lnet/klnds/o2iblnd/o2iblnd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2006 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 "o2iblnd.h"
  25
  26 lnd_t the_kiblnd = {
  27         .lnd_type       = O2IBLND,
  28         .lnd_startup    = kiblnd_startup,
  29         .lnd_shutdown   = kiblnd_shutdown,
  30         .lnd_ctl        = kiblnd_ctl,
  31         .lnd_send       = kiblnd_send,
  32         .lnd_recv       = kiblnd_recv,
  33 };
  34
  35 kib_data_t              kiblnd_data;
  36
  37 __u32
  38 kiblnd_cksum (void *ptr, int nob)
  39 {
  40         char  *c  = ptr;
  41         __u32  sum = 0;
  42
  43         while (nob-- > 0)
  44                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  45
  46         /* ensure I don't return 0 (== no checksum) */
  47         return (sum == 0) ? 1 : sum;
  48 }
  49
  50 void
  51 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
  52 {
  53         msg->ibm_type = type;
  54         msg->ibm_nob  = offsetof(kib_msg_t, ibm_u) + body_nob;
  55 }
  56
  57 void
  58 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg,
  59                  int credits, lnet_nid_t dstnid, __u64 dststamp)
  60 {
  61         kib_net_t *net = ni->ni_data;
  62
  63         /* CAVEAT EMPTOR! all message fields not set here should have been
  64          * initialised previously. */
  65         msg->ibm_magic    = IBLND_MSG_MAGIC;
  66         msg->ibm_version  = IBLND_MSG_VERSION;
  67         /*   ibm_type */
  68         msg->ibm_credits  = credits;
  69         /*   ibm_nob */
  70         msg->ibm_cksum    = 0;
  71         msg->ibm_srcnid   = lnet_ptlcompat_srcnid(ni->ni_nid, dstnid);
  72         msg->ibm_srcstamp = net->ibn_incarnation;
  73         msg->ibm_dstnid   = dstnid;
  74         msg->ibm_dststamp = dststamp;
  75
  76         if (*kiblnd_tunables.kib_cksum) {
  77                 /* NB ibm_cksum zero while computing cksum */
  78                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
  79         }
  80 }
  81
  82 int
  83 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
  84 {
  85         const int hdr_size = offsetof(kib_msg_t, ibm_u);
  86         __u32     msg_cksum;
  87         int       flip;
  88         int       msg_nob;
  89 #if !IBLND_MAP_ON_DEMAND
  90         int       i;
  91         int       n;
  92 #endif
  93         /* 6 bytes are enough to have received magic + version */
  94         if (nob < 6) {
  95                 CERROR("Short message: %d\n", nob);
  96                 return -EPROTO;
  97         }
  98
  99         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 100                 flip = 0;
 101         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 102                 flip = 1;
 103         } else {
 104                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 105                 return -EPROTO;
 106         }
 107
 108         if (msg->ibm_version !=
 109             (flip ? __swab16(IBLND_MSG_VERSION) : IBLND_MSG_VERSION)) {
 110                 CERROR("Bad version: %d\n", msg->ibm_version);
 111                 return -EPROTO;
 112         }
 113
 114         if (nob < hdr_size) {
 115                 CERROR("Short message: %d\n", nob);
 116                 return -EPROTO;
 117         }
 118
 119         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 120         if (msg_nob > nob) {
 121                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 122                 return -EPROTO;
 123         }
 124
 125         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 126          * gets flipped */
 127         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 128         msg->ibm_cksum = 0;
 129         if (msg_cksum != 0 &&
 130             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 131                 CERROR("Bad checksum\n");
 132                 return -EPROTO;
 133         }
 134         msg->ibm_cksum = msg_cksum;
 135
 136         if (flip) {
 137                 /* leave magic unflipped as a clue to peer endianness */
 138                 __swab16s(&msg->ibm_version);
 139                 CLASSERT (sizeof(msg->ibm_type) == 1);
 140                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 141                 msg->ibm_nob = msg_nob;
 142                 __swab64s(&msg->ibm_srcnid);
 143                 __swab64s(&msg->ibm_srcstamp);
 144                 __swab64s(&msg->ibm_dstnid);
 145                 __swab64s(&msg->ibm_dststamp);
 146         }
 147
 148         if (msg->ibm_srcnid == LNET_NID_ANY) {
 149                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 150                 return -EPROTO;
 151         }
 152
 153         switch (msg->ibm_type) {
 154         default:
 155                 CERROR("Unknown message type %x\n", msg->ibm_type);
 156                 return -EPROTO;
 157
 158         case IBLND_MSG_NOOP:
 159                 break;
 160
 161         case IBLND_MSG_IMMEDIATE:
 162                 if (msg_nob < offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0])) {
 163                         CERROR("Short IMMEDIATE: %d(%d)\n", msg_nob,
 164                                (int)offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]));
 165                         return -EPROTO;
 166                 }
 167                 break;
 168
 169         case IBLND_MSG_PUT_REQ:
 170                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.putreq)) {
 171                         CERROR("Short PUT_REQ: %d(%d)\n", msg_nob,
 172                                (int)(hdr_size + sizeof(msg->ibm_u.putreq)));
 173                         return -EPROTO;
 174                 }
 175                 break;
 176
 177         case IBLND_MSG_PUT_ACK:
 178                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.putack)) {
 179                         CERROR("Short PUT_ACK: %d(%d)\n", msg_nob,
 180                                (int)(hdr_size + sizeof(msg->ibm_u.putack)));
 181                         return -EPROTO;
 182                 }
 183 #if IBLND_MAP_ON_DEMAND
 184                 if (flip) {
 185                         __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_addr);
 186                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nob);
 187                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key);
 188                 }
 189 #else
 190                 if (flip) {
 191                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key);
 192                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nfrags);
 193                 }
 194
 195                 n = msg->ibm_u.putack.ibpam_rd.rd_nfrags;
 196                 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 197                         CERROR("Bad PUT_ACK nfrags: %d, should be 0 < n <= %d\n",
 198                                n, IBLND_MAX_RDMA_FRAGS);
 199                         return -EPROTO;
 200                 }
 201
 202                 if (msg_nob < offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n])) {
 203                         CERROR("Short PUT_ACK: %d(%d)\n", msg_nob,
 204                                (int)offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n]));
 205                         return -EPROTO;
 206                 }
 207
 208                 if (flip) {
 209                         for (i = 0; i < n; i++) {
 210                                 __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_nob);
 211                                 __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_addr);
 212                         }
 213                 }
 214 #endif
 215                 break;
 216
 217         case IBLND_MSG_GET_REQ:
 218                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.get)) {
 219                         CERROR("Short GET_REQ: %d(%d)\n", msg_nob,
 220                                (int)(hdr_size + sizeof(msg->ibm_u.get)));
 221                         return -EPROTO;
 222                 }
 223 #if IBLND_MAP_ON_DEMAND
 224                 if (flip) {
 225                         __swab64s(&msg->ibm_u.get.ibgm_rd.rd_addr);
 226                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nob);
 227                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key);
 228                 }
 229 #else
 230                 if (flip) {
 231                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key);
 232                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nfrags);
 233                 }
 234
 235                 n = msg->ibm_u.get.ibgm_rd.rd_nfrags;
 236                 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 237                         CERROR("Bad GET_REQ nfrags: %d, should be 0 < n <= %d\n",
 238                                n, IBLND_MAX_RDMA_FRAGS);
 239                         return -EPROTO;
 240                 }
 241
 242                 if (msg_nob < offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n])) {
 243                         CERROR("Short GET_REQ: %d(%d)\n", msg_nob,
 244                                (int)offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n]));
 245                         return -EPROTO;
 246                 }
 247
 248                 if (flip)
 249                         for (i = 0; i < msg->ibm_u.get.ibgm_rd.rd_nfrags; i++) {
 250                                 __swab32s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_nob);
 251                                 __swab64s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_addr);
 252                         }
 253 #endif
 254                 break;
 255
 256         case IBLND_MSG_PUT_NAK:
 257         case IBLND_MSG_PUT_DONE:
 258         case IBLND_MSG_GET_DONE:
 259                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.completion)) {
 260                         CERROR("Short RDMA completion: %d(%d)\n", msg_nob,
 261                                (int)(hdr_size + sizeof(msg->ibm_u.completion)));
 262                         return -EPROTO;
 263                 }
 264                 if (flip)
 265                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 266                 break;
 267
 268         case IBLND_MSG_CONNREQ:
 269         case IBLND_MSG_CONNACK:
 270                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.connparams)) {
 271                         CERROR("Short connreq/ack: %d(%d)\n", msg_nob,
 272                                (int)(hdr_size + sizeof(msg->ibm_u.connparams)));
 273                         return -EPROTO;
 274                 }
 275                 if (flip) {
 276                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 277                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 278                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 279                 }
 280                 break;
 281         }
 282         return 0;
 283 }
 284
 285 int
 286 kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 287 {
 288         kib_peer_t     *peer;
 289         kib_net_t      *net = ni->ni_data;
 290         unsigned long   flags;
 291
 292         LASSERT (net != NULL);
 293         LASSERT (nid != LNET_NID_ANY);
 294
 295         LIBCFS_ALLOC(peer, sizeof(*peer));
 296         if (peer == NULL) {
 297                 CERROR("Cannot allocate peer\n");
 298                 return -ENOMEM;
 299         }
 300
 301         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 302
 303         peer->ibp_ni = ni;
 304         peer->ibp_nid = nid;
 305         peer->ibp_error = 0;
 306         peer->ibp_last_alive = cfs_time_current();
 307         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 308
 309         INIT_LIST_HEAD(&peer->ibp_list);       /* not in the peer table yet */
 310         INIT_LIST_HEAD(&peer->ibp_conns);
 311         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 312
 313         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 314
 315         /* always called with a ref on ni, which prevents ni being shutdown */
 316         LASSERT (net->ibn_shutdown == 0);
 317
 318         /* npeers only grows with the global lock held */
 319         atomic_inc(&net->ibn_npeers);
 320
 321         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 322
 323         *peerp = peer;
 324         return 0;
 325 }
 326
 327 void
 328 kiblnd_destroy_peer (kib_peer_t *peer)
 329 {
 330         kib_net_t *net = peer->ibp_ni->ni_data;
 331
 332         LASSERT (net != NULL);
 333         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 334         LASSERT (!kiblnd_peer_active(peer));
 335         LASSERT (peer->ibp_connecting == 0);
 336         LASSERT (peer->ibp_accepting == 0);
 337         LASSERT (list_empty(&peer->ibp_conns));
 338         LASSERT (list_empty(&peer->ibp_tx_queue));
 339
 340         LIBCFS_FREE(peer, sizeof(*peer));
 341
 342         /* NB a peer's connections keep a reference on their peer until
 343          * they are destroyed, so we can be assured that _all_ state to do
 344          * with this peer has been cleaned up when its refcount drops to
 345          * zero. */
 346         atomic_dec(&net->ibn_npeers);
 347 }
 348
 349 void
 350 kiblnd_destroy_dev (kib_dev_t *dev)
 351 {
 352         LASSERT (dev->ibd_nnets == 0);
 353
 354         if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
 355                 list_del_init(&dev->ibd_list);
 356
 357         if (dev->ibd_mr != NULL)
 358                 ib_dereg_mr(dev->ibd_mr);
 359
 360         if (dev->ibd_pd != NULL)
 361                 ib_dealloc_pd(dev->ibd_pd);
 362
 363         if (dev->ibd_cmid != NULL)
 364                 rdma_destroy_id(dev->ibd_cmid);
 365
 366         LIBCFS_FREE(dev, sizeof(*dev));
 367 }
 368
 369 kib_peer_t *
 370 kiblnd_find_peer_locked (lnet_nid_t nid)
 371 {
 372         /* the caller is responsible for accounting the additional reference
 373          * that this creates */
 374         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 375         struct list_head *tmp;
 376         kib_peer_t       *peer;
 377
 378         list_for_each (tmp, peer_list) {
 379
 380                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 381
 382                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 383                          peer->ibp_accepting > 0 ||
 384                          !list_empty(&peer->ibp_conns));  /* active conn */
 385
 386                 if (peer->ibp_nid != nid)
 387                         continue;
 388
 389                 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
 390                        peer, libcfs_nid2str(nid),
 391                        atomic_read(&peer->ibp_refcount));
 392                 return peer;
 393         }
 394         return NULL;
 395 }
 396
 397 void
 398 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 399 {
 400         LASSERT (list_empty(&peer->ibp_conns));
 401
 402         LASSERT (kiblnd_peer_active(peer));
 403         list_del_init(&peer->ibp_list);
 404         /* lose peerlist's ref */
 405         kiblnd_peer_decref(peer);
 406 }
 407
 408 int
 409 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 410                       lnet_nid_t *nidp, int *count)
 411 {
 412         kib_peer_t        *peer;
 413         struct list_head  *ptmp;
 414         int                i;
 415         unsigned long      flags;
 416
 417         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 418
 419         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 420
 421                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 422
 423                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 424                         LASSERT (peer->ibp_connecting > 0 ||
 425                                  peer->ibp_accepting > 0 ||
 426                                  !list_empty(&peer->ibp_conns));
 427
 428                         if (peer->ibp_ni != ni)
 429                                 continue;
 430
 431                         if (index-- > 0)
 432                                 continue;
 433
 434                         *nidp = peer->ibp_nid;
 435                         *count = atomic_read(&peer->ibp_refcount);
 436
 437                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 438                                                flags);
 439                         return 0;
 440                 }
 441         }
 442
 443         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 444         return -ENOENT;
 445 }
 446
 447 void
 448 kiblnd_del_peer_locked (kib_peer_t *peer)
 449 {
 450         struct list_head *ctmp;
 451         struct list_head *cnxt;
 452         kib_conn_t       *conn;
 453
 454         if (list_empty(&peer->ibp_conns)) {
 455                 kiblnd_unlink_peer_locked(peer);
 456         } else {
 457                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 458                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 459
 460                         kiblnd_close_conn_locked(conn, 0);
 461                 }
 462                 /* NB closing peer's last conn unlinked it. */
 463         }
 464         /* NB peer now unlinked; might even be freed if the peer table had the
 465          * last ref on it. */
 466 }
 467
 468 int
 469 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 470 {
 471         CFS_LIST_HEAD     (zombies);
 472         struct list_head  *ptmp;
 473         struct list_head  *pnxt;
 474         kib_peer_t        *peer;
 475         int                lo;
 476         int                hi;
 477         int                i;
 478         unsigned long      flags;
 479         int                rc = -ENOENT;
 480
 481         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 482
 483         if (nid != LNET_NID_ANY) {
 484                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 485         } else {
 486                 lo = 0;
 487                 hi = kiblnd_data.kib_peer_hash_size - 1;
 488         }
 489
 490         for (i = lo; i <= hi; i++) {
 491                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 492                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 493                         LASSERT (peer->ibp_connecting > 0 ||
 494                                  peer->ibp_accepting > 0 ||
 495                                  !list_empty(&peer->ibp_conns));
 496
 497                         if (peer->ibp_ni != ni)
 498                                 continue;
 499
 500                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 501                                 continue;
 502
 503                         if (!list_empty(&peer->ibp_tx_queue)) {
 504                                 LASSERT (list_empty(&peer->ibp_conns));
 505
 506                                 list_splice_init(&peer->ibp_tx_queue, &zombies);
 507                         }
 508
 509                         kiblnd_del_peer_locked(peer);
 510                         rc = 0;         /* matched something */
 511                 }
 512         }
 513
 514         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 515
 516         kiblnd_txlist_done(ni, &zombies, -EIO);
 517
 518         return rc;
 519 }
 520
 521 kib_conn_t *
 522 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 523 {
 524         kib_peer_t        *peer;
 525         struct list_head  *ptmp;
 526         kib_conn_t        *conn;
 527         struct list_head  *ctmp;
 528         int                i;
 529         unsigned long      flags;
 530
 531         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 532
 533         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 534                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 535
 536                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 537                         LASSERT (peer->ibp_connecting > 0 ||
 538                                  peer->ibp_accepting > 0 ||
 539                                  !list_empty(&peer->ibp_conns));
 540
 541                         if (peer->ibp_ni != ni)
 542                                 continue;
 543
 544                         list_for_each (ctmp, &peer->ibp_conns) {
 545                                 if (index-- > 0)
 546                                         continue;
 547
 548                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 549                                 kiblnd_conn_addref(conn);
 550                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 551                                                        flags);
 552                                 return conn;
 553                         }
 554                 }
 555         }
 556
 557         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 558         return NULL;
 559 }
 560
 561 void
 562 kiblnd_debug_rx (kib_rx_t *rx)
 563 {
 564         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 565                rx, rx->rx_status, rx->rx_msg->ibm_type,
 566                rx->rx_msg->ibm_credits);
 567 }
 568
 569 void
 570 kiblnd_debug_tx (kib_tx_t *tx)
 571 {
 572         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 573                "cookie "LPX64" msg %s%s type %x cred %d\n",
 574                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 575                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 576                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 577                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 578                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 579 }
 580
 581 void
 582 kiblnd_debug_conn (kib_conn_t *conn)
 583 {
 584         struct list_head *tmp;
 585         int               i;
 586
 587         spin_lock(&conn->ibc_lock);
 588
 589         CDEBUG(D_CONSOLE, "conn[%d] %p -> %s: \n",
 590                atomic_read(&conn->ibc_refcount), conn,
 591                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 592         CDEBUG(D_CONSOLE, "   state %d nposted %d cred %d o_cred %d r_cred %d\n",
 593                conn->ibc_state, conn->ibc_nsends_posted, conn->ibc_credits,
 594                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 595         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 596
 597         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 598         list_for_each(tmp, &conn->ibc_early_rxs)
 599                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 600
 601         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 602         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 603                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 604
 605         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 606         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 607                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 608
 609         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 610         list_for_each(tmp, &conn->ibc_tx_queue)
 611                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 612
 613         CDEBUG(D_CONSOLE, "   active_txs:\n");
 614         list_for_each(tmp, &conn->ibc_active_txs)
 615                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 616
 617         CDEBUG(D_CONSOLE, "   rxs:\n");
 618         for (i = 0; i < IBLND_RX_MSGS; i++)
 619                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 620
 621         spin_unlock(&conn->ibc_lock);
 622 }
 623
 624 kib_conn_t *
 625 kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid, int state)
 626 {
 627         /* CAVEAT EMPTOR:
 628          * If the new conn is created successfully it takes over the caller's
 629          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 630          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 631          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 632          * to destroy 'cmid' here since I'm called from the CM which still has
 633          * its ref on 'cmid'). */
 634         kib_conn_t             *conn;
 635         kib_net_t              *net = peer->ibp_ni->ni_data;
 636         int                     i;
 637         int                     page_offset;
 638         int                     ipage;
 639         int                     rc;
 640         struct ib_cq           *cq;
 641         struct ib_qp_init_attr *init_qp_attr;
 642         unsigned long           flags;
 643
 644         LASSERT (net != NULL);
 645         LASSERT (!in_interrupt());
 646
 647         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 648         if (init_qp_attr == NULL) {
 649                 CERROR("Can't allocate qp_attr for %s\n",
 650                        libcfs_nid2str(peer->ibp_nid));
 651                 goto failed_0;
 652         }
 653
 654         LIBCFS_ALLOC(conn, sizeof(*conn));
 655         if (conn == NULL) {
 656                 CERROR("Can't allocate connection for %s\n",
 657                        libcfs_nid2str(peer->ibp_nid));
 658                 goto failed_1;
 659         }
 660
 661         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 662
 663         conn->ibc_state = IBLND_CONN_INIT;
 664         conn->ibc_peer = peer;                  /* I take the caller's ref */
 665         cmid->context = conn;                   /* for future CM callbacks */
 666         conn->ibc_cmid = cmid;
 667
 668         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 669         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 670         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 671         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 672         INIT_LIST_HEAD(&conn->ibc_active_txs);
 673         spin_lock_init(&conn->ibc_lock);
 674
 675         LIBCFS_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 676         if (conn->ibc_connvars == NULL) {
 677                 CERROR("Can't allocate in-progress connection state\n");
 678                 goto failed_2;
 679         }
 680         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 681
 682         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS * sizeof(kib_rx_t));
 683         if (conn->ibc_rxs == NULL) {
 684                 CERROR("Cannot allocate RX buffers\n");
 685                 goto failed_2;
 686         }
 687         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS * sizeof(kib_rx_t));
 688
 689         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, IBLND_RX_MSG_PAGES);
 690         if (rc != 0)
 691                 goto failed_2;
 692
 693         for (i = ipage = page_offset = 0; i < IBLND_RX_MSGS; i++) {
 694                 struct page *page = conn->ibc_rx_pages->ibp_pages[ipage];
 695                 kib_rx_t    *rx = &conn->ibc_rxs[i];
 696
 697                 rx->rx_conn = conn;
 698                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) +
 699                                            page_offset);
 700                 rx->rx_msgaddr = kiblnd_dma_map_single(cmid->device,
 701                                                        rx->rx_msg, IBLND_MSG_SIZE,
 702                                                        DMA_FROM_DEVICE);
 703                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
 704
 705                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
 706                        i, rx->rx_msg, rx->rx_msgaddr,
 707                        lnet_page2phys(page) + page_offset);
 708
 709                 page_offset += IBLND_MSG_SIZE;
 710                 LASSERT (page_offset <= PAGE_SIZE);
 711
 712                 if (page_offset == PAGE_SIZE) {
 713                         page_offset = 0;
 714                         ipage++;
 715                         LASSERT (ipage <= IBLND_RX_MSG_PAGES);
 716                 }
 717         }
 718
 719 #if (IBLND_OFED_VERSION == 1025)
 720         cq = ib_create_cq(cmid->device,
 721                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 722                           IBLND_CQ_ENTRIES(), 0);
 723 #else
 724         cq = ib_create_cq(cmid->device,
 725                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 726                           IBLND_CQ_ENTRIES());
 727 #endif
 728         if (!IS_ERR(cq)) {
 729                 conn->ibc_cq = cq;
 730         } else {
 731                 CERROR("Can't create CQ: %ld\n", PTR_ERR(cq));
 732                 goto failed_2;
 733         }
 734
 735         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 736         if (rc != 0) {
 737                 CERROR("Can't request completion notificiation: %d\n", rc);
 738                 goto failed_2;
 739         }
 740
 741         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 742         init_qp_attr->event_handler = kiblnd_qp_event;
 743         init_qp_attr->qp_context = conn;
 744         init_qp_attr->cap.max_send_wr = (*kiblnd_tunables.kib_concurrent_sends) *
 745                                         (1 + IBLND_MAX_RDMA_FRAGS);
 746         init_qp_attr->cap.max_recv_wr = IBLND_RX_MSGS;
 747         init_qp_attr->cap.max_send_sge = 1;
 748         init_qp_attr->cap.max_recv_sge = 1;
 749         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 750         init_qp_attr->qp_type = IB_QPT_RC;
 751         init_qp_attr->send_cq = cq;
 752         init_qp_attr->recv_cq = cq;
 753
 754         rc = 0;
 755         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 756         switch (*kiblnd_tunables.kib_ib_mtu) {
 757         default:
 758                 rc = *kiblnd_tunables.kib_ib_mtu;
 759                 /* fall through to... */
 760         case 0: /* set tunable to the default
 761                  * CAVEAT EMPTOR! this assumes the default is one of the MTUs
 762                  * below, otherwise we'll WARN on the next QP create */
 763                 *kiblnd_tunables.kib_ib_mtu =
 764                         ib_mtu_enum_to_int(cmid->route.path_rec->mtu);
 765                 break;
 766         case 256:
 767                 cmid->route.path_rec->mtu = IB_MTU_256;
 768                 break;
 769         case 512:
 770                 cmid->route.path_rec->mtu = IB_MTU_512;
 771                 break;
 772         case 1024:
 773                 cmid->route.path_rec->mtu = IB_MTU_1024;
 774                 break;
 775         case 2048:
 776                 cmid->route.path_rec->mtu = IB_MTU_2048;
 777                 break;
 778         case 4096:
 779                 cmid->route.path_rec->mtu = IB_MTU_4096;
 780                 break;
 781         }
 782         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 783
 784         if (rc != 0)
 785                 CWARN("Invalid IB MTU value %d, using default value %d\n",
 786                       rc, *kiblnd_tunables.kib_ib_mtu);
 787
 788         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 789         if (rc != 0) {
 790                 CERROR("Can't create QP: %d\n", rc);
 791                 goto failed_2;
 792         }
 793
 794         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 795
 796         /* 1 ref for caller and each rxmsg */
 797         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS);
 798         conn->ibc_nrx = IBLND_RX_MSGS;
 799
 800         /* post receives */
 801         for (i = 0; i < IBLND_RX_MSGS; i++) {
 802                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 803                                     IBLND_POSTRX_NO_CREDIT);
 804                 if (rc != 0) {
 805                         CERROR("Can't post rxmsg: %d\n", rc);
 806
 807                         /* Make posted receives complete */
 808                         kiblnd_abort_receives(conn);
 809
 810                         /* correct # of posted buffers
 811                          * NB locking needed now I'm racing with completion */
 812                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
 813                         conn->ibc_nrx -= IBLND_RX_MSGS - i;
 814                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 815                                                flags);
 816
 817                         /* Drop my own and unused rxbuffer refcounts */
 818                         while (i++ <= IBLND_RX_MSGS)
 819                                 kiblnd_conn_decref(conn);
 820
 821                         return NULL;
 822                 }
 823         }
 824
 825         /* Init successful! */
 826         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 827                  state == IBLND_CONN_PASSIVE_WAIT);
 828         conn->ibc_state = state;
 829
 830         /* 1 more conn */
 831         atomic_inc(&net->ibn_nconns);
 832         return conn;
 833
 834  failed_2:
 835         kiblnd_destroy_conn(conn);
 836  failed_1:
 837         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 838  failed_0:
 839         return NULL;
 840 }
 841
 842 void
 843 kiblnd_destroy_conn (kib_conn_t *conn)
 844 {
 845         struct rdma_cm_id *cmid = conn->ibc_cmid;
 846         kib_peer_t        *peer = conn->ibc_peer;
 847         int                rc;
 848         int                i;
 849
 850         LASSERT (!in_interrupt());
 851         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 852         LASSERT (list_empty(&conn->ibc_early_rxs));
 853         LASSERT (list_empty(&conn->ibc_tx_queue));
 854         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 855         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 856         LASSERT (list_empty(&conn->ibc_active_txs));
 857         LASSERT (conn->ibc_nsends_posted == 0);
 858
 859         switch (conn->ibc_state) {
 860         default:
 861                 /* conn must be completely disengaged from the network */
 862                 LBUG();
 863
 864         case IBLND_CONN_DISCONNECTED:
 865                 /* connvars should have been freed already */
 866                 LASSERT (conn->ibc_connvars == NULL);
 867                 break;
 868
 869         case IBLND_CONN_INIT:
 870                 break;
 871         }
 872
 873         if (conn->ibc_cmid->qp != NULL)
 874                 rdma_destroy_qp(conn->ibc_cmid);
 875
 876         if (conn->ibc_cq != NULL) {
 877                 rc = ib_destroy_cq(conn->ibc_cq);
 878                 if (rc != 0)
 879                         CWARN("Error destroying CQ: %d\n", rc);
 880         }
 881
 882         if (conn->ibc_rx_pages != NULL) {
 883                 LASSERT (conn->ibc_rxs != NULL);
 884
 885                 for (i = 0; i < IBLND_RX_MSGS; i++) {
 886                         kib_rx_t *rx = &conn->ibc_rxs[i];
 887
 888                         LASSERT (rx->rx_nob >= 0); /* not posted */
 889
 890                         kiblnd_dma_unmap_single(conn->ibc_cmid->device,
 891                                                 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
 892                                                                   rx->rx_msgaddr),
 893                                                 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
 894                 }
 895
 896                 kiblnd_free_pages(conn->ibc_rx_pages);
 897         }
 898
 899         if (conn->ibc_rxs != NULL) {
 900                 LIBCFS_FREE(conn->ibc_rxs,
 901                             IBLND_RX_MSGS * sizeof(kib_rx_t));
 902         }
 903
 904         if (conn->ibc_connvars != NULL)
 905                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 906
 907         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 908         if (conn->ibc_state != IBLND_CONN_INIT) {
 909                 kib_net_t *net = peer->ibp_ni->ni_data;
 910
 911                 kiblnd_peer_decref(peer);
 912                 rdma_destroy_id(cmid);
 913                 atomic_dec(&net->ibn_nconns);
 914         }
 915
 916         LIBCFS_FREE(conn, sizeof(*conn));
 917 }
 918
 919 int
 920 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 921 {
 922         kib_conn_t         *conn;
 923         struct list_head   *ctmp;
 924         struct list_head   *cnxt;
 925         int                 count = 0;
 926
 927         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 928                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 929
 930                 count++;
 931                 kiblnd_close_conn_locked(conn, why);
 932         }
 933
 934         return count;
 935 }
 936
 937 int
 938 kiblnd_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation)
 939 {
 940         kib_conn_t         *conn;
 941         struct list_head   *ctmp;
 942         struct list_head   *cnxt;
 943         int                 count = 0;
 944
 945         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 946                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 947
 948                 if (conn->ibc_incarnation == incarnation)
 949                         continue;
 950
 951                 CDEBUG(D_NET, "Closing stale conn -> %s incarnation:"LPX64"("LPX64")\n",
 952                        libcfs_nid2str(peer->ibp_nid),
 953                        conn->ibc_incarnation, incarnation);
 954
 955                 count++;
 956                 kiblnd_close_conn_locked(conn, -ESTALE);
 957         }
 958
 959         return count;
 960 }
 961
 962 int
 963 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 964 {
 965         kib_peer_t         *peer;
 966         struct list_head   *ptmp;
 967         struct list_head   *pnxt;
 968         int                 lo;
 969         int                 hi;
 970         int                 i;
 971         unsigned long       flags;
 972         int                 count = 0;
 973
 974         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 975
 976         if (nid != LNET_NID_ANY)
 977                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 978         else {
 979                 lo = 0;
 980                 hi = kiblnd_data.kib_peer_hash_size - 1;
 981         }
 982
 983         for (i = lo; i <= hi; i++) {
 984                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 985
 986                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 987                         LASSERT (peer->ibp_connecting > 0 ||
 988                                  peer->ibp_accepting > 0 ||
 989                                  !list_empty(&peer->ibp_conns));
 990
 991                         if (peer->ibp_ni != ni)
 992                                 continue;
 993
 994                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 995                                 continue;
 996
 997                         count += kiblnd_close_peer_conns_locked(peer, 0);
 998                 }
 999         }
1000
1001         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1002
1003         /* wildcards always succeed */
1004         if (nid == LNET_NID_ANY)
1005                 return 0;
1006
1007         return (count == 0) ? -ENOENT : 0;
1008 }
1009
1010 int
1011 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1012 {
1013         struct libcfs_ioctl_data *data = arg;
1014         int                       rc = -EINVAL;
1015
1016         switch(cmd) {
1017         case IOC_LIBCFS_GET_PEER: {
1018                 lnet_nid_t   nid = 0;
1019                 int          count = 0;
1020
1021                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1022                                           &nid, &count);
1023                 data->ioc_nid    = nid;
1024                 data->ioc_count  = count;
1025                 break;
1026         }
1027
1028         case IOC_LIBCFS_DEL_PEER: {
1029                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1030                 break;
1031         }
1032         case IOC_LIBCFS_GET_CONN: {
1033                 kib_conn_t *conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1034
1035                 if (conn == NULL) {
1036                         rc = -ENOENT;
1037                 } else {
1038                         // kiblnd_debug_conn(conn);
1039                         rc = 0;
1040                         data->ioc_nid = conn->ibc_peer->ibp_nid;
1041                         kiblnd_conn_decref(conn);
1042                 }
1043                 break;
1044         }
1045         case IOC_LIBCFS_CLOSE_CONNECTION: {
1046                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1047                 break;
1048         }
1049
1050         default:
1051                 break;
1052         }
1053
1054         return rc;
1055 }
1056
1057 void
1058 kiblnd_free_pages (kib_pages_t *p)
1059 {
1060         int         npages = p->ibp_npages;
1061         int         i;
1062
1063         for (i = 0; i < npages; i++)
1064                 if (p->ibp_pages[i] != NULL)
1065                         __free_page(p->ibp_pages[i]);
1066
1067         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1068 }
1069
1070 int
1071 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1072 {
1073         kib_pages_t   *p;
1074         int            i;
1075
1076         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1077         if (p == NULL) {
1078                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1079                 return -ENOMEM;
1080         }
1081
1082         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1083         p->ibp_npages = npages;
1084
1085         for (i = 0; i < npages; i++) {
1086                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1087                 if (p->ibp_pages[i] == NULL) {
1088                         CERROR("Can't allocate page %d of %d\n", i, npages);
1089                         kiblnd_free_pages(p);
1090                         return -ENOMEM;
1091                 }
1092         }
1093
1094         *pp = p;
1095         return 0;
1096 }
1097
1098 void
1099 kiblnd_free_tx_descs (lnet_ni_t *ni)
1100 {
1101         int        i;
1102         kib_net_t *net = ni->ni_data;
1103
1104         LASSERT (net != NULL);
1105
1106         if (net->ibn_tx_descs != NULL) {
1107                 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1108                         kib_tx_t *tx = &net->ibn_tx_descs[i];
1109
1110 #if IBLND_MAP_ON_DEMAND
1111                         if (tx->tx_pages != NULL)
1112                                 LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV *
1113                                             sizeof(*tx->tx_pages));
1114 #else
1115                         if (tx->tx_wrq != NULL)
1116                                 LIBCFS_FREE(tx->tx_wrq,
1117                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1118                                             sizeof(*tx->tx_wrq));
1119
1120                         if (tx->tx_sge != NULL)
1121                                 LIBCFS_FREE(tx->tx_sge,
1122                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1123                                             sizeof(*tx->tx_sge));
1124
1125                         if (tx->tx_rd != NULL)
1126                                 LIBCFS_FREE(tx->tx_rd,
1127                                             offsetof(kib_rdma_desc_t,
1128                                                rd_frags[IBLND_MAX_RDMA_FRAGS]));
1129
1130                         if (tx->tx_frags != NULL)
1131                                 LIBCFS_FREE(tx->tx_frags,
1132                                             IBLND_MAX_RDMA_FRAGS *
1133                                             sizeof(*tx->tx_frags));
1134 #endif
1135                 }
1136
1137                 LIBCFS_FREE(net->ibn_tx_descs,
1138                             IBLND_TX_MSGS() * sizeof(kib_tx_t));
1139         }
1140
1141         if (net->ibn_tx_pages != NULL)
1142                 kiblnd_free_pages(net->ibn_tx_pages);
1143 }
1144
1145 int
1146 kiblnd_alloc_tx_descs (lnet_ni_t *ni)
1147 {
1148         int        i;
1149         int        rc;
1150         kib_net_t *net = ni->ni_data;
1151
1152         LASSERT (net != NULL);
1153
1154         rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES());
1155
1156         if (rc != 0) {
1157                 CERROR("Can't allocate tx pages\n");
1158                 return rc;
1159         }
1160
1161         LIBCFS_ALLOC (net->ibn_tx_descs,
1162                       IBLND_TX_MSGS() * sizeof(kib_tx_t));
1163         if (net->ibn_tx_descs == NULL) {
1164                 CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS());
1165                 return -ENOMEM;
1166         }
1167
1168         memset(net->ibn_tx_descs, 0,
1169                IBLND_TX_MSGS() * sizeof(kib_tx_t));
1170
1171         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1172                 kib_tx_t *tx = &net->ibn_tx_descs[i];
1173
1174 #if IBLND_MAP_ON_DEMAND
1175                 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1176                              sizeof(*tx->tx_pages));
1177                 if (tx->tx_pages == NULL) {
1178                         CERROR("Can't allocate phys page vector[%d]\n",
1179                                LNET_MAX_IOV);
1180                         return -ENOMEM;
1181                 }
1182 #else
1183                 LIBCFS_ALLOC(tx->tx_wrq,
1184                              (1 + IBLND_MAX_RDMA_FRAGS) *
1185                              sizeof(*tx->tx_wrq));
1186                 if (tx->tx_wrq == NULL)
1187                         return -ENOMEM;
1188
1189                 LIBCFS_ALLOC(tx->tx_sge,
1190                              (1 + IBLND_MAX_RDMA_FRAGS) *
1191                              sizeof(*tx->tx_sge));
1192                 if (tx->tx_sge == NULL)
1193                         return -ENOMEM;
1194
1195                 LIBCFS_ALLOC(tx->tx_rd,
1196                              offsetof(kib_rdma_desc_t,
1197                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1198                 if (tx->tx_rd == NULL)
1199                         return -ENOMEM;
1200
1201                 LIBCFS_ALLOC(tx->tx_frags,
1202                              IBLND_MAX_RDMA_FRAGS *
1203                              sizeof(*tx->tx_frags));
1204                 if (tx->tx_frags == NULL)
1205                         return -ENOMEM;
1206 #endif
1207         }
1208
1209         return 0;
1210 }
1211
1212 void
1213 kiblnd_unmap_tx_descs (lnet_ni_t *ni)
1214 {
1215         int             i;
1216         kib_tx_t       *tx;
1217         kib_net_t      *net = ni->ni_data;
1218
1219         LASSERT (net != NULL);
1220
1221         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1222                 tx = &net->ibn_tx_descs[i];
1223
1224                 kiblnd_dma_unmap_single(net->ibn_dev->ibd_cmid->device,
1225                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1226                                                           tx->tx_msgaddr),
1227                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1228         }
1229 }
1230
1231 void
1232 kiblnd_map_tx_descs (lnet_ni_t *ni)
1233 {
1234         int             ipage = 0;
1235         int             page_offset = 0;
1236         int             i;
1237         struct page    *page;
1238         kib_tx_t       *tx;
1239         kib_net_t      *net = ni->ni_data;
1240
1241         LASSERT (net != NULL);
1242
1243         /* pre-mapped messages are not bigger than 1 page */
1244         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1245
1246         /* No fancy arithmetic when we do the buffer calculations */
1247         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1248
1249         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1250                 page = net->ibn_tx_pages->ibp_pages[ipage];
1251                 tx = &net->ibn_tx_descs[i];
1252
1253                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1254                                            page_offset);
1255
1256                 tx->tx_msgaddr = kiblnd_dma_map_single(
1257                         net->ibn_dev->ibd_cmid->device,
1258                         tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE);
1259                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1260
1261                 list_add(&tx->tx_list, &net->ibn_idle_txs);
1262
1263                 page_offset += IBLND_MSG_SIZE;
1264                 LASSERT (page_offset <= PAGE_SIZE);
1265
1266                 if (page_offset == PAGE_SIZE) {
1267                         page_offset = 0;
1268                         ipage++;
1269                         LASSERT (ipage <= IBLND_TX_MSG_PAGES());
1270                 }
1271         }
1272 }
1273
1274 void
1275 kiblnd_base_shutdown (void)
1276 {
1277         int i;
1278
1279         LASSERT (list_empty(&kiblnd_data.kib_devs));
1280
1281         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
1282                atomic_read(&libcfs_kmemory));
1283
1284         switch (kiblnd_data.kib_init) {
1285         default:
1286                 LBUG();
1287
1288         case IBLND_INIT_ALL:
1289         case IBLND_INIT_DATA:
1290                 LASSERT (kiblnd_data.kib_peers != NULL);
1291                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
1292                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
1293                 }
1294                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
1295                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
1296
1297                 /* flag threads to terminate; wake and wait for them to die */
1298                 kiblnd_data.kib_shutdown = 1;
1299                 wake_up_all(&kiblnd_data.kib_sched_waitq);
1300                 wake_up_all(&kiblnd_data.kib_connd_waitq);
1301
1302                 i = 2;
1303                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
1304                         i++;
1305                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1306                                "Waiting for %d threads to terminate\n",
1307                                atomic_read(&kiblnd_data.kib_nthreads));
1308                         cfs_pause(cfs_time_seconds(1));
1309                 }
1310
1311                 /* fall through */
1312
1313         case IBLND_INIT_NOTHING:
1314                 break;
1315         }
1316
1317         if (kiblnd_data.kib_peers != NULL)
1318                 LIBCFS_FREE(kiblnd_data.kib_peers,
1319                             sizeof(struct list_head) *
1320                             kiblnd_data.kib_peer_hash_size);
1321
1322         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
1323                atomic_read(&libcfs_kmemory));
1324
1325         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
1326         PORTAL_MODULE_UNUSE;
1327 }
1328
1329 void
1330 kiblnd_shutdown (lnet_ni_t *ni)
1331 {
1332         kib_net_t        *net = ni->ni_data;
1333         rwlock_t         *g_lock = &kiblnd_data.kib_global_lock;
1334         int               i;
1335         unsigned long     flags;
1336
1337         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
1338
1339         if (net == NULL)
1340                 goto out;
1341
1342         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
1343                atomic_read(&libcfs_kmemory));
1344
1345         write_lock_irqsave(g_lock, flags);
1346         net->ibn_shutdown = 1;
1347         write_unlock_irqrestore(g_lock, flags);
1348
1349         switch (net->ibn_init) {
1350         default:
1351                 LBUG();
1352
1353         case IBLND_INIT_ALL:
1354                 /* nuke all existing peers within this net */
1355                 kiblnd_del_peer(ni, LNET_NID_ANY);
1356
1357                 /* Wait for all peer state to clean up */
1358                 i = 2;
1359                 while (atomic_read(&net->ibn_npeers) != 0) {
1360                         i++;
1361                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
1362                                "%s: waiting for %d peers to disconnect\n",
1363                                libcfs_nid2str(ni->ni_nid),
1364                                atomic_read(&net->ibn_npeers));
1365                         cfs_pause(cfs_time_seconds(1));
1366                 }
1367
1368                 kiblnd_unmap_tx_descs(ni);
1369
1370                 LASSERT (net->ibn_dev->ibd_nnets > 0);
1371                 net->ibn_dev->ibd_nnets--;
1372
1373                 /* fall through */
1374
1375         case IBLND_INIT_NOTHING:
1376                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
1377
1378 #if IBLND_MAP_ON_DEMAND
1379                 if (net->ibn_fmrpool != NULL)
1380                         ib_destroy_fmr_pool(net->ibn_fmrpool);
1381 #endif
1382                 if (net->ibn_dev != NULL &&
1383                     net->ibn_dev->ibd_nnets == 0)
1384                         kiblnd_destroy_dev(net->ibn_dev);
1385
1386                 break;
1387         }
1388
1389         kiblnd_free_tx_descs(ni);
1390
1391         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
1392                atomic_read(&libcfs_kmemory));
1393
1394         net->ibn_init = IBLND_INIT_NOTHING;
1395         ni->ni_data = NULL;
1396
1397         LIBCFS_FREE(net, sizeof(*net));
1398
1399 out:
1400         if (list_empty(&kiblnd_data.kib_devs))
1401                 kiblnd_base_shutdown();
1402         return;
1403 }
1404
1405 int
1406 kiblnd_base_startup (void)
1407 {
1408         int               rc;
1409         int               i;
1410
1411         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
1412
1413         if (*kiblnd_tunables.kib_credits > *kiblnd_tunables.kib_ntx) {
1414                 CERROR("Can't set credits(%d) > ntx(%d)\n",
1415                        *kiblnd_tunables.kib_credits,
1416                        *kiblnd_tunables.kib_ntx);
1417                 return -EINVAL;
1418         }
1419
1420         PORTAL_MODULE_USE;
1421         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
1422
1423         rwlock_init(&kiblnd_data.kib_global_lock);
1424
1425         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
1426
1427         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
1428         LIBCFS_ALLOC(kiblnd_data.kib_peers,
1429                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
1430         if (kiblnd_data.kib_peers == NULL) {
1431                 goto failed;
1432         }
1433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
1434                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
1435
1436         spin_lock_init(&kiblnd_data.kib_connd_lock);
1437         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
1438         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
1439         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
1440
1441         spin_lock_init(&kiblnd_data.kib_sched_lock);
1442         INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
1443         init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
1444
1445         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
1446
1447         /* lists/ptrs/locks initialised */
1448         kiblnd_data.kib_init = IBLND_INIT_DATA;
1449         /*****************************************************/
1450
1451         for (i = 0; i < IBLND_N_SCHED; i++) {
1452                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
1453                 if (rc != 0) {
1454                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
1455                                i, rc);
1456                         goto failed;
1457                 }
1458         }
1459
1460         rc = kiblnd_thread_start(kiblnd_connd, NULL);
1461         if (rc != 0) {
1462                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
1463                 goto failed;
1464         }
1465
1466         /* flag everything initialised */
1467         kiblnd_data.kib_init = IBLND_INIT_ALL;
1468         /*****************************************************/
1469
1470         return 0;
1471
1472  failed:
1473         kiblnd_base_shutdown();
1474         return -ENETDOWN;
1475 }
1476
1477 int
1478 kiblnd_startup (lnet_ni_t *ni)
1479 {
1480         char                     *ifname;
1481         kib_net_t                *net;
1482         kib_dev_t                *ibdev;
1483         struct list_head         *tmp;
1484         struct timeval            tv;
1485         int                       rc;
1486
1487         LASSERT (ni->ni_lnd == &the_kiblnd);
1488
1489         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
1490                 rc = kiblnd_base_startup();
1491                 if (rc != 0)
1492                         return rc;
1493         }
1494
1495         LIBCFS_ALLOC(net, sizeof(*net));
1496         ni->ni_data = net;
1497         if (net == NULL)
1498                 goto failed;
1499
1500         memset(net, 0, sizeof(*net));
1501
1502         do_gettimeofday(&tv);
1503         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
1504
1505         ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
1506         ni->ni_peertxcredits = *kiblnd_tunables.kib_peercredits;
1507
1508         spin_lock_init(&net->ibn_tx_lock);
1509         INIT_LIST_HEAD(&net->ibn_idle_txs);
1510
1511         rc = kiblnd_alloc_tx_descs(ni);
1512         if (rc != 0) {
1513                 CERROR("Can't allocate tx descs\n");
1514                 goto failed;
1515         }
1516
1517         if (ni->ni_interfaces[0] != NULL) {
1518                 /* Use the IPoIB interface specified in 'networks=' */
1519
1520                 CLASSERT (LNET_MAX_INTERFACES > 1);
1521                 if (ni->ni_interfaces[1] != NULL) {
1522                         CERROR("Multiple interfaces not supported\n");
1523                         goto failed;
1524                 }
1525
1526                 ifname = ni->ni_interfaces[0];
1527         } else {
1528                 ifname = *kiblnd_tunables.kib_default_ipif;
1529         }
1530
1531         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
1532                 CERROR("IPoIB interface name too long: %s\n", ifname);
1533                 goto failed;
1534         }
1535
1536         ibdev = NULL;
1537         list_for_each (tmp, &kiblnd_data.kib_devs) {
1538                 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
1539
1540                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
1541                         break;
1542
1543                 ibdev = NULL;
1544         }
1545
1546         if (ibdev == NULL) {
1547                 __u32                     ip;
1548                 __u32                     netmask;
1549                 int                       up;
1550                 struct rdma_cm_id        *id;
1551                 struct ib_pd             *pd;
1552                 struct ib_mr             *mr;
1553                 struct sockaddr_in        addr;
1554
1555                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
1556                 if (rc != 0) {
1557                         CERROR("Can't query IPoIB interface %s: %d\n",
1558                                ifname, rc);
1559                         goto failed;
1560                 }
1561
1562                 if (!up) {
1563                         CERROR("Can't query IPoIB interface %s: it's down\n",
1564                                ifname);
1565                         goto failed;
1566                 }
1567
1568                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
1569                 if (ibdev == NULL)
1570                         goto failed;
1571
1572                 memset(ibdev, 0, sizeof(*ibdev));
1573
1574                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
1575                 ibdev->ibd_ifip = ip;
1576                 strcpy(&ibdev->ibd_ifname[0], ifname);
1577
1578                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
1579                 if (!IS_ERR(id)) {
1580                         ibdev->ibd_cmid = id;
1581                 } else {
1582                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
1583                         goto failed;
1584                 }
1585
1586                 memset(&addr, 0, sizeof(addr));
1587                 addr.sin_family      = AF_INET;
1588                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
1589                 addr.sin_addr.s_addr = htonl(ip);
1590
1591                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
1592                 if (rc != 0) {
1593                         CERROR("Can't bind to %s: %d\n", ifname, rc);
1594                         goto failed;
1595                 }
1596
1597                 /* Binding should have assigned me an IB device */
1598                 LASSERT (id->device != NULL);
1599
1600                 pd = ib_alloc_pd(id->device);
1601                 if (!IS_ERR(pd)) {
1602                         ibdev->ibd_pd = pd;
1603                 } else {
1604                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
1605                         goto failed;
1606                 }
1607
1608 #if IBLND_MAP_ON_DEMAND
1609                 /* MR for sends and receives */
1610                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE);
1611 #else
1612                 /* MR for sends, recieves _and_ RDMA...........v */
1613                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE |
1614                                        IB_ACCESS_REMOTE_WRITE);
1615 #endif
1616                 if (!IS_ERR(mr)) {
1617                         ibdev->ibd_mr = mr;
1618                 } else {
1619                         CERROR("Can't get MR: %ld\n", PTR_ERR(mr));
1620                         goto failed;
1621                 }
1622
1623                 rc = rdma_listen(id, 0);
1624                 if (rc != 0) {
1625                         CERROR("Can't start listener: %d\n", rc);
1626                         goto failed;
1627                 }
1628
1629                 list_add_tail(&ibdev->ibd_list,
1630                               &kiblnd_data.kib_devs);
1631         }
1632
1633         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
1634         net->ibn_dev = ibdev;
1635
1636 #if IBLND_MAP_ON_DEMAND
1637         /* FMR pool for RDMA */
1638         {
1639                 struct ib_fmr_pool      *fmrpool;
1640                 struct ib_fmr_pool_param param = {
1641                         .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1642                         .page_shift        = PAGE_SHIFT,
1643                         .access            = (IB_ACCESS_LOCAL_WRITE |
1644                                               IB_ACCESS_REMOTE_WRITE),
1645                         .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1646                         .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1647                         .flush_function    = NULL,
1648                         .flush_arg         = NULL,
1649                         .cache             = *kiblnd_tunables.kib_fmr_cache};
1650
1651                 if (*kiblnd_tunables.kib_fmr_pool_size <
1652                     *kiblnd_tunables.kib_ntx) {
1653                         CERROR("Can't set fmr pool size (%d) < ntx(%d)\n",
1654                                *kiblnd_tunables.kib_fmr_pool_size,
1655                                *kiblnd_tunables.kib_ntx);
1656                         goto failed;
1657                 }
1658
1659                 fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, &param);
1660                 if (!IS_ERR(fmrpool)) {
1661                         net->ibn_fmrpool = fmrpool;
1662                 } else {
1663                         CERROR("Can't create FMR pool: %ld\n",
1664                                PTR_ERR(fmrpool));
1665                         goto failed;
1666                 }
1667         }
1668 #endif
1669
1670         kiblnd_map_tx_descs(ni);
1671
1672         ibdev->ibd_nnets++;
1673         net->ibn_init = IBLND_INIT_ALL;
1674
1675         return 0;
1676
1677 failed:
1678         kiblnd_shutdown(ni);
1679
1680         CDEBUG(D_NET, "kiblnd_startup failed\n");
1681         return -ENETDOWN;
1682 }
1683
1684 void __exit
1685 kiblnd_module_fini (void)
1686 {
1687         lnet_unregister_lnd(&the_kiblnd);
1688         kiblnd_tunables_fini();
1689 }
1690
1691 int __init
1692 kiblnd_module_init (void)
1693 {
1694         int    rc;
1695
1696         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
1697 #if !IBLND_MAP_ON_DEMAND
1698         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1699                   <= IBLND_MSG_SIZE);
1700         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1701                   <= IBLND_MSG_SIZE);
1702 #endif
1703         rc = kiblnd_tunables_init();
1704         if (rc != 0)
1705                 return rc;
1706
1707         lnet_register_lnd(&the_kiblnd);
1708
1709         return 0;
1710 }
1711
1712 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1713 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v1.00");
1714 MODULE_LICENSE("GPL");
1715
1716 module_init(kiblnd_module_init);
1717 module_exit(kiblnd_module_fini);