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 #ifndef __KERNEL__
 619 /* Temporary workaround to allow uOSS and test programs force server
 620  * mode in userspace. See comments near ln_server_mode_flag in
 621  * lnet/lib-types.h */
 622
 623 void
 624 lnet_server_mode() {
 625         the_lnet.ln_server_mode_flag = 1;
 626 }
 627 #endif
 628
 629 int
 630 lnet_prepare(lnet_pid_t requested_pid)
 631 {
 632         /* Prepare to bring up the network */
 633         int               rc = 0;
 634         int               i;
 635
 636         LASSERT (the_lnet.ln_refcount == 0);
 637
 638         the_lnet.ln_routing = 0;
 639
 640 #ifdef __KERNEL__
 641         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 642         the_lnet.ln_pid = requested_pid;
 643 #else
 644         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 645                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 646
 647                 if (cfs_curproc_uid())/* Only root can run user-space server */
 648                         return -EPERM;
 649                 the_lnet.ln_pid = requested_pid;
 650
 651         } else {/* client case (liblustre) */
 652
 653                 /* My PID must be unique on this node and flag I'm userspace */
 654                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 655         }
 656 #endif
 657
 658         rc = lnet_descriptor_setup();
 659         if (rc != 0)
 660                 goto failed0;
 661
 662         memset(&the_lnet.ln_counters, 0,
 663                sizeof(the_lnet.ln_counters));
 664
 665         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
 666         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
 667         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
 668         CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
 669         CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
 670         CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
 671         CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
 672         CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
 673
 674         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 675
 676         lnet_init_rtrpools();
 677
 678         rc = lnet_setup_handle_hash ();
 679         if (rc != 0)
 680                 goto failed0;
 681
 682         rc = lnet_create_peer_table();
 683         if (rc != 0)
 684                 goto failed1;
 685
 686         rc = lnet_init_finalizers();
 687         if (rc != 0)
 688                 goto failed2;
 689
 690         the_lnet.ln_nportals = MAX_PORTALS;
 691         LIBCFS_ALLOC(the_lnet.ln_portals,
 692                      the_lnet.ln_nportals *
 693                      sizeof(*the_lnet.ln_portals));
 694         if (the_lnet.ln_portals == NULL) {
 695                 rc = -ENOMEM;
 696                 goto failed3;
 697         }
 698
 699         for (i = 0; i < the_lnet.ln_nportals; i++) {
 700                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_ml));
 701                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
 702                 the_lnet.ln_portals[i].ptl_options = 0;
 703         }
 704
 705         return 0;
 706
 707  failed3:
 708         lnet_fini_finalizers();
 709  failed2:
 710         lnet_destroy_peer_table();
 711  failed1:
 712         lnet_cleanup_handle_hash();
 713  failed0:
 714         lnet_descriptor_cleanup();
 715         return rc;
 716 }
 717
 718 int
 719 lnet_unprepare (void)
 720 {
 721         int       idx;
 722
 723         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 724          * have shut down already, so it is safe to unlink and free all
 725          * descriptors, even those that appear committed to a network op (eg MD
 726          * with non-zero pending count) */
 727
 728         lnet_fail_nid(LNET_NID_ANY, 0);
 729
 730         LASSERT (list_empty(&the_lnet.ln_test_peers));
 731         LASSERT (the_lnet.ln_refcount == 0);
 732         LASSERT (list_empty(&the_lnet.ln_nis));
 733         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 734         LASSERT (the_lnet.ln_nzombie_nis == 0);
 735
 736         for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
 737                 LASSERT (list_empty(&the_lnet.ln_portals[idx].ptl_msgq));
 738
 739                 while (!list_empty (&the_lnet.ln_portals[idx].ptl_ml)) {
 740                         lnet_me_t *me = list_entry (the_lnet.ln_portals[idx].ptl_ml.next,
 741                                                     lnet_me_t, me_list);
 742
 743                         CERROR ("Active me %p on exit\n", me);
 744                         list_del (&me->me_list);
 745                         lnet_me_free (me);
 746                 }
 747         }
 748
 749         while (!list_empty (&the_lnet.ln_active_mds)) {
 750                 lnet_libmd_t *md = list_entry (the_lnet.ln_active_mds.next,
 751                                                lnet_libmd_t, md_list);
 752
 753                 CERROR ("Active md %p on exit\n", md);
 754                 list_del (&md->md_list);
 755                 lnet_md_free (md);
 756         }
 757
 758         while (!list_empty (&the_lnet.ln_active_eqs)) {
 759                 lnet_eq_t *eq = list_entry (the_lnet.ln_active_eqs.next,
 760                                             lnet_eq_t, eq_list);
 761
 762                 CERROR ("Active eq %p on exit\n", eq);
 763                 list_del (&eq->eq_list);
 764                 lnet_eq_free (eq);
 765         }
 766
 767         while (!list_empty (&the_lnet.ln_active_msgs)) {
 768                 lnet_msg_t *msg = list_entry (the_lnet.ln_active_msgs.next,
 769                                               lnet_msg_t, msg_activelist);
 770
 771                 CERROR ("Active msg %p on exit\n", msg);
 772                 LASSERT (msg->msg_onactivelist);
 773                 msg->msg_onactivelist = 0;
 774                 list_del (&msg->msg_activelist);
 775                 lnet_msg_free (msg);
 776         }
 777
 778         LIBCFS_FREE(the_lnet.ln_portals,
 779                     the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
 780
 781         lnet_free_rtrpools();
 782         lnet_fini_finalizers();
 783         lnet_destroy_peer_table();
 784         lnet_cleanup_handle_hash();
 785         lnet_descriptor_cleanup();
 786
 787         return (0);
 788 }
 789
 790 lnet_ni_t  *
 791 lnet_net2ni_locked (__u32 net)
 792 {
 793         struct list_head *tmp;
 794         lnet_ni_t        *ni;
 795
 796         list_for_each (tmp, &the_lnet.ln_nis) {
 797                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 798
 799                 if (lnet_ptlcompat_matchnet(LNET_NIDNET(ni->ni_nid), net)) {
 800                         lnet_ni_addref_locked(ni);
 801                         return ni;
 802                 }
 803         }
 804
 805         return NULL;
 806 }
 807
 808 int
 809 lnet_islocalnet (__u32 net)
 810 {
 811         lnet_ni_t        *ni;
 812
 813         LNET_LOCK();
 814         ni = lnet_net2ni_locked(net);
 815         if (ni != NULL)
 816                 lnet_ni_decref_locked(ni);
 817         LNET_UNLOCK();
 818
 819         return ni != NULL;
 820 }
 821
 822 lnet_ni_t  *
 823 lnet_nid2ni_locked (lnet_nid_t nid)
 824 {
 825         struct list_head *tmp;
 826         lnet_ni_t        *ni;
 827
 828         list_for_each (tmp, &the_lnet.ln_nis) {
 829                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 830
 831                 if (lnet_ptlcompat_matchnid(ni->ni_nid, nid)) {
 832                         lnet_ni_addref_locked(ni);
 833                         return ni;
 834                 }
 835         }
 836
 837         return NULL;
 838 }
 839
 840 int
 841 lnet_islocalnid (lnet_nid_t nid)
 842 {
 843         lnet_ni_t     *ni;
 844
 845         LNET_LOCK();
 846         ni = lnet_nid2ni_locked(nid);
 847         if (ni != NULL)
 848                 lnet_ni_decref_locked(ni);
 849         LNET_UNLOCK();
 850
 851         return ni != NULL;
 852 }
 853
 854 int
 855 lnet_count_acceptor_nis (lnet_ni_t **first_ni)
 856 {
 857         /* Return the # of NIs that need the acceptor.  Return the first one in
 858          * *first_ni so the acceptor can pass it connections "blind" to retain
 859          * binary compatibility. */
 860         int                count = 0;
 861         struct list_head  *tmp;
 862         lnet_ni_t         *ni;
 863
 864         LNET_LOCK();
 865         list_for_each (tmp, &the_lnet.ln_nis) {
 866                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 867
 868                 if (ni->ni_lnd->lnd_accept != NULL) {
 869                         /* This LND uses the acceptor */
 870                         if (count == 0 && first_ni != NULL) {
 871                                 lnet_ni_addref_locked(ni);
 872                                 *first_ni = ni;
 873                         }
 874                         count++;
 875                 }
 876         }
 877
 878         LNET_UNLOCK();
 879         return count;
 880 }
 881
 882 void
 883 lnet_shutdown_lndnis (void)
 884 {
 885         int                i;
 886         int                islo;
 887         lnet_ni_t         *ni;
 888
 889         /* NB called holding the global mutex */
 890
 891         /* All quiet on the API front */
 892         LASSERT (!the_lnet.ln_shutdown);
 893         LASSERT (the_lnet.ln_refcount == 0);
 894         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 895         LASSERT (the_lnet.ln_nzombie_nis == 0);
 896         LASSERT (list_empty(&the_lnet.ln_remote_nets));
 897
 898         LNET_LOCK();
 899         the_lnet.ln_shutdown = 1;               /* flag shutdown */
 900
 901         /* Unlink NIs from the global table */
 902         while (!list_empty(&the_lnet.ln_nis)) {
 903                 ni = list_entry(the_lnet.ln_nis.next,
 904                                 lnet_ni_t, ni_list);
 905                 list_del (&ni->ni_list);
 906
 907                 the_lnet.ln_nzombie_nis++;
 908                 lnet_ni_decref_locked(ni); /* drop apini's ref */
 909         }
 910
 911         /* Drop the cached eqwait NI. */
 912         if (the_lnet.ln_eqwaitni != NULL) {
 913                 lnet_ni_decref_locked(the_lnet.ln_eqwaitni);
 914                 the_lnet.ln_eqwaitni = NULL;
 915         }
 916
 917         /* Drop the cached loopback NI. */
 918         if (the_lnet.ln_loni != NULL) {
 919                 lnet_ni_decref_locked(the_lnet.ln_loni);
 920                 the_lnet.ln_loni = NULL;
 921         }
 922
 923         LNET_UNLOCK();
 924
 925         /* Clear lazy portals and drop delayed messages which hold refs
 926          * on their lnet_msg_t::msg_rxpeer */
 927         for (i = 0; i < the_lnet.ln_nportals; i++)
 928                 LNetClearLazyPortal(i);
 929
 930         /* Clear the peer table and wait for all peers to go (they hold refs on
 931          * their NIs) */
 932         lnet_clear_peer_table();
 933
 934         LNET_LOCK();
 935         /* Now wait for the NI's I just nuked to show up on apini_zombie_nis
 936          * and shut them down in guaranteed thread context */
 937         i = 2;
 938         while (the_lnet.ln_nzombie_nis != 0) {
 939
 940                 while (list_empty(&the_lnet.ln_zombie_nis)) {
 941                         LNET_UNLOCK();
 942                         ++i;
 943                         if ((i & (-i)) == i)
 944                                 CDEBUG(D_WARNING,"Waiting for %d zombie NIs\n",
 945                                        the_lnet.ln_nzombie_nis);
 946                         cfs_pause(cfs_time_seconds(1));
 947                         LNET_LOCK();
 948                 }
 949
 950                 ni = list_entry(the_lnet.ln_zombie_nis.next,
 951                                 lnet_ni_t, ni_list);
 952                 list_del(&ni->ni_list);
 953                 ni->ni_lnd->lnd_refcount--;
 954
 955                 LNET_UNLOCK();
 956
 957                 islo = ni->ni_lnd->lnd_type == LOLND;
 958
 959                 LASSERT (!in_interrupt ());
 960                 (ni->ni_lnd->lnd_shutdown)(ni);
 961
 962                 /* can't deref lnd anymore now; it might have unregistered
 963                  * itself...  */
 964
 965                 if (!islo)
 966                         CDEBUG(D_LNI, "Removed LNI %s\n",
 967                                libcfs_nid2str(ni->ni_nid));
 968
 969                 LIBCFS_FREE(ni, sizeof(*ni));
 970
 971                 LNET_LOCK();
 972                 the_lnet.ln_nzombie_nis--;
 973         }
 974
 975         the_lnet.ln_shutdown = 0;
 976         LNET_UNLOCK();
 977
 978         if (the_lnet.ln_network_tokens != NULL) {
 979                 LIBCFS_FREE(the_lnet.ln_network_tokens,
 980                             the_lnet.ln_network_tokens_nob);
 981                 the_lnet.ln_network_tokens = NULL;
 982         }
 983 }
 984
 985 int
 986 lnet_startup_lndnis (void)
 987 {
 988         lnd_t             *lnd;
 989         lnet_ni_t         *ni;
 990         struct list_head   nilist;
 991         int                rc = 0;
 992         int                lnd_type;
 993         int                nicount = 0;
 994         char              *nets = lnet_get_networks();
 995
 996         CFS_INIT_LIST_HEAD(&nilist);
 997
 998         if (nets == NULL)
 999                 goto failed;
1000
1001         rc = lnet_parse_networks(&nilist, nets);
1002         if (rc != 0)
1003                 goto failed;
1004
1005         while (!list_empty(&nilist)) {
1006                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1007                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1008
1009                 LASSERT (libcfs_isknown_lnd(lnd_type));
1010
1011                 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1012                 lnd = lnet_find_lnd_by_type(lnd_type);
1013
1014 #ifdef __KERNEL__
1015                 if (lnd == NULL) {
1016                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1017                         rc = request_module(libcfs_lnd2modname(lnd_type));
1018                         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1019
1020                         lnd = lnet_find_lnd_by_type(lnd_type);
1021                         if (lnd == NULL) {
1022                                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1023                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1024                                        libcfs_lnd2str(lnd_type),
1025                                        libcfs_lnd2modname(lnd_type), rc);
1026 #ifndef CONFIG_KMOD
1027                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1028                                          "compiled with CONFIG_KMOD set for "
1029                                          "automatic module loading.");
1030 #endif
1031                                 goto failed;
1032                         }
1033                 }
1034 #else
1035                 if (lnd == NULL) {
1036                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1037                         CERROR("LND %s not supported\n",
1038                                libcfs_lnd2str(lnd_type));
1039                         goto failed;
1040                 }
1041 #endif
1042
1043                 ni->ni_refcount = 1;
1044
1045                 LNET_LOCK();
1046                 lnd->lnd_refcount++;
1047                 LNET_UNLOCK();
1048
1049                 ni->ni_lnd = lnd;
1050
1051                 rc = (lnd->lnd_startup)(ni);
1052
1053                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1054
1055                 if (rc != 0) {
1056                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1057                                            "\n",
1058                                            rc, libcfs_lnd2str(lnd->lnd_type));
1059                         LNET_LOCK();
1060                         lnd->lnd_refcount--;
1061                         LNET_UNLOCK();
1062                         goto failed;
1063                 }
1064
1065                 list_del(&ni->ni_list);
1066
1067                 LNET_LOCK();
1068                 list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1069                 LNET_UNLOCK();
1070
1071                 if (lnd->lnd_type == LOLND) {
1072                         lnet_ni_addref(ni);
1073                         LASSERT (the_lnet.ln_loni == NULL);
1074                         the_lnet.ln_loni = ni;
1075                         continue;
1076                 }
1077
1078 #ifndef __KERNEL__
1079                 if (lnd->lnd_wait != NULL) {
1080                         if (the_lnet.ln_eqwaitni == NULL) {
1081                                 lnet_ni_addref(ni);
1082                                 the_lnet.ln_eqwaitni = ni;
1083                         }
1084                 } else {
1085 # ifndef HAVE_LIBPTHREAD
1086                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1087                                            "single-threaded runtime\n",
1088                                            libcfs_lnd2str(lnd_type));
1089                         goto failed;
1090 # endif
1091                 }
1092 #endif
1093                 if (ni->ni_peertxcredits == 0 ||
1094                     ni->ni_maxtxcredits == 0) {
1095                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1096                                            libcfs_lnd2str(lnd->lnd_type),
1097                                            ni->ni_peertxcredits == 0 ?
1098                                            "" : "per-peer ");
1099                         goto failed;
1100                 }
1101
1102                 ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
1103
1104                 CDEBUG(D_LNI, "Added LNI %s [%d/%d]\n",
1105                        libcfs_nid2str(ni->ni_nid),
1106                        ni->ni_peertxcredits, ni->ni_txcredits);
1107
1108                 /* Handle nidstrings for network 0 just like this one */
1109                 if (the_lnet.ln_ptlcompat > 0) {
1110                         if (nicount > 0) {
1111                                 LCONSOLE_ERROR_MSG(0x108, "Can't run > 1 "
1112                                        "network when portals_compatibility is "
1113                                        "set\n");
1114                                 goto failed;
1115                         }
1116                         libcfs_setnet0alias(lnd->lnd_type);
1117                 }
1118
1119                 nicount++;
1120         }
1121
1122         if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
1123                 lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
1124                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1125                                    "\n",
1126                                    libcfs_lnd2str(lnd_type));
1127                 goto failed;
1128         }
1129
1130         return 0;
1131
1132  failed:
1133         lnet_shutdown_lndnis();
1134
1135         while (!list_empty(&nilist)) {
1136                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1137                 list_del(&ni->ni_list);
1138                 LIBCFS_FREE(ni, sizeof(*ni));
1139         }
1140
1141         return -ENETDOWN;
1142 }
1143
1144 int
1145 LNetInit(void)
1146 {
1147         int    rc;
1148
1149         lnet_assert_wire_constants ();
1150         LASSERT (!the_lnet.ln_init);
1151
1152         memset(&the_lnet, 0, sizeof(the_lnet));
1153
1154         rc = lnet_get_portals_compatibility();
1155         if (rc < 0)
1156                 return rc;
1157
1158         lnet_init_locks();
1159         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1160         the_lnet.ln_ptlcompat = rc;
1161         the_lnet.ln_refcount = 0;
1162         the_lnet.ln_init = 1;
1163
1164 #ifdef __KERNEL__
1165         /* All LNDs apart from the LOLND are in separate modules.  They
1166          * register themselves when their module loads, and unregister
1167          * themselves when their module is unloaded. */
1168 #else
1169         /* Register LNDs
1170          * NB the order here determines default 'networks=' order */
1171 # ifdef CRAY_XT3
1172         LNET_REGISTER_ULND(the_ptllnd);
1173 # endif
1174 # ifdef HAVE_LIBPTHREAD
1175         LNET_REGISTER_ULND(the_tcplnd);
1176 # endif
1177 #endif
1178         lnet_register_lnd(&the_lolnd);
1179         return 0;
1180 }
1181
1182 void
1183 LNetFini(void)
1184 {
1185         LASSERT (the_lnet.ln_init);
1186         LASSERT (the_lnet.ln_refcount == 0);
1187
1188         while (!list_empty(&the_lnet.ln_lnds))
1189                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1190                                                lnd_t, lnd_list));
1191         lnet_fini_locks();
1192
1193         the_lnet.ln_init = 0;
1194 }
1195
1196 int
1197 LNetNIInit(lnet_pid_t requested_pid)
1198 {
1199         int         im_a_router = 0;
1200         int         rc;
1201
1202         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1203
1204         LASSERT (the_lnet.ln_init);
1205         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1206
1207         if (the_lnet.ln_refcount > 0) {
1208                 rc = the_lnet.ln_refcount++;
1209                 goto out;
1210         }
1211
1212         if (requested_pid == LNET_PID_ANY) {
1213                 /* Don't instantiate LNET just for me */
1214                 rc = -ENETDOWN;
1215                 goto failed0;
1216         }
1217
1218         rc = lnet_prepare(requested_pid);
1219         if (rc != 0)
1220                 goto failed0;
1221
1222         rc = lnet_startup_lndnis();
1223         if (rc != 0)
1224                 goto failed1;
1225
1226         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1227         if (rc != 0)
1228                 goto failed2;
1229
1230         rc = lnet_check_routes();
1231         if (rc != 0)
1232                 goto failed2;
1233
1234         rc = lnet_alloc_rtrpools(im_a_router);
1235         if (rc != 0)
1236                 goto failed2;
1237
1238         rc = lnet_acceptor_start();
1239         if (rc != 0)
1240                 goto failed2;
1241
1242         the_lnet.ln_refcount = 1;
1243         /* Now I may use my own API functions... */
1244
1245         rc = lnet_router_checker_start();
1246         if (rc != 0)
1247                 goto failed3;
1248
1249         rc = lnet_ping_target_init();
1250         if (rc != 0)
1251                 goto failed4;
1252
1253         lnet_proc_init();
1254         goto out;
1255
1256  failed4:
1257         lnet_router_checker_stop();
1258  failed3:
1259         the_lnet.ln_refcount = 0;
1260         lnet_acceptor_stop();
1261  failed2:
1262         lnet_destroy_routes();
1263         lnet_shutdown_lndnis();
1264  failed1:
1265         lnet_unprepare();
1266  failed0:
1267         LASSERT (rc < 0);
1268  out:
1269         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1270         return rc;
1271 }
1272
1273 int
1274 LNetNIFini()
1275 {
1276         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1277
1278         LASSERT (the_lnet.ln_init);
1279         LASSERT (the_lnet.ln_refcount > 0);
1280
1281         if (the_lnet.ln_refcount != 1) {
1282                 the_lnet.ln_refcount--;
1283         } else {
1284                 LASSERT (!the_lnet.ln_niinit_self);
1285
1286                 lnet_proc_fini();
1287                 lnet_ping_target_fini();
1288                 lnet_router_checker_stop();
1289
1290                 /* Teardown fns that use my own API functions BEFORE here */
1291                 the_lnet.ln_refcount = 0;
1292
1293                 lnet_acceptor_stop();
1294                 lnet_destroy_routes();
1295                 lnet_shutdown_lndnis();
1296                 lnet_unprepare();
1297         }
1298
1299         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1300         return 0;
1301 }
1302
1303 int
1304 LNetCtl(unsigned int cmd, void *arg)
1305 {
1306         struct libcfs_ioctl_data *data = arg;
1307         lnet_process_id_t         id;
1308         lnet_ni_t                *ni;
1309         int                       rc;
1310
1311         LASSERT (the_lnet.ln_init);
1312         LASSERT (the_lnet.ln_refcount > 0);
1313
1314         switch (cmd) {
1315         case IOC_LIBCFS_GET_NI:
1316                 rc = LNetGetId(data->ioc_count, &id);
1317                 data->ioc_nid = id.nid;
1318                 return rc;
1319
1320         case IOC_LIBCFS_FAIL_NID:
1321                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1322
1323         case IOC_LIBCFS_ADD_ROUTE:
1324                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1325                                     data->ioc_nid);
1326                 return (rc != 0) ? rc : lnet_check_routes();
1327
1328         case IOC_LIBCFS_DEL_ROUTE:
1329                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1330
1331         case IOC_LIBCFS_GET_ROUTE:
1332                 return lnet_get_route(data->ioc_count,
1333                                       &data->ioc_net, &data->ioc_count,
1334                                       &data->ioc_nid, &data->ioc_flags);
1335         case IOC_LIBCFS_NOTIFY_ROUTER:
1336                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1337                                    (time_t)data->ioc_u64[0]);
1338
1339         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1340                 return the_lnet.ln_ptlcompat;
1341
1342         case IOC_LIBCFS_LNET_DIST:
1343                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1344                 if (rc < 0 && rc != -EHOSTUNREACH)
1345                         return rc;
1346
1347                 data->ioc_u32[0] = rc;
1348                 return 0;
1349
1350         case IOC_LIBCFS_TESTPROTOCOMPAT:
1351                 LNET_LOCK();
1352                 the_lnet.ln_testprotocompat = data->ioc_flags;
1353                 LNET_UNLOCK();
1354                 return 0;
1355
1356         case IOC_LIBCFS_PING:
1357                 rc = lnet_ping((lnet_process_id_t) {.nid = data->ioc_nid,
1358                                                     .pid = data->ioc_u32[0]},
1359                                data->ioc_u32[1], /* timeout */
1360                                (lnet_process_id_t *)data->ioc_pbuf1,
1361                                data->ioc_plen1/sizeof(lnet_process_id_t));
1362                 if (rc < 0)
1363                         return rc;
1364                 data->ioc_count = rc;
1365                 return 0;
1366
1367         case IOC_LIBCFS_DEBUG_PEER: {
1368                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1369                  * via an ioctl */
1370                 lnet_process_id_t *id = arg;
1371
1372                 lnet_debug_peer(id->nid);
1373
1374                 ni = lnet_net2ni(LNET_NIDNET(id->nid));
1375                 if (ni == NULL) {
1376                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(*id));
1377                 } else {
1378                         if (ni->ni_lnd->lnd_ctl == NULL) {
1379                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1380                                        libcfs_id2str(*id));
1381                         } else {
1382                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1383                         }
1384
1385                         lnet_ni_decref(ni);
1386                 }
1387                 return 0;
1388         }
1389
1390         default:
1391                 ni = lnet_net2ni(data->ioc_net);
1392                 if (ni == NULL)
1393                         return -EINVAL;
1394
1395                 if (ni->ni_lnd->lnd_ctl == NULL)
1396                         rc = -EINVAL;
1397                 else
1398                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1399
1400                 lnet_ni_decref(ni);
1401                 return rc;
1402         }
1403         /* not reached */
1404 }
1405
1406 int
1407 LNetGetId(unsigned int index, lnet_process_id_t *id)
1408 {
1409         lnet_ni_t        *ni;
1410         struct list_head *tmp;
1411         int               rc = -ENOENT;
1412
1413         LASSERT (the_lnet.ln_init);
1414         LASSERT (the_lnet.ln_refcount > 0);
1415
1416         LNET_LOCK();
1417
1418         list_for_each(tmp, &the_lnet.ln_nis) {
1419                 if (index-- != 0)
1420                         continue;
1421
1422                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1423
1424                 id->nid = ni->ni_nid;
1425                 id->pid = the_lnet.ln_pid;
1426                 rc = 0;
1427                 break;
1428         }
1429
1430         LNET_UNLOCK();
1431
1432         return rc;
1433 }
1434
1435 void
1436 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1437 {
1438         snprintf(str, len, LPX64, h.cookie);
1439 }
1440
1441
1442 int
1443 lnet_ping_target_init(void)
1444 {
1445         lnet_handle_me_t  meh;
1446         lnet_process_id_t id;
1447         int               rc;
1448         int               rc2;
1449         int               n;
1450         int               infosz;
1451         int               i;
1452
1453         for (n = 0; ; n++) {
1454                 rc = LNetGetId(n, &id);
1455                 if (rc == -ENOENT)
1456                         break;
1457
1458                 LASSERT (rc == 0);
1459         }
1460
1461         infosz = offsetof(lnet_ping_info_t, pi_nid[n]);
1462         LIBCFS_ALLOC(the_lnet.ln_ping_info, infosz);
1463         if (the_lnet.ln_ping_info == NULL) {
1464                 CERROR("Can't allocate ping info[%d]\n", n);
1465                 return -ENOMEM;
1466         }
1467
1468         the_lnet.ln_ping_info->pi_magic   = LNET_PROTO_PING_MAGIC;
1469         the_lnet.ln_ping_info->pi_version = LNET_PROTO_PING_VERSION;
1470         the_lnet.ln_ping_info->pi_pid     = the_lnet.ln_pid;
1471         the_lnet.ln_ping_info->pi_nnids   = n;
1472
1473         for (i = 0; i < n; i++) {
1474                 rc = LNetGetId(i, &id);
1475                 LASSERT (rc == 0);
1476                 the_lnet.ln_ping_info->pi_nid[i] = id.nid;
1477         }
1478
1479         /* We can have a tiny EQ since we only need to see the unlink event on
1480          * teardown, which by definition is the last one! */
1481         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1482         if (rc != 0) {
1483                 CERROR("Can't allocate ping EQ: %d\n", rc);
1484                 goto failed_0;
1485         }
1486
1487         rc = LNetMEAttach(LNET_RESERVED_PORTAL,
1488                           (lnet_process_id_t){.nid = LNET_NID_ANY,
1489                                               .pid = LNET_PID_ANY},
1490                           LNET_PROTO_PING_MATCHBITS, 0LL,
1491                           LNET_UNLINK, LNET_INS_AFTER,
1492                           &meh);
1493         if (rc != 0) {
1494                 CERROR("Can't create ping ME: %d\n", rc);
1495                 goto failed_1;
1496         }
1497
1498         rc = LNetMDAttach(meh,
1499                           (lnet_md_t){.start = the_lnet.ln_ping_info,
1500                                       .length = infosz,
1501                                       .threshold = LNET_MD_THRESH_INF,
1502                                       .options = (LNET_MD_OP_GET |
1503                                                   LNET_MD_TRUNCATE |
1504                                                   LNET_MD_MANAGE_REMOTE),
1505                                       .eq_handle = the_lnet.ln_ping_target_eq},
1506                           LNET_RETAIN,
1507                           &the_lnet.ln_ping_target_md);
1508         if (rc != 0) {
1509                 CERROR("Can't attach ping MD: %d\n", rc);
1510                 goto failed_2;
1511         }
1512
1513         return 0;
1514
1515  failed_2:
1516         rc2 = LNetMEUnlink(meh);
1517         LASSERT (rc2 == 0);
1518  failed_1:
1519         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1520         LASSERT (rc2 == 0);
1521  failed_0:
1522         LIBCFS_FREE(the_lnet.ln_ping_info, infosz);
1523
1524         return rc;
1525 }
1526
1527 void
1528 lnet_ping_target_fini(void)
1529 {
1530         lnet_event_t    event;
1531         int             rc;
1532         int             which;
1533         int             timeout_ms = 1000;
1534         cfs_sigset_t    blocked = cfs_block_allsigs();
1535
1536         LNetMDUnlink(the_lnet.ln_ping_target_md);
1537         /* NB md could be busy; this just starts the unlink */
1538
1539         for (;;) {
1540                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1541                                 timeout_ms, &event, &which);
1542
1543                 /* I expect overflow... */
1544                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1545
1546                 if (rc == 0) {
1547                         /* timed out: provide a diagnostic */
1548                         CWARN("Still waiting for ping MD to unlink\n");
1549                         timeout_ms *= 2;
1550                         continue;
1551                 }
1552
1553                 /* Got a valid event */
1554                 if (event.unlinked)
1555                         break;
1556         }
1557
1558         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1559         LASSERT (rc == 0);
1560
1561         LIBCFS_FREE(the_lnet.ln_ping_info,
1562                     offsetof(lnet_ping_info_t,
1563                              pi_nid[the_lnet.ln_ping_info->pi_nnids]));
1564
1565         cfs_restore_sigs(blocked);
1566 }
1567
1568 int
1569 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1570 {
1571         lnet_handle_eq_t     eqh;
1572         lnet_handle_md_t     mdh;
1573         lnet_event_t         event;
1574         int                  which;
1575         int                  unlinked = 0;
1576         int                  replied = 0;
1577         const int            a_long_time = 60000; /* mS */
1578         int                  infosz = offsetof(lnet_ping_info_t, pi_nid[n_ids]);
1579         lnet_ping_info_t    *info;
1580         lnet_process_id_t    tmpid;
1581         int                  i;
1582         int                  nob;
1583         int                  rc;
1584         int                  rc2;
1585         cfs_sigset_t         blocked;
1586
1587         if (n_ids <= 0 ||
1588             id.nid == LNET_NID_ANY ||
1589             timeout_ms > 500000 ||              /* arbitrary limit! */
1590             n_ids > 20)                         /* arbitrary limit! */
1591                 return -EINVAL;
1592
1593         if (id.pid == LNET_PID_ANY)
1594                 id.pid = LUSTRE_SRV_LNET_PID;
1595
1596         LIBCFS_ALLOC(info, infosz);
1597         if (info == NULL)
1598                 return -ENOMEM;
1599
1600         /* NB 2 events max (including any unlink event) */
1601         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1602         if (rc != 0) {
1603                 CERROR("Can't allocate EQ: %d\n", rc);
1604                 goto out_0;
1605         }
1606
1607         rc = LNetMDBind((lnet_md_t){.start = info,
1608                                     .length = infosz,
1609                                     .threshold = 2, /* GET/REPLY */
1610                                     .options = LNET_MD_TRUNCATE,
1611                                     .eq_handle = eqh},
1612                         LNET_UNLINK,
1613                         &mdh);
1614         if (rc != 0) {
1615                 CERROR("Can't bind MD: %d\n", rc);
1616                 goto out_1;
1617         }
1618
1619         rc = LNetGet(LNET_NID_ANY, mdh, id,
1620                      LNET_RESERVED_PORTAL,
1621                      LNET_PROTO_PING_MATCHBITS, 0);
1622
1623         if (rc != 0) {
1624                 /* Don't CERROR; this could be deliberate! */
1625
1626                 rc2 = LNetMDUnlink(mdh);
1627                 LASSERT (rc2 == 0);
1628
1629                 /* NB must wait for the UNLINK event below... */
1630                 unlinked = 1;
1631                 timeout_ms = a_long_time;
1632         }
1633
1634         do {
1635                 /* MUST block for unlink to complete */
1636                 if (unlinked)
1637                         blocked = cfs_block_allsigs();
1638
1639                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1640
1641                 if (unlinked)
1642                         cfs_restore_sigs(blocked);
1643
1644                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1645                        (rc2 <= 0) ? -1 : event.type,
1646                        (rc2 <= 0) ? -1 : event.status,
1647                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1648
1649                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1650
1651                 if (rc2 <= 0 || event.status != 0) {
1652                         /* timeout or error */
1653                         if (!replied && rc == 0)
1654                                 rc = (rc2 < 0) ? rc2 :
1655                                      (rc2 == 0) ? -ETIMEDOUT :
1656                                      event.status;
1657
1658                         if (!unlinked) {
1659                                 /* Ensure completion in finite time... */
1660                                 LNetMDUnlink(mdh);
1661                                 /* No assertion (racing with network) */
1662                                 unlinked = 1;
1663                                 timeout_ms = a_long_time;
1664                         } else if (rc2 == 0) {
1665                                 /* timed out waiting for unlink */
1666                                 CWARN("ping %s: late network completion\n",
1667                                       libcfs_id2str(id));
1668                         }
1669
1670                 } else if (event.type == LNET_EVENT_REPLY) {
1671                         replied = 1;
1672                         rc = event.mlength;
1673                 }
1674
1675         } while (rc2 <= 0 || !event.unlinked);
1676
1677         if (!replied) {
1678                 if (rc >= 0)
1679                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1680                               libcfs_id2str(id));
1681                 rc = -EIO;
1682                 goto out_1;
1683         }
1684
1685         nob = rc;
1686         LASSERT (nob >= 0 && nob <= infosz);
1687
1688         rc = -EPROTO;                           /* if I can't parse... */
1689
1690         if (nob < 8) {
1691                 /* can't check magic/version */
1692                 CERROR("%s: ping info too short %d\n",
1693                        libcfs_id2str(id), nob);
1694                 goto out_1;
1695         }
1696
1697         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1698                 /* NB I might be swabbing garbage until I check below, but it
1699                  * doesn't matter */
1700                 __swab32s(&info->pi_version);
1701                 __swab32s(&info->pi_pid);
1702                 __swab32s(&info->pi_nnids);
1703                 for (i = 0; i < info->pi_nnids && i < n_ids; i++)
1704                         __swab64s(&info->pi_nid[i]);
1705
1706         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1707                 CERROR("%s: Unexpected magic %08x\n",
1708                        libcfs_id2str(id), info->pi_magic);
1709                 goto out_1;
1710         }
1711
1712         if (info->pi_version != LNET_PROTO_PING_VERSION) {
1713                 CERROR("%s: Unexpected version 0x%x\n",
1714                        libcfs_id2str(id), info->pi_version);
1715                 goto out_1;
1716         }
1717
1718         if (nob < offsetof(lnet_ping_info_t, pi_nid[0])) {
1719                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1720                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[0]));
1721                 goto out_1;
1722         }
1723
1724         if (info->pi_nnids < n_ids)
1725                 n_ids = info->pi_nnids;
1726
1727         if (nob < offsetof(lnet_ping_info_t, pi_nid[n_ids])) {
1728                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1729                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[n_ids]));
1730                 goto out_1;
1731         }
1732
1733         rc = -EFAULT;                           /* If I SEGV... */
1734
1735         for (i = 0; i < n_ids; i++) {
1736                 tmpid.pid = info->pi_pid;
1737                 tmpid.nid = info->pi_nid[i];
1738 #ifdef __KERNEL__
1739                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1740                         goto out_1;
1741 #else
1742                 ids[i] = tmpid;
1743 #endif
1744         }
1745         rc = info->pi_nnids;
1746
1747  out_1:
1748         rc2 = LNetEQFree(eqh);
1749         if (rc2 != 0)
1750                 CERROR("rc2 %d\n", rc2);
1751         LASSERT (rc2 == 0);
1752
1753  out_0:
1754         LIBCFS_FREE(info, infosz);
1755         return rc;
1756 }