lnet/klnds/iiblnd/iiblnd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004 Cluster File Systems, Inc.
   5  *   Author: Eric Barton <eric@bartonsoftware.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #include "iibnal.h"
  25
  26 nal_t                   kibnal_api;
  27 ptl_handle_ni_t         kibnal_ni;
  28 kib_tunables_t          kibnal_tunables;
  29
  30 kib_data_t              kibnal_data = {
  31         .kib_service_id = IBNAL_SERVICE_NUMBER,
  32 };
  33
  34 #ifdef CONFIG_SYSCTL
  35 #define IBNAL_SYSCTL             202
  36
  37 #define IBNAL_SYSCTL_TIMEOUT     1
  38
  39 static ctl_table kibnal_ctl_table[] = {
  40         {IBNAL_SYSCTL_TIMEOUT, "timeout",
  41          &kibnal_tunables.kib_io_timeout, sizeof (int),
  42          0644, NULL, &proc_dointvec},
  43         { 0 }
  44 };
  45
  46 static ctl_table kibnal_top_ctl_table[] = {
  47         {IBNAL_SYSCTL, "iibnal", NULL, 0, 0555, kibnal_ctl_table},
  48         { 0 }
  49 };
  50 #endif
  51
  52 #ifdef unused
  53 void
  54 print_service(IB_SERVICE_RECORD *service, char *tag, int rc)
  55 {
  56         char name[32];
  57
  58         if (service == NULL)
  59         {
  60                 CWARN("tag       : %s\n"
  61                       "status    : %d (NULL)\n", tag, rc);
  62                 return;
  63         }
  64         strncpy (name, service->ServiceName, sizeof(name)-1);
  65         name[sizeof(name)-1] = 0;
  66
  67         CWARN("tag       : %s\n"
  68               "status    : %d\n"
  69               "service id: "LPX64"\n"
  70               "name      : %s\n"
  71               "NID       : "LPX64"\n", tag, rc,
  72               service->RID.ServiceID, name,
  73               *kibnal_service_nid_field(service));
  74 }
  75 #endif
  76
  77 static void
  78 kibnal_service_setunset_done (void *arg, FABRIC_OPERATION_DATA *fod,
  79                               FSTATUS frc, uint32 madrc)
  80 {
  81         *(FSTATUS *)arg = frc;
  82         up (&kibnal_data.kib_nid_signal);
  83 }
  84
  85 #if IBNAL_CHECK_ADVERT
  86 static void
  87 kibnal_service_query_done (void *arg, QUERY *qry,
  88                            QUERY_RESULT_VALUES *qry_result)
  89 {
  90         FSTATUS frc = qry_result->Status;
  91
  92         if (frc != FSUCCESS &&
  93             qry_result->ResultDataSize == 0)
  94                 frc = FERROR;
  95
  96         *(FSTATUS *)arg = frc;
  97         up (&kibnal_data.kib_nid_signal);
  98 }
  99
 100 static void
 101 kibnal_check_advert (void)
 102 {
 103         QUERY                  *qry;
 104         IB_SERVICE_RECORD      *svc;
 105         FSTATUS                 frc;
 106         FSTATUS                 frc2;
 107
 108         PORTAL_ALLOC(qry, sizeof(*qry));
 109         if (qry == NULL)
 110                 return;
 111
 112         memset (qry, 0, sizeof(*qry));
 113         qry->InputType = InputTypeServiceRecord;
 114         qry->OutputType = OutputTypeServiceRecord;
 115         qry->InputValue.ServiceRecordValue.ComponentMask = KIBNAL_SERVICE_KEY_MASK;
 116         svc = &qry->InputValue.ServiceRecordValue.ServiceRecord;
 117         kibnal_set_service_keys(svc, kibnal_data.kib_nid);
 118
 119         frc = iibt_sd_query_port_fabric_information(kibnal_data.kib_sd,
 120                                                     kibnal_data.kib_port_guid,
 121                                                     qry,
 122                                                     kibnal_service_query_done,
 123                                                     NULL, &frc2);
 124         if (frc != FSUCCESS && frc != FPENDING) {
 125                 CERROR ("Immediate error %d checking SM service\n", frc);
 126         } else {
 127                 down (&kibnal_data.kib_nid_signal);
 128                 frc = frc2;
 129
 130                 if (frc != 0)
 131                         CERROR ("Error %d checking SM service\n", rc);
 132         }
 133
 134         return (rc);
 135 }
 136 #endif
 137
 138 static void fill_fod(FABRIC_OPERATION_DATA *fod, FABRIC_OPERATION_TYPE type)
 139 {
 140         IB_SERVICE_RECORD     *svc;
 141
 142         memset (fod, 0, sizeof(*fod));
 143         fod->Type = type;
 144
 145         svc = &fod->Value.ServiceRecordValue.ServiceRecord;
 146         svc->RID.ServiceID = kibnal_data.kib_service_id;
 147         svc->RID.ServiceGID.Type.Global.InterfaceID = kibnal_data.kib_port_guid;
 148         svc->RID.ServiceGID.Type.Global.SubnetPrefix = DEFAULT_SUBNET_PREFIX;
 149         svc->RID.ServiceP_Key = kibnal_data.kib_port_pkey;
 150         svc->ServiceLease = 0xffffffff;
 151
 152         kibnal_set_service_keys(svc, kibnal_data.kib_nid);
 153 }
 154
 155 static int
 156 kibnal_advertise (void)
 157 {
 158         FABRIC_OPERATION_DATA *fod;
 159         IB_SERVICE_RECORD     *svc;
 160         FSTATUS                frc;
 161         FSTATUS                frc2;
 162
 163         LASSERT (kibnal_data.kib_nid != PTL_NID_ANY);
 164
 165         PORTAL_ALLOC(fod, sizeof(*fod));
 166         if (fod == NULL)
 167                 return (-ENOMEM);
 168
 169         fill_fod(fod, FabOpSetServiceRecord);
 170         svc = &fod->Value.ServiceRecordValue.ServiceRecord;
 171
 172         CDEBUG(D_NET, "Advertising service id "LPX64" %s:"LPX64"\n",
 173                svc->RID.ServiceID,
 174                svc->ServiceName, *kibnal_service_nid_field(svc));
 175
 176         frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd,
 177                                             kibnal_data.kib_port_guid,
 178                                             fod, kibnal_service_setunset_done,
 179                                             NULL, &frc2);
 180
 181         if (frc != FSUCCESS && frc != FPENDING) {
 182                 CERROR ("Immediate error %d advertising NID "LPX64"\n",
 183                         frc, kibnal_data.kib_nid);
 184                 goto out;
 185         }
 186
 187         down (&kibnal_data.kib_nid_signal);
 188
 189         frc = frc2;
 190         if (frc != FSUCCESS)
 191                 CERROR ("Error %d advertising BUD "LPX64"\n",
 192                         frc, kibnal_data.kib_nid);
 193 out:
 194         PORTAL_FREE(fod, sizeof(*fod));
 195         return (frc == FSUCCESS) ? 0 : -EINVAL;
 196 }
 197
 198 static void
 199 kibnal_unadvertise (int expect_success)
 200 {
 201         FABRIC_OPERATION_DATA *fod;
 202         IB_SERVICE_RECORD     *svc;
 203         FSTATUS                frc;
 204         FSTATUS                frc2;
 205
 206         LASSERT (kibnal_data.kib_nid != PTL_NID_ANY);
 207
 208         PORTAL_ALLOC(fod, sizeof(*fod));
 209         if (fod == NULL)
 210                 return;
 211
 212         fill_fod(fod, FabOpDeleteServiceRecord);
 213         svc = &fod->Value.ServiceRecordValue.ServiceRecord;
 214
 215         CDEBUG(D_NET, "Unadvertising service %s:"LPX64"\n",
 216                svc->ServiceName, *kibnal_service_nid_field(svc));
 217
 218         frc = iibt_sd_port_fabric_operation(kibnal_data.kib_sd,
 219                                             kibnal_data.kib_port_guid,
 220                                             fod, kibnal_service_setunset_done,
 221                                             NULL, &frc2);
 222
 223         if (frc != FSUCCESS && frc != FPENDING) {
 224                 CERROR ("Immediate error %d unadvertising NID "LPX64"\n",
 225                         frc, kibnal_data.kib_nid);
 226                 goto out;
 227         }
 228
 229         down (&kibnal_data.kib_nid_signal);
 230
 231         if ((frc2 == FSUCCESS) == !!expect_success)
 232                 goto out;
 233
 234         if (expect_success)
 235                 CERROR("Error %d unadvertising NID "LPX64"\n",
 236                        frc2, kibnal_data.kib_nid);
 237         else
 238                 CWARN("Removed conflicting NID "LPX64"\n",
 239                       kibnal_data.kib_nid);
 240  out:
 241         PORTAL_FREE(fod, sizeof(*fod));
 242 }
 243
 244 static int
 245 kibnal_set_mynid(ptl_nid_t nid)
 246 {
 247         struct timeval tv;
 248         lib_ni_t      *ni = &kibnal_lib.libnal_ni;
 249         int            rc;
 250         FSTATUS        frc;
 251
 252         CDEBUG(D_IOCTL, "setting mynid to "LPX64" (old nid="LPX64")\n",
 253                nid, ni->ni_pid.nid);
 254
 255         do_gettimeofday(&tv);
 256
 257         down (&kibnal_data.kib_nid_mutex);
 258
 259         if (nid == kibnal_data.kib_nid) {
 260                 /* no change of NID */
 261                 up (&kibnal_data.kib_nid_mutex);
 262                 return (0);
 263         }
 264
 265         CDEBUG(D_NET, "NID "LPX64"("LPX64")\n",
 266                kibnal_data.kib_nid, nid);
 267
 268         if (kibnal_data.kib_nid != PTL_NID_ANY) {
 269
 270                 kibnal_unadvertise (1);
 271
 272                 frc = iibt_cm_cancel(kibnal_data.kib_cep);
 273                 if (frc != FSUCCESS && frc != FPENDING)
 274                         CERROR ("Error %d stopping listener\n", frc);
 275
 276                 frc = iibt_cm_destroy_cep(kibnal_data.kib_cep);
 277                 if (frc != FSUCCESS)
 278                         CERROR ("Error %d destroying CEP\n", frc);
 279
 280                 kibnal_data.kib_cep = NULL;
 281         }
 282
 283         kibnal_data.kib_nid = ni->ni_pid.nid = nid;
 284         kibnal_data.kib_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
 285
 286         /* Delete all existing peers and their connections after new
 287          * NID/incarnation set to ensure no old connections in our brave
 288          * new world. */
 289         kibnal_del_peer (PTL_NID_ANY, 0);
 290
 291         if (kibnal_data.kib_nid == PTL_NID_ANY) {
 292                 /* No new NID to install */
 293                 up (&kibnal_data.kib_nid_mutex);
 294                 return (0);
 295         }
 296
 297         /* remove any previous advert (crashed node etc) */
 298         kibnal_unadvertise(0);
 299
 300         kibnal_data.kib_cep = iibt_cm_create_cep(CM_RC_TYPE);
 301         if (kibnal_data.kib_cep == NULL) {
 302                 CERROR ("Can't create CEP\n");
 303                 rc = -ENOMEM;
 304         } else {
 305                 CM_LISTEN_INFO info;
 306                 memset (&info, 0, sizeof(info));
 307                 info.ListenAddr.EndPt.SID = kibnal_data.kib_service_id;
 308
 309                 frc = iibt_cm_listen(kibnal_data.kib_cep, &info,
 310                                      kibnal_listen_callback, NULL);
 311                 if (frc != FSUCCESS && frc != FPENDING) {
 312                         CERROR ("iibt_cm_listen error: %d\n", frc);
 313                         rc = -EINVAL;
 314                 } else {
 315                         rc = 0;
 316                 }
 317         }
 318
 319         if (rc == 0) {
 320                 rc = kibnal_advertise();
 321                 if (rc == 0) {
 322 #if IBNAL_CHECK_ADVERT
 323                         kibnal_check_advert();
 324 #endif
 325                         up (&kibnal_data.kib_nid_mutex);
 326                         return (0);
 327                 }
 328
 329                 iibt_cm_cancel (kibnal_data.kib_cep);
 330                 iibt_cm_destroy_cep (kibnal_data.kib_cep);
 331                 /* remove any peers that sprung up while I failed to
 332                  * advertise myself */
 333                 kibnal_del_peer (PTL_NID_ANY, 0);
 334         }
 335
 336         kibnal_data.kib_nid = PTL_NID_ANY;
 337         up (&kibnal_data.kib_nid_mutex);
 338         return (rc);
 339 }
 340
 341 kib_peer_t *
 342 kibnal_create_peer (ptl_nid_t nid)
 343 {
 344         kib_peer_t *peer;
 345
 346         LASSERT (nid != PTL_NID_ANY);
 347
 348         PORTAL_ALLOC (peer, sizeof (*peer));
 349         if (peer == NULL)
 350                 return (NULL);
 351
 352         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 353
 354         peer->ibp_nid = nid;
 355         atomic_set (&peer->ibp_refcount, 1);    /* 1 ref for caller */
 356
 357         INIT_LIST_HEAD (&peer->ibp_list);       /* not in the peer table yet */
 358         INIT_LIST_HEAD (&peer->ibp_conns);
 359         INIT_LIST_HEAD (&peer->ibp_tx_queue);
 360
 361         peer->ibp_reconnect_time = jiffies;
 362         peer->ibp_reconnect_interval = IBNAL_MIN_RECONNECT_INTERVAL;
 363
 364         atomic_inc (&kibnal_data.kib_npeers);
 365         return (peer);
 366 }
 367
 368 void
 369 kibnal_destroy_peer (kib_peer_t *peer)
 370 {
 371
 372         LASSERT (atomic_read (&peer->ibp_refcount) == 0);
 373         LASSERT (peer->ibp_persistence == 0);
 374         LASSERT (!kibnal_peer_active(peer));
 375         LASSERT (peer->ibp_connecting == 0);
 376         LASSERT (list_empty (&peer->ibp_conns));
 377         LASSERT (list_empty (&peer->ibp_tx_queue));
 378
 379         PORTAL_FREE (peer, sizeof (*peer));
 380
 381         /* NB a peer's connections keep a reference on their peer until
 382          * they are destroyed, so we can be assured that _all_ state to do
 383          * with this peer has been cleaned up when its refcount drops to
 384          * zero. */
 385         atomic_dec (&kibnal_data.kib_npeers);
 386 }
 387
 388 /* the caller is responsible for accounting for the additional reference
 389  * that this creates */
 390 kib_peer_t *
 391 kibnal_find_peer_locked (ptl_nid_t nid)
 392 {
 393         struct list_head *peer_list = kibnal_nid2peerlist (nid);
 394         struct list_head *tmp;
 395         kib_peer_t       *peer;
 396
 397         list_for_each (tmp, peer_list) {
 398
 399                 peer = list_entry (tmp, kib_peer_t, ibp_list);
 400
 401                 LASSERT (peer->ibp_persistence != 0 || /* persistent peer */
 402                          peer->ibp_connecting != 0 || /* creating conns */
 403                          !list_empty (&peer->ibp_conns));  /* active conn */
 404
 405                 if (peer->ibp_nid != nid)
 406                         continue;
 407
 408                 CDEBUG(D_NET, "got peer [%p] -> "LPX64" (%d)\n",
 409                        peer, nid, atomic_read (&peer->ibp_refcount));
 410                 return (peer);
 411         }
 412         return (NULL);
 413 }
 414
 415 kib_peer_t *
 416 kibnal_get_peer (ptl_nid_t nid)
 417 {
 418         kib_peer_t     *peer;
 419
 420         read_lock (&kibnal_data.kib_global_lock);
 421         peer = kibnal_find_peer_locked (nid);
 422         if (peer != NULL)                       /* +1 ref for caller? */
 423                 kib_peer_addref(peer);
 424         read_unlock (&kibnal_data.kib_global_lock);
 425
 426         return (peer);
 427 }
 428
 429 void
 430 kibnal_unlink_peer_locked (kib_peer_t *peer)
 431 {
 432         LASSERT (peer->ibp_persistence == 0);
 433         LASSERT (list_empty(&peer->ibp_conns));
 434
 435         LASSERT (kibnal_peer_active(peer));
 436         list_del_init (&peer->ibp_list);
 437         /* lose peerlist's ref */
 438         kib_peer_decref(peer);
 439 }
 440
 441 static int
 442 kibnal_get_peer_info (int index, ptl_nid_t *nidp, int *persistencep)
 443 {
 444         kib_peer_t        *peer;
 445         struct list_head  *ptmp;
 446         int                i;
 447
 448         read_lock (&kibnal_data.kib_global_lock);
 449
 450         for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) {
 451
 452                 list_for_each (ptmp, &kibnal_data.kib_peers[i]) {
 453
 454                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 455                         LASSERT (peer->ibp_persistence != 0 ||
 456                                  peer->ibp_connecting != 0 ||
 457                                  !list_empty (&peer->ibp_conns));
 458
 459                         if (index-- > 0)
 460                                 continue;
 461
 462                         *nidp = peer->ibp_nid;
 463                         *persistencep = peer->ibp_persistence;
 464
 465                         read_unlock (&kibnal_data.kib_global_lock);
 466                         return (0);
 467                 }
 468         }
 469
 470         read_unlock (&kibnal_data.kib_global_lock);
 471         return (-ENOENT);
 472 }
 473
 474 static int
 475 kibnal_add_persistent_peer (ptl_nid_t nid)
 476 {
 477         unsigned long      flags;
 478         kib_peer_t        *peer;
 479         kib_peer_t        *peer2;
 480
 481         if (nid == PTL_NID_ANY)
 482                 return (-EINVAL);
 483
 484         peer = kibnal_create_peer (nid);
 485         if (peer == NULL)
 486                 return (-ENOMEM);
 487
 488         write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
 489
 490         peer2 = kibnal_find_peer_locked (nid);
 491         if (peer2 != NULL) {
 492                 kib_peer_decref (peer);
 493                 peer = peer2;
 494         } else {
 495                 /* peer table takes existing ref on peer */
 496                 list_add_tail (&peer->ibp_list,
 497                                kibnal_nid2peerlist (nid));
 498         }
 499
 500         peer->ibp_persistence++;
 501
 502         write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
 503         return (0);
 504 }
 505
 506 static void
 507 kibnal_del_peer_locked (kib_peer_t *peer, int single_share)
 508 {
 509         struct list_head *ctmp;
 510         struct list_head *cnxt;
 511         kib_conn_t       *conn;
 512
 513         if (!single_share)
 514                 peer->ibp_persistence = 0;
 515         else if (peer->ibp_persistence > 0)
 516                 peer->ibp_persistence--;
 517
 518         if (peer->ibp_persistence != 0)
 519                 return;
 520
 521         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 522                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 523
 524                 kibnal_close_conn_locked (conn, 0);
 525         }
 526
 527         /* NB peer unlinks itself when last conn is closed */
 528 }
 529
 530 int
 531 kibnal_del_peer (ptl_nid_t nid, int single_share)
 532 {
 533         unsigned long      flags;
 534         struct list_head  *ptmp;
 535         struct list_head  *pnxt;
 536         kib_peer_t        *peer;
 537         int                lo;
 538         int                hi;
 539         int                i;
 540         int                rc = -ENOENT;
 541
 542         write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
 543
 544         if (nid != PTL_NID_ANY)
 545                 lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers;
 546         else {
 547                 lo = 0;
 548                 hi = kibnal_data.kib_peer_hash_size - 1;
 549         }
 550
 551         for (i = lo; i <= hi; i++) {
 552                 list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) {
 553                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 554                         LASSERT (peer->ibp_persistence != 0 ||
 555                                  peer->ibp_connecting != 0 ||
 556                                  !list_empty (&peer->ibp_conns));
 557
 558                         if (!(nid == PTL_NID_ANY || peer->ibp_nid == nid))
 559                                 continue;
 560
 561                         kibnal_del_peer_locked (peer, single_share);
 562                         rc = 0;         /* matched something */
 563
 564                         if (single_share)
 565                                 goto out;
 566                 }
 567         }
 568  out:
 569         write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
 570
 571         return (rc);
 572 }
 573
 574 static kib_conn_t *
 575 kibnal_get_conn_by_idx (int index)
 576 {
 577         kib_peer_t        *peer;
 578         struct list_head  *ptmp;
 579         kib_conn_t        *conn;
 580         struct list_head  *ctmp;
 581         int                i;
 582
 583         read_lock (&kibnal_data.kib_global_lock);
 584
 585         for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) {
 586                 list_for_each (ptmp, &kibnal_data.kib_peers[i]) {
 587
 588                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 589                         LASSERT (peer->ibp_persistence > 0 ||
 590                                  peer->ibp_connecting != 0 ||
 591                                  !list_empty (&peer->ibp_conns));
 592
 593                         list_for_each (ctmp, &peer->ibp_conns) {
 594                                 if (index-- > 0)
 595                                         continue;
 596
 597                                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 598                                 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
 599                                        conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
 600                                        atomic_read (&conn->ibc_refcount));
 601                                 atomic_inc (&conn->ibc_refcount);
 602                                 read_unlock (&kibnal_data.kib_global_lock);
 603                                 return (conn);
 604                         }
 605                 }
 606         }
 607
 608         read_unlock (&kibnal_data.kib_global_lock);
 609         return (NULL);
 610 }
 611
 612 kib_conn_t *
 613 kibnal_create_conn (void)
 614 {
 615         kib_conn_t  *conn;
 616         int          i;
 617         __u64        vaddr = 0;
 618         __u64        vaddr_base;
 619         int          page_offset;
 620         int          ipage;
 621         int          rc;
 622         FSTATUS      frc;
 623         union {
 624                 IB_QP_ATTRIBUTES_CREATE    qp_create;
 625                 IB_QP_ATTRIBUTES_MODIFY    qp_attr;
 626         } params;
 627
 628         PORTAL_ALLOC (conn, sizeof (*conn));
 629         if (conn == NULL) {
 630                 CERROR ("Can't allocate connection\n");
 631                 return (NULL);
 632         }
 633
 634         /* zero flags, NULL pointers etc... */
 635         memset (conn, 0, sizeof (*conn));
 636
 637         INIT_LIST_HEAD (&conn->ibc_tx_queue);
 638         INIT_LIST_HEAD (&conn->ibc_active_txs);
 639         spin_lock_init (&conn->ibc_lock);
 640
 641         atomic_inc (&kibnal_data.kib_nconns);
 642         /* well not really, but I call destroy() on failure, which decrements */
 643
 644         PORTAL_ALLOC (conn->ibc_rxs, IBNAL_RX_MSGS * sizeof (kib_rx_t));
 645         if (conn->ibc_rxs == NULL)
 646                 goto failed;
 647         memset (conn->ibc_rxs, 0, IBNAL_RX_MSGS * sizeof(kib_rx_t));
 648
 649         rc = kibnal_alloc_pages(&conn->ibc_rx_pages, IBNAL_RX_MSG_PAGES, 1);
 650         if (rc != 0)
 651                 goto failed;
 652
 653         vaddr_base = vaddr = conn->ibc_rx_pages->ibp_vaddr;
 654
 655         for (i = ipage = page_offset = 0; i < IBNAL_RX_MSGS; i++) {
 656                 struct page *page = conn->ibc_rx_pages->ibp_pages[ipage];
 657                 kib_rx_t   *rx = &conn->ibc_rxs[i];
 658
 659                 rx->rx_conn = conn;
 660                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) +
 661                              page_offset);
 662
 663                 if (kibnal_whole_mem())
 664                         rx->rx_vaddr = kibnal_page2phys(page) +
 665                                        page_offset +
 666                                        kibnal_data.kib_md.md_addr;
 667                 else
 668                         rx->rx_vaddr = vaddr;
 669
 670                 vaddr += IBNAL_MSG_SIZE;
 671                 LASSERT (vaddr <= vaddr_base + IBNAL_RX_MSG_BYTES);
 672
 673                 page_offset += IBNAL_MSG_SIZE;
 674                 LASSERT (page_offset <= PAGE_SIZE);
 675
 676                 if (page_offset == PAGE_SIZE) {
 677                         page_offset = 0;
 678                         ipage++;
 679                         LASSERT (ipage <= IBNAL_RX_MSG_PAGES);
 680                 }
 681         }
 682
 683         params.qp_create = (IB_QP_ATTRIBUTES_CREATE) {
 684                 .Type                    = QPTypeReliableConnected,
 685                 .SendQDepth              = IBNAL_TX_MAX_SG *
 686                                            IBNAL_MSG_QUEUE_SIZE,
 687                 .RecvQDepth              = IBNAL_MSG_QUEUE_SIZE,
 688                 .SendDSListDepth         = 1,
 689                 .RecvDSListDepth         = 1,
 690                 .SendCQHandle            = kibnal_data.kib_cq,
 691                 .RecvCQHandle            = kibnal_data.kib_cq,
 692                 .PDHandle                = kibnal_data.kib_pd,
 693                 .SendSignaledCompletions = TRUE,
 694         };
 695         frc = iibt_qp_create(kibnal_data.kib_hca, &params.qp_create, NULL,
 696                              &conn->ibc_qp, &conn->ibc_qp_attrs);
 697         if (rc != 0) {
 698                 CERROR ("Failed to create queue pair: %d\n", rc);
 699                 goto failed;
 700         }
 701
 702         /* Mark QP created */
 703         conn->ibc_state = IBNAL_CONN_INIT_QP;
 704
 705         params.qp_attr = (IB_QP_ATTRIBUTES_MODIFY) {
 706                 .RequestState             = QPStateInit,
 707                 .Attrs                    = (IB_QP_ATTR_PORTGUID |
 708                                              IB_QP_ATTR_PKEYINDEX |
 709                                              IB_QP_ATTR_ACCESSCONTROL),
 710                 .PortGUID                 = kibnal_data.kib_port_guid,
 711                 .PkeyIndex                = 0,
 712                 .AccessControl = {
 713                         .s = {
 714                                 .RdmaWrite = 1,
 715                                 .RdmaRead  = 1,
 716                         },
 717                 },
 718         };
 719         rc = iibt_qp_modify(conn->ibc_qp, &params.qp_attr, NULL);
 720         if (rc != 0) {
 721                 CERROR ("Failed to modify queue pair: %d\n", rc);
 722                 goto failed;
 723         }
 724
 725         /* 1 ref for caller */
 726         atomic_set (&conn->ibc_refcount, 1);
 727         return (conn);
 728
 729  failed:
 730         kibnal_destroy_conn (conn);
 731         return (NULL);
 732 }
 733
 734 void
 735 kibnal_destroy_conn (kib_conn_t *conn)
 736 {
 737         int    rc;
 738         FSTATUS frc;
 739
 740         CDEBUG (D_NET, "connection %p\n", conn);
 741
 742         LASSERT (atomic_read (&conn->ibc_refcount) == 0);
 743         LASSERT (list_empty(&conn->ibc_tx_queue));
 744         LASSERT (list_empty(&conn->ibc_active_txs));
 745         LASSERT (conn->ibc_nsends_posted == 0);
 746         LASSERT (conn->ibc_connreq == NULL);
 747
 748         switch (conn->ibc_state) {
 749         case IBNAL_CONN_DISCONNECTED:
 750                 /* called after connection sequence initiated */
 751                 /* fall through */
 752
 753         case IBNAL_CONN_INIT_QP:
 754                 /* _destroy includes an implicit Reset of the QP which
 755                  * discards posted work */
 756                 rc = iibt_qp_destroy(conn->ibc_qp);
 757                 if (rc != 0)
 758                         CERROR("Can't destroy QP: %d\n", rc);
 759                 /* fall through */
 760
 761         case IBNAL_CONN_INIT_NOTHING:
 762                 break;
 763
 764         default:
 765                 LASSERT (0);
 766         }
 767
 768         if (conn->ibc_cep != NULL) {
 769                 frc = iibt_cm_destroy_cep(conn->ibc_cep);
 770                 if (frc != 0)
 771                         CERROR("Can't destroy CEP %p: %d\n", conn->ibc_cep,
 772                                frc);
 773         }
 774
 775         if (conn->ibc_rx_pages != NULL)
 776                 kibnal_free_pages(conn->ibc_rx_pages);
 777
 778         if (conn->ibc_rxs != NULL)
 779                 PORTAL_FREE(conn->ibc_rxs,
 780                             IBNAL_RX_MSGS * sizeof(kib_rx_t));
 781
 782         if (conn->ibc_peer != NULL)
 783                 kib_peer_decref(conn->ibc_peer);
 784
 785         PORTAL_FREE(conn, sizeof (*conn));
 786
 787         atomic_dec(&kibnal_data.kib_nconns);
 788
 789         if (atomic_read (&kibnal_data.kib_nconns) == 0 &&
 790             kibnal_data.kib_shutdown) {
 791                 /* I just nuked the last connection on shutdown; wake up
 792                  * everyone so they can exit. */
 793                 wake_up_all(&kibnal_data.kib_sched_waitq);
 794                 wake_up_all(&kibnal_data.kib_connd_waitq);
 795         }
 796 }
 797
 798 void
 799 kibnal_put_conn (kib_conn_t *conn)
 800 {
 801         unsigned long flags;
 802
 803         CDEBUG (D_NET, "putting conn[%p] state %d -> "LPX64" (%d)\n",
 804                 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
 805                 atomic_read (&conn->ibc_refcount));
 806
 807         LASSERT (atomic_read (&conn->ibc_refcount) > 0);
 808         if (!atomic_dec_and_test (&conn->ibc_refcount))
 809                 return;
 810
 811         /* must disconnect before dropping the final ref */
 812         LASSERT (conn->ibc_state == IBNAL_CONN_DISCONNECTED);
 813
 814         spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags);
 815
 816         list_add (&conn->ibc_list, &kibnal_data.kib_connd_conns);
 817         wake_up (&kibnal_data.kib_connd_waitq);
 818
 819         spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
 820 }
 821
 822 static int
 823 kibnal_close_peer_conns_locked (kib_peer_t *peer, int why)
 824 {
 825         kib_conn_t         *conn;
 826         struct list_head   *ctmp;
 827         struct list_head   *cnxt;
 828         int                 count = 0;
 829
 830         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 831                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 832
 833                 count++;
 834                 kibnal_close_conn_locked (conn, why);
 835         }
 836
 837         return (count);
 838 }
 839
 840 int
 841 kibnal_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation)
 842 {
 843         kib_conn_t         *conn;
 844         struct list_head   *ctmp;
 845         struct list_head   *cnxt;
 846         int                 count = 0;
 847
 848         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 849                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 850
 851                 if (conn->ibc_incarnation == incarnation)
 852                         continue;
 853
 854                 CDEBUG(D_NET, "Closing stale conn nid:"LPX64" incarnation:"LPX64"("LPX64")\n",
 855                        peer->ibp_nid, conn->ibc_incarnation, incarnation);
 856
 857                 count++;
 858                 kibnal_close_conn_locked (conn, -ESTALE);
 859         }
 860
 861         return (count);
 862 }
 863
 864 static int
 865 kibnal_close_matching_conns (ptl_nid_t nid)
 866 {
 867         unsigned long       flags;
 868         kib_peer_t         *peer;
 869         struct list_head   *ptmp;
 870         struct list_head   *pnxt;
 871         int                 lo;
 872         int                 hi;
 873         int                 i;
 874         int                 count = 0;
 875
 876         write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
 877
 878         if (nid != PTL_NID_ANY)
 879                 lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers;
 880         else {
 881                 lo = 0;
 882                 hi = kibnal_data.kib_peer_hash_size - 1;
 883         }
 884
 885         for (i = lo; i <= hi; i++) {
 886                 list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) {
 887
 888                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 889                         LASSERT (peer->ibp_persistence != 0 ||
 890                                  peer->ibp_connecting != 0 ||
 891                                  !list_empty (&peer->ibp_conns));
 892
 893                         if (!(nid == PTL_NID_ANY || nid == peer->ibp_nid))
 894                                 continue;
 895
 896                         count += kibnal_close_peer_conns_locked (peer, 0);
 897                 }
 898         }
 899
 900         write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
 901
 902         /* wildcards always succeed */
 903         if (nid == PTL_NID_ANY)
 904                 return (0);
 905
 906         return (count == 0 ? -ENOENT : 0);
 907 }
 908
 909 static int
 910 kibnal_cmd(struct portals_cfg *pcfg, void * private)
 911 {
 912         int rc = -EINVAL;
 913         ENTRY;
 914
 915         LASSERT (pcfg != NULL);
 916
 917         switch(pcfg->pcfg_command) {
 918         case NAL_CMD_GET_PEER: {
 919                 ptl_nid_t   nid = 0;
 920                 int         share_count = 0;
 921
 922                 rc = kibnal_get_peer_info(pcfg->pcfg_count,
 923                                           &nid, &share_count);
 924                 pcfg->pcfg_nid   = nid;
 925                 pcfg->pcfg_size  = 0;
 926                 pcfg->pcfg_id    = 0;
 927                 pcfg->pcfg_misc  = 0;
 928                 pcfg->pcfg_count = 0;
 929                 pcfg->pcfg_wait  = share_count;
 930                 break;
 931         }
 932         case NAL_CMD_ADD_PEER: {
 933                 rc = kibnal_add_persistent_peer (pcfg->pcfg_nid);
 934                 break;
 935         }
 936         case NAL_CMD_DEL_PEER: {
 937                 rc = kibnal_del_peer (pcfg->pcfg_nid,
 938                                        /* flags == single_share */
 939                                        pcfg->pcfg_flags != 0);
 940                 break;
 941         }
 942         case NAL_CMD_GET_CONN: {
 943                 kib_conn_t *conn = kibnal_get_conn_by_idx (pcfg->pcfg_count);
 944
 945                 if (conn == NULL)
 946                         rc = -ENOENT;
 947                 else {
 948                         rc = 0;
 949                         pcfg->pcfg_nid   = conn->ibc_peer->ibp_nid;
 950                         pcfg->pcfg_id    = 0;
 951                         pcfg->pcfg_misc  = 0;
 952                         pcfg->pcfg_flags = 0;
 953                         kibnal_put_conn (conn);
 954                 }
 955                 break;
 956         }
 957         case NAL_CMD_CLOSE_CONNECTION: {
 958                 rc = kibnal_close_matching_conns (pcfg->pcfg_nid);
 959                 break;
 960         }
 961         case NAL_CMD_REGISTER_MYNID: {
 962                 if (pcfg->pcfg_nid == PTL_NID_ANY)
 963                         rc = -EINVAL;
 964                 else
 965                         rc = kibnal_set_mynid (pcfg->pcfg_nid);
 966                 break;
 967         }
 968         }
 969
 970         RETURN(rc);
 971 }
 972
 973 void
 974 kibnal_free_pages (kib_pages_t *p)
 975 {
 976         int     npages = p->ibp_npages;
 977         int     rc;
 978         int     i;
 979
 980         if (p->ibp_mapped) {
 981                 rc = iibt_deregister_memory(p->ibp_handle);
 982                 if (rc != 0)
 983                         CERROR ("Deregister error: %d\n", rc);
 984         }
 985
 986         for (i = 0; i < npages; i++)
 987                 if (p->ibp_pages[i] != NULL)
 988                         __free_page(p->ibp_pages[i]);
 989
 990         PORTAL_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
 991 }
 992
 993 int
 994 kibnal_alloc_pages (kib_pages_t **pp, int npages, int allow_write)
 995 {
 996         kib_pages_t                *p;
 997         __u64                      *phys_pages;
 998         int                         i;
 999         FSTATUS                     frc;
1000         IB_ACCESS_CONTROL           access;
1001
1002         memset(&access, 0, sizeof(access));
1003         access.s.MWBindable = 1;
1004         access.s.LocalWrite = 1;
1005         access.s.RdmaRead = 1;
1006         access.s.RdmaWrite = 1;
1007
1008         PORTAL_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1009         if (p == NULL) {
1010                 CERROR ("Can't allocate buffer %d\n", npages);
1011                 return (-ENOMEM);
1012         }
1013
1014         memset (p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1015         p->ibp_npages = npages;
1016
1017         for (i = 0; i < npages; i++) {
1018                 p->ibp_pages[i] = alloc_page (GFP_KERNEL);
1019                 if (p->ibp_pages[i] == NULL) {
1020                         CERROR ("Can't allocate page %d of %d\n", i, npages);
1021                         kibnal_free_pages(p);
1022                         return (-ENOMEM);
1023                 }
1024         }
1025
1026         if (kibnal_whole_mem())
1027                 goto out;
1028
1029         PORTAL_ALLOC(phys_pages, npages * sizeof(*phys_pages));
1030         if (phys_pages == NULL) {
1031                 CERROR ("Can't allocate physarray for %d pages\n", npages);
1032                 /* XXX free ibp_pages? */
1033                 kibnal_free_pages(p);
1034                 return (-ENOMEM);
1035         }
1036
1037         /* if we were using the _contig_ registration variant we would have
1038          * an array of PhysAddr/Length pairs, but the discontiguous variant
1039          * just takes the PhysAddr */
1040         for (i = 0; i < npages; i++)
1041                 phys_pages[i] = kibnal_page2phys(p->ibp_pages[i]);
1042
1043         frc = iibt_register_physical_memory(kibnal_data.kib_hca,
1044                                             0,          /* requested vaddr */
1045                                             phys_pages, npages,
1046                                             0,          /* offset */
1047                                             kibnal_data.kib_pd,
1048                                             access,
1049                                             &p->ibp_handle, &p->ibp_vaddr,
1050                                             &p->ibp_lkey, &p->ibp_rkey);
1051
1052         PORTAL_FREE(phys_pages, npages * sizeof(*phys_pages));
1053
1054         if (frc != FSUCCESS) {
1055                 CERROR ("Error %d mapping %d pages\n", frc, npages);
1056                 kibnal_free_pages(p);
1057                 return (-ENOMEM);
1058         }
1059
1060         CDEBUG(D_NET, "registered %d pages; handle: %p vaddr "LPX64" "
1061                       "lkey %x rkey %x\n", npages, p->ibp_handle,
1062                       p->ibp_vaddr, p->ibp_lkey, p->ibp_rkey);
1063
1064         p->ibp_mapped = 1;
1065 out:
1066         *pp = p;
1067         return (0);
1068 }
1069
1070 static int
1071 kibnal_setup_tx_descs (void)
1072 {
1073         int           ipage = 0;
1074         int           page_offset = 0;
1075         __u64         vaddr;
1076         __u64         vaddr_base;
1077         struct page  *page;
1078         kib_tx_t     *tx;
1079         int           i;
1080         int           rc;
1081
1082         /* pre-mapped messages are not bigger than 1 page */
1083         LASSERT (IBNAL_MSG_SIZE <= PAGE_SIZE);
1084
1085         /* No fancy arithmetic when we do the buffer calculations */
1086         LASSERT (PAGE_SIZE % IBNAL_MSG_SIZE == 0);
1087
1088         rc = kibnal_alloc_pages(&kibnal_data.kib_tx_pages, IBNAL_TX_MSG_PAGES,
1089                                 0);
1090         if (rc != 0)
1091                 return (rc);
1092
1093         /* ignored for the whole_mem case */
1094         vaddr = vaddr_base = kibnal_data.kib_tx_pages->ibp_vaddr;
1095
1096         for (i = 0; i < IBNAL_TX_MSGS; i++) {
1097                 page = kibnal_data.kib_tx_pages->ibp_pages[ipage];
1098                 tx = &kibnal_data.kib_tx_descs[i];
1099
1100                 memset (tx, 0, sizeof(*tx));    /* zero flags etc */
1101
1102                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1103                                             page_offset);
1104
1105                 if (kibnal_whole_mem())
1106                         tx->tx_vaddr = kibnal_page2phys(page) +
1107                                        page_offset +
1108                                        kibnal_data.kib_md.md_addr;
1109                 else
1110                         tx->tx_vaddr = vaddr;
1111
1112                 tx->tx_isnblk = (i >= IBNAL_NTX);
1113                 tx->tx_mapped = KIB_TX_UNMAPPED;
1114
1115                 CDEBUG(D_NET, "Tx[%d] %p->%p - "LPX64"\n",
1116                        i, tx, tx->tx_msg, tx->tx_vaddr);
1117
1118                 if (tx->tx_isnblk)
1119                         list_add (&tx->tx_list,
1120                                   &kibnal_data.kib_idle_nblk_txs);
1121                 else
1122                         list_add (&tx->tx_list,
1123                                   &kibnal_data.kib_idle_txs);
1124
1125                 vaddr += IBNAL_MSG_SIZE;
1126                 LASSERT (vaddr <= vaddr_base + IBNAL_TX_MSG_BYTES);
1127
1128                 page_offset += IBNAL_MSG_SIZE;
1129                 LASSERT (page_offset <= PAGE_SIZE);
1130
1131                 if (page_offset == PAGE_SIZE) {
1132                         page_offset = 0;
1133                         ipage++;
1134                         LASSERT (ipage <= IBNAL_TX_MSG_PAGES);
1135                 }
1136         }
1137
1138         return (0);
1139 }
1140
1141 static void
1142 kibnal_api_shutdown (nal_t *nal)
1143 {
1144         int   i;
1145         int   rc;
1146
1147         if (nal->nal_refct != 0) {
1148                 /* This module got the first ref */
1149                 PORTAL_MODULE_UNUSE;
1150                 return;
1151         }
1152
1153         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
1154                atomic_read (&portal_kmemory));
1155
1156         LASSERT(nal == &kibnal_api);
1157
1158         switch (kibnal_data.kib_init) {
1159         default:
1160                 CERROR ("Unexpected state %d\n", kibnal_data.kib_init);
1161                 LBUG();
1162
1163         case IBNAL_INIT_ALL:
1164                 /* stop calls to nal_cmd */
1165                 libcfs_nal_cmd_unregister(IIBNAL);
1166                 /* No new peers */
1167
1168                 /* resetting my NID to unadvertises me, removes my
1169                  * listener and nukes all current peers */
1170                 kibnal_set_mynid (PTL_NID_ANY);
1171
1172                 /* Wait for all peer state to clean up (crazy) */
1173                 i = 2;
1174                 while (atomic_read (&kibnal_data.kib_npeers) != 0) {
1175                         i++;
1176                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1177                                "waiting for %d peers to disconnect (can take a few seconds)\n",
1178                                atomic_read (&kibnal_data.kib_npeers));
1179                         set_current_state (TASK_UNINTERRUPTIBLE);
1180                         schedule_timeout (HZ);
1181                 }
1182                 /* fall through */
1183
1184         case IBNAL_INIT_CQ:
1185                 rc = iibt_cq_destroy(kibnal_data.kib_cq);
1186                 if (rc != 0)
1187                         CERROR ("Destroy CQ error: %d\n", rc);
1188                 /* fall through */
1189
1190         case IBNAL_INIT_TXD:
1191                 kibnal_free_pages (kibnal_data.kib_tx_pages);
1192                 /* fall through */
1193
1194         case IBNAL_INIT_MR:
1195                 if (kibnal_data.kib_md.md_handle != NULL) {
1196                         rc = iibt_deregister_memory(kibnal_data.kib_md.md_handle);
1197                         if (rc != FSUCCESS)
1198                                 CERROR ("Deregister memory: %d\n", rc);
1199                 }
1200                 /* fall through */
1201
1202 #if IBNAL_FMR
1203         case IBNAL_INIT_FMR:
1204                 rc = ib_fmr_pool_destroy (kibnal_data.kib_fmr_pool);
1205                 if (rc != 0)
1206                         CERROR ("Destroy FMR pool error: %d\n", rc);
1207                 /* fall through */
1208 #endif
1209         case IBNAL_INIT_PD:
1210                 rc = iibt_pd_free(kibnal_data.kib_pd);
1211                 if (rc != 0)
1212                         CERROR ("Destroy PD error: %d\n", rc);
1213                 /* fall through */
1214
1215         case IBNAL_INIT_SD:
1216                 rc = iibt_sd_deregister(kibnal_data.kib_sd);
1217                 if (rc != 0)
1218                         CERROR ("Deregister SD error: %d\n", rc);
1219                 /* fall through */
1220
1221         case IBNAL_INIT_PORT:
1222                 /* XXX ??? */
1223                 /* fall through */
1224
1225         case IBNAL_INIT_PORTATTRS:
1226                 PORTAL_FREE(kibnal_data.kib_hca_attrs.PortAttributesList,
1227                             kibnal_data.kib_hca_attrs.PortAttributesListSize);
1228                 /* fall through */
1229
1230         case IBNAL_INIT_HCA:
1231                 rc = iibt_close_hca(kibnal_data.kib_hca);
1232                 if (rc != 0)
1233                         CERROR ("Close HCA  error: %d\n", rc);
1234                 /* fall through */
1235
1236         case IBNAL_INIT_LIB:
1237                 lib_fini(&kibnal_lib);
1238                 /* fall through */
1239
1240         case IBNAL_INIT_DATA:
1241                 /* Module refcount only gets to zero when all peers
1242                  * have been closed so all lists must be empty */
1243                 LASSERT (atomic_read (&kibnal_data.kib_npeers) == 0);
1244                 LASSERT (kibnal_data.kib_peers != NULL);
1245                 for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) {
1246                         LASSERT (list_empty (&kibnal_data.kib_peers[i]));
1247                 }
1248                 LASSERT (atomic_read (&kibnal_data.kib_nconns) == 0);
1249                 LASSERT (list_empty (&kibnal_data.kib_sched_rxq));
1250                 LASSERT (list_empty (&kibnal_data.kib_sched_txq));
1251                 LASSERT (list_empty (&kibnal_data.kib_connd_conns));
1252                 LASSERT (list_empty (&kibnal_data.kib_connd_peers));
1253
1254                 /* flag threads to terminate; wake and wait for them to die */
1255                 kibnal_data.kib_shutdown = 1;
1256                 wake_up_all (&kibnal_data.kib_sched_waitq);
1257                 wake_up_all (&kibnal_data.kib_connd_waitq);
1258
1259                 i = 2;
1260                 while (atomic_read (&kibnal_data.kib_nthreads) != 0) {
1261                         i++;
1262                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1263                                "Waiting for %d threads to terminate\n",
1264                                atomic_read (&kibnal_data.kib_nthreads));
1265                         set_current_state (TASK_INTERRUPTIBLE);
1266                         schedule_timeout (HZ);
1267                 }
1268                 /* fall through */
1269
1270         case IBNAL_INIT_NOTHING:
1271                 break;
1272         }
1273
1274         if (kibnal_data.kib_tx_descs != NULL)
1275                 PORTAL_FREE (kibnal_data.kib_tx_descs,
1276                              IBNAL_TX_MSGS * sizeof(kib_tx_t));
1277
1278         if (kibnal_data.kib_peers != NULL)
1279                 PORTAL_FREE (kibnal_data.kib_peers,
1280                              sizeof (struct list_head) *
1281                              kibnal_data.kib_peer_hash_size);
1282
1283         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
1284                atomic_read (&portal_kmemory));
1285         printk(KERN_INFO "Lustre: Infinicon IB NAL unloaded (final mem %d)\n",
1286                atomic_read(&portal_kmemory));
1287
1288         kibnal_data.kib_init = IBNAL_INIT_NOTHING;
1289 }
1290
1291 #define roundup_power(val, power) \
1292         ( (val + (__u64)(power - 1)) & ~((__u64)(power - 1)) )
1293
1294 /* this isn't very portable or sturdy in the face of funny mem/bus configs */
1295 static __u64 max_phys_mem(IB_CA_ATTRIBUTES *ca_attr)
1296 {
1297         struct sysinfo si;
1298         __u64 ret;
1299
1300         /* XXX we don't bother with first-gen cards */
1301         if (ca_attr->VendorId == 0xd0b7 && ca_attr->DeviceId == 0x3101)
1302                 return 0ULL;
1303
1304         si_meminfo(&si);
1305         ret = (__u64)max(si.totalram, max_mapnr) * si.mem_unit;
1306         return roundup_power(ret, 128 * 1024 * 1024);
1307 }
1308 #undef roundup_power
1309
1310 static int
1311 kibnal_api_startup (nal_t *nal, ptl_pid_t requested_pid,
1312                      ptl_ni_limits_t *requested_limits,
1313                      ptl_ni_limits_t *actual_limits)
1314 {
1315         ptl_process_id_t    process_id;
1316         int                 pkmem = atomic_read(&portal_kmemory);
1317         IB_PORT_ATTRIBUTES *pattr;
1318         FSTATUS             frc;
1319         int                 rc;
1320         int                 n;
1321         int                 i;
1322
1323         LASSERT (nal == &kibnal_api);
1324
1325         if (nal->nal_refct != 0) {
1326                 if (actual_limits != NULL)
1327                         *actual_limits = kibnal_lib.libnal_ni.ni_actual_limits;
1328                 /* This module got the first ref */
1329                 PORTAL_MODULE_USE;
1330                 return (PTL_OK);
1331         }
1332
1333         LASSERT (kibnal_data.kib_init == IBNAL_INIT_NOTHING);
1334
1335         frc = IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2,
1336                                        &kibnal_data.kib_interfaces);
1337         if (frc != FSUCCESS) {
1338                 CERROR("IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2) = %d\n",
1339                         frc);
1340                 return -ENOSYS;
1341         }
1342
1343         init_MUTEX (&kibnal_data.kib_nid_mutex);
1344         init_MUTEX_LOCKED (&kibnal_data.kib_nid_signal);
1345         kibnal_data.kib_nid = PTL_NID_ANY;
1346
1347         rwlock_init(&kibnal_data.kib_global_lock);
1348
1349         kibnal_data.kib_peer_hash_size = IBNAL_PEER_HASH_SIZE;
1350         PORTAL_ALLOC (kibnal_data.kib_peers,
1351                       sizeof (struct list_head) * kibnal_data.kib_peer_hash_size);
1352         if (kibnal_data.kib_peers == NULL) {
1353                 goto failed;
1354         }
1355         for (i = 0; i < kibnal_data.kib_peer_hash_size; i++)
1356                 INIT_LIST_HEAD(&kibnal_data.kib_peers[i]);
1357
1358         spin_lock_init (&kibnal_data.kib_connd_lock);
1359         INIT_LIST_HEAD (&kibnal_data.kib_connd_peers);
1360         INIT_LIST_HEAD (&kibnal_data.kib_connd_conns);
1361         init_waitqueue_head (&kibnal_data.kib_connd_waitq);
1362
1363         spin_lock_init (&kibnal_data.kib_sched_lock);
1364         INIT_LIST_HEAD (&kibnal_data.kib_sched_txq);
1365         INIT_LIST_HEAD (&kibnal_data.kib_sched_rxq);
1366         init_waitqueue_head (&kibnal_data.kib_sched_waitq);
1367
1368         spin_lock_init (&kibnal_data.kib_tx_lock);
1369         INIT_LIST_HEAD (&kibnal_data.kib_idle_txs);
1370         INIT_LIST_HEAD (&kibnal_data.kib_idle_nblk_txs);
1371         init_waitqueue_head(&kibnal_data.kib_idle_tx_waitq);
1372
1373         PORTAL_ALLOC (kibnal_data.kib_tx_descs,
1374                       IBNAL_TX_MSGS * sizeof(kib_tx_t));
1375         if (kibnal_data.kib_tx_descs == NULL) {
1376                 CERROR ("Can't allocate tx descs\n");
1377                 goto failed;
1378         }
1379
1380         /* lists/ptrs/locks initialised */
1381         kibnal_data.kib_init = IBNAL_INIT_DATA;
1382         /*****************************************************/
1383
1384         process_id.pid = 0;
1385         process_id.nid = kibnal_data.kib_nid;
1386
1387         rc = lib_init(&kibnal_lib, nal, process_id,
1388                       requested_limits, actual_limits);
1389         if (rc != PTL_OK) {
1390                 CERROR("lib_init failed: error %d\n", rc);
1391                 goto failed;
1392         }
1393
1394         /* lib interface initialised */
1395         kibnal_data.kib_init = IBNAL_INIT_LIB;
1396         /*****************************************************/
1397
1398         for (i = 0; i < IBNAL_N_SCHED; i++) {
1399                 rc = kibnal_thread_start (kibnal_scheduler, (void *)i);
1400                 if (rc != 0) {
1401                         CERROR("Can't spawn iibnal scheduler[%d]: %d\n",
1402                                i, rc);
1403                         goto failed;
1404                 }
1405         }
1406
1407         rc = kibnal_thread_start (kibnal_connd, NULL);
1408         if (rc != 0) {
1409                 CERROR ("Can't spawn iibnal connd: %d\n", rc);
1410                 goto failed;
1411         }
1412
1413         n = sizeof(kibnal_data.kib_hca_guids) /
1414             sizeof(kibnal_data.kib_hca_guids[0]);
1415         frc = iibt_get_hca_guids(&n, kibnal_data.kib_hca_guids);
1416         if (frc != FSUCCESS) {
1417                 CERROR ("Can't get channel adapter guids: %d\n", frc);
1418                 goto failed;
1419         }
1420         if (n == 0) {
1421                 CERROR ("No channel adapters found\n");
1422                 goto failed;
1423         }
1424
1425         /* Infinicon has per-HCA rather than per CQ completion handlers */
1426         frc = iibt_open_hca(kibnal_data.kib_hca_guids[0],
1427                             kibnal_ca_callback,
1428                             kibnal_ca_async_callback,
1429                             &kibnal_data.kib_hca,
1430                             &kibnal_data.kib_hca);
1431         if (frc != FSUCCESS) {
1432                 CERROR ("Can't open CA[0]: %d\n", frc);
1433                 goto failed;
1434         }
1435
1436         /* Channel Adapter opened */
1437         kibnal_data.kib_init = IBNAL_INIT_HCA;
1438         /*****************************************************/
1439
1440         kibnal_data.kib_hca_attrs.PortAttributesList = NULL;
1441         kibnal_data.kib_hca_attrs.PortAttributesListSize = 0;
1442         frc = iibt_query_hca(kibnal_data.kib_hca,
1443                              &kibnal_data.kib_hca_attrs, NULL);
1444         if (frc != FSUCCESS) {
1445                 CERROR ("Can't size port attrs: %d\n", frc);
1446                 goto failed;
1447         }
1448
1449         PORTAL_ALLOC(kibnal_data.kib_hca_attrs.PortAttributesList,
1450                      kibnal_data.kib_hca_attrs.PortAttributesListSize);
1451         if (kibnal_data.kib_hca_attrs.PortAttributesList == NULL)
1452                 goto failed;
1453
1454         /* Port attrs allocated */
1455         kibnal_data.kib_init = IBNAL_INIT_PORTATTRS;
1456         /*****************************************************/
1457
1458         frc = iibt_query_hca(kibnal_data.kib_hca, &kibnal_data.kib_hca_attrs,
1459                              NULL);
1460         if (frc != FSUCCESS) {
1461                 CERROR ("Can't get port attrs for CA 0: %d\n", frc);
1462                 goto failed;
1463         }
1464
1465         for (i = 0, pattr = kibnal_data.kib_hca_attrs.PortAttributesList;
1466              pattr != NULL;
1467              i++, pattr = pattr->Next) {
1468                 switch (pattr->PortState) {
1469                 default:
1470                         CERROR("Unexpected port[%d] state %d\n",
1471                                i, pattr->PortState);
1472                         continue;
1473                 case PortStateDown:
1474                         CDEBUG(D_NET, "port[%d] Down\n", i);
1475                         continue;
1476                 case PortStateInit:
1477                         CDEBUG(D_NET, "port[%d] Init\n", i);
1478                         continue;
1479                 case PortStateArmed:
1480                         CDEBUG(D_NET, "port[%d] Armed\n", i);
1481                         continue;
1482
1483                 case PortStateActive:
1484                         CDEBUG(D_NET, "port[%d] Active\n", i);
1485                         kibnal_data.kib_port = i;
1486                         kibnal_data.kib_port_guid = pattr->GUID;
1487                         kibnal_data.kib_port_pkey = pattr->PkeyTable[0];
1488                         break;
1489                 }
1490                 break;
1491         }
1492
1493         if (pattr == NULL) {
1494                 CERROR ("Can't find an active port\n");
1495                 goto failed;
1496         }
1497
1498         CDEBUG(D_NET, "got guid "LPX64"\n", kibnal_data.kib_port_guid);
1499
1500         /* Active port found */
1501         kibnal_data.kib_init = IBNAL_INIT_PORT;
1502         /*****************************************************/
1503
1504         frc = iibt_sd_register(&kibnal_data.kib_sd, NULL);
1505         if (frc != FSUCCESS) {
1506                 CERROR ("Can't register with SD: %d\n", frc);
1507                 goto failed;
1508         }
1509
1510         /* Registered with SD OK */
1511         kibnal_data.kib_init = IBNAL_INIT_SD;
1512         /*****************************************************/
1513
1514         frc = iibt_pd_allocate(kibnal_data.kib_hca, 0, &kibnal_data.kib_pd);
1515         if (frc != FSUCCESS) {
1516                 CERROR ("Can't create PD: %d\n", rc);
1517                 goto failed;
1518         }
1519
1520         /* flag PD initialised */
1521         kibnal_data.kib_init = IBNAL_INIT_PD;
1522         /*****************************************************/
1523
1524 #if IBNAL_FMR
1525         {
1526                 const int pool_size = IBNAL_NTX + IBNAL_NTX_NBLK;
1527                 struct ib_fmr_pool_param params = {
1528                         .max_pages_per_fmr = PTL_MTU/PAGE_SIZE,
1529                         .access            = (IB_ACCESS_LOCAL_WRITE |
1530                                               IB_ACCESS_REMOTE_WRITE |
1531                                               IB_ACCESS_REMOTE_READ),
1532                         .pool_size         = pool_size,
1533                         .dirty_watermark   = (pool_size * 3)/4,
1534                         .flush_function    = NULL,
1535                         .flush_arg         = NULL,
1536                         .cache             = 1,
1537                 };
1538                 rc = ib_fmr_pool_create(kibnal_data.kib_pd, &params,
1539                                         &kibnal_data.kib_fmr_pool);
1540                 if (rc != 0) {
1541                         CERROR ("Can't create FMR pool size %d: %d\n",
1542                                 pool_size, rc);
1543                         goto failed;
1544                 }
1545         }
1546
1547         /* flag FMR pool initialised */
1548         kibnal_data.kib_init = IBNAL_INIT_FMR;
1549 #endif
1550         /*****************************************************/
1551         if (IBNAL_WHOLE_MEM) {
1552                 IB_MR_PHYS_BUFFER phys;
1553                 IB_ACCESS_CONTROL access;
1554                 kib_md_t *md = &kibnal_data.kib_md;
1555
1556                 memset(&access, 0, sizeof(access));
1557                 access.s.MWBindable = 1;
1558                 access.s.LocalWrite = 1;
1559                 access.s.RdmaRead = 1;
1560                 access.s.RdmaWrite = 1;
1561
1562                 phys.PhysAddr = 0;
1563                 phys.Length = max_phys_mem(&kibnal_data.kib_hca_attrs);
1564                 if (phys.Length == 0) {
1565                         CERROR ("couldn't determine the end of phys mem\n");
1566                         goto failed;
1567                 }
1568
1569                 rc = iibt_register_contig_physical_memory(kibnal_data.kib_hca,
1570                                                           0,
1571                                                           &phys, 1,
1572                                                           0,
1573                                                           kibnal_data.kib_pd,
1574                                                           access,
1575                                                           &md->md_handle,
1576                                                           &md->md_addr,
1577                                                           &md->md_lkey,
1578                                                           &md->md_rkey);
1579                 if (rc != FSUCCESS) {
1580                         CERROR("registering physical memory failed: %d\n",
1581                                rc);
1582                         CERROR("falling back to registration per-rdma\n");
1583                         md->md_handle = NULL;
1584                 } else {
1585                         CDEBUG(D_NET, "registered "LPU64" bytes of mem\n",
1586                                phys.Length);
1587                         kibnal_data.kib_init = IBNAL_INIT_MR;
1588                 }
1589         }
1590
1591         /*****************************************************/
1592
1593         rc = kibnal_setup_tx_descs();
1594         if (rc != 0) {
1595                 CERROR ("Can't register tx descs: %d\n", rc);
1596                 goto failed;
1597         }
1598
1599         /* flag TX descs initialised */
1600         kibnal_data.kib_init = IBNAL_INIT_TXD;
1601         /*****************************************************/
1602
1603         {
1604                 uint32 nentries;
1605
1606                 frc = iibt_cq_create(kibnal_data.kib_hca, IBNAL_CQ_ENTRIES,
1607                                      &kibnal_data.kib_cq, &kibnal_data.kib_cq,
1608                                      &nentries);
1609                 if (frc != FSUCCESS) {
1610                         CERROR ("Can't create RX CQ: %d\n", frc);
1611                         goto failed;
1612                 }
1613
1614                 /* flag CQ initialised */
1615                 kibnal_data.kib_init = IBNAL_INIT_CQ;
1616
1617                 if (nentries < IBNAL_CQ_ENTRIES) {
1618                         CERROR ("CQ only has %d entries, need %d\n",
1619                                 nentries, IBNAL_CQ_ENTRIES);
1620                         goto failed;
1621                 }
1622
1623                 rc = iibt_cq_rearm(kibnal_data.kib_cq, CQEventSelNextWC);
1624                 if (rc != 0) {
1625                         CERROR ("Failed to re-arm completion queue: %d\n", rc);
1626                         goto failed;
1627                 }
1628         }
1629
1630         /*****************************************************/
1631
1632         rc = libcfs_nal_cmd_register(IIBNAL, &kibnal_cmd, NULL);
1633         if (rc != 0) {
1634                 CERROR ("Can't initialise command interface (rc = %d)\n", rc);
1635                 goto failed;
1636         }
1637
1638         /* flag everything initialised */
1639         kibnal_data.kib_init = IBNAL_INIT_ALL;
1640         /*****************************************************/
1641
1642         printk(KERN_INFO "Lustre: Infinicon IB NAL loaded "
1643                "(initial mem %d)\n", pkmem);
1644
1645         return (PTL_OK);
1646
1647  failed:
1648         kibnal_api_shutdown (&kibnal_api);
1649         return (PTL_FAIL);
1650 }
1651
1652 void __exit
1653 kibnal_module_fini (void)
1654 {
1655 #ifdef CONFIG_SYSCTL
1656         if (kibnal_tunables.kib_sysctl != NULL)
1657                 unregister_sysctl_table (kibnal_tunables.kib_sysctl);
1658 #endif
1659         PtlNIFini(kibnal_ni);
1660
1661         ptl_unregister_nal(IIBNAL);
1662 }
1663
1664 int __init
1665 kibnal_module_init (void)
1666 {
1667         int    rc;
1668
1669         if (sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN) {
1670                 CERROR("sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN\n");
1671                 return -EINVAL;
1672         }
1673
1674         /* the following must be sizeof(int) for proc_dointvec() */
1675         if (sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)) {
1676                 CERROR("sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)\n");
1677                 return -EINVAL;
1678         }
1679
1680         kibnal_api.nal_ni_init = kibnal_api_startup;
1681         kibnal_api.nal_ni_fini = kibnal_api_shutdown;
1682
1683         /* Initialise dynamic tunables to defaults once only */
1684         kibnal_tunables.kib_io_timeout = IBNAL_IO_TIMEOUT;
1685
1686         rc = ptl_register_nal(IIBNAL, &kibnal_api);
1687         if (rc != PTL_OK) {
1688                 CERROR("Can't register IBNAL: %d\n", rc);
1689                 return (-ENOMEM);               /* or something... */
1690         }
1691
1692         /* Pure gateways want the NAL started up at module load time... */
1693         rc = PtlNIInit(IIBNAL, 0, NULL, NULL, &kibnal_ni);
1694         if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
1695                 ptl_unregister_nal(IIBNAL);
1696                 return (-ENODEV);
1697         }
1698
1699 #ifdef CONFIG_SYSCTL
1700         /* Press on regardless even if registering sysctl doesn't work */
1701         kibnal_tunables.kib_sysctl =
1702                 register_sysctl_table (kibnal_top_ctl_table, 0);
1703 #endif
1704         return (0);
1705 }
1706
1707 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1708 MODULE_DESCRIPTION("Kernel Infinicon IB NAL v0.01");
1709 MODULE_LICENSE("GPL");
1710
1711 module_init(kibnal_module_init);
1712 module_exit(kibnal_module_fini);
1713