 139 #ifdef NOT_YET
 140         if (networks != NULL && ip2nets != NULL) {
 141                 LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
 142                                    " 'LNET_IP2NETS' but not both at once\n");
 143                 return NULL;
 144         }
 145
 146         if (ip2nets != NULL) {
 147                 rc = lnet_parse_ip2nets(&networks, ip2nets);
 148                 return (rc == 0) ? networks : NULL;
 149         }
 150 #else
 151         ip2nets = NULL;
 152         rc = 0;
 153 #endif
 154         if (networks != NULL)
 155                 return networks;
 156
 157         /* In userland, the default 'networks=' is the list of known net types */
 158
 159         len = sizeof(default_networks);
 160         str = default_networks;
 161         *str = 0;
 162         sep = "";
 163
 164         list_for_each (tmp, &the_lnet.ln_lnds) {
 165                         lnd_t *lnd = list_entry(tmp, lnd_t, lnd_list);
 166
 167                         nob = snprintf(str, len, "%s%s", sep,
 168                                        libcfs_lnd2str(lnd->lnd_type));
 169                         len -= nob;
 170                         if (len < 0) {
 171                                 /* overflowed the string; leave it where it was */
 172                                 *str = 0;
 173                                 break;
 174                         }
 175
 176                         str += nob;
 177                         sep = ",";
 178         }
 179
 180         return default_networks;
 181 }
 182
 183 int
 184 lnet_get_portals_compatibility(void)
 185 {
 186         return 0;
 187 }
 188
 189 # ifndef HAVE_LIBPTHREAD
 190
 191 void lnet_init_locks(void)
 192 {
 193         the_lnet.ln_lock = 0;
 194         the_lnet.ln_lnd_mutex = 0;
 195         the_lnet.ln_api_mutex = 0;
 196 }
 197
 198 void lnet_fini_locks(void)
 199 {
 200         LASSERT (the_lnet.ln_api_mutex == 0);
 201         LASSERT (the_lnet.ln_lnd_mutex == 0);
 202         LASSERT (the_lnet.ln_lock == 0);
 203 }
 204
 205 # else
 206
 207 void lnet_init_locks(void)
 208 {
 209         pthread_cond_init(&the_lnet.ln_cond, NULL);
 210         pthread_mutex_init(&the_lnet.ln_lock, NULL);
 211         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 212         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 213 }
 214
 215 void lnet_fini_locks(void)
 216 {
 217         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 218         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 219         pthread_mutex_destroy(&the_lnet.ln_lock);
 220         pthread_cond_destroy(&the_lnet.ln_cond);
 221 }
 222
 223 # endif
 224 #endif
 225
 226 void lnet_assert_wire_constants (void)
 227 {
 228         /* Wire protocol assertions generated by 'wirecheck'
 229          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 230          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 231          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 232
 233         /* Constants... */
 234         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 235         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 236         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 237         CLASSERT (LNET_MSG_ACK == 0);
 238         CLASSERT (LNET_MSG_PUT == 1);
 239         CLASSERT (LNET_MSG_GET == 2);
 240         CLASSERT (LNET_MSG_REPLY == 3);
 241         CLASSERT (LNET_MSG_HELLO == 4);
 242
 243         /* Checks for struct ptl_handle_wire_t */
 244         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 245         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 246         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 247         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 248         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 249
 250         /* Checks for struct lnet_magicversion_t */
 251         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 252         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 253         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 254         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 255         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 256         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 257         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 258
 259         /* Checks for struct lnet_hdr_t */
 260         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 261         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 262         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 263         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 264         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 265         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 266         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 267         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 268         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 269         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 270         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 271         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 272         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 273         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 274         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 275
 276         /* Ack */
 277         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 278         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 279         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 280         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 281         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 282         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 283
 284         /* Put */
 285         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 286         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 287         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 288         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 289         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 290         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 291         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 292         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 293         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 294         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 295
 296         /* Get */
 297         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 298         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 299         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 300         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 301         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 302         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 303         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 304         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 305         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 306         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 307
 308         /* Reply */
 309         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 310         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 311
 312         /* Hello */
 313         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 314         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 315         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 316         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 317 }
 318
 319 lnd_t *
 320 lnet_find_lnd_by_type (int type)
 321 {
 322         lnd_t              *lnd;
 323         struct list_head   *tmp;
 324
 325         /* holding lnd mutex */
 326         list_for_each (tmp, &the_lnet.ln_lnds) {
 327                 lnd = list_entry(tmp, lnd_t, lnd_list);
 328
 329                 if (lnd->lnd_type == type)
 330                         return lnd;
 331         }
 332
 333         return NULL;
 334 }
 335
 336 void
 337 lnet_register_lnd (lnd_t *lnd)
 338 {
 339         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 340
 341         LASSERT (the_lnet.ln_init);
 342         LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
 343         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 344
 345         list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
 346         lnd->lnd_refcount = 0;
 347
 348         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 349
 350         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 351 }
 352
 353 void
 354 lnet_unregister_lnd (lnd_t *lnd)
 355 {
 356         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 357
 358         LASSERT (the_lnet.ln_init);
 359         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 360         LASSERT (lnd->lnd_refcount == 0);
 361
 362         list_del (&lnd->lnd_list);
 363         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 364
 365         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 366 }
 367
 368 #ifndef LNET_USE_LIB_FREELIST
 369
 370 int
 371 lnet_descriptor_setup (void)
 372 {
 373         return 0;
 374 }
 375
 376 void
 377 lnet_descriptor_cleanup (void)
 378 {
 379 }
 380
 381 #else
 382
 383 int
 384 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
 385 {
 386         char *space;
 387
 388         LASSERT (n > 0);
 389
 390         size += offsetof (lnet_freeobj_t, fo_contents);
 391
 392         LIBCFS_ALLOC(space, n * size);
 393         if (space == NULL)
 394                 return (-ENOMEM);
 395
 396         CFS_INIT_LIST_HEAD (&fl->fl_list);
 397         fl->fl_objs = space;
 398         fl->fl_nobjs = n;
 399         fl->fl_objsize = size;
 400
 401         do
 402         {
 403                 memset (space, 0, size);
 404                 list_add ((struct list_head *)space, &fl->fl_list);
 405                 space += size;
 406         } while (--n != 0);
 407
 408         return (0);
 409 }
 410
 411 void
 412 lnet_freelist_fini (lnet_freelist_t *fl)
 413 {
 414         struct list_head *el;
 415         int               count;
 416
 417         if (fl->fl_nobjs == 0)
 418                 return;
 419
 420         count = 0;
 421         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 422                 count++;
 423
 424         LASSERT (count == fl->fl_nobjs);
 425
 426         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 427         memset (fl, 0, sizeof (fl));
 428 }
 429
 430 int
 431 lnet_descriptor_setup (void)
 432 {
 433         /* NB on failure caller must still call lnet_descriptor_cleanup */
 434         /*               ******                                         */
 435         int        rc;
 436
 437         memset (&the_lnet.ln_free_mes,  0, sizeof (the_lnet.ln_free_mes));
 438         memset (&the_lnet.ln_free_msgs, 0, sizeof (the_lnet.ln_free_msgs));
 439         memset (&the_lnet.ln_free_mds,  0, sizeof (the_lnet.ln_free_mds));
 440         memset (&the_lnet.ln_free_eqs,  0, sizeof (the_lnet.ln_free_eqs));
 441
 442         rc = lnet_freelist_init(&the_lnet.ln_free_mes,
 443                                 MAX_MES, sizeof (lnet_me_t));
 444         if (rc != 0)
 445                 return (rc);
 446
 447         rc = lnet_freelist_init(&the_lnet.ln_free_msgs,
 448                                 MAX_MSGS, sizeof (lnet_msg_t));
 449         if (rc != 0)
 450                 return (rc);
 451
 452         rc = lnet_freelist_init(&the_lnet.ln_free_mds,
 453                                 MAX_MDS, sizeof (lnet_libmd_t));
 454         if (rc != 0)
 455                 return (rc);
 456
 457         rc = lnet_freelist_init(&the_lnet.ln_free_eqs,
 458                                 MAX_EQS, sizeof (lnet_eq_t));
 459         return (rc);
 460 }
 461
 462 void
 463 lnet_descriptor_cleanup (void)
 464 {
 465         lnet_freelist_fini (&the_lnet.ln_free_mes);
 466         lnet_freelist_fini (&the_lnet.ln_free_msgs);
 467         lnet_freelist_fini (&the_lnet.ln_free_mds);
 468         lnet_freelist_fini (&the_lnet.ln_free_eqs);
 469 }
 470
 471 #endif
 472
 473 __u64
 474 lnet_create_interface_cookie (void)
 475 {
 476         /* NB the interface cookie in wire handles guards against delayed
 477          * replies and ACKs appearing valid after reboot. Initialisation time,
 478          * even if it's only implemented to millisecond resolution is probably
 479          * easily good enough. */
 480         struct timeval tv;
 481         __u64          cookie;
 482 #ifndef __KERNEL__
 483         int            rc = gettimeofday (&tv, NULL);
 484         LASSERT (rc == 0);
 485 #else
 486         do_gettimeofday(&tv);
 487 #endif
 488         cookie = tv.tv_sec;
 489         cookie *= 1000000;
 490         cookie += tv.tv_usec;
 491         return cookie;
 492 }
 493
 494 int
 495 lnet_setup_handle_hash (void)
 496 {
 497         int       i;
 498
 499         /* Arbitrary choice of hash table size */
 500 #ifdef __KERNEL__
 501         the_lnet.ln_lh_hash_size = CFS_PAGE_SIZE / sizeof (struct list_head);
 502 #else
 503         the_lnet.ln_lh_hash_size = (MAX_MES + MAX_MDS + MAX_EQS)/4;
 504 #endif
 505         LIBCFS_ALLOC(the_lnet.ln_lh_hash_table,
 506                      the_lnet.ln_lh_hash_size * sizeof (struct list_head));
 507         if (the_lnet.ln_lh_hash_table == NULL)
 508                 return (-ENOMEM);
 509
 510         for (i = 0; i < the_lnet.ln_lh_hash_size; i++)
 511                 CFS_INIT_LIST_HEAD (&the_lnet.ln_lh_hash_table[i]);
 512
 513         the_lnet.ln_next_object_cookie = LNET_COOKIE_TYPES;
 514
 515         return (0);
 516 }
 517
 518 void
 519 lnet_cleanup_handle_hash (void)
 520 {
 521         if (the_lnet.ln_lh_hash_table == NULL)
 522                 return;
 523
 524         LIBCFS_FREE(the_lnet.ln_lh_hash_table,
 525                     the_lnet.ln_lh_hash_size * sizeof (struct list_head));
 526 }
 527
 528 lnet_libhandle_t *
 529 lnet_lookup_cookie (__u64 cookie, int type)
 530 {
 531         /* ALWAYS called with LNET_LOCK held */
 532         struct list_head    *list;
 533         struct list_head    *el;
 534         unsigned int         hash;
 535
 536         if ((cookie & (LNET_COOKIE_TYPES - 1)) != type)
 537                 return (NULL);
 538
 539         hash = ((unsigned int)cookie) % the_lnet.ln_lh_hash_size;
 540         list = &the_lnet.ln_lh_hash_table[hash];
 541
 542         list_for_each (el, list) {
 543                 lnet_libhandle_t *lh = list_entry (el, lnet_libhandle_t,
 544                                                   lh_hash_chain);
 545
 546                 if (lh->lh_cookie == cookie)
 547                         return (lh);
 548         }
 549
 550         return (NULL);
 551 }
 552
 553 void
 554 lnet_initialise_handle (lnet_libhandle_t *lh, int type)
 555 {
 556         /* ALWAYS called with LNET_LOCK held */
 557         unsigned int    hash;
 558
 559         LASSERT (type >= 0 && type < LNET_COOKIE_TYPES);
 560         lh->lh_cookie = the_lnet.ln_next_object_cookie | type;
 561         the_lnet.ln_next_object_cookie += LNET_COOKIE_TYPES;
 562
 563         hash = ((unsigned int)lh->lh_cookie) % the_lnet.ln_lh_hash_size;
 564         list_add (&lh->lh_hash_chain, &the_lnet.ln_lh_hash_table[hash]);
 565 }
 566
 567 void
 568 lnet_invalidate_handle (lnet_libhandle_t *lh)
 569 {
 570         /* ALWAYS called with LNET_LOCK held */
 571         list_del (&lh->lh_hash_chain);
 572 }
 573
 574 int
 575 lnet_init_finalizers(void)
 576 {
 577 #ifdef __KERNEL__
 578         int    i;
 579
 580         the_lnet.ln_nfinalizers = num_online_cpus();
 581
 582         LIBCFS_ALLOC(the_lnet.ln_finalizers,
 583                      the_lnet.ln_nfinalizers *
 584                      sizeof(*the_lnet.ln_finalizers));
 585         if (the_lnet.ln_finalizers == NULL) {
 586                 CERROR("Can't allocate ln_finalizers\n");
 587                 return -ENOMEM;
 588         }
 589
 590         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 591                 the_lnet.ln_finalizers[i] = NULL;
 592 #else
 593         the_lnet.ln_finalizing = 0;
 594 #endif
 595
 596         CFS_INIT_LIST_HEAD(&the_lnet.ln_finalizeq);
 597         return 0;
 598 }
 599
 600 void
 601 lnet_fini_finalizers(void)
 602 {
 603 #ifdef __KERNEL__
 604         int    i;
 605
 606         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 607                 LASSERT (the_lnet.ln_finalizers[i] == NULL);
 608
 609         LIBCFS_FREE(the_lnet.ln_finalizers,
 610                     the_lnet.ln_nfinalizers *
 611                     sizeof(*the_lnet.ln_finalizers));
 612 #else
 613         LASSERT (!the_lnet.ln_finalizing);
 614 #endif
 615         LASSERT (list_empty(&the_lnet.ln_finalizeq));
 616 }
 617
 618 int
 619 lnet_prepare(lnet_pid_t requested_pid)
 620 {
 621         /* Prepare to bring up the network */
 622         int               rc = 0;
 623         int               i;
 624
 625         LASSERT (the_lnet.ln_refcount == 0);
 626
 627         the_lnet.ln_routing = 0;
 628
 629 #ifdef __KERNEL__
 630         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 631         the_lnet.ln_pid = requested_pid;
 632 #else
 633         /* My PID must be unique on this node and flag I'm userspace */
 634         the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 635 #endif
 636
 637         rc = lnet_descriptor_setup();
 638         if (rc != 0)
 639                 goto failed0;
 640
 641         memset(&the_lnet.ln_counters, 0,
 642                sizeof(the_lnet.ln_counters));
 643
 644         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
 645         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
 646         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
 647         CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
 648         CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
 649         CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
 650         CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
 651         CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
 652
 653         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 654
 655         lnet_init_rtrpools();
 656
 657         rc = lnet_setup_handle_hash ();
 658         if (rc != 0)
 659                 goto failed0;
 660
 661         rc = lnet_create_peer_table();
 662         if (rc != 0)
 663                 goto failed1;
 664
 665         rc = lnet_init_finalizers();
 666         if (rc != 0)
 667                 goto failed2;
 668
 669         the_lnet.ln_nportals = MAX_PORTALS;
 670         LIBCFS_ALLOC(the_lnet.ln_portals,
 671                      the_lnet.ln_nportals *
 672                      sizeof(*the_lnet.ln_portals));
 673         if (the_lnet.ln_portals == NULL) {
 674                 rc = -ENOMEM;
 675                 goto failed3;
 676         }
 677
 678         for (i = 0; i < the_lnet.ln_nportals; i++) {
 679                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_ml));
 680                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
 681                 the_lnet.ln_portals[i].ptl_options = 0;
 682         }
 683
 684         return 0;
 685
 686  failed3:
 687         lnet_fini_finalizers();
 688  failed2:
 689         lnet_destroy_peer_table();
 690  failed1:
 691         lnet_cleanup_handle_hash();
 692  failed0:
 693         lnet_descriptor_cleanup();
 694         return rc;
 695 }
 696
 697 int
 698 lnet_unprepare (void)
 699 {
 700         int       idx;
 701
 702         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 703          * have shut down already, so it is safe to unlink and free all
 704          * descriptors, even those that appear committed to a network op (eg MD
 705          * with non-zero pending count) */
 706
 707         lnet_fail_nid(LNET_NID_ANY, 0);
 708
 709         LASSERT (list_empty(&the_lnet.ln_test_peers));
 710         LASSERT (the_lnet.ln_refcount == 0);
 711         LASSERT (list_empty(&the_lnet.ln_nis));
 712         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 713         LASSERT (the_lnet.ln_nzombie_nis == 0);
 714
 715         for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
 716                 LASSERT (list_empty(&the_lnet.ln_portals[idx].ptl_msgq));
 717
 718                 while (!list_empty (&the_lnet.ln_portals[idx].ptl_ml)) {
 719                         lnet_me_t *me = list_entry (the_lnet.ln_portals[idx].ptl_ml.next,
 720                                                     lnet_me_t, me_list);
 721
 722                         CERROR ("Active me %p on exit\n", me);
 723                         list_del (&me->me_list);
 724                         lnet_me_free (me);
 725                 }
 726         }
 727
 728         while (!list_empty (&the_lnet.ln_active_mds)) {
 729                 lnet_libmd_t *md = list_entry (the_lnet.ln_active_mds.next,
 730                                                lnet_libmd_t, md_list);
 731
 732                 CERROR ("Active md %p on exit\n", md);
 733                 list_del (&md->md_list);
 734                 lnet_md_free (md);
 735         }
 736
 737         while (!list_empty (&the_lnet.ln_active_eqs)) {
 738                 lnet_eq_t *eq = list_entry (the_lnet.ln_active_eqs.next,
 739                                             lnet_eq_t, eq_list);
 740
 741                 CERROR ("Active eq %p on exit\n", eq);
 742                 list_del (&eq->eq_list);
 743                 lnet_eq_free (eq);
 744         }
 745
 746         while (!list_empty (&the_lnet.ln_active_msgs)) {
 747                 lnet_msg_t *msg = list_entry (the_lnet.ln_active_msgs.next,
 748                                               lnet_msg_t, msg_activelist);
 749
 750                 CERROR ("Active msg %p on exit\n", msg);
 751                 LASSERT (msg->msg_onactivelist);
 752                 msg->msg_onactivelist = 0;
 753                 list_del (&msg->msg_activelist);
 754                 lnet_msg_free (msg);
 755         }
 756
 757         LIBCFS_FREE(the_lnet.ln_portals,
 758                     the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
 759
 760         lnet_free_rtrpools();
 761         lnet_fini_finalizers();
 762         lnet_destroy_peer_table();
 763         lnet_cleanup_handle_hash();
 764         lnet_descriptor_cleanup();
 765
 766         return (0);
 767 }
 768
 769 lnet_ni_t  *
 770 lnet_net2ni_locked (__u32 net)
 771 {
 772         struct list_head *tmp;
 773         lnet_ni_t        *ni;
 774
 775         list_for_each (tmp, &the_lnet.ln_nis) {
 776                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 777
 778                 if (lnet_ptlcompat_matchnet(LNET_NIDNET(ni->ni_nid), net)) {
 779                         lnet_ni_addref_locked(ni);
 780                         return ni;
 781                 }
 782         }
 783
 784         return NULL;
 785 }
 786
 787 int
 788 lnet_islocalnet (__u32 net)
 789 {
 790         lnet_ni_t        *ni;
 791
 792         LNET_LOCK();
 793         ni = lnet_net2ni_locked(net);
 794         if (ni != NULL)
 795                 lnet_ni_decref_locked(ni);
 796         LNET_UNLOCK();
 797
 798         return ni != NULL;
 799 }
 800
 801 lnet_ni_t  *
 802 lnet_nid2ni_locked (lnet_nid_t nid)
 803 {
 804         struct list_head *tmp;
 805         lnet_ni_t        *ni;
 806
 807         list_for_each (tmp, &the_lnet.ln_nis) {
 808                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 809
 810                 if (lnet_ptlcompat_matchnid(ni->ni_nid, nid)) {
 811                         lnet_ni_addref_locked(ni);
 812                         return ni;
 813                 }
 814         }
 815
 816         return NULL;
 817 }
 818
 819 int
 820 lnet_islocalnid (lnet_nid_t nid)
 821 {
 822         lnet_ni_t     *ni;
 823
 824         LNET_LOCK();
 825         ni = lnet_nid2ni_locked(nid);
 826         if (ni != NULL)
 827                 lnet_ni_decref_locked(ni);
 828         LNET_UNLOCK();
 829
 830         return ni != NULL;
 831 }
 832
 833 int
 834 lnet_count_acceptor_nis (lnet_ni_t **first_ni)
 835 {
 836         /* Return the # of NIs that need the acceptor.  Return the first one in
 837          * *first_ni so the acceptor can pass it connections "blind" to retain
 838          * binary compatibility. */
 839         int                count = 0;
 840 #ifdef __KERNEL__
 841         struct list_head  *tmp;
 842         lnet_ni_t         *ni;
 843
 844         LNET_LOCK();
 845         list_for_each (tmp, &the_lnet.ln_nis) {
 846                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 847
 848                 if (ni->ni_lnd->lnd_accept != NULL) {
 849                         /* This LND uses the acceptor */
 850                         if (count == 0 && first_ni != NULL) {
 851                                 lnet_ni_addref_locked(ni);
 852                                 *first_ni = ni;
 853                         }
 854                         count++;
 855                 }
 856         }
 857
 858         LNET_UNLOCK();
 859 #endif
 860         return count;
 861 }
 862
 863 void
 864 lnet_shutdown_lndnis (void)
 865 {
 866         int                i;
 867         int                islo;
 868         lnet_ni_t         *ni;
 869
 870         /* NB called holding the global mutex */
 871
 872         /* All quiet on the API front */
 873         LASSERT (!the_lnet.ln_shutdown);
 874         LASSERT (the_lnet.ln_refcount == 0);
 875         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 876         LASSERT (the_lnet.ln_nzombie_nis == 0);
 877         LASSERT (list_empty(&the_lnet.ln_remote_nets));
 878
 879         LNET_LOCK();
 880         the_lnet.ln_shutdown = 1;               /* flag shutdown */
 881
 882         /* Unlink NIs from the global table */
 883         while (!list_empty(&the_lnet.ln_nis)) {
 884                 ni = list_entry(the_lnet.ln_nis.next,
 885                                 lnet_ni_t, ni_list);
 886                 list_del (&ni->ni_list);
 887
 888                 the_lnet.ln_nzombie_nis++;
 889                 lnet_ni_decref_locked(ni); /* drop apini's ref */
 890         }
 891
 892         /* Drop the cached eqwait NI. */
 893         if (the_lnet.ln_eqwaitni != NULL) {
 894                 lnet_ni_decref_locked(the_lnet.ln_eqwaitni);
 895                 the_lnet.ln_eqwaitni = NULL;
 896         }
 897
 898         /* Drop the cached loopback NI. */
 899         if (the_lnet.ln_loni != NULL) {
 900                 lnet_ni_decref_locked(the_lnet.ln_loni);
 901                 the_lnet.ln_loni = NULL;
 902         }
 903
 904         LNET_UNLOCK();
 905
 906         /* Clear lazy portals and drop delayed messages which hold refs
 907          * on their lnet_msg_t::msg_rxpeer */
 908         for (i = 0; i < the_lnet.ln_nportals; i++)
 909                 LNetClearLazyPortal(i);
 910
 911         /* Clear the peer table and wait for all peers to go (they hold refs on
 912          * their NIs) */
 913         lnet_clear_peer_table();
 914
 915         LNET_LOCK();
 916         /* Now wait for the NI's I just nuked to show up on apini_zombie_nis
 917          * and shut them down in guaranteed thread context */
 918         i = 2;
 919         while (the_lnet.ln_nzombie_nis != 0) {
 920
 921                 while (list_empty(&the_lnet.ln_zombie_nis)) {
 922                         LNET_UNLOCK();
 923                         ++i;
 924                         if ((i & (-i)) == i)
 925                                 CDEBUG(D_WARNING,"Waiting for %d zombie NIs\n",
 926                                        the_lnet.ln_nzombie_nis);
 927                         cfs_pause(cfs_time_seconds(1));
 928                         LNET_LOCK();
 929                 }
 930
 931                 ni = list_entry(the_lnet.ln_zombie_nis.next,
 932                                 lnet_ni_t, ni_list);
 933                 list_del(&ni->ni_list);
 934                 ni->ni_lnd->lnd_refcount--;
 935
 936                 LNET_UNLOCK();
 937
 938                 islo = ni->ni_lnd->lnd_type == LOLND;
 939
 940                 LASSERT (!in_interrupt ());
 941                 (ni->ni_lnd->lnd_shutdown)(ni);
 942
 943                 /* can't deref lnd anymore now; it might have unregistered
 944                  * itself...  */
 945
 946                 if (!islo)
 947                         CDEBUG(D_LNI, "Removed LNI %s\n",
 948                                libcfs_nid2str(ni->ni_nid));
 949
 950                 LIBCFS_FREE(ni, sizeof(*ni));
 951
 952                 LNET_LOCK();
 953                 the_lnet.ln_nzombie_nis--;
 954         }
 955
 956         the_lnet.ln_shutdown = 0;
 957         LNET_UNLOCK();
 958
 959         if (the_lnet.ln_network_tokens != NULL) {
 960                 LIBCFS_FREE(the_lnet.ln_network_tokens,
 961                             the_lnet.ln_network_tokens_nob);
 962                 the_lnet.ln_network_tokens = NULL;
 963         }
 964 }
 965
 966 int
 967 lnet_startup_lndnis (void)
 968 {
 969         lnd_t             *lnd;
 970         lnet_ni_t         *ni;
 971         struct list_head   nilist;
 972         int                rc = 0;
 973         int                lnd_type;
 974         int                nicount = 0;
 975         char              *nets = lnet_get_networks();
 976
 977         INIT_LIST_HEAD(&nilist);
 978
 979         if (nets == NULL)
 980                 goto failed;
 981
 982         rc = lnet_parse_networks(&nilist, nets);
 983         if (rc != 0)
 984                 goto failed;
 985
 986         while (!list_empty(&nilist)) {
 987                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
 988                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 989
 990                 LASSERT (libcfs_isknown_lnd(lnd_type));
 991
 992                 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 993                 lnd = lnet_find_lnd_by_type(lnd_type);
 994
 995 #ifdef __KERNEL__
 996                 if (lnd == NULL) {
 997                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 998                         rc = request_module(libcfs_lnd2modname(lnd_type));
 999                         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1000
1001                         lnd = lnet_find_lnd_by_type(lnd_type);
1002                         if (lnd == NULL) {
1003                                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1004                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1005                                        libcfs_lnd2str(lnd_type),
1006                                        libcfs_lnd2modname(lnd_type), rc);
1007 #ifndef CONFIG_KMOD
1008                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1009                                          "compiled with CONFIG_KMOD set for "
1010                                          "automatic module loading.");
1011 #endif
1012                                 goto failed;
1013                         }
1014                 }
1015 #else
1016                 if (lnd == NULL) {
1017                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1018                         CERROR("LND %s not supported\n",
1019                                libcfs_lnd2str(lnd_type));
1020                         goto failed;
1021                 }
1022 #endif
1023
1024                 ni->ni_refcount = 1;
1025
1026                 LNET_LOCK();
1027                 lnd->lnd_refcount++;
1028                 LNET_UNLOCK();
1029
1030                 ni->ni_lnd = lnd;
1031
1032                 rc = (lnd->lnd_startup)(ni);
1033
1034                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1035
1036                 if (rc != 0) {
1037                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1038                                            "\n",
1039                                            rc, libcfs_lnd2str(lnd->lnd_type));
1040                         LNET_LOCK();
1041                         lnd->lnd_refcount--;
1042                         LNET_UNLOCK();
1043                         goto failed;
1044                 }
1045
1046                 list_del(&ni->ni_list);
1047
1048                 LNET_LOCK();
1049                 list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1050                 LNET_UNLOCK();
1051
1052                 if (lnd->lnd_type == LOLND) {
1053                         lnet_ni_addref(ni);
1054                         LASSERT (the_lnet.ln_loni == NULL);
1055                         the_lnet.ln_loni = ni;
1056                         continue;
1057                 }
1058
1059 #ifndef __KERNEL__
1060                 if (lnd->lnd_wait != NULL) {
1061                         if (the_lnet.ln_eqwaitni == NULL) {
1062                                 lnet_ni_addref(ni);
1063                                 the_lnet.ln_eqwaitni = ni;
1064                         }
1065                 } else {
1066 # ifndef HAVE_LIBPTHREAD
1067                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1068                                            "single-threaded runtime\n",
1069                                            libcfs_lnd2str(lnd_type));
1070                         goto failed;
1071 # endif
1072                 }
1073 #endif
1074                 if (ni->ni_peertxcredits == 0 ||
1075                     ni->ni_maxtxcredits == 0) {
1076                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1077                                            libcfs_lnd2str(lnd->lnd_type),
1078                                            ni->ni_peertxcredits == 0 ?
1079                                            "" : "per-peer ");
1080                         goto failed;
1081                 }
1082
1083                 ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
1084
1085                 CDEBUG(D_LNI, "Added LNI %s [%d/%d]\n",
1086                        libcfs_nid2str(ni->ni_nid),
1087                        ni->ni_peertxcredits, ni->ni_txcredits);
1088
1089                 /* Handle nidstrings for network 0 just like this one */
1090                 if (the_lnet.ln_ptlcompat > 0) {
1091                         if (nicount > 0) {
1092                                 LCONSOLE_ERROR_MSG(0x108, "Can't run > 1 "
1093                                        "network when portals_compatibility is "
1094                                        "set\n");
1095                                 goto failed;
1096                         }
1097                         libcfs_setnet0alias(lnd->lnd_type);
1098                 }
1099
1100                 nicount++;
1101         }
1102
1103         if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
1104                 lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
1105                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1106                                    "\n",
1107                                    libcfs_lnd2str(lnd_type));
1108                 goto failed;
1109         }
1110
1111         return 0;
1112
1113  failed:
1114         lnet_shutdown_lndnis();
1115
1116         while (!list_empty(&nilist)) {
1117                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1118                 list_del(&ni->ni_list);
1119                 LIBCFS_FREE(ni, sizeof(*ni));
1120         }
1121
1122         return -ENETDOWN;
1123 }
1124
1125 int
1126 LNetInit(void)
1127 {
1128         int    rc;
1129
1130         lnet_assert_wire_constants ();
1131         LASSERT (!the_lnet.ln_init);
1132
1133         memset(&the_lnet, 0, sizeof(the_lnet));
1134
1135         rc = lnet_get_portals_compatibility();
1136         if (rc < 0)
1137                 return rc;
1138
1139         lnet_init_locks();
1140         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1141         the_lnet.ln_ptlcompat = rc;
1142         the_lnet.ln_refcount = 0;
1143         the_lnet.ln_init = 1;
1144
1145 #ifdef __KERNEL__
1146         /* All LNDs apart from the LOLND are in separate modules.  They
1147          * register themselves when their module loads, and unregister
1148          * themselves when their module is unloaded. */
1149 #else
1150         /* Register LNDs
1151          * NB the order here determines default 'networks=' order */
1152 # ifdef CRAY_XT3
1153         LNET_REGISTER_ULND(the_ptllnd);
1154 # endif
1155 # ifdef HAVE_LIBPTHREAD
1156         LNET_REGISTER_ULND(the_tcplnd);
1157 # endif
1158 #endif
1159         lnet_register_lnd(&the_lolnd);
1160         return 0;
1161 }
1162
1163 void
1164 LNetFini(void)
1165 {
1166         LASSERT (the_lnet.ln_init);
1167         LASSERT (the_lnet.ln_refcount == 0);
1168
1169         while (!list_empty(&the_lnet.ln_lnds))
1170                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1171                                                lnd_t, lnd_list));
1172         lnet_fini_locks();
1173
1174         the_lnet.ln_init = 0;
1175 }
1176
1177 int
1178 LNetNIInit(lnet_pid_t requested_pid)
1179 {
1180         int         im_a_router = 0;
1181         int         rc;
1182
1183         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1184
1185         LASSERT (the_lnet.ln_init);
1186         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1187
1188         if (the_lnet.ln_refcount > 0) {
1189                 rc = the_lnet.ln_refcount++;
1190                 goto out;
1191         }
1192
1193         if (requested_pid == LNET_PID_ANY) {
1194                 /* Don't instantiate LNET just for me */
1195                 rc = -ENETDOWN;
1196                 goto failed0;
1197         }
1198
1199         rc = lnet_prepare(requested_pid);
1200         if (rc != 0)
1201                 goto failed0;
1202
1203         rc = lnet_startup_lndnis();
1204         if (rc != 0)
1205                 goto failed1;
1206
1207         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1208         if (rc != 0)
1209                 goto failed2;
1210
1211         rc = lnet_check_routes();
1212         if (rc != 0)
1213                 goto failed2;
1214
1215         rc = lnet_alloc_rtrpools(im_a_router);
1216         if (rc != 0)
1217                 goto failed2;
1218
1219         rc = lnet_acceptor_start();
1220         if (rc != 0)
1221                 goto failed2;
1222
1223         the_lnet.ln_refcount = 1;
1224         /* Now I may use my own API functions... */
1225
1226         rc = lnet_router_checker_start();
1227         if (rc != 0)
1228                 goto failed3;
1229
1230         rc = lnet_ping_target_init();
1231         if (rc != 0)
1232                 goto failed4;
1233
1234         lnet_proc_init();
1235         goto out;
1236
1237  failed4:
1238         lnet_router_checker_stop();
1239  failed3:
1240         the_lnet.ln_refcount = 0;
1241         lnet_acceptor_stop();
1242  failed2:
1243         lnet_destroy_routes();
1244         lnet_shutdown_lndnis();
1245  failed1:
1246         lnet_unprepare();
1247  failed0:
1248         LASSERT (rc < 0);
1249  out:
1250         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1251         return rc;
1252 }
1253
1254 int
1255 LNetNIFini()
1256 {
1257         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1258
1259         LASSERT (the_lnet.ln_init);
1260         LASSERT (the_lnet.ln_refcount > 0);
1261
1262         if (the_lnet.ln_refcount != 1) {
1263                 the_lnet.ln_refcount--;
1264         } else {
1265                 LASSERT (!the_lnet.ln_niinit_self);
1266
1267                 lnet_proc_fini();
1268                 lnet_ping_target_fini();
1269                 lnet_router_checker_stop();
1270
1271                 /* Teardown fns that use my own API functions BEFORE here */
1272                 the_lnet.ln_refcount = 0;
1273
1274                 lnet_acceptor_stop();
1275                 lnet_destroy_routes();
1276                 lnet_shutdown_lndnis();
1277                 lnet_unprepare();
1278         }
1279
1280         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1281         return 0;
1282 }
1283
1284 int
1285 LNetCtl(unsigned int cmd, void *arg)
1286 {
1287         struct libcfs_ioctl_data *data = arg;
1288         lnet_process_id_t         id;
1289         lnet_ni_t                *ni;
1290         int                       rc;
1291
1292         LASSERT (the_lnet.ln_init);
1293         LASSERT (the_lnet.ln_refcount > 0);
1294
1295         switch (cmd) {
1296         case IOC_LIBCFS_GET_NI:
1297                 rc = LNetGetId(data->ioc_count, &id);
1298                 data->ioc_nid = id.nid;
1299                 return rc;
1300
1301         case IOC_LIBCFS_FAIL_NID:
1302                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1303
1304         case IOC_LIBCFS_ADD_ROUTE:
1305                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1306                                     data->ioc_nid);
1307                 return (rc != 0) ? rc : lnet_check_routes();
1308
1309         case IOC_LIBCFS_DEL_ROUTE:
1310                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1311
1312         case IOC_LIBCFS_GET_ROUTE:
1313                 return lnet_get_route(data->ioc_count,
1314                                       &data->ioc_net, &data->ioc_count,
1315                                       &data->ioc_nid, &data->ioc_flags);
1316         case IOC_LIBCFS_NOTIFY_ROUTER:
1317                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1318                                    (time_t)data->ioc_u64[0]);
1319
1320         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1321                 return the_lnet.ln_ptlcompat;
1322
1323         case IOC_LIBCFS_LNET_DIST:
1324                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1325                 if (rc < 0 && rc != -EHOSTUNREACH)
1326                         return rc;
1327
1328                 data->ioc_u32[0] = rc;
1329                 return 0;
1330
1331         case IOC_LIBCFS_TESTPROTOCOMPAT:
1332                 LNET_LOCK();
1333                 the_lnet.ln_testprotocompat = data->ioc_flags;
1334                 LNET_UNLOCK();
1335                 return 0;
1336
1337         case IOC_LIBCFS_PING:
1338                 rc = lnet_ping((lnet_process_id_t) {.nid = data->ioc_nid,
1339                                                     .pid = data->ioc_u32[0]},
1340                                data->ioc_u32[1], /* timeout */
1341                                (lnet_process_id_t *)data->ioc_pbuf1,
1342                                data->ioc_plen1/sizeof(lnet_process_id_t));
1343                 if (rc < 0)
1344                         return rc;
1345                 data->ioc_count = rc;
1346                 return 0;
1347
1348         case IOC_LIBCFS_DEBUG_PEER: {
1349                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1350                  * via an ioctl */
1351                 lnet_process_id_t *id = arg;
1352
1353                 lnet_debug_peer(id->nid);
1354
1355                 ni = lnet_net2ni(LNET_NIDNET(id->nid));
1356                 if (ni == NULL) {
1357                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(*id));
1358                 } else {
1359                         if (ni->ni_lnd->lnd_ctl == NULL) {
1360                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1361                                        libcfs_id2str(*id));
1362                         } else {
1363                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1364                         }
1365
1366                         lnet_ni_decref(ni);
1367                 }
1368                 return 0;
1369         }
1370
1371         default:
1372                 ni = lnet_net2ni(data->ioc_net);
1373                 if (ni == NULL)
1374                         return -EINVAL;
1375
1376                 if (ni->ni_lnd->lnd_ctl == NULL)
1377                         rc = -EINVAL;
1378                 else
1379                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1380
1381                 lnet_ni_decref(ni);
1382                 return rc;
1383         }
1384         /* not reached */
1385 }
1386
1387 int
1388 LNetGetId(unsigned int index, lnet_process_id_t *id)
1389 {
1390         lnet_ni_t        *ni;
1391         struct list_head *tmp;
1392         int               rc = -ENOENT;
1393
1394         LASSERT (the_lnet.ln_init);
1395         LASSERT (the_lnet.ln_refcount > 0);
1396
1397         LNET_LOCK();
1398
1399         list_for_each(tmp, &the_lnet.ln_nis) {
1400                 if (index-- != 0)
1401                         continue;
1402
1403                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1404
1405                 id->nid = ni->ni_nid;
1406                 id->pid = the_lnet.ln_pid;
1407                 rc = 0;
1408                 break;
1409         }
1410
1411         LNET_UNLOCK();
1412
1413         return rc;
1414 }
1415
1416 void
1417 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1418 {
1419         snprintf(str, len, LPX64, h.cookie);
1420 }
1421
1422
1423 int
1424 lnet_ping_target_init(void)
1425 {
1426         lnet_handle_me_t  meh;
1427         lnet_process_id_t id;
1428         int               rc;
1429         int               rc2;
1430         int               n;
1431         int               infosz;
1432         int               i;
1433
1434         for (n = 0; ; n++) {
1435                 rc = LNetGetId(n, &id);
1436                 if (rc == -ENOENT)
1437                         break;
1438
1439                 LASSERT (rc == 0);
1440         }
1441
1442         infosz = offsetof(lnet_ping_info_t, pi_nid[n]);
1443         LIBCFS_ALLOC(the_lnet.ln_ping_info, infosz);
1444         if (the_lnet.ln_ping_info == NULL) {
1445                 CERROR("Can't allocate ping info[%d]\n", n);
1446                 return -ENOMEM;
1447         }
1448
1449         the_lnet.ln_ping_info->pi_magic   = LNET_PROTO_PING_MAGIC;
1450         the_lnet.ln_ping_info->pi_version = LNET_PROTO_PING_VERSION;
1451         the_lnet.ln_ping_info->pi_pid     = the_lnet.ln_pid;
1452         the_lnet.ln_ping_info->pi_nnids   = n;
1453
1454         for (i = 0; i < n; i++) {
1455                 rc = LNetGetId(i, &id);
1456                 LASSERT (rc == 0);
1457                 the_lnet.ln_ping_info->pi_nid[i] = id.nid;
1458         }
1459
1460         /* We can have a tiny EQ since we only need to see the unlink event on
1461          * teardown, which by definition is the last one! */
1462         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1463         if (rc != 0) {
1464                 CERROR("Can't allocate ping EQ: %d\n", rc);
1465                 goto failed_0;
1466         }
1467
1468         rc = LNetMEAttach(LNET_RESERVED_PORTAL,
1469                           (lnet_process_id_t){.nid = LNET_NID_ANY,
1470                                               .pid = LNET_PID_ANY},
1471                           LNET_PROTO_PING_MATCHBITS, 0LL,
1472                           LNET_UNLINK, LNET_INS_AFTER,
1473                           &meh);
1474         if (rc != 0) {
1475                 CERROR("Can't create ping ME: %d\n", rc);
1476                 goto failed_1;
1477         }
1478
1479         rc = LNetMDAttach(meh,
1480                           (lnet_md_t){.start = the_lnet.ln_ping_info,
1481                                       .length = infosz,
1482                                       .threshold = LNET_MD_THRESH_INF,
1483                                       .options = (LNET_MD_OP_GET |
1484                                                   LNET_MD_TRUNCATE |
1485                                                   LNET_MD_MANAGE_REMOTE),
1486                                       .eq_handle = the_lnet.ln_ping_target_eq},
1487                           LNET_RETAIN,
1488                           &the_lnet.ln_ping_target_md);
1489         if (rc != 0) {
1490                 CERROR("Can't attach ping MD: %d\n", rc);
1491                 goto failed_2;
1492         }
1493
1494         return 0;
1495
1496  failed_2:
1497         rc2 = LNetMEUnlink(meh);
1498         LASSERT (rc2 == 0);
1499  failed_1:
1500         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1501         LASSERT (rc2 == 0);
1502  failed_0:
1503         LIBCFS_FREE(the_lnet.ln_ping_info, infosz);
1504
1505         return rc;
1506 }
1507
1508 void
1509 lnet_ping_target_fini(void)
1510 {
1511         lnet_event_t    event;
1512         int             rc;
1513         int             which;
1514         int             timeout_ms = 1000;
1515         cfs_sigset_t    blocked = cfs_block_allsigs();
1516
1517         LNetMDUnlink(the_lnet.ln_ping_target_md);
1518         /* NB md could be busy; this just starts the unlink */
1519
1520         for (;;) {
1521                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1522                                 timeout_ms, &event, &which);
1523
1524                 /* I expect overflow... */
1525                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1526
1527                 if (rc == 0) {
1528                         /* timed out: provide a diagnostic */
1529                         CWARN("Still waiting for ping MD to unlink\n");
1530                         timeout_ms *= 2;
1531                         continue;
1532                 }
1533
1534                 /* Got a valid event */
1535                 if (event.unlinked)
1536                         break;
1537         }
1538
1539         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1540         LASSERT (rc == 0);
1541
1542         LIBCFS_FREE(the_lnet.ln_ping_info,
1543                     offsetof(lnet_ping_info_t,
1544                              pi_nid[the_lnet.ln_ping_info->pi_nnids]));
1545
1546         cfs_restore_sigs(blocked);
1547 }
1548
1549 int
1550 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1551 {
1552         lnet_handle_eq_t     eqh;
1553         lnet_handle_md_t     mdh;
1554         lnet_event_t         event;
1555         int                  which;
1556         int                  unlinked = 0;
1557         int                  replied = 0;
1558         const int            a_long_time = 60000; /* mS */
1559         int                  infosz = offsetof(lnet_ping_info_t, pi_nid[n_ids]);
1560         lnet_ping_info_t    *info;
1561         lnet_process_id_t    tmpid;
1562         int                  i;
1563         int                  nob;
1564         int                  rc;
1565         int                  rc2;
1566         cfs_sigset_t         blocked;
1567
1568         if (n_ids <= 0 ||
1569             id.nid == LNET_NID_ANY ||
1570             timeout_ms > 500000 ||              /* arbitrary limit! */
1571             n_ids > 20)                         /* arbitrary limit! */
1572                 return -EINVAL;
1573
1574         if (id.pid == LNET_PID_ANY)
1575                 id.pid = LUSTRE_SRV_LNET_PID;
1576
1577         LIBCFS_ALLOC(info, infosz);
1578         if (info == NULL)
1579                 return -ENOMEM;
1580
1581         /* NB 2 events max (including any unlink event) */
1582         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1583         if (rc != 0) {
1584                 CERROR("Can't allocate EQ: %d\n", rc);
1585                 goto out_0;
1586         }
1587
1588         rc = LNetMDBind((lnet_md_t){.start = info,
1589                                     .length = infosz,
1590                                     .threshold = 2, /* GET/REPLY */
1591                                     .options = LNET_MD_TRUNCATE,
1592                                     .eq_handle = eqh},
1593                         LNET_UNLINK,
1594                         &mdh);
1595         if (rc != 0) {
1596                 CERROR("Can't bind MD: %d\n", rc);
1597                 goto out_1;
1598         }
1599
1600         rc = LNetGet(LNET_NID_ANY, mdh, id,
1601                      LNET_RESERVED_PORTAL,
1602                      LNET_PROTO_PING_MATCHBITS, 0);
1603
1604         if (rc != 0) {
1605                 /* Don't CERROR; this could be deliberate! */
1606
1607                 rc2 = LNetMDUnlink(mdh);
1608                 LASSERT (rc2 == 0);
1609
1610                 /* NB must wait for the UNLINK event below... */
1611                 unlinked = 1;
1612                 timeout_ms = a_long_time;
1613         }
1614
1615         do {
1616                 /* MUST block for unlink to complete */
1617                 if (unlinked)
1618                         blocked = cfs_block_allsigs();
1619
1620                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1621
1622                 if (unlinked)
1623                         cfs_restore_sigs(blocked);
1624
1625                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1626                        (rc2 <= 0) ? -1 : event.type,
1627                        (rc2 <= 0) ? -1 : event.status,
1628                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1629
1630                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1631
1632                 if (rc2 <= 0 || event.status != 0) {
1633                         /* timeout or error */
1634                         if (!replied && rc == 0)
1635                                 rc = (rc2 < 0) ? rc2 :
1636                                      (rc2 == 0) ? -ETIMEDOUT :
1637                                      event.status;
1638
1639                         if (!unlinked) {
1640                                 /* Ensure completion in finite time... */
1641                                 LNetMDUnlink(mdh);
1642                                 /* No assertion (racing with network) */
1643                                 unlinked = 1;
1644                                 timeout_ms = a_long_time;
1645                         } else if (rc2 == 0) {
1646                                 /* timed out waiting for unlink */
1647                                 CWARN("ping %s: late network completion\n",
1648                                       libcfs_id2str(id));
1649                         }
1650
1651                 } else if (event.type == LNET_EVENT_REPLY) {
1652                         replied = 1;
1653                         rc = event.mlength;
1654                 }
1655
1656         } while (rc2 <= 0 || !event.unlinked);
1657
1658         if (!replied) {
1659                 if (rc >= 0)
1660                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1661                               libcfs_id2str(id));
1662                 rc = -EIO;
1663                 goto out_1;
1664         }
1665
1666         nob = rc;
1667         LASSERT (nob >= 0 && nob <= infosz);
1668
1669         rc = -EPROTO;                           /* if I can't parse... */
1670
1671         if (nob < 8) {
1672                 /* can't check magic/version */
1673                 CERROR("%s: ping info too short %d\n",
1674                        libcfs_id2str(id), nob);
1675                 goto out_1;
1676         }
1677
1678         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1679                 /* NB I might be swabbing garbage until I check below, but it
1680                  * doesn't matter */
1681                 __swab32s(&info->pi_version);
1682                 __swab32s(&info->pi_pid);
1683                 __swab32s(&info->pi_nnids);
1684                 for (i = 0; i < info->pi_nnids && i < n_ids; i++)
1685                         __swab64s(&info->pi_nid[i]);
1686
1687         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1688                 CERROR("%s: Unexpected magic %08x\n",
1689                        libcfs_id2str(id), info->pi_magic);
1690                 goto out_1;
1691         }
1692
1693         if (info->pi_version != LNET_PROTO_PING_VERSION) {
1694                 CERROR("%s: Unexpected version 0x%x\n",
1695                        libcfs_id2str(id), info->pi_version);
1696                 goto out_1;
1697         }
1698
1699         if (nob < offsetof(lnet_ping_info_t, pi_nid[0])) {
1700                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1701                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[0]));
1702                 goto out_1;
1703         }
1704
1705         if (info->pi_nnids < n_ids)
1706                 n_ids = info->pi_nnids;
1707
1708         if (nob < offsetof(lnet_ping_info_t, pi_nid[n_ids])) {
1709                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1710                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[n_ids]));
1711                 goto out_1;
1712         }
1713
1714         rc = -EFAULT;                           /* If I SEGV... */
1715
1716         for (i = 0; i < n_ids; i++) {
1717                 tmpid.pid = info->pi_pid;
1718                 tmpid.nid = info->pi_nid[i];
1719 #ifdef __KERNEL__
1720                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1721                         goto out_1;
1722 #else
1723                 ids[i] = tmpid;
1724 #endif
1725         }
1726         rc = info->pi_nnids;
1727
1728  out_1:
1729         rc2 = LNetEQFree(eqh);
1730         if (rc2 != 0)
1731                 CERROR("rc2 %d\n", rc2);
1732         LASSERT (rc2 == 0);
1733
1734  out_0:
1735         LIBCFS_FREE(info, infosz);
1736         return rc;
1737 }