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