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         if (list_empty(&peer->ibp_conns)) {
 522                 kibnal_unlink_peer_locked(peer);
 523         } else {
 524                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 525                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 526
 527                         kibnal_close_conn_locked (conn, 0);
 528                 }
 529                 /* NB peer is no longer persistent; closing its last conn
 530                  * unlinked it. */
 531         }
 532         /* NB peer now unlinked; might even be freed if the peer table had the
 533          * last ref on it. */
 534 }
 535
 536 int
 537 kibnal_del_peer (ptl_nid_t nid, int single_share)
 538 {
 539         unsigned long      flags;
 540         struct list_head  *ptmp;
 541         struct list_head  *pnxt;
 542         kib_peer_t        *peer;
 543         int                lo;
 544         int                hi;
 545         int                i;
 546         int                rc = -ENOENT;
 547
 548         write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
 549
 550         if (nid != PTL_NID_ANY)
 551                 lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers;
 552         else {
 553                 lo = 0;
 554                 hi = kibnal_data.kib_peer_hash_size - 1;
 555         }
 556
 557         for (i = lo; i <= hi; i++) {
 558                 list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) {
 559                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 560                         LASSERT (peer->ibp_persistence != 0 ||
 561                                  peer->ibp_connecting != 0 ||
 562                                  !list_empty (&peer->ibp_conns));
 563
 564                         if (!(nid == PTL_NID_ANY || peer->ibp_nid == nid))
 565                                 continue;
 566
 567                         kibnal_del_peer_locked (peer, single_share);
 568                         rc = 0;         /* matched something */
 569
 570                         if (single_share)
 571                                 goto out;
 572                 }
 573         }
 574  out:
 575         write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
 576
 577         return (rc);
 578 }
 579
 580 static kib_conn_t *
 581 kibnal_get_conn_by_idx (int index)
 582 {
 583         kib_peer_t        *peer;
 584         struct list_head  *ptmp;
 585         kib_conn_t        *conn;
 586         struct list_head  *ctmp;
 587         int                i;
 588
 589         read_lock (&kibnal_data.kib_global_lock);
 590
 591         for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) {
 592                 list_for_each (ptmp, &kibnal_data.kib_peers[i]) {
 593
 594                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 595                         LASSERT (peer->ibp_persistence > 0 ||
 596                                  peer->ibp_connecting != 0 ||
 597                                  !list_empty (&peer->ibp_conns));
 598
 599                         list_for_each (ctmp, &peer->ibp_conns) {
 600                                 if (index-- > 0)
 601                                         continue;
 602
 603                                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 604                                 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
 605                                        conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
 606                                        atomic_read (&conn->ibc_refcount));
 607                                 atomic_inc (&conn->ibc_refcount);
 608                                 read_unlock (&kibnal_data.kib_global_lock);
 609                                 return (conn);
 610                         }
 611                 }
 612         }
 613
 614         read_unlock (&kibnal_data.kib_global_lock);
 615         return (NULL);
 616 }
 617
 618 kib_conn_t *
 619 kibnal_create_conn (void)
 620 {
 621         kib_conn_t  *conn;
 622         int          i;
 623         __u64        vaddr = 0;
 624         __u64        vaddr_base;
 625         int          page_offset;
 626         int          ipage;
 627         int          rc;
 628         FSTATUS      frc;
 629         union {
 630                 IB_QP_ATTRIBUTES_CREATE    qp_create;
 631                 IB_QP_ATTRIBUTES_MODIFY    qp_attr;
 632         } params;
 633
 634         PORTAL_ALLOC (conn, sizeof (*conn));
 635         if (conn == NULL) {
 636                 CERROR ("Can't allocate connection\n");
 637                 return (NULL);
 638         }
 639
 640         /* zero flags, NULL pointers etc... */
 641         memset (conn, 0, sizeof (*conn));
 642
 643         INIT_LIST_HEAD (&conn->ibc_tx_queue);
 644         INIT_LIST_HEAD (&conn->ibc_active_txs);
 645         spin_lock_init (&conn->ibc_lock);
 646
 647         atomic_inc (&kibnal_data.kib_nconns);
 648         /* well not really, but I call destroy() on failure, which decrements */
 649
 650         PORTAL_ALLOC (conn->ibc_rxs, IBNAL_RX_MSGS * sizeof (kib_rx_t));
 651         if (conn->ibc_rxs == NULL)
 652                 goto failed;
 653         memset (conn->ibc_rxs, 0, IBNAL_RX_MSGS * sizeof(kib_rx_t));
 654
 655         rc = kibnal_alloc_pages(&conn->ibc_rx_pages, IBNAL_RX_MSG_PAGES, 1);
 656         if (rc != 0)
 657                 goto failed;
 658
 659         vaddr_base = vaddr = conn->ibc_rx_pages->ibp_vaddr;
 660
 661         for (i = ipage = page_offset = 0; i < IBNAL_RX_MSGS; i++) {
 662                 struct page *page = conn->ibc_rx_pages->ibp_pages[ipage];
 663                 kib_rx_t   *rx = &conn->ibc_rxs[i];
 664
 665                 rx->rx_conn = conn;
 666                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) +
 667                              page_offset);
 668
 669                 if (kibnal_whole_mem())
 670                         rx->rx_vaddr = kibnal_page2phys(page) +
 671                                        page_offset +
 672                                        kibnal_data.kib_md.md_addr;
 673                 else
 674                         rx->rx_vaddr = vaddr;
 675
 676                 vaddr += IBNAL_MSG_SIZE;
 677                 LASSERT (vaddr <= vaddr_base + IBNAL_RX_MSG_BYTES);
 678
 679                 page_offset += IBNAL_MSG_SIZE;
 680                 LASSERT (page_offset <= PAGE_SIZE);
 681
 682                 if (page_offset == PAGE_SIZE) {
 683                         page_offset = 0;
 684                         ipage++;
 685                         LASSERT (ipage <= IBNAL_RX_MSG_PAGES);
 686                 }
 687         }
 688
 689         params.qp_create = (IB_QP_ATTRIBUTES_CREATE) {
 690                 .Type                    = QPTypeReliableConnected,
 691                 .SendQDepth              = IBNAL_TX_MAX_SG *
 692                                            IBNAL_MSG_QUEUE_SIZE,
 693                 .RecvQDepth              = IBNAL_MSG_QUEUE_SIZE,
 694                 .SendDSListDepth         = 1,
 695                 .RecvDSListDepth         = 1,
 696                 .SendCQHandle            = kibnal_data.kib_cq,
 697                 .RecvCQHandle            = kibnal_data.kib_cq,
 698                 .PDHandle                = kibnal_data.kib_pd,
 699                 .SendSignaledCompletions = TRUE,
 700         };
 701         frc = iibt_qp_create(kibnal_data.kib_hca, &params.qp_create, NULL,
 702                              &conn->ibc_qp, &conn->ibc_qp_attrs);
 703         if (rc != 0) {
 704                 CERROR ("Failed to create queue pair: %d\n", rc);
 705                 goto failed;
 706         }
 707
 708         /* Mark QP created */
 709         conn->ibc_state = IBNAL_CONN_INIT_QP;
 710
 711         params.qp_attr = (IB_QP_ATTRIBUTES_MODIFY) {
 712                 .RequestState             = QPStateInit,
 713                 .Attrs                    = (IB_QP_ATTR_PORTGUID |
 714                                              IB_QP_ATTR_PKEYINDEX |
 715                                              IB_QP_ATTR_ACCESSCONTROL),
 716                 .PortGUID                 = kibnal_data.kib_port_guid,
 717                 .PkeyIndex                = 0,
 718                 .AccessControl = {
 719                         .s = {
 720                                 .RdmaWrite = 1,
 721                                 .RdmaRead  = 1,
 722                         },
 723                 },
 724         };
 725         rc = iibt_qp_modify(conn->ibc_qp, &params.qp_attr, NULL);
 726         if (rc != 0) {
 727                 CERROR ("Failed to modify queue pair: %d\n", rc);
 728                 goto failed;
 729         }
 730
 731         /* 1 ref for caller */
 732         atomic_set (&conn->ibc_refcount, 1);
 733         return (conn);
 734
 735  failed:
 736         kibnal_destroy_conn (conn);
 737         return (NULL);
 738 }
 739
 740 void
 741 kibnal_destroy_conn (kib_conn_t *conn)
 742 {
 743         int    rc;
 744         FSTATUS frc;
 745
 746         CDEBUG (D_NET, "connection %p\n", conn);
 747
 748         LASSERT (atomic_read (&conn->ibc_refcount) == 0);
 749         LASSERT (list_empty(&conn->ibc_tx_queue));
 750         LASSERT (list_empty(&conn->ibc_active_txs));
 751         LASSERT (conn->ibc_nsends_posted == 0);
 752         LASSERT (conn->ibc_connreq == NULL);
 753
 754         switch (conn->ibc_state) {
 755         case IBNAL_CONN_DISCONNECTED:
 756                 /* called after connection sequence initiated */
 757                 /* fall through */
 758
 759         case IBNAL_CONN_INIT_QP:
 760                 /* _destroy includes an implicit Reset of the QP which
 761                  * discards posted work */
 762                 rc = iibt_qp_destroy(conn->ibc_qp);
 763                 if (rc != 0)
 764                         CERROR("Can't destroy QP: %d\n", rc);
 765                 /* fall through */
 766
 767         case IBNAL_CONN_INIT_NOTHING:
 768                 break;
 769
 770         default:
 771                 LASSERT (0);
 772         }
 773
 774         if (conn->ibc_cep != NULL) {
 775                 frc = iibt_cm_destroy_cep(conn->ibc_cep);
 776                 if (frc != 0)
 777                         CERROR("Can't destroy CEP %p: %d\n", conn->ibc_cep,
 778                                frc);
 779         }
 780
 781         if (conn->ibc_rx_pages != NULL)
 782                 kibnal_free_pages(conn->ibc_rx_pages);
 783
 784         if (conn->ibc_rxs != NULL)
 785                 PORTAL_FREE(conn->ibc_rxs,
 786                             IBNAL_RX_MSGS * sizeof(kib_rx_t));
 787
 788         if (conn->ibc_peer != NULL)
 789                 kib_peer_decref(conn->ibc_peer);
 790
 791         PORTAL_FREE(conn, sizeof (*conn));
 792
 793         atomic_dec(&kibnal_data.kib_nconns);
 794
 795         if (atomic_read (&kibnal_data.kib_nconns) == 0 &&
 796             kibnal_data.kib_shutdown) {
 797                 /* I just nuked the last connection on shutdown; wake up
 798                  * everyone so they can exit. */
 799                 wake_up_all(&kibnal_data.kib_sched_waitq);
 800                 wake_up_all(&kibnal_data.kib_connd_waitq);
 801         }
 802 }
 803
 804 void
 805 kibnal_put_conn (kib_conn_t *conn)
 806 {
 807         unsigned long flags;
 808
 809         CDEBUG (D_NET, "putting conn[%p] state %d -> "LPX64" (%d)\n",
 810                 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
 811                 atomic_read (&conn->ibc_refcount));
 812
 813         LASSERT (atomic_read (&conn->ibc_refcount) > 0);
 814         if (!atomic_dec_and_test (&conn->ibc_refcount))
 815                 return;
 816
 817         /* must disconnect before dropping the final ref */
 818         LASSERT (conn->ibc_state == IBNAL_CONN_DISCONNECTED);
 819
 820         spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags);
 821
 822         list_add (&conn->ibc_list, &kibnal_data.kib_connd_conns);
 823         wake_up (&kibnal_data.kib_connd_waitq);
 824
 825         spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
 826 }
 827
 828 static int
 829 kibnal_close_peer_conns_locked (kib_peer_t *peer, int why)
 830 {
 831         kib_conn_t         *conn;
 832         struct list_head   *ctmp;
 833         struct list_head   *cnxt;
 834         int                 count = 0;
 835
 836         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 837                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 838
 839                 count++;
 840                 kibnal_close_conn_locked (conn, why);
 841         }
 842
 843         return (count);
 844 }
 845
 846 int
 847 kibnal_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation)
 848 {
 849         kib_conn_t         *conn;
 850         struct list_head   *ctmp;
 851         struct list_head   *cnxt;
 852         int                 count = 0;
 853
 854         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 855                 conn = list_entry (ctmp, kib_conn_t, ibc_list);
 856
 857                 if (conn->ibc_incarnation == incarnation)
 858                         continue;
 859
 860                 CDEBUG(D_NET, "Closing stale conn nid:"LPX64" incarnation:"LPX64"("LPX64")\n",
 861                        peer->ibp_nid, conn->ibc_incarnation, incarnation);
 862
 863                 count++;
 864                 kibnal_close_conn_locked (conn, -ESTALE);
 865         }
 866
 867         return (count);
 868 }
 869
 870 static int
 871 kibnal_close_matching_conns (ptl_nid_t nid)
 872 {
 873         unsigned long       flags;
 874         kib_peer_t         *peer;
 875         struct list_head   *ptmp;
 876         struct list_head   *pnxt;
 877         int                 lo;
 878         int                 hi;
 879         int                 i;
 880         int                 count = 0;
 881
 882         write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
 883
 884         if (nid != PTL_NID_ANY)
 885                 lo = hi = kibnal_nid2peerlist(nid) - kibnal_data.kib_peers;
 886         else {
 887                 lo = 0;
 888                 hi = kibnal_data.kib_peer_hash_size - 1;
 889         }
 890
 891         for (i = lo; i <= hi; i++) {
 892                 list_for_each_safe (ptmp, pnxt, &kibnal_data.kib_peers[i]) {
 893
 894                         peer = list_entry (ptmp, kib_peer_t, ibp_list);
 895                         LASSERT (peer->ibp_persistence != 0 ||
 896                                  peer->ibp_connecting != 0 ||
 897                                  !list_empty (&peer->ibp_conns));
 898
 899                         if (!(nid == PTL_NID_ANY || nid == peer->ibp_nid))
 900                                 continue;
 901
 902                         count += kibnal_close_peer_conns_locked (peer, 0);
 903                 }
 904         }
 905
 906         write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
 907
 908         /* wildcards always succeed */
 909         if (nid == PTL_NID_ANY)
 910                 return (0);
 911
 912         return (count == 0 ? -ENOENT : 0);
 913 }
 914
 915 static int
 916 kibnal_cmd(struct portals_cfg *pcfg, void * private)
 917 {
 918         int rc = -EINVAL;
 919         ENTRY;
 920
 921         LASSERT (pcfg != NULL);
 922
 923         switch(pcfg->pcfg_command) {
 924         case NAL_CMD_GET_PEER: {
 925                 ptl_nid_t   nid = 0;
 926                 int         share_count = 0;
 927
 928                 rc = kibnal_get_peer_info(pcfg->pcfg_count,
 929                                           &nid, &share_count);
 930                 pcfg->pcfg_nid   = nid;
 931                 pcfg->pcfg_size  = 0;
 932                 pcfg->pcfg_id    = 0;
 933                 pcfg->pcfg_misc  = 0;
 934                 pcfg->pcfg_count = 0;
 935                 pcfg->pcfg_wait  = share_count;
 936                 break;
 937         }
 938         case NAL_CMD_ADD_PEER: {
 939                 rc = kibnal_add_persistent_peer (pcfg->pcfg_nid);
 940                 break;
 941         }
 942         case NAL_CMD_DEL_PEER: {
 943                 rc = kibnal_del_peer (pcfg->pcfg_nid,
 944                                        /* flags == single_share */
 945                                        pcfg->pcfg_flags != 0);
 946                 break;
 947         }
 948         case NAL_CMD_GET_CONN: {
 949                 kib_conn_t *conn = kibnal_get_conn_by_idx (pcfg->pcfg_count);
 950
 951                 if (conn == NULL)
 952                         rc = -ENOENT;
 953                 else {
 954                         rc = 0;
 955                         pcfg->pcfg_nid   = conn->ibc_peer->ibp_nid;
 956                         pcfg->pcfg_id    = 0;
 957                         pcfg->pcfg_misc  = 0;
 958                         pcfg->pcfg_flags = 0;
 959                         kibnal_put_conn (conn);
 960                 }
 961                 break;
 962         }
 963         case NAL_CMD_CLOSE_CONNECTION: {
 964                 rc = kibnal_close_matching_conns (pcfg->pcfg_nid);
 965                 break;
 966         }
 967         case NAL_CMD_REGISTER_MYNID: {
 968                 if (pcfg->pcfg_nid == PTL_NID_ANY)
 969                         rc = -EINVAL;
 970                 else
 971                         rc = kibnal_set_mynid (pcfg->pcfg_nid);
 972                 break;
 973         }
 974         }
 975
 976         RETURN(rc);
 977 }
 978
 979 void
 980 kibnal_free_pages (kib_pages_t *p)
 981 {
 982         int     npages = p->ibp_npages;
 983         int     rc;
 984         int     i;
 985
 986         if (p->ibp_mapped) {
 987                 rc = iibt_deregister_memory(p->ibp_handle);
 988                 if (rc != 0)
 989                         CERROR ("Deregister error: %d\n", rc);
 990         }
 991
 992         for (i = 0; i < npages; i++)
 993                 if (p->ibp_pages[i] != NULL)
 994                         __free_page(p->ibp_pages[i]);
 995
 996         PORTAL_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
 997 }
 998
 999 int
1000 kibnal_alloc_pages (kib_pages_t **pp, int npages, int allow_write)
1001 {
1002         kib_pages_t                *p;
1003         __u64                      *phys_pages;
1004         int                         i;
1005         FSTATUS                     frc;
1006         IB_ACCESS_CONTROL           access;
1007
1008         memset(&access, 0, sizeof(access));
1009         access.s.MWBindable = 1;
1010         access.s.LocalWrite = 1;
1011         access.s.RdmaRead = 1;
1012         access.s.RdmaWrite = 1;
1013
1014         PORTAL_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1015         if (p == NULL) {
1016                 CERROR ("Can't allocate buffer %d\n", npages);
1017                 return (-ENOMEM);
1018         }
1019
1020         memset (p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1021         p->ibp_npages = npages;
1022
1023         for (i = 0; i < npages; i++) {
1024                 p->ibp_pages[i] = alloc_page (GFP_KERNEL);
1025                 if (p->ibp_pages[i] == NULL) {
1026                         CERROR ("Can't allocate page %d of %d\n", i, npages);
1027                         kibnal_free_pages(p);
1028                         return (-ENOMEM);
1029                 }
1030         }
1031
1032         if (kibnal_whole_mem())
1033                 goto out;
1034
1035         PORTAL_ALLOC(phys_pages, npages * sizeof(*phys_pages));
1036         if (phys_pages == NULL) {
1037                 CERROR ("Can't allocate physarray for %d pages\n", npages);
1038                 /* XXX free ibp_pages? */
1039                 kibnal_free_pages(p);
1040                 return (-ENOMEM);
1041         }
1042
1043         /* if we were using the _contig_ registration variant we would have
1044          * an array of PhysAddr/Length pairs, but the discontiguous variant
1045          * just takes the PhysAddr */
1046         for (i = 0; i < npages; i++)
1047                 phys_pages[i] = kibnal_page2phys(p->ibp_pages[i]);
1048
1049         frc = iibt_register_physical_memory(kibnal_data.kib_hca,
1050                                             0,          /* requested vaddr */
1051                                             phys_pages, npages,
1052                                             0,          /* offset */
1053                                             kibnal_data.kib_pd,
1054                                             access,
1055                                             &p->ibp_handle, &p->ibp_vaddr,
1056                                             &p->ibp_lkey, &p->ibp_rkey);
1057
1058         PORTAL_FREE(phys_pages, npages * sizeof(*phys_pages));
1059
1060         if (frc != FSUCCESS) {
1061                 CERROR ("Error %d mapping %d pages\n", frc, npages);
1062                 kibnal_free_pages(p);
1063                 return (-ENOMEM);
1064         }
1065
1066         CDEBUG(D_NET, "registered %d pages; handle: %p vaddr "LPX64" "
1067                       "lkey %x rkey %x\n", npages, p->ibp_handle,
1068                       p->ibp_vaddr, p->ibp_lkey, p->ibp_rkey);
1069
1070         p->ibp_mapped = 1;
1071 out:
1072         *pp = p;
1073         return (0);
1074 }
1075
1076 static int
1077 kibnal_setup_tx_descs (void)
1078 {
1079         int           ipage = 0;
1080         int           page_offset = 0;
1081         __u64         vaddr;
1082         __u64         vaddr_base;
1083         struct page  *page;
1084         kib_tx_t     *tx;
1085         int           i;
1086         int           rc;
1087
1088         /* pre-mapped messages are not bigger than 1 page */
1089         LASSERT (IBNAL_MSG_SIZE <= PAGE_SIZE);
1090
1091         /* No fancy arithmetic when we do the buffer calculations */
1092         LASSERT (PAGE_SIZE % IBNAL_MSG_SIZE == 0);
1093
1094         rc = kibnal_alloc_pages(&kibnal_data.kib_tx_pages, IBNAL_TX_MSG_PAGES,
1095                                 0);
1096         if (rc != 0)
1097                 return (rc);
1098
1099         /* ignored for the whole_mem case */
1100         vaddr = vaddr_base = kibnal_data.kib_tx_pages->ibp_vaddr;
1101
1102         for (i = 0; i < IBNAL_TX_MSGS; i++) {
1103                 page = kibnal_data.kib_tx_pages->ibp_pages[ipage];
1104                 tx = &kibnal_data.kib_tx_descs[i];
1105
1106                 memset (tx, 0, sizeof(*tx));    /* zero flags etc */
1107
1108                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1109                                             page_offset);
1110
1111                 if (kibnal_whole_mem())
1112                         tx->tx_vaddr = kibnal_page2phys(page) +
1113                                        page_offset +
1114                                        kibnal_data.kib_md.md_addr;
1115                 else
1116                         tx->tx_vaddr = vaddr;
1117
1118                 tx->tx_isnblk = (i >= IBNAL_NTX);
1119                 tx->tx_mapped = KIB_TX_UNMAPPED;
1120
1121                 CDEBUG(D_NET, "Tx[%d] %p->%p - "LPX64"\n",
1122                        i, tx, tx->tx_msg, tx->tx_vaddr);
1123
1124                 if (tx->tx_isnblk)
1125                         list_add (&tx->tx_list,
1126                                   &kibnal_data.kib_idle_nblk_txs);
1127                 else
1128                         list_add (&tx->tx_list,
1129                                   &kibnal_data.kib_idle_txs);
1130
1131                 vaddr += IBNAL_MSG_SIZE;
1132                 LASSERT (vaddr <= vaddr_base + IBNAL_TX_MSG_BYTES);
1133
1134                 page_offset += IBNAL_MSG_SIZE;
1135                 LASSERT (page_offset <= PAGE_SIZE);
1136
1137                 if (page_offset == PAGE_SIZE) {
1138                         page_offset = 0;
1139                         ipage++;
1140                         LASSERT (ipage <= IBNAL_TX_MSG_PAGES);
1141                 }
1142         }
1143
1144         return (0);
1145 }
1146
1147 static void
1148 kibnal_api_shutdown (nal_t *nal)
1149 {
1150         int   i;
1151         int   rc;
1152
1153         if (nal->nal_refct != 0) {
1154                 /* This module got the first ref */
1155                 PORTAL_MODULE_UNUSE;
1156                 return;
1157         }
1158
1159         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
1160                atomic_read (&portal_kmemory));
1161
1162         LASSERT(nal == &kibnal_api);
1163
1164         switch (kibnal_data.kib_init) {
1165         default:
1166                 CERROR ("Unexpected state %d\n", kibnal_data.kib_init);
1167                 LBUG();
1168
1169         case IBNAL_INIT_ALL:
1170                 /* stop calls to nal_cmd */
1171                 libcfs_nal_cmd_unregister(IIBNAL);
1172                 /* No new peers */
1173
1174                 /* resetting my NID to unadvertises me, removes my
1175                  * listener and nukes all current peers */
1176                 kibnal_set_mynid (PTL_NID_ANY);
1177
1178                 /* Wait for all peer state to clean up (crazy) */
1179                 i = 2;
1180                 while (atomic_read (&kibnal_data.kib_npeers) != 0) {
1181                         i++;
1182                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1183                                "waiting for %d peers to disconnect (can take a few seconds)\n",
1184                                atomic_read (&kibnal_data.kib_npeers));
1185                         set_current_state (TASK_UNINTERRUPTIBLE);
1186                         schedule_timeout (HZ);
1187                 }
1188                 /* fall through */
1189
1190         case IBNAL_INIT_CQ:
1191                 rc = iibt_cq_destroy(kibnal_data.kib_cq);
1192                 if (rc != 0)
1193                         CERROR ("Destroy CQ error: %d\n", rc);
1194                 /* fall through */
1195
1196         case IBNAL_INIT_TXD:
1197                 kibnal_free_pages (kibnal_data.kib_tx_pages);
1198                 /* fall through */
1199
1200         case IBNAL_INIT_MR:
1201                 if (kibnal_data.kib_md.md_handle != NULL) {
1202                         rc = iibt_deregister_memory(kibnal_data.kib_md.md_handle);
1203                         if (rc != FSUCCESS)
1204                                 CERROR ("Deregister memory: %d\n", rc);
1205                 }
1206                 /* fall through */
1207
1208 #if IBNAL_FMR
1209         case IBNAL_INIT_FMR:
1210                 rc = ib_fmr_pool_destroy (kibnal_data.kib_fmr_pool);
1211                 if (rc != 0)
1212                         CERROR ("Destroy FMR pool error: %d\n", rc);
1213                 /* fall through */
1214 #endif
1215         case IBNAL_INIT_PD:
1216                 rc = iibt_pd_free(kibnal_data.kib_pd);
1217                 if (rc != 0)
1218                         CERROR ("Destroy PD error: %d\n", rc);
1219                 /* fall through */
1220
1221         case IBNAL_INIT_SD:
1222                 rc = iibt_sd_deregister(kibnal_data.kib_sd);
1223                 if (rc != 0)
1224                         CERROR ("Deregister SD error: %d\n", rc);
1225                 /* fall through */
1226
1227         case IBNAL_INIT_PORT:
1228                 /* XXX ??? */
1229                 /* fall through */
1230
1231         case IBNAL_INIT_PORTATTRS:
1232                 PORTAL_FREE(kibnal_data.kib_hca_attrs.PortAttributesList,
1233                             kibnal_data.kib_hca_attrs.PortAttributesListSize);
1234                 /* fall through */
1235
1236         case IBNAL_INIT_HCA:
1237                 rc = iibt_close_hca(kibnal_data.kib_hca);
1238                 if (rc != 0)
1239                         CERROR ("Close HCA  error: %d\n", rc);
1240                 /* fall through */
1241
1242         case IBNAL_INIT_LIB:
1243                 lib_fini(&kibnal_lib);
1244                 /* fall through */
1245
1246         case IBNAL_INIT_DATA:
1247                 /* Module refcount only gets to zero when all peers
1248                  * have been closed so all lists must be empty */
1249                 LASSERT (atomic_read (&kibnal_data.kib_npeers) == 0);
1250                 LASSERT (kibnal_data.kib_peers != NULL);
1251                 for (i = 0; i < kibnal_data.kib_peer_hash_size; i++) {
1252                         LASSERT (list_empty (&kibnal_data.kib_peers[i]));
1253                 }
1254                 LASSERT (atomic_read (&kibnal_data.kib_nconns) == 0);
1255                 LASSERT (list_empty (&kibnal_data.kib_sched_rxq));
1256                 LASSERT (list_empty (&kibnal_data.kib_sched_txq));
1257                 LASSERT (list_empty (&kibnal_data.kib_connd_conns));
1258                 LASSERT (list_empty (&kibnal_data.kib_connd_peers));
1259
1260                 /* flag threads to terminate; wake and wait for them to die */
1261                 kibnal_data.kib_shutdown = 1;
1262                 wake_up_all (&kibnal_data.kib_sched_waitq);
1263                 wake_up_all (&kibnal_data.kib_connd_waitq);
1264
1265                 i = 2;
1266                 while (atomic_read (&kibnal_data.kib_nthreads) != 0) {
1267                         i++;
1268                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1269                                "Waiting for %d threads to terminate\n",
1270                                atomic_read (&kibnal_data.kib_nthreads));
1271                         set_current_state (TASK_INTERRUPTIBLE);
1272                         schedule_timeout (HZ);
1273                 }
1274                 /* fall through */
1275
1276         case IBNAL_INIT_NOTHING:
1277                 break;
1278         }
1279
1280         if (kibnal_data.kib_tx_descs != NULL)
1281                 PORTAL_FREE (kibnal_data.kib_tx_descs,
1282                              IBNAL_TX_MSGS * sizeof(kib_tx_t));
1283
1284         if (kibnal_data.kib_peers != NULL)
1285                 PORTAL_FREE (kibnal_data.kib_peers,
1286                              sizeof (struct list_head) *
1287                              kibnal_data.kib_peer_hash_size);
1288
1289         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
1290                atomic_read (&portal_kmemory));
1291         printk(KERN_INFO "Lustre: Infinicon IB NAL unloaded (final mem %d)\n",
1292                atomic_read(&portal_kmemory));
1293
1294         kibnal_data.kib_init = IBNAL_INIT_NOTHING;
1295 }
1296
1297 #define roundup_power(val, power) \
1298         ( (val + (__u64)(power - 1)) & ~((__u64)(power - 1)) )
1299
1300 /* this isn't very portable or sturdy in the face of funny mem/bus configs */
1301 static __u64 max_phys_mem(IB_CA_ATTRIBUTES *ca_attr)
1302 {
1303         struct sysinfo si;
1304         __u64 ret;
1305
1306         /* XXX we don't bother with first-gen cards */
1307         if (ca_attr->VendorId == 0xd0b7 && ca_attr->DeviceId == 0x3101)
1308                 return 0ULL;
1309
1310         si_meminfo(&si);
1311         ret = (__u64)max(si.totalram, max_mapnr) * si.mem_unit;
1312         return roundup_power(ret, 128 * 1024 * 1024);
1313 }
1314 #undef roundup_power
1315
1316 static int
1317 kibnal_api_startup (nal_t *nal, ptl_pid_t requested_pid,
1318                      ptl_ni_limits_t *requested_limits,
1319                      ptl_ni_limits_t *actual_limits)
1320 {
1321         ptl_process_id_t    process_id;
1322         int                 pkmem = atomic_read(&portal_kmemory);
1323         IB_PORT_ATTRIBUTES *pattr;
1324         FSTATUS             frc;
1325         int                 rc;
1326         int                 n;
1327         int                 i;
1328
1329         LASSERT (nal == &kibnal_api);
1330
1331         if (nal->nal_refct != 0) {
1332                 if (actual_limits != NULL)
1333                         *actual_limits = kibnal_lib.libnal_ni.ni_actual_limits;
1334                 /* This module got the first ref */
1335                 PORTAL_MODULE_USE;
1336                 return (PTL_OK);
1337         }
1338
1339         LASSERT (kibnal_data.kib_init == IBNAL_INIT_NOTHING);
1340
1341         frc = IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2,
1342                                        &kibnal_data.kib_interfaces);
1343         if (frc != FSUCCESS) {
1344                 CERROR("IbtGetInterfaceByVersion(IBT_INTERFACE_VERSION_2) = %d\n",
1345                         frc);
1346                 return -ENOSYS;
1347         }
1348
1349         init_MUTEX (&kibnal_data.kib_nid_mutex);
1350         init_MUTEX_LOCKED (&kibnal_data.kib_nid_signal);
1351         kibnal_data.kib_nid = PTL_NID_ANY;
1352
1353         rwlock_init(&kibnal_data.kib_global_lock);
1354
1355         kibnal_data.kib_peer_hash_size = IBNAL_PEER_HASH_SIZE;
1356         PORTAL_ALLOC (kibnal_data.kib_peers,
1357                       sizeof (struct list_head) * kibnal_data.kib_peer_hash_size);
1358         if (kibnal_data.kib_peers == NULL) {
1359                 goto failed;
1360         }
1361         for (i = 0; i < kibnal_data.kib_peer_hash_size; i++)
1362                 INIT_LIST_HEAD(&kibnal_data.kib_peers[i]);
1363
1364         spin_lock_init (&kibnal_data.kib_connd_lock);
1365         INIT_LIST_HEAD (&kibnal_data.kib_connd_peers);
1366         INIT_LIST_HEAD (&kibnal_data.kib_connd_conns);
1367         init_waitqueue_head (&kibnal_data.kib_connd_waitq);
1368
1369         spin_lock_init (&kibnal_data.kib_sched_lock);
1370         INIT_LIST_HEAD (&kibnal_data.kib_sched_txq);
1371         INIT_LIST_HEAD (&kibnal_data.kib_sched_rxq);
1372         init_waitqueue_head (&kibnal_data.kib_sched_waitq);
1373
1374         spin_lock_init (&kibnal_data.kib_tx_lock);
1375         INIT_LIST_HEAD (&kibnal_data.kib_idle_txs);
1376         INIT_LIST_HEAD (&kibnal_data.kib_idle_nblk_txs);
1377         init_waitqueue_head(&kibnal_data.kib_idle_tx_waitq);
1378
1379         PORTAL_ALLOC (kibnal_data.kib_tx_descs,
1380                       IBNAL_TX_MSGS * sizeof(kib_tx_t));
1381         if (kibnal_data.kib_tx_descs == NULL) {
1382                 CERROR ("Can't allocate tx descs\n");
1383                 goto failed;
1384         }
1385
1386         /* lists/ptrs/locks initialised */
1387         kibnal_data.kib_init = IBNAL_INIT_DATA;
1388         /*****************************************************/
1389
1390         process_id.pid = requested_pid;
1391         process_id.nid = kibnal_data.kib_nid;
1392
1393         rc = lib_init(&kibnal_lib, nal, process_id,
1394                       requested_limits, actual_limits);
1395         if (rc != PTL_OK) {
1396                 CERROR("lib_init failed: error %d\n", rc);
1397                 goto failed;
1398         }
1399
1400         /* lib interface initialised */
1401         kibnal_data.kib_init = IBNAL_INIT_LIB;
1402         /*****************************************************/
1403
1404         for (i = 0; i < IBNAL_N_SCHED; i++) {
1405                 rc = kibnal_thread_start (kibnal_scheduler, (void *)i);
1406                 if (rc != 0) {
1407                         CERROR("Can't spawn iibnal scheduler[%d]: %d\n",
1408                                i, rc);
1409                         goto failed;
1410                 }
1411         }
1412
1413         rc = kibnal_thread_start (kibnal_connd, NULL);
1414         if (rc != 0) {
1415                 CERROR ("Can't spawn iibnal connd: %d\n", rc);
1416                 goto failed;
1417         }
1418
1419         n = sizeof(kibnal_data.kib_hca_guids) /
1420             sizeof(kibnal_data.kib_hca_guids[0]);
1421         frc = iibt_get_hca_guids(&n, kibnal_data.kib_hca_guids);
1422         if (frc != FSUCCESS) {
1423                 CERROR ("Can't get channel adapter guids: %d\n", frc);
1424                 goto failed;
1425         }
1426         if (n == 0) {
1427                 CERROR ("No channel adapters found\n");
1428                 goto failed;
1429         }
1430
1431         /* Infinicon has per-HCA rather than per CQ completion handlers */
1432         frc = iibt_open_hca(kibnal_data.kib_hca_guids[0],
1433                             kibnal_ca_callback,
1434                             kibnal_ca_async_callback,
1435                             &kibnal_data.kib_hca,
1436                             &kibnal_data.kib_hca);
1437         if (frc != FSUCCESS) {
1438                 CERROR ("Can't open CA[0]: %d\n", frc);
1439                 goto failed;
1440         }
1441
1442         /* Channel Adapter opened */
1443         kibnal_data.kib_init = IBNAL_INIT_HCA;
1444         /*****************************************************/
1445
1446         kibnal_data.kib_hca_attrs.PortAttributesList = NULL;
1447         kibnal_data.kib_hca_attrs.PortAttributesListSize = 0;
1448         frc = iibt_query_hca(kibnal_data.kib_hca,
1449                              &kibnal_data.kib_hca_attrs, NULL);
1450         if (frc != FSUCCESS) {
1451                 CERROR ("Can't size port attrs: %d\n", frc);
1452                 goto failed;
1453         }
1454
1455         PORTAL_ALLOC(kibnal_data.kib_hca_attrs.PortAttributesList,
1456                      kibnal_data.kib_hca_attrs.PortAttributesListSize);
1457         if (kibnal_data.kib_hca_attrs.PortAttributesList == NULL)
1458                 goto failed;
1459
1460         /* Port attrs allocated */
1461         kibnal_data.kib_init = IBNAL_INIT_PORTATTRS;
1462         /*****************************************************/
1463
1464         frc = iibt_query_hca(kibnal_data.kib_hca, &kibnal_data.kib_hca_attrs,
1465                              NULL);
1466         if (frc != FSUCCESS) {
1467                 CERROR ("Can't get port attrs for CA 0: %d\n", frc);
1468                 goto failed;
1469         }
1470
1471         for (i = 0, pattr = kibnal_data.kib_hca_attrs.PortAttributesList;
1472              pattr != NULL;
1473              i++, pattr = pattr->Next) {
1474                 switch (pattr->PortState) {
1475                 default:
1476                         CERROR("Unexpected port[%d] state %d\n",
1477                                i, pattr->PortState);
1478                         continue;
1479                 case PortStateDown:
1480                         CDEBUG(D_NET, "port[%d] Down\n", i);
1481                         continue;
1482                 case PortStateInit:
1483                         CDEBUG(D_NET, "port[%d] Init\n", i);
1484                         continue;
1485                 case PortStateArmed:
1486                         CDEBUG(D_NET, "port[%d] Armed\n", i);
1487                         continue;
1488
1489                 case PortStateActive:
1490                         CDEBUG(D_NET, "port[%d] Active\n", i);
1491                         kibnal_data.kib_port = i;
1492                         kibnal_data.kib_port_guid = pattr->GUID;
1493                         kibnal_data.kib_port_pkey = pattr->PkeyTable[0];
1494                         break;
1495                 }
1496                 break;
1497         }
1498
1499         if (pattr == NULL) {
1500                 CERROR ("Can't find an active port\n");
1501                 goto failed;
1502         }
1503
1504         CDEBUG(D_NET, "got guid "LPX64"\n", kibnal_data.kib_port_guid);
1505
1506         /* Active port found */
1507         kibnal_data.kib_init = IBNAL_INIT_PORT;
1508         /*****************************************************/
1509
1510         frc = iibt_sd_register(&kibnal_data.kib_sd, NULL);
1511         if (frc != FSUCCESS) {
1512                 CERROR ("Can't register with SD: %d\n", frc);
1513                 goto failed;
1514         }
1515
1516         /* Registered with SD OK */
1517         kibnal_data.kib_init = IBNAL_INIT_SD;
1518         /*****************************************************/
1519
1520         frc = iibt_pd_allocate(kibnal_data.kib_hca, 0, &kibnal_data.kib_pd);
1521         if (frc != FSUCCESS) {
1522                 CERROR ("Can't create PD: %d\n", rc);
1523                 goto failed;
1524         }
1525
1526         /* flag PD initialised */
1527         kibnal_data.kib_init = IBNAL_INIT_PD;
1528         /*****************************************************/
1529
1530 #if IBNAL_FMR
1531         {
1532                 const int pool_size = IBNAL_NTX + IBNAL_NTX_NBLK;
1533                 struct ib_fmr_pool_param params = {
1534                         .max_pages_per_fmr = PTL_MTU/PAGE_SIZE,
1535                         .access            = (IB_ACCESS_LOCAL_WRITE |
1536                                               IB_ACCESS_REMOTE_WRITE |
1537                                               IB_ACCESS_REMOTE_READ),
1538                         .pool_size         = pool_size,
1539                         .dirty_watermark   = (pool_size * 3)/4,
1540                         .flush_function    = NULL,
1541                         .flush_arg         = NULL,
1542                         .cache             = 1,
1543                 };
1544                 rc = ib_fmr_pool_create(kibnal_data.kib_pd, &params,
1545                                         &kibnal_data.kib_fmr_pool);
1546                 if (rc != 0) {
1547                         CERROR ("Can't create FMR pool size %d: %d\n",
1548                                 pool_size, rc);
1549                         goto failed;
1550                 }
1551         }
1552
1553         /* flag FMR pool initialised */
1554         kibnal_data.kib_init = IBNAL_INIT_FMR;
1555 #endif
1556         /*****************************************************/
1557         if (IBNAL_WHOLE_MEM) {
1558                 IB_MR_PHYS_BUFFER phys;
1559                 IB_ACCESS_CONTROL access;
1560                 kib_md_t *md = &kibnal_data.kib_md;
1561
1562                 memset(&access, 0, sizeof(access));
1563                 access.s.MWBindable = 1;
1564                 access.s.LocalWrite = 1;
1565                 access.s.RdmaRead = 1;
1566                 access.s.RdmaWrite = 1;
1567
1568                 phys.PhysAddr = 0;
1569                 phys.Length = max_phys_mem(&kibnal_data.kib_hca_attrs);
1570                 if (phys.Length == 0) {
1571                         CERROR ("couldn't determine the end of phys mem\n");
1572                         goto failed;
1573                 }
1574
1575                 rc = iibt_register_contig_physical_memory(kibnal_data.kib_hca,
1576                                                           0,
1577                                                           &phys, 1,
1578                                                           0,
1579                                                           kibnal_data.kib_pd,
1580                                                           access,
1581                                                           &md->md_handle,
1582                                                           &md->md_addr,
1583                                                           &md->md_lkey,
1584                                                           &md->md_rkey);
1585                 if (rc != FSUCCESS) {
1586                         CERROR("registering physical memory failed: %d\n",
1587                                rc);
1588                         CERROR("falling back to registration per-rdma\n");
1589                         md->md_handle = NULL;
1590                 } else {
1591                         CDEBUG(D_NET, "registered "LPU64" bytes of mem\n",
1592                                phys.Length);
1593                         kibnal_data.kib_init = IBNAL_INIT_MR;
1594                 }
1595         }
1596
1597         /*****************************************************/
1598
1599         rc = kibnal_setup_tx_descs();
1600         if (rc != 0) {
1601                 CERROR ("Can't register tx descs: %d\n", rc);
1602                 goto failed;
1603         }
1604
1605         /* flag TX descs initialised */
1606         kibnal_data.kib_init = IBNAL_INIT_TXD;
1607         /*****************************************************/
1608
1609         {
1610                 uint32 nentries;
1611
1612                 frc = iibt_cq_create(kibnal_data.kib_hca, IBNAL_CQ_ENTRIES,
1613                                      &kibnal_data.kib_cq, &kibnal_data.kib_cq,
1614                                      &nentries);
1615                 if (frc != FSUCCESS) {
1616                         CERROR ("Can't create RX CQ: %d\n", frc);
1617                         goto failed;
1618                 }
1619
1620                 /* flag CQ initialised */
1621                 kibnal_data.kib_init = IBNAL_INIT_CQ;
1622
1623                 if (nentries < IBNAL_CQ_ENTRIES) {
1624                         CERROR ("CQ only has %d entries, need %d\n",
1625                                 nentries, IBNAL_CQ_ENTRIES);
1626                         goto failed;
1627                 }
1628
1629                 rc = iibt_cq_rearm(kibnal_data.kib_cq, CQEventSelNextWC);
1630                 if (rc != 0) {
1631                         CERROR ("Failed to re-arm completion queue: %d\n", rc);
1632                         goto failed;
1633                 }
1634         }
1635
1636         /*****************************************************/
1637
1638         rc = libcfs_nal_cmd_register(IIBNAL, &kibnal_cmd, NULL);
1639         if (rc != 0) {
1640                 CERROR ("Can't initialise command interface (rc = %d)\n", rc);
1641                 goto failed;
1642         }
1643
1644         /* flag everything initialised */
1645         kibnal_data.kib_init = IBNAL_INIT_ALL;
1646         /*****************************************************/
1647
1648         printk(KERN_INFO "Lustre: Infinicon IB NAL loaded "
1649                "(initial mem %d)\n", pkmem);
1650
1651         return (PTL_OK);
1652
1653  failed:
1654         kibnal_api_shutdown (&kibnal_api);
1655         return (PTL_FAIL);
1656 }
1657
1658 void __exit
1659 kibnal_module_fini (void)
1660 {
1661 #ifdef CONFIG_SYSCTL
1662         if (kibnal_tunables.kib_sysctl != NULL)
1663                 unregister_sysctl_table (kibnal_tunables.kib_sysctl);
1664 #endif
1665         PtlNIFini(kibnal_ni);
1666
1667         ptl_unregister_nal(IIBNAL);
1668 }
1669
1670 int __init
1671 kibnal_module_init (void)
1672 {
1673         int    rc;
1674
1675         if (sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN) {
1676                 CERROR("sizeof(kib_wire_connreq_t) > CM_REQUEST_INFO_USER_LEN\n");
1677                 return -EINVAL;
1678         }
1679
1680         /* the following must be sizeof(int) for proc_dointvec() */
1681         if (sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)) {
1682                 CERROR("sizeof (kibnal_tunables.kib_io_timeout) != sizeof (int)\n");
1683                 return -EINVAL;
1684         }
1685
1686         kibnal_api.nal_ni_init = kibnal_api_startup;
1687         kibnal_api.nal_ni_fini = kibnal_api_shutdown;
1688
1689         /* Initialise dynamic tunables to defaults once only */
1690         kibnal_tunables.kib_io_timeout = IBNAL_IO_TIMEOUT;
1691
1692         rc = ptl_register_nal(IIBNAL, &kibnal_api);
1693         if (rc != PTL_OK) {
1694                 CERROR("Can't register IBNAL: %d\n", rc);
1695                 return (-ENOMEM);               /* or something... */
1696         }
1697
1698         /* Pure gateways want the NAL started up at module load time... */
1699         rc = PtlNIInit(IIBNAL, LUSTRE_SRV_PTL_PID, NULL, NULL, &kibnal_ni);
1700         if (rc != PTL_OK && rc != PTL_IFACE_DUP) {
1701                 ptl_unregister_nal(IIBNAL);
1702                 return (-ENODEV);
1703         }
1704
1705 #ifdef CONFIG_SYSCTL
1706         /* Press on regardless even if registering sysctl doesn't work */
1707         kibnal_tunables.kib_sysctl =
1708                 register_sysctl_table (kibnal_top_ctl_table, 0);
1709 #endif
1710         return (0);
1711 }
1712
1713 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1714 MODULE_DESCRIPTION("Kernel Infinicon IB NAL v0.01");
1715 MODULE_LICENSE("GPL");
1716
1717 module_init(kibnal_module_init);
1718 module_exit(kibnal_module_fini);
1719