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 = dma_map_single(cmid->device->dma_device,
 701                                                 rx->rx_msg,
 702                                                 IBLND_MSG_SIZE,
 703                                                 DMA_FROM_DEVICE);
 704                 pci_unmap_addr_set(rx, rx_msgunmap, rx->rx_msgaddr);
 705
 706                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
 707                        i, rx->rx_msg, rx->rx_msgaddr,
 708                        lnet_page2phys(page) + page_offset);
 709
 710                 page_offset += IBLND_MSG_SIZE;
 711                 LASSERT (page_offset <= PAGE_SIZE);
 712
 713                 if (page_offset == PAGE_SIZE) {
 714                         page_offset = 0;
 715                         ipage++;
 716                         LASSERT (ipage <= IBLND_RX_MSG_PAGES);
 717                 }
 718         }
 719
 720         cq = ib_create_cq(cmid->device,
 721                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 722                           IBLND_CQ_ENTRIES());
 723         if (!IS_ERR(cq)) {
 724                 conn->ibc_cq = cq;
 725         } else {
 726                 CERROR("Can't create CQ: %ld\n", PTR_ERR(cq));
 727                 goto failed_2;
 728         }
 729
 730         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 731         if (rc != 0) {
 732                 CERROR("Can't request completion notificiation: %d\n", rc);
 733                 goto failed_2;
 734         }
 735
 736         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 737         init_qp_attr->event_handler = kiblnd_qp_event;
 738         init_qp_attr->qp_context = conn;
 739         init_qp_attr->cap.max_send_wr = (*kiblnd_tunables.kib_concurrent_sends) *
 740                                         (1 + IBLND_MAX_RDMA_FRAGS);
 741         init_qp_attr->cap.max_recv_wr = IBLND_RX_MSGS;
 742         init_qp_attr->cap.max_send_sge = 1;
 743         init_qp_attr->cap.max_recv_sge = 1;
 744         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 745         init_qp_attr->qp_type = IB_QPT_RC;
 746         init_qp_attr->send_cq = cq;
 747         init_qp_attr->recv_cq = cq;
 748
 749         rc = 0;
 750         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 751         switch (*kiblnd_tunables.kib_ib_mtu) {
 752         default:
 753                 rc = *kiblnd_tunables.kib_ib_mtu;
 754                 /* fall through to... */
 755         case 0: /* set tunable to the default
 756                  * CAVEAT EMPTOR! this assumes the default is one of the MTUs
 757                  * below, otherwise we'll WARN on the next QP create */
 758                 *kiblnd_tunables.kib_ib_mtu =
 759                         ib_mtu_enum_to_int(cmid->route.path_rec->mtu);
 760                 break;
 761         case 256:
 762                 cmid->route.path_rec->mtu = IB_MTU_256;
 763                 break;
 764         case 512:
 765                 cmid->route.path_rec->mtu = IB_MTU_512;
 766                 break;
 767         case 1024:
 768                 cmid->route.path_rec->mtu = IB_MTU_1024;
 769                 break;
 770         case 2048:
 771                 cmid->route.path_rec->mtu = IB_MTU_2048;
 772                 break;
 773         case 4096:
 774                 cmid->route.path_rec->mtu = IB_MTU_4096;
 775                 break;
 776         }
 777         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 778
 779         if (rc != 0)
 780                 CWARN("Invalid IB MTU value %d, using default value %d\n",
 781                       rc, *kiblnd_tunables.kib_ib_mtu);
 782
 783         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 784         if (rc != 0) {
 785                 CERROR("Can't create QP: %d\n", rc);
 786                 goto failed_2;
 787         }
 788
 789         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 790
 791         /* 1 ref for caller and each rxmsg */
 792         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS);
 793         conn->ibc_nrx = IBLND_RX_MSGS;
 794
 795         /* post receives */
 796         for (i = 0; i < IBLND_RX_MSGS; i++) {
 797                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 798                                     IBLND_POSTRX_NO_CREDIT);
 799                 if (rc != 0) {
 800                         CERROR("Can't post rxmsg: %d\n", rc);
 801
 802                         /* Make posted receives complete */
 803                         kiblnd_abort_receives(conn);
 804
 805                         /* correct # of posted buffers
 806                          * NB locking needed now I'm racing with completion */
 807                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
 808                         conn->ibc_nrx -= IBLND_RX_MSGS - i;
 809                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 810                                                flags);
 811
 812                         /* Drop my own and unused rxbuffer refcounts */
 813                         while (i++ <= IBLND_RX_MSGS)
 814                                 kiblnd_conn_decref(conn);
 815
 816                         return NULL;
 817                 }
 818         }
 819
 820         /* Init successful! */
 821         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 822                  state == IBLND_CONN_PASSIVE_WAIT);
 823         conn->ibc_state = state;
 824
 825         /* 1 more conn */
 826         atomic_inc(&net->ibn_nconns);
 827         return conn;
 828
 829  failed_2:
 830         kiblnd_destroy_conn(conn);
 831  failed_1:
 832         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 833  failed_0:
 834         return NULL;
 835 }
 836
 837 void
 838 kiblnd_destroy_conn (kib_conn_t *conn)
 839 {
 840         struct rdma_cm_id *cmid = conn->ibc_cmid;
 841         kib_peer_t        *peer = conn->ibc_peer;
 842         int                rc;
 843         int                i;
 844
 845         LASSERT (!in_interrupt());
 846         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 847         LASSERT (list_empty(&conn->ibc_early_rxs));
 848         LASSERT (list_empty(&conn->ibc_tx_queue));
 849         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 850         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 851         LASSERT (list_empty(&conn->ibc_active_txs));
 852         LASSERT (conn->ibc_nsends_posted == 0);
 853
 854         switch (conn->ibc_state) {
 855         default:
 856                 /* conn must be completely disengaged from the network */
 857                 LBUG();
 858
 859         case IBLND_CONN_DISCONNECTED:
 860                 /* connvars should have been freed already */
 861                 LASSERT (conn->ibc_connvars == NULL);
 862                 break;
 863
 864         case IBLND_CONN_INIT:
 865                 break;
 866         }
 867
 868         if (conn->ibc_cmid->qp != NULL)
 869                 rdma_destroy_qp(conn->ibc_cmid);
 870
 871         if (conn->ibc_cq != NULL) {
 872                 rc = ib_destroy_cq(conn->ibc_cq);
 873                 if (rc != 0)
 874                         CWARN("Error destroying CQ: %d\n", rc);
 875         }
 876
 877         if (conn->ibc_rx_pages != NULL) {
 878                 LASSERT (conn->ibc_rxs != NULL);
 879
 880                 for (i = 0; i < IBLND_RX_MSGS; i++) {
 881                         kib_rx_t *rx = &conn->ibc_rxs[i];
 882
 883                         LASSERT (rx->rx_nob >= 0); /* not posted */
 884
 885                         dma_unmap_single(conn->ibc_cmid->device->dma_device,
 886                                          pci_unmap_addr(rx, rx_msgunmap),
 887                                          IBLND_MSG_SIZE, DMA_FROM_DEVICE);
 888                 }
 889
 890                 kiblnd_free_pages(conn->ibc_rx_pages);
 891         }
 892
 893         if (conn->ibc_rxs != NULL) {
 894                 LIBCFS_FREE(conn->ibc_rxs,
 895                             IBLND_RX_MSGS * sizeof(kib_rx_t));
 896         }
 897
 898         if (conn->ibc_connvars != NULL)
 899                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 900
 901         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 902         if (conn->ibc_state != IBLND_CONN_INIT) {
 903                 kib_net_t *net = peer->ibp_ni->ni_data;
 904
 905                 kiblnd_peer_decref(peer);
 906                 rdma_destroy_id(cmid);
 907                 atomic_dec(&net->ibn_nconns);
 908         }
 909
 910         LIBCFS_FREE(conn, sizeof(*conn));
 911 }
 912
 913 int
 914 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 915 {
 916         kib_conn_t         *conn;
 917         struct list_head   *ctmp;
 918         struct list_head   *cnxt;
 919         int                 count = 0;
 920
 921         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 922                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 923
 924                 count++;
 925                 kiblnd_close_conn_locked(conn, why);
 926         }
 927
 928         return count;
 929 }
 930
 931 int
 932 kiblnd_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation)
 933 {
 934         kib_conn_t         *conn;
 935         struct list_head   *ctmp;
 936         struct list_head   *cnxt;
 937         int                 count = 0;
 938
 939         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 940                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 941
 942                 if (conn->ibc_incarnation == incarnation)
 943                         continue;
 944
 945                 CDEBUG(D_NET, "Closing stale conn -> %s incarnation:"LPX64"("LPX64")\n",
 946                        libcfs_nid2str(peer->ibp_nid),
 947                        conn->ibc_incarnation, incarnation);
 948
 949                 count++;
 950                 kiblnd_close_conn_locked(conn, -ESTALE);
 951         }
 952
 953         return count;
 954 }
 955
 956 int
 957 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 958 {
 959         kib_peer_t         *peer;
 960         struct list_head   *ptmp;
 961         struct list_head   *pnxt;
 962         int                 lo;
 963         int                 hi;
 964         int                 i;
 965         unsigned long       flags;
 966         int                 count = 0;
 967
 968         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 969
 970         if (nid != LNET_NID_ANY)
 971                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 972         else {
 973                 lo = 0;
 974                 hi = kiblnd_data.kib_peer_hash_size - 1;
 975         }
 976
 977         for (i = lo; i <= hi; i++) {
 978                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 979
 980                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 981                         LASSERT (peer->ibp_connecting > 0 ||
 982                                  peer->ibp_accepting > 0 ||
 983                                  !list_empty(&peer->ibp_conns));
 984
 985                         if (peer->ibp_ni != ni)
 986                                 continue;
 987
 988                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 989                                 continue;
 990
 991                         count += kiblnd_close_peer_conns_locked(peer, 0);
 992                 }
 993         }
 994
 995         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 996
 997         /* wildcards always succeed */
 998         if (nid == LNET_NID_ANY)
 999                 return 0;
1000
1001         return (count == 0) ? -ENOENT : 0;
1002 }
1003
1004 int
1005 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1006 {
1007         struct libcfs_ioctl_data *data = arg;
1008         int                       rc = -EINVAL;
1009
1010         switch(cmd) {
1011         case IOC_LIBCFS_GET_PEER: {
1012                 lnet_nid_t   nid = 0;
1013                 int          count = 0;
1014
1015                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1016                                           &nid, &count);
1017                 data->ioc_nid    = nid;
1018                 data->ioc_count  = count;
1019                 break;
1020         }
1021
1022         case IOC_LIBCFS_DEL_PEER: {
1023                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1024                 break;
1025         }
1026         case IOC_LIBCFS_GET_CONN: {
1027                 kib_conn_t *conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1028
1029                 if (conn == NULL) {
1030                         rc = -ENOENT;
1031                 } else {
1032                         // kiblnd_debug_conn(conn);
1033                         rc = 0;
1034                         data->ioc_nid = conn->ibc_peer->ibp_nid;
1035                         kiblnd_conn_decref(conn);
1036                 }
1037                 break;
1038         }
1039         case IOC_LIBCFS_CLOSE_CONNECTION: {
1040                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1041                 break;
1042         }
1043
1044         default:
1045                 break;
1046         }
1047
1048         return rc;
1049 }
1050
1051 void
1052 kiblnd_free_pages (kib_pages_t *p)
1053 {
1054         int         npages = p->ibp_npages;
1055         int         i;
1056
1057         for (i = 0; i < npages; i++)
1058                 if (p->ibp_pages[i] != NULL)
1059                         __free_page(p->ibp_pages[i]);
1060
1061         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1062 }
1063
1064 int
1065 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1066 {
1067         kib_pages_t   *p;
1068         int            i;
1069
1070         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1071         if (p == NULL) {
1072                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1073                 return -ENOMEM;
1074         }
1075
1076         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1077         p->ibp_npages = npages;
1078
1079         for (i = 0; i < npages; i++) {
1080                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1081                 if (p->ibp_pages[i] == NULL) {
1082                         CERROR("Can't allocate page %d of %d\n", i, npages);
1083                         kiblnd_free_pages(p);
1084                         return -ENOMEM;
1085                 }
1086         }
1087
1088         *pp = p;
1089         return 0;
1090 }
1091
1092 void
1093 kiblnd_free_tx_descs (lnet_ni_t *ni)
1094 {
1095         int        i;
1096         kib_net_t *net = ni->ni_data;
1097
1098         LASSERT (net != NULL);
1099
1100         if (net->ibn_tx_descs != NULL) {
1101                 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1102                         kib_tx_t *tx = &net->ibn_tx_descs[i];
1103
1104 #if IBLND_MAP_ON_DEMAND
1105                         if (tx->tx_pages != NULL)
1106                                 LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV *
1107                                             sizeof(*tx->tx_pages));
1108 #else
1109                         if (tx->tx_wrq != NULL)
1110                                 LIBCFS_FREE(tx->tx_wrq,
1111                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1112                                             sizeof(*tx->tx_wrq));
1113
1114                         if (tx->tx_sge != NULL)
1115                                 LIBCFS_FREE(tx->tx_sge,
1116                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1117                                             sizeof(*tx->tx_sge));
1118
1119                         if (tx->tx_rd != NULL)
1120                                 LIBCFS_FREE(tx->tx_rd,
1121                                             offsetof(kib_rdma_desc_t,
1122                                                rd_frags[IBLND_MAX_RDMA_FRAGS]));
1123
1124                         if (tx->tx_frags != NULL)
1125                                 LIBCFS_FREE(tx->tx_frags,
1126                                             IBLND_MAX_RDMA_FRAGS *
1127                                             sizeof(*tx->tx_frags));
1128 #endif
1129                 }
1130
1131                 LIBCFS_FREE(net->ibn_tx_descs,
1132                             IBLND_TX_MSGS() * sizeof(kib_tx_t));
1133         }
1134
1135         if (net->ibn_tx_pages != NULL)
1136                 kiblnd_free_pages(net->ibn_tx_pages);
1137 }
1138
1139 int
1140 kiblnd_alloc_tx_descs (lnet_ni_t *ni)
1141 {
1142         int        i;
1143         int        rc;
1144         kib_net_t *net = ni->ni_data;
1145
1146         LASSERT (net != NULL);
1147
1148         rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES());
1149
1150         if (rc != 0) {
1151                 CERROR("Can't allocate tx pages\n");
1152                 return rc;
1153         }
1154
1155         LIBCFS_ALLOC (net->ibn_tx_descs,
1156                       IBLND_TX_MSGS() * sizeof(kib_tx_t));
1157         if (net->ibn_tx_descs == NULL) {
1158                 CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS());
1159                 return -ENOMEM;
1160         }
1161
1162         memset(net->ibn_tx_descs, 0,
1163                IBLND_TX_MSGS() * sizeof(kib_tx_t));
1164
1165         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1166                 kib_tx_t *tx = &net->ibn_tx_descs[i];
1167
1168 #if IBLND_MAP_ON_DEMAND
1169                 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1170                              sizeof(*tx->tx_pages));
1171                 if (tx->tx_pages == NULL) {
1172                         CERROR("Can't allocate phys page vector[%d]\n",
1173                                LNET_MAX_IOV);
1174                         return -ENOMEM;
1175                 }
1176 #else
1177                 LIBCFS_ALLOC(tx->tx_wrq,
1178                              (1 + IBLND_MAX_RDMA_FRAGS) *
1179                              sizeof(*tx->tx_wrq));
1180                 if (tx->tx_wrq == NULL)
1181                         return -ENOMEM;
1182
1183                 LIBCFS_ALLOC(tx->tx_sge,
1184                              (1 + IBLND_MAX_RDMA_FRAGS) *
1185                              sizeof(*tx->tx_sge));
1186                 if (tx->tx_sge == NULL)
1187                         return -ENOMEM;
1188
1189                 LIBCFS_ALLOC(tx->tx_rd,
1190                              offsetof(kib_rdma_desc_t,
1191                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1192                 if (tx->tx_rd == NULL)
1193                         return -ENOMEM;
1194
1195                 LIBCFS_ALLOC(tx->tx_frags,
1196                              IBLND_MAX_RDMA_FRAGS *
1197                              sizeof(*tx->tx_frags));
1198                 if (tx->tx_frags == NULL)
1199                         return -ENOMEM;
1200 #endif
1201         }
1202
1203         return 0;
1204 }
1205
1206 void
1207 kiblnd_unmap_tx_descs (lnet_ni_t *ni)
1208 {
1209         int             i;
1210         kib_tx_t       *tx;
1211         kib_net_t      *net = ni->ni_data;
1212
1213         LASSERT (net != NULL);
1214
1215         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1216                 tx = &net->ibn_tx_descs[i];
1217
1218                 dma_unmap_single(net->ibn_dev->ibd_cmid->device->dma_device,
1219                                  pci_unmap_addr(tx, tx_msgunmap),
1220                                  IBLND_MSG_SIZE, DMA_TO_DEVICE);
1221         }
1222 }
1223
1224 void
1225 kiblnd_map_tx_descs (lnet_ni_t *ni)
1226 {
1227         int             ipage = 0;
1228         int             page_offset = 0;
1229         int             i;
1230         struct page    *page;
1231         kib_tx_t       *tx;
1232         kib_net_t      *net = ni->ni_data;
1233
1234         LASSERT (net != NULL);
1235
1236         /* pre-mapped messages are not bigger than 1 page */
1237         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1238
1239         /* No fancy arithmetic when we do the buffer calculations */
1240         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1241
1242         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1243                 page = net->ibn_tx_pages->ibp_pages[ipage];
1244                 tx = &net->ibn_tx_descs[i];
1245
1246                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1247                                            page_offset);
1248
1249                 tx->tx_msgaddr = dma_map_single(
1250                         net->ibn_dev->ibd_cmid->device->dma_device,
1251                         tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE);
1252                 pci_unmap_addr_set(tx, tx_msgunmap, tx->tx_msgaddr);
1253
1254                 list_add(&tx->tx_list, &net->ibn_idle_txs);
1255
1256                 page_offset += IBLND_MSG_SIZE;
1257                 LASSERT (page_offset <= PAGE_SIZE);
1258
1259                 if (page_offset == PAGE_SIZE) {
1260                         page_offset = 0;
1261                         ipage++;
1262                         LASSERT (ipage <= IBLND_TX_MSG_PAGES());
1263                 }
1264         }
1265 }
1266
1267 void
1268 kiblnd_base_shutdown (void)
1269 {
1270         int i;
1271
1272         LASSERT (list_empty(&kiblnd_data.kib_devs));
1273
1274         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
1275                atomic_read(&libcfs_kmemory));
1276
1277         switch (kiblnd_data.kib_init) {
1278         default:
1279                 LBUG();
1280
1281         case IBLND_INIT_ALL:
1282         case IBLND_INIT_DATA:
1283                 LASSERT (kiblnd_data.kib_peers != NULL);
1284                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
1285                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
1286                 }
1287                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
1288                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
1289
1290                 /* flag threads to terminate; wake and wait for them to die */
1291                 kiblnd_data.kib_shutdown = 1;
1292                 wake_up_all(&kiblnd_data.kib_sched_waitq);
1293                 wake_up_all(&kiblnd_data.kib_connd_waitq);
1294
1295                 i = 2;
1296                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
1297                         i++;
1298                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1299                                "Waiting for %d threads to terminate\n",
1300                                atomic_read(&kiblnd_data.kib_nthreads));
1301                         cfs_pause(cfs_time_seconds(1));
1302                 }
1303
1304                 /* fall through */
1305
1306         case IBLND_INIT_NOTHING:
1307                 break;
1308         }
1309
1310         if (kiblnd_data.kib_peers != NULL)
1311                 LIBCFS_FREE(kiblnd_data.kib_peers,
1312                             sizeof(struct list_head) *
1313                             kiblnd_data.kib_peer_hash_size);
1314
1315         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
1316                atomic_read(&libcfs_kmemory));
1317
1318         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
1319         PORTAL_MODULE_UNUSE;
1320 }
1321
1322 void
1323 kiblnd_shutdown (lnet_ni_t *ni)
1324 {
1325         kib_net_t        *net = ni->ni_data;
1326         rwlock_t         *g_lock = &kiblnd_data.kib_global_lock;
1327         int               i;
1328         unsigned long     flags;
1329
1330         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
1331
1332         if (net == NULL)
1333                 goto out;
1334
1335         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
1336                atomic_read(&libcfs_kmemory));
1337
1338         write_lock_irqsave(g_lock, flags);
1339         net->ibn_shutdown = 1;
1340         write_unlock_irqrestore(g_lock, flags);
1341
1342         switch (net->ibn_init) {
1343         default:
1344                 LBUG();
1345
1346         case IBLND_INIT_ALL:
1347                 /* nuke all existing peers within this net */
1348                 kiblnd_del_peer(ni, LNET_NID_ANY);
1349
1350                 /* Wait for all peer state to clean up */
1351                 i = 2;
1352                 while (atomic_read(&net->ibn_npeers) != 0) {
1353                         i++;
1354                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
1355                                "%s: waiting for %d peers to disconnect\n",
1356                                libcfs_nid2str(ni->ni_nid),
1357                                atomic_read(&net->ibn_npeers));
1358                         cfs_pause(cfs_time_seconds(1));
1359                 }
1360
1361                 kiblnd_unmap_tx_descs(ni);
1362
1363                 LASSERT (net->ibn_dev->ibd_nnets > 0);
1364                 net->ibn_dev->ibd_nnets--;
1365
1366                 /* fall through */
1367
1368         case IBLND_INIT_NOTHING:
1369                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
1370
1371 #if IBLND_MAP_ON_DEMAND
1372                 if (net->ibn_fmrpool != NULL)
1373                         ib_destroy_fmr_pool(net->ibn_fmrpool);
1374 #endif
1375                 if (net->ibn_dev != NULL &&
1376                     net->ibn_dev->ibd_nnets == 0)
1377                         kiblnd_destroy_dev(net->ibn_dev);
1378
1379                 break;
1380         }
1381
1382         kiblnd_free_tx_descs(ni);
1383
1384         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
1385                atomic_read(&libcfs_kmemory));
1386
1387         net->ibn_init = IBLND_INIT_NOTHING;
1388         ni->ni_data = NULL;
1389
1390         LIBCFS_FREE(net, sizeof(*net));
1391
1392 out:
1393         if (list_empty(&kiblnd_data.kib_devs))
1394                 kiblnd_base_shutdown();
1395         return;
1396 }
1397
1398 int
1399 kiblnd_base_startup (void)
1400 {
1401         int               rc;
1402         int               i;
1403
1404         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
1405
1406         if (*kiblnd_tunables.kib_credits > *kiblnd_tunables.kib_ntx) {
1407                 CERROR("Can't set credits(%d) > ntx(%d)\n",
1408                        *kiblnd_tunables.kib_credits,
1409                        *kiblnd_tunables.kib_ntx);
1410                 return -EINVAL;
1411         }
1412
1413         PORTAL_MODULE_USE;
1414         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
1415
1416         rwlock_init(&kiblnd_data.kib_global_lock);
1417
1418         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
1419
1420         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
1421         LIBCFS_ALLOC(kiblnd_data.kib_peers,
1422                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
1423         if (kiblnd_data.kib_peers == NULL) {
1424                 goto failed;
1425         }
1426         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
1427                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
1428
1429         spin_lock_init(&kiblnd_data.kib_connd_lock);
1430         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
1431         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
1432         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
1433
1434         spin_lock_init(&kiblnd_data.kib_sched_lock);
1435         INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
1436         init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
1437
1438         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
1439
1440         /* lists/ptrs/locks initialised */
1441         kiblnd_data.kib_init = IBLND_INIT_DATA;
1442         /*****************************************************/
1443
1444         for (i = 0; i < IBLND_N_SCHED; i++) {
1445                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
1446                 if (rc != 0) {
1447                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
1448                                i, rc);
1449                         goto failed;
1450                 }
1451         }
1452
1453         rc = kiblnd_thread_start(kiblnd_connd, NULL);
1454         if (rc != 0) {
1455                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
1456                 goto failed;
1457         }
1458
1459         /* flag everything initialised */
1460         kiblnd_data.kib_init = IBLND_INIT_ALL;
1461         /*****************************************************/
1462
1463         return 0;
1464
1465  failed:
1466         kiblnd_base_shutdown();
1467         return -ENETDOWN;
1468 }
1469
1470 int
1471 kiblnd_startup (lnet_ni_t *ni)
1472 {
1473         char                     *ifname;
1474         kib_net_t                *net;
1475         kib_dev_t                *ibdev;
1476         struct list_head         *tmp;
1477         struct timeval            tv;
1478         int                       rc;
1479
1480         LASSERT (ni->ni_lnd == &the_kiblnd);
1481
1482         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
1483                 rc = kiblnd_base_startup();
1484                 if (rc != 0)
1485                         return rc;
1486         }
1487
1488         LIBCFS_ALLOC(net, sizeof(*net));
1489         ni->ni_data = net;
1490         if (net == NULL)
1491                 goto failed;
1492
1493         memset(net, 0, sizeof(*net));
1494
1495         do_gettimeofday(&tv);
1496         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
1497
1498         ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
1499         ni->ni_peertxcredits = *kiblnd_tunables.kib_peercredits;
1500
1501         spin_lock_init(&net->ibn_tx_lock);
1502         INIT_LIST_HEAD(&net->ibn_idle_txs);
1503
1504         rc = kiblnd_alloc_tx_descs(ni);
1505         if (rc != 0) {
1506                 CERROR("Can't allocate tx descs\n");
1507                 goto failed;
1508         }
1509
1510         if (ni->ni_interfaces[0] != NULL) {
1511                 /* Use the IPoIB interface specified in 'networks=' */
1512
1513                 CLASSERT (LNET_MAX_INTERFACES > 1);
1514                 if (ni->ni_interfaces[1] != NULL) {
1515                         CERROR("Multiple interfaces not supported\n");
1516                         goto failed;
1517                 }
1518
1519                 ifname = ni->ni_interfaces[0];
1520         } else {
1521                 ifname = *kiblnd_tunables.kib_default_ipif;
1522         }
1523
1524         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
1525                 CERROR("IPoIB interface name too long: %s\n", ifname);
1526                 goto failed;
1527         }
1528
1529         ibdev = NULL;
1530         list_for_each (tmp, &kiblnd_data.kib_devs) {
1531                 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
1532
1533                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
1534                         break;
1535
1536                 ibdev = NULL;
1537         }
1538
1539         if (ibdev == NULL) {
1540                 __u32                     ip;
1541                 __u32                     netmask;
1542                 int                       up;
1543                 struct rdma_cm_id        *id;
1544                 struct ib_pd             *pd;
1545                 struct ib_mr             *mr;
1546                 struct sockaddr_in        addr;
1547
1548                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
1549                 if (rc != 0) {
1550                         CERROR("Can't query IPoIB interface %s: %d\n",
1551                                ifname, rc);
1552                         goto failed;
1553                 }
1554
1555                 if (!up) {
1556                         CERROR("Can't query IPoIB interface %s: it's down\n",
1557                                ifname);
1558                         goto failed;
1559                 }
1560
1561                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
1562                 if (ibdev == NULL)
1563                         goto failed;
1564
1565                 memset(ibdev, 0, sizeof(*ibdev));
1566
1567                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
1568                 ibdev->ibd_ifip = ip;
1569                 strcpy(&ibdev->ibd_ifname[0], ifname);
1570
1571                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
1572                 if (!IS_ERR(id)) {
1573                         ibdev->ibd_cmid = id;
1574                 } else {
1575                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
1576                         goto failed;
1577                 }
1578
1579                 memset(&addr, 0, sizeof(addr));
1580                 addr.sin_family      = AF_INET;
1581                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
1582                 addr.sin_addr.s_addr = htonl(ip);
1583
1584                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
1585                 if (rc != 0) {
1586                         CERROR("Can't bind to %s: %d\n", ifname, rc);
1587                         goto failed;
1588                 }
1589
1590                 /* Binding should have assigned me an IB device */
1591                 LASSERT (id->device != NULL);
1592
1593                 pd = ib_alloc_pd(id->device);
1594                 if (!IS_ERR(pd)) {
1595                         ibdev->ibd_pd = pd;
1596                 } else {
1597                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
1598                         goto failed;
1599                 }
1600
1601 #if IBLND_MAP_ON_DEMAND
1602                 /* MR for sends and receives */
1603                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE);
1604 #else
1605                 /* MR for sends, recieves _and_ RDMA...........v */
1606                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE |
1607                                        IB_ACCESS_REMOTE_WRITE);
1608 #endif
1609                 if (!IS_ERR(mr)) {
1610                         ibdev->ibd_mr = mr;
1611                 } else {
1612                         CERROR("Can't get MR: %ld\n", PTR_ERR(pd));
1613                         goto failed;
1614                 }
1615
1616                 rc = rdma_listen(id, 0);
1617                 if (rc != 0) {
1618                         CERROR("Can't start listener: %d\n", rc);
1619                         goto failed;
1620                 }
1621
1622                 list_add_tail(&ibdev->ibd_list,
1623                               &kiblnd_data.kib_devs);
1624         }
1625
1626         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
1627         net->ibn_dev = ibdev;
1628
1629 #if IBLND_MAP_ON_DEMAND
1630         /* FMR pool for RDMA */
1631         {
1632                 struct ib_fmr_pool      *fmrpool;
1633                 struct ib_fmr_pool_param param = {
1634                         .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1635                         .page_shift        = PAGE_SHIFT,
1636                         .access            = (IB_ACCESS_LOCAL_WRITE |
1637                                               IB_ACCESS_REMOTE_WRITE),
1638                         .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1639                         .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1640                         .flush_function    = NULL,
1641                         .flush_arg         = NULL,
1642                         .cache             = *kiblnd_tunables.kib_fmr_cache};
1643
1644                 if (*kiblnd_tunables.kib_fmr_pool_size <
1645                     *kiblnd_tunables.kib_ntx) {
1646                         CERROR("Can't set fmr pool size (%d) < ntx(%d)\n",
1647                                *kiblnd_tunables.kib_fmr_pool_size,
1648                                *kiblnd_tunables.kib_ntx);
1649                         goto failed;
1650                 }
1651
1652                 fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, &param);
1653                 if (!IS_ERR(fmrpool)) {
1654                         net->ibn_fmrpool = fmrpool;
1655                 } else {
1656                         CERROR("Can't create FMR pool: %ld\n",
1657                                PTR_ERR(fmrpool));
1658                         goto failed;
1659                 }
1660         }
1661 #endif
1662
1663         kiblnd_map_tx_descs(ni);
1664
1665         ibdev->ibd_nnets++;
1666         net->ibn_init = IBLND_INIT_ALL;
1667
1668         return 0;
1669
1670 failed:
1671         kiblnd_shutdown(ni);
1672
1673         CDEBUG(D_NET, "kiblnd_startup failed\n");
1674         return -ENETDOWN;
1675 }
1676
1677 void __exit
1678 kiblnd_module_fini (void)
1679 {
1680         lnet_unregister_lnd(&the_kiblnd);
1681         kiblnd_tunables_fini();
1682 }
1683
1684 int __init
1685 kiblnd_module_init (void)
1686 {
1687         int    rc;
1688
1689         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
1690 #if !IBLND_MAP_ON_DEMAND
1691         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1692                   <= IBLND_MSG_SIZE);
1693         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1694                   <= IBLND_MSG_SIZE);
1695 #endif
1696         rc = kiblnd_tunables_init();
1697         if (rc != 0)
1698                 return rc;
1699
1700         lnet_register_lnd(&the_kiblnd);
1701
1702         return 0;
1703 }
1704
1705 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1706 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v1.00");
1707 MODULE_LICENSE("GPL");
1708
1709 module_init(kiblnd_module_init);
1710 module_exit(kiblnd_module_fini);