lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include <asm/page.h>
  38 #include "o2iblnd.h"
  39
  40 static struct lnet_lnd the_o2iblnd;
  41
  42 kib_data_t              kiblnd_data;
  43
  44 static __u32
  45 kiblnd_cksum (void *ptr, int nob)
  46 {
  47         char  *c  = ptr;
  48         __u32  sum = 0;
  49
  50         while (nob-- > 0)
  51                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  52
  53         /* ensure I don't return 0 (== no checksum) */
  54         return (sum == 0) ? 1 : sum;
  55 }
  56
  57 static char *
  58 kiblnd_msgtype2str(int type)
  59 {
  60         switch (type) {
  61         case IBLND_MSG_CONNREQ:
  62                 return "CONNREQ";
  63
  64         case IBLND_MSG_CONNACK:
  65                 return "CONNACK";
  66
  67         case IBLND_MSG_NOOP:
  68                 return "NOOP";
  69
  70         case IBLND_MSG_IMMEDIATE:
  71                 return "IMMEDIATE";
  72
  73         case IBLND_MSG_PUT_REQ:
  74                 return "PUT_REQ";
  75
  76         case IBLND_MSG_PUT_NAK:
  77                 return "PUT_NAK";
  78
  79         case IBLND_MSG_PUT_ACK:
  80                 return "PUT_ACK";
  81
  82         case IBLND_MSG_PUT_DONE:
  83                 return "PUT_DONE";
  84
  85         case IBLND_MSG_GET_REQ:
  86                 return "GET_REQ";
  87
  88         case IBLND_MSG_GET_DONE:
  89                 return "GET_DONE";
  90
  91         default:
  92                 return "???";
  93         }
  94 }
  95
  96 static int
  97 kiblnd_msgtype2size(int type)
  98 {
  99         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 100
 101         switch (type) {
 102         case IBLND_MSG_CONNREQ:
 103         case IBLND_MSG_CONNACK:
 104                 return hdr_size + sizeof(kib_connparams_t);
 105
 106         case IBLND_MSG_NOOP:
 107                 return hdr_size;
 108
 109         case IBLND_MSG_IMMEDIATE:
 110                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 111
 112         case IBLND_MSG_PUT_REQ:
 113                 return hdr_size + sizeof(kib_putreq_msg_t);
 114
 115         case IBLND_MSG_PUT_ACK:
 116                 return hdr_size + sizeof(kib_putack_msg_t);
 117
 118         case IBLND_MSG_GET_REQ:
 119                 return hdr_size + sizeof(kib_get_msg_t);
 120
 121         case IBLND_MSG_PUT_NAK:
 122         case IBLND_MSG_PUT_DONE:
 123         case IBLND_MSG_GET_DONE:
 124                 return hdr_size + sizeof(kib_completion_msg_t);
 125         default:
 126                 return -1;
 127         }
 128 }
 129
 130 static int
 131 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 132 {
 133         kib_rdma_desc_t   *rd;
 134         int                nob;
 135         int                n;
 136         int                i;
 137
 138         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 139                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 140
 141         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 142                               &msg->ibm_u.get.ibgm_rd :
 143                               &msg->ibm_u.putack.ibpam_rd;
 144
 145         if (flip) {
 146                 __swab32s(&rd->rd_key);
 147                 __swab32s(&rd->rd_nfrags);
 148         }
 149
 150         n = rd->rd_nfrags;
 151
 152         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 153                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 154                        n, IBLND_MAX_RDMA_FRAGS);
 155                 return 1;
 156         }
 157
 158         nob = offsetof (kib_msg_t, ibm_u) +
 159               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 160
 161         if (msg->ibm_nob < nob) {
 162                 CERROR("Short %s: %d(%d)\n",
 163                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 164                 return 1;
 165         }
 166
 167         if (!flip)
 168                 return 0;
 169
 170         for (i = 0; i < n; i++) {
 171                 __swab32s(&rd->rd_frags[i].rf_nob);
 172                 __swab64s(&rd->rd_frags[i].rf_addr);
 173         }
 174
 175         return 0;
 176 }
 177
 178 void
 179 kiblnd_pack_msg(struct lnet_ni *ni, kib_msg_t *msg, int version,
 180                 int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         kib_net_t *net = ni->ni_data;
 183
 184         /* CAVEAT EMPTOR! all message fields not set here should have been
 185          * initialised previously. */
 186         msg->ibm_magic    = IBLND_MSG_MAGIC;
 187         msg->ibm_version  = version;
 188         /*   ibm_type */
 189         msg->ibm_credits  = credits;
 190         /*   ibm_nob */
 191         msg->ibm_cksum    = 0;
 192         msg->ibm_srcnid   = ni->ni_nid;
 193         msg->ibm_srcstamp = net->ibn_incarnation;
 194         msg->ibm_dstnid   = dstnid;
 195         msg->ibm_dststamp = dststamp;
 196
 197         if (*kiblnd_tunables.kib_cksum) {
 198                 /* NB ibm_cksum zero while computing cksum */
 199                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 200         }
 201 }
 202
 203 int
 204 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 205 {
 206         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 207         __u32     msg_cksum;
 208         __u16     version;
 209         int       msg_nob;
 210         int       flip;
 211
 212         /* 6 bytes are enough to have received magic + version */
 213         if (nob < 6) {
 214                 CERROR("Short message: %d\n", nob);
 215                 return -EPROTO;
 216         }
 217
 218         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 219                 flip = 0;
 220         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 221                 flip = 1;
 222         } else {
 223                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 224                 return -EPROTO;
 225         }
 226
 227         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 228         if (version != IBLND_MSG_VERSION &&
 229             version != IBLND_MSG_VERSION_1) {
 230                 CERROR("Bad version: %x\n", version);
 231                 return -EPROTO;
 232         }
 233
 234         if (nob < hdr_size) {
 235                 CERROR("Short message: %d\n", nob);
 236                 return -EPROTO;
 237         }
 238
 239         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 240         if (msg_nob > nob) {
 241                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 242                 return -EPROTO;
 243         }
 244
 245         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 246          * gets flipped */
 247         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 248         msg->ibm_cksum = 0;
 249         if (msg_cksum != 0 &&
 250             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 251                 CERROR("Bad checksum\n");
 252                 return -EPROTO;
 253         }
 254
 255         msg->ibm_cksum = msg_cksum;
 256
 257         if (flip) {
 258                 /* leave magic unflipped as a clue to peer_ni endianness */
 259                 msg->ibm_version = version;
 260                 CLASSERT (sizeof(msg->ibm_type) == 1);
 261                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 262                 msg->ibm_nob     = msg_nob;
 263                 __swab64s(&msg->ibm_srcnid);
 264                 __swab64s(&msg->ibm_srcstamp);
 265                 __swab64s(&msg->ibm_dstnid);
 266                 __swab64s(&msg->ibm_dststamp);
 267         }
 268
 269         if (msg->ibm_srcnid == LNET_NID_ANY) {
 270                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 271                 return -EPROTO;
 272         }
 273
 274         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 275                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 276                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 277                 return -EPROTO;
 278         }
 279
 280         switch (msg->ibm_type) {
 281         default:
 282                 CERROR("Unknown message type %x\n", msg->ibm_type);
 283                 return -EPROTO;
 284
 285         case IBLND_MSG_NOOP:
 286         case IBLND_MSG_IMMEDIATE:
 287         case IBLND_MSG_PUT_REQ:
 288                 break;
 289
 290         case IBLND_MSG_PUT_ACK:
 291         case IBLND_MSG_GET_REQ:
 292                 if (kiblnd_unpack_rd(msg, flip))
 293                         return -EPROTO;
 294                 break;
 295
 296         case IBLND_MSG_PUT_NAK:
 297         case IBLND_MSG_PUT_DONE:
 298         case IBLND_MSG_GET_DONE:
 299                 if (flip)
 300                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 301                 break;
 302
 303         case IBLND_MSG_CONNREQ:
 304         case IBLND_MSG_CONNACK:
 305                 if (flip) {
 306                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 307                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 308                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 309                 }
 310                 break;
 311         }
 312         return 0;
 313 }
 314
 315 int
 316 kiblnd_create_peer(struct lnet_ni *ni, kib_peer_ni_t **peerp, lnet_nid_t nid)
 317 {
 318         kib_peer_ni_t   *peer_ni;
 319         kib_net_t       *net = ni->ni_data;
 320         int             cpt = lnet_cpt_of_nid(nid, ni);
 321         unsigned long   flags;
 322
 323         LASSERT(net != NULL);
 324         LASSERT(nid != LNET_NID_ANY);
 325
 326         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 327         if (peer_ni == NULL) {
 328                 CERROR("Cannot allocate peer_ni\n");
 329                 return -ENOMEM;
 330         }
 331
 332         peer_ni->ibp_ni = ni;
 333         peer_ni->ibp_nid = nid;
 334         peer_ni->ibp_error = 0;
 335         peer_ni->ibp_last_alive = 0;
 336         peer_ni->ibp_max_frags = kiblnd_cfg_rdma_frags(peer_ni->ibp_ni);
 337         peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
 338         atomic_set(&peer_ni->ibp_refcount, 1);  /* 1 ref for caller */
 339
 340         INIT_LIST_HEAD(&peer_ni->ibp_list);     /* not in the peer_ni table yet */
 341         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 342         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 343
 344         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 345
 346         /* always called with a ref on ni, which prevents ni being shutdown */
 347         LASSERT(net->ibn_shutdown == 0);
 348
 349         /* npeers only grows with the global lock held */
 350         atomic_inc(&net->ibn_npeers);
 351
 352         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 353
 354         *peerp = peer_ni;
 355         return 0;
 356 }
 357
 358 void
 359 kiblnd_destroy_peer (kib_peer_ni_t *peer_ni)
 360 {
 361         kib_net_t *net = peer_ni->ibp_ni->ni_data;
 362
 363         LASSERT(net != NULL);
 364         LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
 365         LASSERT(!kiblnd_peer_active(peer_ni));
 366         LASSERT(kiblnd_peer_idle(peer_ni));
 367         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 368
 369         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 370
 371         /* NB a peer_ni's connections keep a reference on their peer_ni until
 372          * they are destroyed, so we can be assured that _all_ state to do
 373          * with this peer_ni has been cleaned up when its refcount drops to
 374          * zero. */
 375         atomic_dec(&net->ibn_npeers);
 376 }
 377
 378 kib_peer_ni_t *
 379 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 380 {
 381         /* the caller is responsible for accounting the additional reference
 382          * that this creates */
 383         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 384         struct list_head        *tmp;
 385         kib_peer_ni_t           *peer_ni;
 386
 387         list_for_each(tmp, peer_list) {
 388
 389                 peer_ni = list_entry(tmp, kib_peer_ni_t, ibp_list);
 390                 LASSERT(!kiblnd_peer_idle(peer_ni));
 391
 392                 /*
 393                  * Match a peer if its NID and the NID of the local NI it
 394                  * communicates over are the same. Otherwise don't match
 395                  * the peer, which will result in a new lnd peer being
 396                  * created.
 397                  */
 398                 if (peer_ni->ibp_nid != nid ||
 399                     peer_ni->ibp_ni->ni_nid != ni->ni_nid)
 400                         continue;
 401
 402                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 403                        peer_ni, libcfs_nid2str(nid),
 404                        atomic_read(&peer_ni->ibp_refcount),
 405                        peer_ni->ibp_version);
 406                 return peer_ni;
 407         }
 408         return NULL;
 409 }
 410
 411 void
 412 kiblnd_unlink_peer_locked (kib_peer_ni_t *peer_ni)
 413 {
 414         LASSERT(list_empty(&peer_ni->ibp_conns));
 415
 416         LASSERT (kiblnd_peer_active(peer_ni));
 417         list_del_init(&peer_ni->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer_ni);
 420 }
 421
 422 static int
 423 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
 424                      lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_ni_t           *peer_ni;
 427         struct list_head        *ptmp;
 428         int                      i;
 429         unsigned long            flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 438                         LASSERT(!kiblnd_peer_idle(peer_ni));
 439
 440                         if (peer_ni->ibp_ni != ni)
 441                                 continue;
 442
 443                         if (index-- > 0)
 444                                 continue;
 445
 446                         *nidp = peer_ni->ibp_nid;
 447                         *count = atomic_read(&peer_ni->ibp_refcount);
 448
 449                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 450                                                flags);
 451                         return 0;
 452                 }
 453         }
 454
 455         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 456         return -ENOENT;
 457 }
 458
 459 static void
 460 kiblnd_del_peer_locked (kib_peer_ni_t *peer_ni)
 461 {
 462         struct list_head        *ctmp;
 463         struct list_head        *cnxt;
 464         kib_conn_t              *conn;
 465
 466         if (list_empty(&peer_ni->ibp_conns)) {
 467                 kiblnd_unlink_peer_locked(peer_ni);
 468         } else {
 469                 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
 470                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 471
 472                         kiblnd_close_conn_locked(conn, 0);
 473                 }
 474                 /* NB closing peer_ni's last conn unlinked it. */
 475         }
 476         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 477          * last ref on it. */
 478 }
 479
 480 static int
 481 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 482 {
 483         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 484         struct list_head        *ptmp;
 485         struct list_head        *pnxt;
 486         kib_peer_ni_t           *peer_ni;
 487         int                     lo;
 488         int                     hi;
 489         int                     i;
 490         unsigned long           flags;
 491         int                     rc = -ENOENT;
 492
 493         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 494
 495         if (nid != LNET_NID_ANY) {
 496                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 497         } else {
 498                 lo = 0;
 499                 hi = kiblnd_data.kib_peer_hash_size - 1;
 500         }
 501
 502         for (i = lo; i <= hi; i++) {
 503                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 504                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 505                         LASSERT(!kiblnd_peer_idle(peer_ni));
 506
 507                         if (peer_ni->ibp_ni != ni)
 508                                 continue;
 509
 510                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 511                                 continue;
 512
 513                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 514                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 515
 516                                 list_splice_init(&peer_ni->ibp_tx_queue,
 517                                                  &zombies);
 518                         }
 519
 520                         kiblnd_del_peer_locked(peer_ni);
 521                         rc = 0;         /* matched something */
 522                 }
 523         }
 524
 525         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 526
 527         kiblnd_txlist_done(&zombies, -EIO);
 528
 529         return rc;
 530 }
 531
 532 static kib_conn_t *
 533 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 534 {
 535         kib_peer_ni_t           *peer_ni;
 536         struct list_head        *ptmp;
 537         kib_conn_t              *conn;
 538         struct list_head        *ctmp;
 539         int                     i;
 540         unsigned long           flags;
 541
 542         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 543
 544         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 545                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 546
 547                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 548                         LASSERT(!kiblnd_peer_idle(peer_ni));
 549
 550                         if (peer_ni->ibp_ni != ni)
 551                                 continue;
 552
 553                         list_for_each(ctmp, &peer_ni->ibp_conns) {
 554                                 if (index-- > 0)
 555                                         continue;
 556
 557                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 558                                 kiblnd_conn_addref(conn);
 559                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 560                                                        flags);
 561                                 return conn;
 562                         }
 563                 }
 564         }
 565
 566         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 567         return NULL;
 568 }
 569
 570 static void
 571 kiblnd_debug_rx (kib_rx_t *rx)
 572 {
 573         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 574                rx, rx->rx_status, rx->rx_msg->ibm_type,
 575                rx->rx_msg->ibm_credits);
 576 }
 577
 578 static void
 579 kiblnd_debug_tx (kib_tx_t *tx)
 580 {
 581         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 582                "cookie %#llx msg %s%s type %x cred %d\n",
 583                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 584                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 585                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 586                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 587                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 588 }
 589
 590 void
 591 kiblnd_debug_conn (kib_conn_t *conn)
 592 {
 593         struct list_head        *tmp;
 594         int                     i;
 595
 596         spin_lock(&conn->ibc_lock);
 597
 598         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 599                atomic_read(&conn->ibc_refcount), conn,
 600                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 601         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 602                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 603                conn->ibc_nsends_posted, conn->ibc_credits,
 604                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 605         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 606
 607         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 608         list_for_each(tmp, &conn->ibc_early_rxs)
 609                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 610
 611         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 612         list_for_each(tmp, &conn->ibc_tx_noops)
 613                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 614
 615         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 616         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 617                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 620         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   active_txs:\n");
 628         list_for_each(tmp, &conn->ibc_active_txs)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   rxs:\n");
 632         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 633                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 634
 635         spin_unlock(&conn->ibc_lock);
 636 }
 637
 638 int
 639 kiblnd_translate_mtu(int value)
 640 {
 641         switch (value) {
 642         default:
 643                 return -1;
 644         case 0:
 645                 return 0;
 646         case 256:
 647                 return IB_MTU_256;
 648         case 512:
 649                 return IB_MTU_512;
 650         case 1024:
 651                 return IB_MTU_1024;
 652         case 2048:
 653                 return IB_MTU_2048;
 654         case 4096:
 655                 return IB_MTU_4096;
 656         }
 657 }
 658
 659 static void
 660 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 661 {
 662         int           mtu;
 663
 664         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 665         if (cmid->route.path_rec == NULL)
 666                 return;
 667
 668         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 669         LASSERT (mtu >= 0);
 670         if (mtu != 0)
 671                 cmid->route.path_rec->mtu = mtu;
 672 }
 673
 674 static int
 675 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 676 {
 677         cpumask_t       *mask;
 678         int             vectors;
 679         int             off;
 680         int             i;
 681         lnet_nid_t      ibp_nid;
 682
 683         vectors = conn->ibc_cmid->device->num_comp_vectors;
 684         if (vectors <= 1)
 685                 return 0;
 686
 687         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 688
 689         /* hash NID to CPU id in this partition... */
 690         ibp_nid = conn->ibc_peer->ibp_nid;
 691         off = do_div(ibp_nid, cpumask_weight(mask));
 692         for_each_cpu(i, mask) {
 693                 if (off-- == 0)
 694                         return i % vectors;
 695         }
 696
 697         LBUG();
 698         return 1;
 699 }
 700
 701 kib_conn_t *
 702 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 703                    int state, int version)
 704 {
 705         /* CAVEAT EMPTOR:
 706          * If the new conn is created successfully it takes over the caller's
 707          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 708          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 709          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 710          * to destroy 'cmid' here since I'm called from the CM which still has
 711          * its ref on 'cmid'). */
 712         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 713         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 714         kib_dev_t              *dev;
 715         struct ib_qp_init_attr *init_qp_attr;
 716         struct kib_sched_info   *sched;
 717 #ifdef HAVE_IB_CQ_INIT_ATTR
 718         struct ib_cq_init_attr  cq_attr = {};
 719 #endif
 720         kib_conn_t              *conn;
 721         struct ib_cq            *cq;
 722         unsigned long           flags;
 723         int                     cpt;
 724         int                     rc;
 725         int                     i;
 726
 727         LASSERT(net != NULL);
 728         LASSERT(!in_interrupt());
 729
 730         dev = net->ibn_dev;
 731
 732         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 733         sched = kiblnd_data.kib_scheds[cpt];
 734
 735         LASSERT(sched->ibs_nthreads > 0);
 736
 737         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 738                          sizeof(*init_qp_attr));
 739         if (init_qp_attr == NULL) {
 740                 CERROR("Can't allocate qp_attr for %s\n",
 741                        libcfs_nid2str(peer_ni->ibp_nid));
 742                 goto failed_0;
 743         }
 744
 745         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 746         if (conn == NULL) {
 747                 CERROR("Can't allocate connection for %s\n",
 748                        libcfs_nid2str(peer_ni->ibp_nid));
 749                 goto failed_1;
 750         }
 751
 752         conn->ibc_state = IBLND_CONN_INIT;
 753         conn->ibc_version = version;
 754         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 755         cmid->context = conn;                   /* for future CM callbacks */
 756         conn->ibc_cmid = cmid;
 757         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 758         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 759
 760         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 761         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 762         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 763         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 764         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 765         INIT_LIST_HEAD(&conn->ibc_active_txs);
 766         spin_lock_init(&conn->ibc_lock);
 767
 768         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 769                          sizeof(*conn->ibc_connvars));
 770         if (conn->ibc_connvars == NULL) {
 771                 CERROR("Can't allocate in-progress connection state\n");
 772                 goto failed_2;
 773         }
 774
 775         write_lock_irqsave(glock, flags);
 776         if (dev->ibd_failover) {
 777                 write_unlock_irqrestore(glock, flags);
 778                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 779                 goto failed_2;
 780         }
 781
 782         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 783                 /* wakeup failover thread and teardown connection */
 784                 if (kiblnd_dev_can_failover(dev)) {
 785                         list_add_tail(&dev->ibd_fail_list,
 786                                       &kiblnd_data.kib_failed_devs);
 787                         wake_up(&kiblnd_data.kib_failover_waitq);
 788                 }
 789
 790                 write_unlock_irqrestore(glock, flags);
 791                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 792                        cmid->device->name, dev->ibd_ifname);
 793                 goto failed_2;
 794         }
 795
 796         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 797         conn->ibc_hdev = dev->ibd_hdev;
 798
 799         kiblnd_setup_mtu_locked(cmid);
 800
 801         write_unlock_irqrestore(glock, flags);
 802
 803         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 804                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 805         if (conn->ibc_rxs == NULL) {
 806                 CERROR("Cannot allocate RX buffers\n");
 807                 goto failed_2;
 808         }
 809
 810         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 811                                 IBLND_RX_MSG_PAGES(conn));
 812         if (rc != 0)
 813                 goto failed_2;
 814
 815         kiblnd_map_rx_descs(conn);
 816
 817 #ifdef HAVE_IB_CQ_INIT_ATTR
 818         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 819         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 820         cq = ib_create_cq(cmid->device,
 821                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 822                           &cq_attr);
 823 #else
 824         cq = ib_create_cq(cmid->device,
 825                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 826                           IBLND_CQ_ENTRIES(conn),
 827                           kiblnd_get_completion_vector(conn, cpt));
 828 #endif
 829         if (IS_ERR(cq)) {
 830                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 831                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 832                 goto failed_2;
 833         }
 834
 835         conn->ibc_cq = cq;
 836
 837         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 838         if (rc != 0) {
 839                 CERROR("Can't request completion notification: %d\n", rc);
 840                 goto failed_2;
 841         }
 842
 843         init_qp_attr->event_handler = kiblnd_qp_event;
 844         init_qp_attr->qp_context = conn;
 845         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 846         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 847         init_qp_attr->cap.max_send_sge = 1;
 848         init_qp_attr->cap.max_recv_sge = 1;
 849         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 850         init_qp_attr->qp_type = IB_QPT_RC;
 851         init_qp_attr->send_cq = cq;
 852         init_qp_attr->recv_cq = cq;
 853
 854         conn->ibc_sched = sched;
 855
 856         do {
 857                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 858                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 859                         break;
 860
 861                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 862         } while (rc);
 863
 864         if (rc) {
 865                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 866                         rc, init_qp_attr->cap.max_send_wr,
 867                         init_qp_attr->cap.max_recv_wr);
 868                 goto failed_2;
 869         }
 870
 871         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 872                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 873                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 874
 875         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 876
 877         /* 1 ref for caller and each rxmsg */
 878         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 879         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 880
 881         /* post receives */
 882         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 883                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 884                 if (rc != 0) {
 885                         CERROR("Can't post rxmsg: %d\n", rc);
 886
 887                         /* Make posted receives complete */
 888                         kiblnd_abort_receives(conn);
 889
 890                         /* correct # of posted buffers
 891                          * NB locking needed now I'm racing with completion */
 892                         spin_lock_irqsave(&sched->ibs_lock, flags);
 893                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 894                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 895
 896                         /* cmid will be destroyed by CM(ofed) after cm_callback
 897                          * returned, so we can't refer it anymore
 898                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 899                         rdma_destroy_qp(conn->ibc_cmid);
 900                         conn->ibc_cmid = NULL;
 901
 902                         /* Drop my own and unused rxbuffer refcounts */
 903                         while (i++ <= IBLND_RX_MSGS(conn))
 904                                 kiblnd_conn_decref(conn);
 905
 906                         return NULL;
 907                 }
 908         }
 909
 910         /* Init successful! */
 911         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 912                  state == IBLND_CONN_PASSIVE_WAIT);
 913         conn->ibc_state = state;
 914
 915         /* 1 more conn */
 916         atomic_inc(&net->ibn_nconns);
 917         return conn;
 918
 919  failed_2:
 920         kiblnd_destroy_conn(conn, true);
 921  failed_1:
 922         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 923  failed_0:
 924         return NULL;
 925 }
 926
 927 void
 928 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 929 {
 930         struct rdma_cm_id *cmid = conn->ibc_cmid;
 931         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
 932         int                rc;
 933
 934         LASSERT (!in_interrupt());
 935         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 936         LASSERT(list_empty(&conn->ibc_early_rxs));
 937         LASSERT(list_empty(&conn->ibc_tx_noops));
 938         LASSERT(list_empty(&conn->ibc_tx_queue));
 939         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 940         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 941         LASSERT(list_empty(&conn->ibc_active_txs));
 942         LASSERT (conn->ibc_noops_posted == 0);
 943         LASSERT (conn->ibc_nsends_posted == 0);
 944
 945         switch (conn->ibc_state) {
 946         default:
 947                 /* conn must be completely disengaged from the network */
 948                 LBUG();
 949
 950         case IBLND_CONN_DISCONNECTED:
 951                 /* connvars should have been freed already */
 952                 LASSERT (conn->ibc_connvars == NULL);
 953                 break;
 954
 955         case IBLND_CONN_INIT:
 956                 break;
 957         }
 958
 959         /* conn->ibc_cmid might be destroyed by CM already */
 960         if (cmid != NULL && cmid->qp != NULL)
 961                 rdma_destroy_qp(cmid);
 962
 963         if (conn->ibc_cq != NULL) {
 964                 rc = ib_destroy_cq(conn->ibc_cq);
 965                 if (rc != 0)
 966                         CWARN("Error destroying CQ: %d\n", rc);
 967         }
 968
 969         if (conn->ibc_rx_pages != NULL)
 970                 kiblnd_unmap_rx_descs(conn);
 971
 972         if (conn->ibc_rxs != NULL) {
 973                 LIBCFS_FREE(conn->ibc_rxs,
 974                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 975         }
 976
 977         if (conn->ibc_connvars != NULL)
 978                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 979
 980         if (conn->ibc_hdev != NULL)
 981                 kiblnd_hdev_decref(conn->ibc_hdev);
 982
 983         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 984         if (conn->ibc_state != IBLND_CONN_INIT) {
 985                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
 986
 987                 kiblnd_peer_decref(peer_ni);
 988                 rdma_destroy_id(cmid);
 989                 atomic_dec(&net->ibn_nconns);
 990         }
 991
 992         if (free_conn)
 993                 LIBCFS_FREE(conn, sizeof(*conn));
 994 }
 995
 996 int
 997 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
 998 {
 999         kib_conn_t              *conn;
1000         struct list_head        *ctmp;
1001         struct list_head        *cnxt;
1002         int                     count = 0;
1003
1004         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1005                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1006
1007                 CDEBUG(D_NET, "Closing conn -> %s, "
1008                               "version: %x, reason: %d\n",
1009                        libcfs_nid2str(peer_ni->ibp_nid),
1010                        conn->ibc_version, why);
1011
1012                 kiblnd_close_conn_locked(conn, why);
1013                 count++;
1014         }
1015
1016         return count;
1017 }
1018
1019 int
1020 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1021                                 int version, __u64 incarnation)
1022 {
1023         kib_conn_t              *conn;
1024         struct list_head        *ctmp;
1025         struct list_head        *cnxt;
1026         int                     count = 0;
1027
1028         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1029                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1030
1031                 if (conn->ibc_version     == version &&
1032                     conn->ibc_incarnation == incarnation)
1033                         continue;
1034
1035                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1036                               "incarnation:%#llx(%x, %#llx)\n",
1037                        libcfs_nid2str(peer_ni->ibp_nid),
1038                        conn->ibc_version, conn->ibc_incarnation,
1039                        version, incarnation);
1040
1041                 kiblnd_close_conn_locked(conn, -ESTALE);
1042                 count++;
1043         }
1044
1045         return count;
1046 }
1047
1048 static int
1049 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1050 {
1051         kib_peer_ni_t           *peer_ni;
1052         struct list_head        *ptmp;
1053         struct list_head        *pnxt;
1054         int                     lo;
1055         int                     hi;
1056         int                     i;
1057         unsigned long           flags;
1058         int                     count = 0;
1059
1060         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1061
1062         if (nid != LNET_NID_ANY)
1063                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1064         else {
1065                 lo = 0;
1066                 hi = kiblnd_data.kib_peer_hash_size - 1;
1067         }
1068
1069         for (i = lo; i <= hi; i++) {
1070                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1071
1072                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1073                         LASSERT(!kiblnd_peer_idle(peer_ni));
1074
1075                         if (peer_ni->ibp_ni != ni)
1076                                 continue;
1077
1078                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1079                                 continue;
1080
1081                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1082                 }
1083         }
1084
1085         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1086
1087         /* wildcards always succeed */
1088         if (nid == LNET_NID_ANY)
1089                 return 0;
1090
1091         return (count == 0) ? -ENOENT : 0;
1092 }
1093
1094 static int
1095 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1096 {
1097         struct libcfs_ioctl_data *data = arg;
1098         int                       rc = -EINVAL;
1099
1100         switch(cmd) {
1101         case IOC_LIBCFS_GET_PEER: {
1102                 lnet_nid_t   nid = 0;
1103                 int          count = 0;
1104
1105                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1106                                           &nid, &count);
1107                 data->ioc_nid    = nid;
1108                 data->ioc_count  = count;
1109                 break;
1110         }
1111
1112         case IOC_LIBCFS_DEL_PEER: {
1113                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1114                 break;
1115         }
1116         case IOC_LIBCFS_GET_CONN: {
1117                 kib_conn_t *conn;
1118
1119                 rc = 0;
1120                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1121                 if (conn == NULL) {
1122                         rc = -ENOENT;
1123                         break;
1124                 }
1125
1126                 LASSERT (conn->ibc_cmid != NULL);
1127                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1128                 if (conn->ibc_cmid->route.path_rec == NULL)
1129                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1130                 else
1131                         data->ioc_u32[0] =
1132                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1133                 kiblnd_conn_decref(conn);
1134                 break;
1135         }
1136         case IOC_LIBCFS_CLOSE_CONNECTION: {
1137                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1138                 break;
1139         }
1140
1141         default:
1142                 break;
1143         }
1144
1145         return rc;
1146 }
1147
1148 static void
1149 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1150 {
1151         cfs_time_t      last_alive = 0;
1152         cfs_time_t      now = cfs_time_current();
1153         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1154         kib_peer_ni_t   *peer_ni;
1155         unsigned long   flags;
1156
1157         read_lock_irqsave(glock, flags);
1158
1159         peer_ni = kiblnd_find_peer_locked(ni, nid);
1160         if (peer_ni != NULL)
1161                 last_alive = peer_ni->ibp_last_alive;
1162
1163         read_unlock_irqrestore(glock, flags);
1164
1165         if (last_alive != 0)
1166                 *when = last_alive;
1167
1168         /* peer_ni is not persistent in hash, trigger peer_ni creation
1169          * and connection establishment with a NULL tx */
1170         if (peer_ni == NULL)
1171                 kiblnd_launch_tx(ni, NULL, nid);
1172
1173         CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1174                libcfs_nid2str(nid), peer_ni,
1175                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1176         return;
1177 }
1178
1179 static void
1180 kiblnd_free_pages(kib_pages_t *p)
1181 {
1182         int     npages = p->ibp_npages;
1183         int     i;
1184
1185         for (i = 0; i < npages; i++) {
1186                 if (p->ibp_pages[i] != NULL)
1187                         __free_page(p->ibp_pages[i]);
1188         }
1189
1190         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1191 }
1192
1193 int
1194 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1195 {
1196         kib_pages_t     *p;
1197         int             i;
1198
1199         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1200                          offsetof(kib_pages_t, ibp_pages[npages]));
1201         if (p == NULL) {
1202                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1203                 return -ENOMEM;
1204         }
1205
1206         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1207         p->ibp_npages = npages;
1208
1209         for (i = 0; i < npages; i++) {
1210                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1211                                                      GFP_NOFS);
1212                 if (p->ibp_pages[i] == NULL) {
1213                         CERROR("Can't allocate page %d of %d\n", i, npages);
1214                         kiblnd_free_pages(p);
1215                         return -ENOMEM;
1216                 }
1217         }
1218
1219         *pp = p;
1220         return 0;
1221 }
1222
1223 void
1224 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1225 {
1226         kib_rx_t *rx;
1227         int       i;
1228
1229         LASSERT (conn->ibc_rxs != NULL);
1230         LASSERT (conn->ibc_hdev != NULL);
1231
1232         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1233                 rx = &conn->ibc_rxs[i];
1234
1235                 LASSERT(rx->rx_nob >= 0); /* not posted */
1236
1237                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1238                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1239                                                           rx->rx_msgaddr),
1240                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1241         }
1242
1243         kiblnd_free_pages(conn->ibc_rx_pages);
1244
1245         conn->ibc_rx_pages = NULL;
1246 }
1247
1248 void
1249 kiblnd_map_rx_descs(kib_conn_t *conn)
1250 {
1251         kib_rx_t       *rx;
1252         struct page    *pg;
1253         int             pg_off;
1254         int             ipg;
1255         int             i;
1256
1257         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1258                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1259                 rx = &conn->ibc_rxs[i];
1260
1261                 rx->rx_conn = conn;
1262                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1263
1264                 rx->rx_msgaddr =
1265                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1266                                               rx->rx_msg, IBLND_MSG_SIZE,
1267                                               DMA_FROM_DEVICE);
1268                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1269                                                   rx->rx_msgaddr));
1270                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1271
1272                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1273                        i, rx->rx_msg, rx->rx_msgaddr,
1274                        (__u64)(page_to_phys(pg) + pg_off));
1275
1276                 pg_off += IBLND_MSG_SIZE;
1277                 LASSERT(pg_off <= PAGE_SIZE);
1278
1279                 if (pg_off == PAGE_SIZE) {
1280                         pg_off = 0;
1281                         ipg++;
1282                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1283                 }
1284         }
1285 }
1286
1287 static void
1288 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1289 {
1290         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1291         kib_tx_t       *tx;
1292         int             i;
1293
1294         LASSERT (tpo->tpo_pool.po_allocated == 0);
1295
1296         if (hdev == NULL)
1297                 return;
1298
1299         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1300                 tx = &tpo->tpo_tx_descs[i];
1301                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1302                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1303                                                           tx->tx_msgaddr),
1304                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1305         }
1306
1307         kiblnd_hdev_decref(hdev);
1308         tpo->tpo_hdev = NULL;
1309 }
1310
1311 static kib_hca_dev_t *
1312 kiblnd_current_hdev(kib_dev_t *dev)
1313 {
1314         kib_hca_dev_t *hdev;
1315         unsigned long  flags;
1316         int            i = 0;
1317
1318         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1319         while (dev->ibd_failover) {
1320                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1321                 if (i++ % 50 == 0)
1322                         CDEBUG(D_NET, "%s: Wait for failover\n",
1323                                dev->ibd_ifname);
1324                 set_current_state(TASK_INTERRUPTIBLE);
1325                 schedule_timeout(cfs_time_seconds(1) / 100);
1326
1327                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1328         }
1329
1330         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1331         hdev = dev->ibd_hdev;
1332
1333         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1334
1335         return hdev;
1336 }
1337
1338 static void
1339 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1340 {
1341         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1342         kib_pool_t     *pool  = &tpo->tpo_pool;
1343         kib_net_t      *net   = pool->po_owner->ps_net;
1344         kib_dev_t      *dev;
1345         struct page    *page;
1346         kib_tx_t       *tx;
1347         int             page_offset;
1348         int             ipage;
1349         int             i;
1350
1351         LASSERT (net != NULL);
1352
1353         dev = net->ibn_dev;
1354
1355         /* pre-mapped messages are not bigger than 1 page */
1356         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1357
1358         /* No fancy arithmetic when we do the buffer calculations */
1359         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1360
1361         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1362
1363         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1364                 page = txpgs->ibp_pages[ipage];
1365                 tx = &tpo->tpo_tx_descs[i];
1366
1367                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1368                                            page_offset);
1369
1370                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1371                                                        tx->tx_msg,
1372                                                        IBLND_MSG_SIZE,
1373                                                        DMA_TO_DEVICE);
1374                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1375                                                   tx->tx_msgaddr));
1376                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1377
1378                 list_add(&tx->tx_list, &pool->po_free_list);
1379
1380                 page_offset += IBLND_MSG_SIZE;
1381                 LASSERT(page_offset <= PAGE_SIZE);
1382
1383                 if (page_offset == PAGE_SIZE) {
1384                         page_offset = 0;
1385                         ipage++;
1386                         LASSERT(ipage <= txpgs->ibp_npages);
1387                 }
1388         }
1389 }
1390
1391 #ifdef HAVE_IB_GET_DMA_MR
1392 struct ib_mr *
1393 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1394                       int negotiated_nfrags)
1395 {
1396         kib_net_t     *net   = ni->ni_data;
1397         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
1398         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1399         int     mod;
1400         __u16   nfrags;
1401
1402         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1403         mod = tunables->lnd_map_on_demand;
1404         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1405
1406         LASSERT(hdev->ibh_mrs != NULL);
1407
1408         if (mod > 0 && nfrags <= rd->rd_nfrags)
1409                 return NULL;
1410
1411         return hdev->ibh_mrs;
1412 }
1413 #endif
1414
1415 static void
1416 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1417 {
1418         LASSERT(fpo->fpo_map_count == 0);
1419
1420         if (fpo->fpo_is_fmr) {
1421                 if (fpo->fmr.fpo_fmr_pool)
1422                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1423         } else {
1424                 struct kib_fast_reg_descriptor *frd, *tmp;
1425                 int i = 0;
1426
1427                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1428                                          frd_list) {
1429                         list_del(&frd->frd_list);
1430 #ifndef HAVE_IB_MAP_MR_SG
1431                         ib_free_fast_reg_page_list(frd->frd_frpl);
1432 #endif
1433                         ib_dereg_mr(frd->frd_mr);
1434                         LIBCFS_FREE(frd, sizeof(*frd));
1435                         i++;
1436                 }
1437                 if (i < fpo->fast_reg.fpo_pool_size)
1438                         CERROR("FastReg pool still has %d regions registered\n",
1439                                 fpo->fast_reg.fpo_pool_size - i);
1440         }
1441
1442         if (fpo->fpo_hdev)
1443                 kiblnd_hdev_decref(fpo->fpo_hdev);
1444
1445         LIBCFS_FREE(fpo, sizeof(*fpo));
1446 }
1447
1448 static void
1449 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1450 {
1451         kib_fmr_pool_t *fpo, *tmp;
1452
1453         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1454                 list_del(&fpo->fpo_list);
1455                 kiblnd_destroy_fmr_pool(fpo);
1456         }
1457 }
1458
1459 static int
1460 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1461                      int ncpts)
1462 {
1463         int size = tunables->lnd_fmr_pool_size / ncpts;
1464
1465         return max(IBLND_FMR_POOL, size);
1466 }
1467
1468 static int
1469 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1470                          int ncpts)
1471 {
1472         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1473
1474         return max(IBLND_FMR_POOL_FLUSH, size);
1475 }
1476
1477 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1478 {
1479         struct ib_fmr_pool_param param = {
1480                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1481                 .page_shift        = PAGE_SHIFT,
1482                 .access            = (IB_ACCESS_LOCAL_WRITE |
1483                                       IB_ACCESS_REMOTE_WRITE),
1484                 .pool_size         = fps->fps_pool_size,
1485                 .dirty_watermark   = fps->fps_flush_trigger,
1486                 .flush_function    = NULL,
1487                 .flush_arg         = NULL,
1488                 .cache             = !!fps->fps_cache };
1489         int rc = 0;
1490
1491         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1492                                                    &param);
1493         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1494                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1495                 if (rc != -ENOSYS)
1496                         CERROR("Failed to create FMR pool: %d\n", rc);
1497                 else
1498                         CERROR("FMRs are not supported\n");
1499         }
1500
1501         return rc;
1502 }
1503
1504 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1505 {
1506         struct kib_fast_reg_descriptor *frd, *tmp;
1507         int i, rc;
1508
1509         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1510         fpo->fast_reg.fpo_pool_size = 0;
1511         for (i = 0; i < fps->fps_pool_size; i++) {
1512                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1513                                  sizeof(*frd));
1514                 if (!frd) {
1515                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1516                         rc = -ENOMEM;
1517                         goto out;
1518                 }
1519                 frd->frd_mr = NULL;
1520
1521 #ifndef HAVE_IB_MAP_MR_SG
1522                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1523                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1524                 if (IS_ERR(frd->frd_frpl)) {
1525                         rc = PTR_ERR(frd->frd_frpl);
1526                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1527                                 rc);
1528                         frd->frd_frpl = NULL;
1529                         goto out_middle;
1530                 }
1531 #endif
1532
1533 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1534                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1535                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1536 #else
1537                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1538                                           IB_MR_TYPE_MEM_REG,
1539                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1540 #endif
1541                 if (IS_ERR(frd->frd_mr)) {
1542                         rc = PTR_ERR(frd->frd_mr);
1543                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1544                         frd->frd_mr = NULL;
1545                         goto out_middle;
1546                 }
1547
1548                 /* There appears to be a bug in MLX5 code where you must
1549                  * invalidate the rkey of a new FastReg pool before first
1550                  * using it. Thus, I am marking the FRD invalid here. */
1551                 frd->frd_valid = false;
1552
1553                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1554                 fpo->fast_reg.fpo_pool_size++;
1555         }
1556
1557         return 0;
1558
1559 out_middle:
1560         if (frd->frd_mr)
1561                 ib_dereg_mr(frd->frd_mr);
1562 #ifndef HAVE_IB_MAP_MR_SG
1563         if (frd->frd_frpl)
1564                 ib_free_fast_reg_page_list(frd->frd_frpl);
1565 #endif
1566         LIBCFS_FREE(frd, sizeof(*frd));
1567
1568 out:
1569         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1570                                  frd_list) {
1571                 list_del(&frd->frd_list);
1572 #ifndef HAVE_IB_MAP_MR_SG
1573                 ib_free_fast_reg_page_list(frd->frd_frpl);
1574 #endif
1575                 ib_dereg_mr(frd->frd_mr);
1576                 LIBCFS_FREE(frd, sizeof(*frd));
1577         }
1578
1579         return rc;
1580 }
1581
1582 static int
1583 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1584 {
1585         struct ib_device_attr *dev_attr;
1586         kib_dev_t *dev = fps->fps_net->ibn_dev;
1587         kib_fmr_pool_t *fpo;
1588         int rc;
1589
1590 #ifndef HAVE_IB_DEVICE_ATTRS
1591         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1592         if (!dev_attr)
1593                 return -ENOMEM;
1594 #endif
1595
1596         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1597         if (!fpo) {
1598                 rc = -ENOMEM;
1599                 goto out_dev_attr;
1600         }
1601
1602         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1603
1604 #ifdef HAVE_IB_DEVICE_ATTRS
1605         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1606 #else
1607         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1608         if (rc) {
1609                 CERROR("Query device failed for %s: %d\n",
1610                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1611                 goto out_dev_attr;
1612         }
1613 #endif
1614
1615         /* Check for FMR or FastReg support */
1616         fpo->fpo_is_fmr = 0;
1617         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1618             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1619             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1620             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1621                 LCONSOLE_INFO("Using FMR for registration\n");
1622                 fpo->fpo_is_fmr = 1;
1623         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1624                 LCONSOLE_INFO("Using FastReg for registration\n");
1625         } else {
1626                 rc = -ENOSYS;
1627                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1628                 goto out_dev_attr;
1629         }
1630
1631         if (fpo->fpo_is_fmr)
1632                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1633         else
1634                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1635         if (rc)
1636                 goto out_fpo;
1637
1638 #ifndef HAVE_IB_DEVICE_ATTRS
1639         kfree(dev_attr);
1640 #endif
1641         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1642         fpo->fpo_owner    = fps;
1643         *pp_fpo = fpo;
1644
1645         return 0;
1646
1647 out_fpo:
1648         kiblnd_hdev_decref(fpo->fpo_hdev);
1649         LIBCFS_FREE(fpo, sizeof(*fpo));
1650
1651 out_dev_attr:
1652 #ifndef HAVE_IB_DEVICE_ATTRS
1653         kfree(dev_attr);
1654 #endif
1655
1656         return rc;
1657 }
1658
1659 static void
1660 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1661 {
1662         if (fps->fps_net == NULL) /* intialized? */
1663                 return;
1664
1665         spin_lock(&fps->fps_lock);
1666
1667         while (!list_empty(&fps->fps_pool_list)) {
1668                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1669                                                  kib_fmr_pool_t, fpo_list);
1670                 fpo->fpo_failed = 1;
1671                 list_del(&fpo->fpo_list);
1672                 if (fpo->fpo_map_count == 0)
1673                         list_add(&fpo->fpo_list, zombies);
1674                 else
1675                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1676         }
1677
1678         spin_unlock(&fps->fps_lock);
1679 }
1680
1681 static void
1682 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1683 {
1684         if (fps->fps_net != NULL) { /* initialized? */
1685                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1686                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1687         }
1688 }
1689
1690 static int
1691 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1692                         kib_net_t *net,
1693                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1694 {
1695         kib_fmr_pool_t *fpo;
1696         int             rc;
1697
1698         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1699
1700         fps->fps_net = net;
1701         fps->fps_cpt = cpt;
1702
1703         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1704         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1705         fps->fps_cache = tunables->lnd_fmr_cache;
1706
1707         spin_lock_init(&fps->fps_lock);
1708         INIT_LIST_HEAD(&fps->fps_pool_list);
1709         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1710
1711         rc = kiblnd_create_fmr_pool(fps, &fpo);
1712         if (rc == 0)
1713                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1714
1715         return rc;
1716 }
1717
1718 static int
1719 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1720 {
1721         if (fpo->fpo_map_count != 0) /* still in use */
1722                 return 0;
1723         if (fpo->fpo_failed)
1724                 return 1;
1725         return cfs_time_aftereq(now, fpo->fpo_deadline);
1726 }
1727
1728 static int
1729 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1730 {
1731         kib_hca_dev_t   *hdev;
1732         __u64           *pages = tx->tx_pages;
1733         int             npages;
1734         int             size;
1735         int             i;
1736
1737         hdev = tx->tx_pool->tpo_hdev;
1738
1739         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1740                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1741                         size += hdev->ibh_page_size) {
1742                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1743                                            hdev->ibh_page_mask) + size;
1744                 }
1745         }
1746
1747         return npages;
1748 }
1749
1750 void
1751 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1752 {
1753         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1754         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1755         kib_fmr_poolset_t *fps;
1756         cfs_time_t         now = cfs_time_current();
1757         kib_fmr_pool_t    *tmp;
1758         int                rc;
1759
1760         if (!fpo)
1761                 return;
1762
1763         fps = fpo->fpo_owner;
1764         if (fpo->fpo_is_fmr) {
1765                 if (fmr->fmr_pfmr) {
1766                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1767                         LASSERT(!rc);
1768                         fmr->fmr_pfmr = NULL;
1769                 }
1770
1771                 if (status) {
1772                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1773                         LASSERT(!rc);
1774                 }
1775         } else {
1776                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1777
1778                 if (frd) {
1779                         frd->frd_valid = false;
1780                         spin_lock(&fps->fps_lock);
1781                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1782                         spin_unlock(&fps->fps_lock);
1783                         fmr->fmr_frd = NULL;
1784                 }
1785         }
1786         fmr->fmr_pool = NULL;
1787
1788         spin_lock(&fps->fps_lock);
1789         fpo->fpo_map_count--;   /* decref the pool */
1790
1791         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1792                 /* the first pool is persistent */
1793                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1794                         continue;
1795
1796                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1797                         list_move(&fpo->fpo_list, &zombies);
1798                         fps->fps_version++;
1799                 }
1800         }
1801         spin_unlock(&fps->fps_lock);
1802
1803         if (!list_empty(&zombies))
1804                 kiblnd_destroy_fmr_pool_list(&zombies);
1805 }
1806
1807 int
1808 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1809                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1810 {
1811         kib_fmr_pool_t *fpo;
1812         __u64 *pages = tx->tx_pages;
1813         __u64 version;
1814         bool is_rx = (rd != tx->tx_rd);
1815         bool tx_pages_mapped = 0;
1816         int npages = 0;
1817         int rc;
1818
1819 again:
1820         spin_lock(&fps->fps_lock);
1821         version = fps->fps_version;
1822         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1823                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1824                 fpo->fpo_map_count++;
1825
1826                 if (fpo->fpo_is_fmr) {
1827                         struct ib_pool_fmr *pfmr;
1828
1829                         spin_unlock(&fps->fps_lock);
1830
1831                         if (!tx_pages_mapped) {
1832                                 npages = kiblnd_map_tx_pages(tx, rd);
1833                                 tx_pages_mapped = 1;
1834                         }
1835
1836                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1837                                                     pages, npages, iov);
1838                         if (likely(!IS_ERR(pfmr))) {
1839                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1840                                                       : pfmr->fmr->lkey;
1841                                 fmr->fmr_frd  = NULL;
1842                                 fmr->fmr_pfmr = pfmr;
1843                                 fmr->fmr_pool = fpo;
1844                                 return 0;
1845                         }
1846                         rc = PTR_ERR(pfmr);
1847                 } else {
1848                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1849                                 struct kib_fast_reg_descriptor *frd;
1850 #ifdef HAVE_IB_MAP_MR_SG
1851                                 struct ib_reg_wr *wr;
1852                                 int n;
1853 #else
1854                                 struct ib_rdma_wr *wr;
1855                                 struct ib_fast_reg_page_list *frpl;
1856 #endif
1857                                 struct ib_mr *mr;
1858
1859                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1860                                                         struct kib_fast_reg_descriptor,
1861                                                         frd_list);
1862                                 list_del(&frd->frd_list);
1863                                 spin_unlock(&fps->fps_lock);
1864
1865 #ifndef HAVE_IB_MAP_MR_SG
1866                                 frpl = frd->frd_frpl;
1867 #endif
1868                                 mr   = frd->frd_mr;
1869
1870                                 if (!frd->frd_valid) {
1871                                         struct ib_rdma_wr *inv_wr;
1872                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1873
1874                                         inv_wr = &frd->frd_inv_wr;
1875                                         memset(inv_wr, 0, sizeof(*inv_wr));
1876
1877                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1878                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1879                                         inv_wr->wr.ex.invalidate_rkey = key;
1880
1881                                         /* Bump the key */
1882                                         key = ib_inc_rkey(key);
1883                                         ib_update_fast_reg_key(mr, key);
1884                                 }
1885
1886 #ifdef HAVE_IB_MAP_MR_SG
1887 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1888                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1889                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1890 #else
1891                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1892                                                  tx->tx_nfrags, PAGE_SIZE);
1893 #endif
1894                                 if (unlikely(n != tx->tx_nfrags)) {
1895                                         CERROR("Failed to map mr %d/%d "
1896                                                "elements\n", n, tx->tx_nfrags);
1897                                         return n < 0 ? n : -EINVAL;
1898                                 }
1899
1900                                 mr->iova = iov;
1901
1902                                 wr = &frd->frd_fastreg_wr;
1903                                 memset(wr, 0, sizeof(*wr));
1904
1905                                 wr->wr.opcode = IB_WR_REG_MR;
1906                                 wr->wr.wr_id  = IBLND_WID_MR;
1907                                 wr->wr.num_sge = 0;
1908                                 wr->wr.send_flags = 0;
1909                                 wr->mr = mr;
1910                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1911                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1912                                               IB_ACCESS_REMOTE_WRITE);
1913 #else
1914                                 if (!tx_pages_mapped) {
1915                                         npages = kiblnd_map_tx_pages(tx, rd);
1916                                         tx_pages_mapped = 1;
1917                                 }
1918
1919                                 LASSERT(npages <= frpl->max_page_list_len);
1920                                 memcpy(frpl->page_list, pages,
1921                                         sizeof(*pages) * npages);
1922
1923                                 /* Prepare FastReg WR */
1924                                 wr = &frd->frd_fastreg_wr;
1925                                 memset(wr, 0, sizeof(*wr));
1926
1927                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1928                                 wr->wr.wr_id  = IBLND_WID_MR;
1929
1930                                 wr->wr.wr.fast_reg.iova_start = iov;
1931                                 wr->wr.wr.fast_reg.page_list  = frpl;
1932                                 wr->wr.wr.fast_reg.page_list_len = npages;
1933                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1934                                 wr->wr.wr.fast_reg.length = nob;
1935                                 wr->wr.wr.fast_reg.rkey =
1936                                                 is_rx ? mr->rkey : mr->lkey;
1937                                 wr->wr.wr.fast_reg.access_flags =
1938                                                 (IB_ACCESS_LOCAL_WRITE |
1939                                                  IB_ACCESS_REMOTE_WRITE);
1940 #endif
1941
1942                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1943                                 fmr->fmr_frd  = frd;
1944                                 fmr->fmr_pfmr = NULL;
1945                                 fmr->fmr_pool = fpo;
1946                                 return 0;
1947                         }
1948                         spin_unlock(&fps->fps_lock);
1949                         rc = -EBUSY;
1950                 }
1951
1952                 spin_lock(&fps->fps_lock);
1953                 fpo->fpo_map_count--;
1954                 if (rc != -EAGAIN) {
1955                         spin_unlock(&fps->fps_lock);
1956                         return rc;
1957                 }
1958
1959                 /* EAGAIN and ... */
1960                 if (version != fps->fps_version) {
1961                         spin_unlock(&fps->fps_lock);
1962                         goto again;
1963                 }
1964         }
1965
1966         if (fps->fps_increasing) {
1967                 spin_unlock(&fps->fps_lock);
1968                 CDEBUG(D_NET, "Another thread is allocating new "
1969                        "FMR pool, waiting for her to complete\n");
1970                 schedule();
1971                 goto again;
1972
1973         }
1974
1975         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1976                 /* someone failed recently */
1977                 spin_unlock(&fps->fps_lock);
1978                 return -EAGAIN;
1979         }
1980
1981         fps->fps_increasing = 1;
1982         spin_unlock(&fps->fps_lock);
1983
1984         CDEBUG(D_NET, "Allocate new FMR pool\n");
1985         rc = kiblnd_create_fmr_pool(fps, &fpo);
1986         spin_lock(&fps->fps_lock);
1987         fps->fps_increasing = 0;
1988         if (rc == 0) {
1989                 fps->fps_version++;
1990                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1991         } else {
1992                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1993         }
1994         spin_unlock(&fps->fps_lock);
1995
1996         goto again;
1997 }
1998
1999 static void
2000 kiblnd_fini_pool(kib_pool_t *pool)
2001 {
2002         LASSERT(list_empty(&pool->po_free_list));
2003         LASSERT(pool->po_allocated == 0);
2004
2005         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2006 }
2007
2008 static void
2009 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2010 {
2011         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2012
2013         memset(pool, 0, sizeof(kib_pool_t));
2014         INIT_LIST_HEAD(&pool->po_free_list);
2015         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2016         pool->po_owner    = ps;
2017         pool->po_size     = size;
2018 }
2019
2020 static void
2021 kiblnd_destroy_pool_list(struct list_head *head)
2022 {
2023         kib_pool_t *pool;
2024
2025         while (!list_empty(head)) {
2026                 pool = list_entry(head->next, kib_pool_t, po_list);
2027                 list_del(&pool->po_list);
2028
2029                 LASSERT(pool->po_owner != NULL);
2030                 pool->po_owner->ps_pool_destroy(pool);
2031         }
2032 }
2033
2034 static void
2035 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2036 {
2037         if (ps->ps_net == NULL) /* intialized? */
2038                 return;
2039
2040         spin_lock(&ps->ps_lock);
2041         while (!list_empty(&ps->ps_pool_list)) {
2042                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2043                                             kib_pool_t, po_list);
2044                 po->po_failed = 1;
2045                 list_del(&po->po_list);
2046                 if (po->po_allocated == 0)
2047                         list_add(&po->po_list, zombies);
2048                 else
2049                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2050         }
2051         spin_unlock(&ps->ps_lock);
2052 }
2053
2054 static void
2055 kiblnd_fini_poolset(kib_poolset_t *ps)
2056 {
2057         if (ps->ps_net != NULL) { /* initialized? */
2058                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2059                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2060         }
2061 }
2062
2063 static int
2064 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2065                     kib_net_t *net, char *name, int size,
2066                     kib_ps_pool_create_t po_create,
2067                     kib_ps_pool_destroy_t po_destroy,
2068                     kib_ps_node_init_t nd_init,
2069                     kib_ps_node_fini_t nd_fini)
2070 {
2071         kib_pool_t      *pool;
2072         int             rc;
2073
2074         memset(ps, 0, sizeof(kib_poolset_t));
2075
2076         ps->ps_cpt          = cpt;
2077         ps->ps_net          = net;
2078         ps->ps_pool_create  = po_create;
2079         ps->ps_pool_destroy = po_destroy;
2080         ps->ps_node_init    = nd_init;
2081         ps->ps_node_fini    = nd_fini;
2082         ps->ps_pool_size    = size;
2083         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2084             >= sizeof(ps->ps_name))
2085                 return -E2BIG;
2086         spin_lock_init(&ps->ps_lock);
2087         INIT_LIST_HEAD(&ps->ps_pool_list);
2088         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2089
2090         rc = ps->ps_pool_create(ps, size, &pool);
2091         if (rc == 0)
2092                 list_add(&pool->po_list, &ps->ps_pool_list);
2093         else
2094                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2095
2096         return rc;
2097 }
2098
2099 static int
2100 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2101 {
2102         if (pool->po_allocated != 0) /* still in use */
2103                 return 0;
2104         if (pool->po_failed)
2105                 return 1;
2106         return cfs_time_aftereq(now, pool->po_deadline);
2107 }
2108
2109 void
2110 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2111 {
2112         struct list_head zombies = LIST_HEAD_INIT(zombies);
2113         kib_poolset_t   *ps = pool->po_owner;
2114         kib_pool_t      *tmp;
2115         cfs_time_t       now = cfs_time_current();
2116
2117         spin_lock(&ps->ps_lock);
2118
2119         if (ps->ps_node_fini != NULL)
2120                 ps->ps_node_fini(pool, node);
2121
2122         LASSERT(pool->po_allocated > 0);
2123         list_add(node, &pool->po_free_list);
2124         pool->po_allocated--;
2125
2126         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2127                 /* the first pool is persistent */
2128                 if (ps->ps_pool_list.next == &pool->po_list)
2129                         continue;
2130
2131                 if (kiblnd_pool_is_idle(pool, now))
2132                         list_move(&pool->po_list, &zombies);
2133         }
2134         spin_unlock(&ps->ps_lock);
2135
2136         if (!list_empty(&zombies))
2137                 kiblnd_destroy_pool_list(&zombies);
2138 }
2139
2140 struct list_head *
2141 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2142 {
2143         struct list_head        *node;
2144         kib_pool_t              *pool;
2145         int                     rc;
2146         unsigned int            interval = 1;
2147         cfs_time_t              time_before;
2148         unsigned int            trips = 0;
2149
2150 again:
2151         spin_lock(&ps->ps_lock);
2152         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2153                 if (list_empty(&pool->po_free_list))
2154                         continue;
2155
2156                 pool->po_allocated++;
2157                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2158                 node = pool->po_free_list.next;
2159                 list_del(node);
2160
2161                 if (ps->ps_node_init != NULL) {
2162                         /* still hold the lock */
2163                         ps->ps_node_init(pool, node);
2164                 }
2165                 spin_unlock(&ps->ps_lock);
2166                 return node;
2167         }
2168
2169         /* no available tx pool and ... */
2170         if (ps->ps_increasing) {
2171                 /* another thread is allocating a new pool */
2172                 spin_unlock(&ps->ps_lock);
2173                 trips++;
2174                 CDEBUG(D_NET, "Another thread is allocating new "
2175                        "%s pool, waiting %d HZs for her to complete."
2176                        "trips = %d\n",
2177                        ps->ps_name, interval, trips);
2178
2179                 set_current_state(TASK_INTERRUPTIBLE);
2180                 schedule_timeout(interval);
2181                 if (interval < cfs_time_seconds(1))
2182                         interval *= 2;
2183
2184                 goto again;
2185         }
2186
2187         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2188                 /* someone failed recently */
2189                 spin_unlock(&ps->ps_lock);
2190                 return NULL;
2191         }
2192
2193         ps->ps_increasing = 1;
2194         spin_unlock(&ps->ps_lock);
2195
2196         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2197         time_before = cfs_time_current();
2198         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2199         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2200                cfs_time_current() - time_before);
2201
2202         spin_lock(&ps->ps_lock);
2203         ps->ps_increasing = 0;
2204         if (rc == 0) {
2205                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2206         } else {
2207                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2208                 CERROR("Can't allocate new %s pool because out of memory\n",
2209                        ps->ps_name);
2210         }
2211         spin_unlock(&ps->ps_lock);
2212
2213         goto again;
2214 }
2215
2216 static void
2217 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2218 {
2219         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2220         int             i;
2221
2222         LASSERT (pool->po_allocated == 0);
2223
2224         if (tpo->tpo_tx_pages != NULL) {
2225                 kiblnd_unmap_tx_pool(tpo);
2226                 kiblnd_free_pages(tpo->tpo_tx_pages);
2227         }
2228
2229         if (tpo->tpo_tx_descs == NULL)
2230                 goto out;
2231
2232         for (i = 0; i < pool->po_size; i++) {
2233                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2234
2235                 list_del(&tx->tx_list);
2236                 if (tx->tx_pages != NULL)
2237                         LIBCFS_FREE(tx->tx_pages,
2238                                     LNET_MAX_IOV *
2239                                     sizeof(*tx->tx_pages));
2240                 if (tx->tx_frags != NULL)
2241                         LIBCFS_FREE(tx->tx_frags,
2242                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2243                                     sizeof(*tx->tx_frags));
2244                 if (tx->tx_wrq != NULL)
2245                         LIBCFS_FREE(tx->tx_wrq,
2246                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2247                                     sizeof(*tx->tx_wrq));
2248                 if (tx->tx_sge != NULL)
2249                         LIBCFS_FREE(tx->tx_sge,
2250                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2251                                     sizeof(*tx->tx_sge));
2252                 if (tx->tx_rd != NULL)
2253                         LIBCFS_FREE(tx->tx_rd,
2254                                     offsetof(kib_rdma_desc_t,
2255                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2256         }
2257
2258         LIBCFS_FREE(tpo->tpo_tx_descs,
2259                     pool->po_size * sizeof(kib_tx_t));
2260 out:
2261         kiblnd_fini_pool(pool);
2262         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2263 }
2264
2265 static int kiblnd_tx_pool_size(int ncpts)
2266 {
2267         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2268
2269         return max(IBLND_TX_POOL, ntx);
2270 }
2271
2272 static int
2273 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2274 {
2275         int            i;
2276         int            npg;
2277         kib_pool_t    *pool;
2278         kib_tx_pool_t *tpo;
2279
2280         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2281         if (tpo == NULL) {
2282                 CERROR("Failed to allocate TX pool\n");
2283                 return -ENOMEM;
2284         }
2285
2286         pool = &tpo->tpo_pool;
2287         kiblnd_init_pool(ps, pool, size);
2288         tpo->tpo_tx_descs = NULL;
2289         tpo->tpo_tx_pages = NULL;
2290
2291         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2292         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2293                 CERROR("Can't allocate tx pages: %d\n", npg);
2294                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2295                 return -ENOMEM;
2296         }
2297
2298         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2299                          size * sizeof(kib_tx_t));
2300         if (tpo->tpo_tx_descs == NULL) {
2301                 CERROR("Can't allocate %d tx descriptors\n", size);
2302                 ps->ps_pool_destroy(pool);
2303                 return -ENOMEM;
2304         }
2305
2306         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2307
2308         for (i = 0; i < size; i++) {
2309                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2310
2311                 tx->tx_pool = tpo;
2312                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2313                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2314                                          lnet_cpt_table(), ps->ps_cpt,
2315                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2316                         if (tx->tx_pages == NULL)
2317                                 break;
2318                 }
2319
2320                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2321                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2322                                  sizeof(*tx->tx_frags));
2323                 if (tx->tx_frags == NULL)
2324                         break;
2325
2326                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2327
2328                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2329                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2330                                  sizeof(*tx->tx_wrq));
2331                 if (tx->tx_wrq == NULL)
2332                         break;
2333
2334                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2335                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2336                                  sizeof(*tx->tx_sge));
2337                 if (tx->tx_sge == NULL)
2338                         break;
2339
2340                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2341                                  offsetof(kib_rdma_desc_t,
2342                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2343                 if (tx->tx_rd == NULL)
2344                         break;
2345         }
2346
2347         if (i == size) {
2348                 kiblnd_map_tx_pool(tpo);
2349                 *pp_po = pool;
2350                 return 0;
2351         }
2352
2353         ps->ps_pool_destroy(pool);
2354         return -ENOMEM;
2355 }
2356
2357 static void
2358 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2359 {
2360         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2361                                              tps_poolset);
2362         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2363
2364         tx->tx_cookie = tps->tps_next_tx_cookie++;
2365 }
2366
2367 static void
2368 kiblnd_net_fini_pools(kib_net_t *net)
2369 {
2370         int     i;
2371
2372         cfs_cpt_for_each(i, lnet_cpt_table()) {
2373                 kib_tx_poolset_t        *tps;
2374                 kib_fmr_poolset_t       *fps;
2375
2376                 if (net->ibn_tx_ps != NULL) {
2377                         tps = net->ibn_tx_ps[i];
2378                         kiblnd_fini_poolset(&tps->tps_poolset);
2379                 }
2380
2381                 if (net->ibn_fmr_ps != NULL) {
2382                         fps = net->ibn_fmr_ps[i];
2383                         kiblnd_fini_fmr_poolset(fps);
2384                 }
2385         }
2386
2387         if (net->ibn_tx_ps != NULL) {
2388                 cfs_percpt_free(net->ibn_tx_ps);
2389                 net->ibn_tx_ps = NULL;
2390         }
2391
2392         if (net->ibn_fmr_ps != NULL) {
2393                 cfs_percpt_free(net->ibn_fmr_ps);
2394                 net->ibn_fmr_ps = NULL;
2395         }
2396 }
2397
2398 static int
2399 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2400                       int ncpts)
2401 {
2402         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2403 #ifdef HAVE_IB_GET_DMA_MR
2404         unsigned long   flags;
2405 #endif
2406         int             cpt;
2407         int             rc;
2408         int             i;
2409
2410         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2411
2412 #ifdef HAVE_IB_GET_DMA_MR
2413         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2414         if (tunables->lnd_map_on_demand == 0) {
2415                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2416                                            flags);
2417                 goto create_tx_pool;
2418         }
2419
2420         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2421 #endif
2422
2423         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2424                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2425                        tunables->lnd_fmr_pool_size,
2426                        *kiblnd_tunables.kib_ntx / 4);
2427                 rc = -EINVAL;
2428                 goto failed;
2429         }
2430
2431         /* TX pool must be created later than FMR, see LU-2268
2432          * for details */
2433         LASSERT(net->ibn_tx_ps == NULL);
2434
2435         /* premapping can fail if ibd_nmr > 1, so we always create
2436          * FMR pool and map-on-demand if premapping failed */
2437
2438         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2439                                            sizeof(kib_fmr_poolset_t));
2440         if (net->ibn_fmr_ps == NULL) {
2441                 CERROR("Failed to allocate FMR pool array\n");
2442                 rc = -ENOMEM;
2443                 goto failed;
2444         }
2445
2446         for (i = 0; i < ncpts; i++) {
2447                 cpt = (cpts == NULL) ? i : cpts[i];
2448                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2449                                              net, tunables);
2450                 if (rc != 0) {
2451                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2452                                cpt, rc);
2453                         goto failed;
2454                 }
2455         }
2456
2457         if (i > 0)
2458                 LASSERT(i == ncpts);
2459
2460 #ifdef HAVE_IB_GET_DMA_MR
2461  create_tx_pool:
2462 #endif
2463         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2464                                           sizeof(kib_tx_poolset_t));
2465         if (net->ibn_tx_ps == NULL) {
2466                 CERROR("Failed to allocate tx pool array\n");
2467                 rc = -ENOMEM;
2468                 goto failed;
2469         }
2470
2471         for (i = 0; i < ncpts; i++) {
2472                 cpt = (cpts == NULL) ? i : cpts[i];
2473                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2474                                          cpt, net, "TX",
2475                                          kiblnd_tx_pool_size(ncpts),
2476                                          kiblnd_create_tx_pool,
2477                                          kiblnd_destroy_tx_pool,
2478                                          kiblnd_tx_init, NULL);
2479                 if (rc != 0) {
2480                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2481                                cpt, rc);
2482                         goto failed;
2483                 }
2484         }
2485
2486         return 0;
2487  failed:
2488         kiblnd_net_fini_pools(net);
2489         LASSERT(rc != 0);
2490         return rc;
2491 }
2492
2493 static int
2494 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2495 {
2496 #ifndef HAVE_IB_DEVICE_ATTRS
2497         struct ib_device_attr *attr;
2498         int                    rc;
2499 #endif
2500
2501         /* It's safe to assume a HCA can handle a page size
2502          * matching that of the native system */
2503         hdev->ibh_page_shift = PAGE_SHIFT;
2504         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2505         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2506
2507 #ifdef HAVE_IB_DEVICE_ATTRS
2508         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2509 #else
2510         LIBCFS_ALLOC(attr, sizeof(*attr));
2511         if (attr == NULL) {
2512                 CERROR("Out of memory\n");
2513                 return -ENOMEM;
2514         }
2515
2516         rc = ib_query_device(hdev->ibh_ibdev, attr);
2517         if (rc == 0)
2518                 hdev->ibh_mr_size = attr->max_mr_size;
2519
2520         LIBCFS_FREE(attr, sizeof(*attr));
2521
2522         if (rc != 0) {
2523                 CERROR("Failed to query IB device: %d\n", rc);
2524                 return rc;
2525         }
2526 #endif
2527
2528         if (hdev->ibh_mr_size == ~0ULL) {
2529                 hdev->ibh_mr_shift = 64;
2530                 return 0;
2531         }
2532
2533         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2534         return -EINVAL;
2535 }
2536
2537 #ifdef HAVE_IB_GET_DMA_MR
2538 static void
2539 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2540 {
2541         if (hdev->ibh_mrs == NULL)
2542                 return;
2543
2544         ib_dereg_mr(hdev->ibh_mrs);
2545
2546         hdev->ibh_mrs = NULL;
2547 }
2548 #endif
2549
2550 void
2551 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2552 {
2553 #ifdef HAVE_IB_GET_DMA_MR
2554         kiblnd_hdev_cleanup_mrs(hdev);
2555 #endif
2556
2557         if (hdev->ibh_pd != NULL)
2558                 ib_dealloc_pd(hdev->ibh_pd);
2559
2560         if (hdev->ibh_cmid != NULL)
2561                 rdma_destroy_id(hdev->ibh_cmid);
2562
2563         LIBCFS_FREE(hdev, sizeof(*hdev));
2564 }
2565
2566 #ifdef HAVE_IB_GET_DMA_MR
2567 static int
2568 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2569 {
2570         struct ib_mr *mr;
2571         int           rc;
2572         int           acflags = IB_ACCESS_LOCAL_WRITE |
2573                                 IB_ACCESS_REMOTE_WRITE;
2574
2575         rc = kiblnd_hdev_get_attr(hdev);
2576         if (rc != 0)
2577                 return rc;
2578
2579         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2580         if (IS_ERR(mr)) {
2581                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2582                 kiblnd_hdev_cleanup_mrs(hdev);
2583                 return PTR_ERR(mr);
2584         }
2585
2586         hdev->ibh_mrs = mr;
2587
2588         return 0;
2589 }
2590 #endif
2591
2592 static int
2593 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2594 {       /* DUMMY */
2595         return 0;
2596 }
2597
2598 static int
2599 kiblnd_dev_need_failover(kib_dev_t *dev)
2600 {
2601         struct rdma_cm_id  *cmid;
2602         struct sockaddr_in  srcaddr;
2603         struct sockaddr_in  dstaddr;
2604         int                 rc;
2605
2606         if (dev->ibd_hdev == NULL || /* initializing */
2607             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2608             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2609                 return 1;
2610
2611         /* XXX: it's UGLY, but I don't have better way to find
2612          * ib-bonding HCA failover because:
2613          *
2614          * a. no reliable CM event for HCA failover...
2615          * b. no OFED API to get ib_device for current net_device...
2616          *
2617          * We have only two choices at this point:
2618          *
2619          * a. rdma_bind_addr(), it will conflict with listener cmid
2620          * b. rdma_resolve_addr() to zero addr */
2621         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2622                                      IB_QPT_RC);
2623         if (IS_ERR(cmid)) {
2624                 rc = PTR_ERR(cmid);
2625                 CERROR("Failed to create cmid for failover: %d\n", rc);
2626                 return rc;
2627         }
2628
2629         memset(&srcaddr, 0, sizeof(srcaddr));
2630         srcaddr.sin_family      = AF_INET;
2631         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2632
2633         memset(&dstaddr, 0, sizeof(dstaddr));
2634         dstaddr.sin_family = AF_INET;
2635         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2636                                (struct sockaddr *)&dstaddr, 1);
2637         if (rc != 0 || cmid->device == NULL) {
2638                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2639                        dev->ibd_ifname, &dev->ibd_ifip,
2640                        cmid->device, rc);
2641                 rdma_destroy_id(cmid);
2642                 return rc;
2643         }
2644
2645         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2646         rdma_destroy_id(cmid);
2647         return rc;
2648 }
2649
2650 int
2651 kiblnd_dev_failover(kib_dev_t *dev)
2652 {
2653         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2654         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2655         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2656         struct rdma_cm_id  *cmid  = NULL;
2657         kib_hca_dev_t      *hdev  = NULL;
2658         kib_hca_dev_t      *old;
2659         struct ib_pd       *pd;
2660         kib_net_t          *net;
2661         struct sockaddr_in  addr;
2662         unsigned long       flags;
2663         int                 rc = 0;
2664         int                 i;
2665
2666         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2667                  dev->ibd_can_failover ||
2668                  dev->ibd_hdev == NULL);
2669
2670         rc = kiblnd_dev_need_failover(dev);
2671         if (rc <= 0)
2672                 goto out;
2673
2674         if (dev->ibd_hdev != NULL &&
2675             dev->ibd_hdev->ibh_cmid != NULL) {
2676                 /* XXX it's not good to close old listener at here,
2677                  * because we can fail to create new listener.
2678                  * But we have to close it now, otherwise rdma_bind_addr
2679                  * will return EADDRINUSE... How crap! */
2680                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2681
2682                 cmid = dev->ibd_hdev->ibh_cmid;
2683                 /* make next schedule of kiblnd_dev_need_failover()
2684                  * return 1 for me */
2685                 dev->ibd_hdev->ibh_cmid  = NULL;
2686                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2687
2688                 rdma_destroy_id(cmid);
2689         }
2690
2691         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2692                                      IB_QPT_RC);
2693         if (IS_ERR(cmid)) {
2694                 rc = PTR_ERR(cmid);
2695                 CERROR("Failed to create cmid for failover: %d\n", rc);
2696                 goto out;
2697         }
2698
2699         memset(&addr, 0, sizeof(addr));
2700         addr.sin_family      = AF_INET;
2701         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2702         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2703
2704         /* Bind to failover device or port */
2705         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2706         if (rc != 0 || cmid->device == NULL) {
2707                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2708                        dev->ibd_ifname, &dev->ibd_ifip,
2709                        cmid->device, rc);
2710                 rdma_destroy_id(cmid);
2711                 goto out;
2712         }
2713
2714         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2715         if (hdev == NULL) {
2716                 CERROR("Failed to allocate kib_hca_dev\n");
2717                 rdma_destroy_id(cmid);
2718                 rc = -ENOMEM;
2719                 goto out;
2720         }
2721
2722         atomic_set(&hdev->ibh_ref, 1);
2723         hdev->ibh_dev   = dev;
2724         hdev->ibh_cmid  = cmid;
2725         hdev->ibh_ibdev = cmid->device;
2726
2727 #ifdef HAVE_IB_GET_DMA_MR
2728         pd = ib_alloc_pd(cmid->device);
2729 #else
2730         pd = ib_alloc_pd(cmid->device, 0);
2731 #endif
2732         if (IS_ERR(pd)) {
2733                 rc = PTR_ERR(pd);
2734                 CERROR("Can't allocate PD: %d\n", rc);
2735                 goto out;
2736         }
2737
2738         hdev->ibh_pd = pd;
2739
2740         rc = rdma_listen(cmid, 0);
2741         if (rc != 0) {
2742                 CERROR("Can't start new listener: %d\n", rc);
2743                 goto out;
2744         }
2745
2746 #ifdef HAVE_IB_GET_DMA_MR
2747         rc = kiblnd_hdev_setup_mrs(hdev);
2748         if (rc != 0) {
2749                 CERROR("Can't setup device: %d\n", rc);
2750                 goto out;
2751         }
2752 #else
2753         rc = kiblnd_hdev_get_attr(hdev);
2754         if (rc != 0) {
2755                 CERROR("Can't get device attributes: %d\n", rc);
2756                 goto out;
2757         }
2758 #endif
2759
2760         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2761
2762         old = dev->ibd_hdev;
2763         dev->ibd_hdev = hdev;   /* take over the refcount */
2764         hdev = old;
2765
2766         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2767                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2768                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2769                                             &zombie_tpo);
2770
2771                         if (net->ibn_fmr_ps != NULL)
2772                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2773                                                         &zombie_fpo);
2774                 }
2775         }
2776
2777         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2778  out:
2779         if (!list_empty(&zombie_tpo))
2780                 kiblnd_destroy_pool_list(&zombie_tpo);
2781         if (!list_empty(&zombie_ppo))
2782                 kiblnd_destroy_pool_list(&zombie_ppo);
2783         if (!list_empty(&zombie_fpo))
2784                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2785         if (hdev != NULL)
2786                 kiblnd_hdev_decref(hdev);
2787
2788         if (rc != 0)
2789                 dev->ibd_failed_failover++;
2790         else
2791                 dev->ibd_failed_failover = 0;
2792
2793         return rc;
2794 }
2795
2796 void
2797 kiblnd_destroy_dev (kib_dev_t *dev)
2798 {
2799         LASSERT (dev->ibd_nnets == 0);
2800         LASSERT(list_empty(&dev->ibd_nets));
2801
2802         list_del(&dev->ibd_fail_list);
2803         list_del(&dev->ibd_list);
2804
2805         if (dev->ibd_hdev != NULL)
2806                 kiblnd_hdev_decref(dev->ibd_hdev);
2807
2808         LIBCFS_FREE(dev, sizeof(*dev));
2809 }
2810
2811 static kib_dev_t *
2812 kiblnd_create_dev(char *ifname)
2813 {
2814         struct net_device *netdev;
2815         kib_dev_t         *dev;
2816         __u32              netmask;
2817         __u32              ip;
2818         int                up;
2819         int                rc;
2820
2821         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2822         if (rc != 0) {
2823                 CERROR("Can't query IPoIB interface %s: %d\n",
2824                        ifname, rc);
2825                 return NULL;
2826         }
2827
2828         if (!up) {
2829                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2830                 return NULL;
2831         }
2832
2833         LIBCFS_ALLOC(dev, sizeof(*dev));
2834         if (dev == NULL)
2835                 return NULL;
2836
2837         netdev = dev_get_by_name(&init_net, ifname);
2838         if (netdev == NULL) {
2839                 dev->ibd_can_failover = 0;
2840         } else {
2841                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2842                 dev_put(netdev);
2843         }
2844
2845         INIT_LIST_HEAD(&dev->ibd_nets);
2846         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2847         INIT_LIST_HEAD(&dev->ibd_fail_list);
2848         dev->ibd_ifip = ip;
2849         strcpy(&dev->ibd_ifname[0], ifname);
2850
2851         /* initialize the device */
2852         rc = kiblnd_dev_failover(dev);
2853         if (rc != 0) {
2854                 CERROR("Can't initialize device: %d\n", rc);
2855                 LIBCFS_FREE(dev, sizeof(*dev));
2856                 return NULL;
2857         }
2858
2859         list_add_tail(&dev->ibd_list,
2860                           &kiblnd_data.kib_devs);
2861         return dev;
2862 }
2863
2864 static void
2865 kiblnd_base_shutdown(void)
2866 {
2867         struct kib_sched_info   *sched;
2868         int                     i;
2869
2870         LASSERT(list_empty(&kiblnd_data.kib_devs));
2871
2872         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2873                atomic_read(&libcfs_kmemory));
2874
2875         switch (kiblnd_data.kib_init) {
2876         default:
2877                 LBUG();
2878
2879         case IBLND_INIT_ALL:
2880         case IBLND_INIT_DATA:
2881                 LASSERT (kiblnd_data.kib_peers != NULL);
2882                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2883                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2884                 }
2885                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2886                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2887                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2888                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2889
2890                 /* flag threads to terminate; wake and wait for them to die */
2891                 kiblnd_data.kib_shutdown = 1;
2892
2893                 /* NB: we really want to stop scheduler threads net by net
2894                  * instead of the whole module, this should be improved
2895                  * with dynamic configuration LNet */
2896                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2897                         wake_up_all(&sched->ibs_waitq);
2898
2899                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2900                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2901
2902                 i = 2;
2903                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2904                         i++;
2905                         /* power of 2? */
2906                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2907                                "Waiting for %d threads to terminate\n",
2908                                atomic_read(&kiblnd_data.kib_nthreads));
2909                         set_current_state(TASK_UNINTERRUPTIBLE);
2910                         schedule_timeout(cfs_time_seconds(1));
2911                 }
2912
2913                 /* fall through */
2914
2915         case IBLND_INIT_NOTHING:
2916                 break;
2917         }
2918
2919         if (kiblnd_data.kib_peers != NULL) {
2920                 LIBCFS_FREE(kiblnd_data.kib_peers,
2921                             sizeof(struct list_head) *
2922                             kiblnd_data.kib_peer_hash_size);
2923         }
2924
2925         if (kiblnd_data.kib_scheds != NULL)
2926                 cfs_percpt_free(kiblnd_data.kib_scheds);
2927
2928         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2929                atomic_read(&libcfs_kmemory));
2930
2931         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2932         module_put(THIS_MODULE);
2933 }
2934
2935 static void
2936 kiblnd_shutdown(struct lnet_ni *ni)
2937 {
2938         kib_net_t        *net = ni->ni_data;
2939         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2940         int               i;
2941         unsigned long     flags;
2942
2943         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2944
2945         if (net == NULL)
2946                 goto out;
2947
2948         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2949                atomic_read(&libcfs_kmemory));
2950
2951         write_lock_irqsave(g_lock, flags);
2952         net->ibn_shutdown = 1;
2953         write_unlock_irqrestore(g_lock, flags);
2954
2955         switch (net->ibn_init) {
2956         default:
2957                 LBUG();
2958
2959         case IBLND_INIT_ALL:
2960                 /* nuke all existing peers within this net */
2961                 kiblnd_del_peer(ni, LNET_NID_ANY);
2962
2963                 /* Wait for all peer_ni state to clean up */
2964                 i = 2;
2965                 while (atomic_read(&net->ibn_npeers) != 0) {
2966                         i++;
2967                         /* power of 2? */
2968                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2969                                "%s: waiting for %d peers to disconnect\n",
2970                                libcfs_nid2str(ni->ni_nid),
2971                                atomic_read(&net->ibn_npeers));
2972                         set_current_state(TASK_UNINTERRUPTIBLE);
2973                         schedule_timeout(cfs_time_seconds(1));
2974                 }
2975
2976                 kiblnd_net_fini_pools(net);
2977
2978                 write_lock_irqsave(g_lock, flags);
2979                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2980                 net->ibn_dev->ibd_nnets--;
2981                 list_del(&net->ibn_list);
2982                 write_unlock_irqrestore(g_lock, flags);
2983
2984                 /* fall through */
2985
2986         case IBLND_INIT_NOTHING:
2987                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
2988
2989                 if (net->ibn_dev != NULL &&
2990                     net->ibn_dev->ibd_nnets == 0)
2991                         kiblnd_destroy_dev(net->ibn_dev);
2992
2993                 break;
2994         }
2995
2996         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
2997                atomic_read(&libcfs_kmemory));
2998
2999         net->ibn_init = IBLND_INIT_NOTHING;
3000         ni->ni_data = NULL;
3001
3002         LIBCFS_FREE(net, sizeof(*net));
3003
3004 out:
3005         if (list_empty(&kiblnd_data.kib_devs))
3006                 kiblnd_base_shutdown();
3007         return;
3008 }
3009
3010 static int
3011 kiblnd_base_startup(void)
3012 {
3013         struct kib_sched_info   *sched;
3014         int                     rc;
3015         int                     i;
3016
3017         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3018
3019         try_module_get(THIS_MODULE);
3020         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3021
3022         rwlock_init(&kiblnd_data.kib_global_lock);
3023
3024         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3025         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3026
3027         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3028         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3029                      sizeof(struct list_head) *
3030                      kiblnd_data.kib_peer_hash_size);
3031         if (kiblnd_data.kib_peers == NULL)
3032                 goto failed;
3033
3034         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3035                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3036
3037         spin_lock_init(&kiblnd_data.kib_connd_lock);
3038         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3039         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3040         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3041         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3042
3043         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3044         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3045
3046         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3047                                                   sizeof(*sched));
3048         if (kiblnd_data.kib_scheds == NULL)
3049                 goto failed;
3050
3051         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3052                 int     nthrs;
3053
3054                 spin_lock_init(&sched->ibs_lock);
3055                 INIT_LIST_HEAD(&sched->ibs_conns);
3056                 init_waitqueue_head(&sched->ibs_waitq);
3057
3058                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3059                 if (*kiblnd_tunables.kib_nscheds > 0) {
3060                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3061                 } else {
3062                         /* max to half of CPUs, another half is reserved for
3063                          * upper layer modules */
3064                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3065                 }
3066
3067                 sched->ibs_nthreads_max = nthrs;
3068                 sched->ibs_cpt = i;
3069         }
3070
3071         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3072
3073         /* lists/ptrs/locks initialised */
3074         kiblnd_data.kib_init = IBLND_INIT_DATA;
3075         /*****************************************************/
3076
3077         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3078         if (rc != 0) {
3079                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3080                 goto failed;
3081         }
3082
3083         if (*kiblnd_tunables.kib_dev_failover != 0)
3084                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3085                                          "kiblnd_failover");
3086
3087         if (rc != 0) {
3088                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3089                 goto failed;
3090         }
3091
3092         /* flag everything initialised */
3093         kiblnd_data.kib_init = IBLND_INIT_ALL;
3094         /*****************************************************/
3095
3096         return 0;
3097
3098  failed:
3099         kiblnd_base_shutdown();
3100         return -ENETDOWN;
3101 }
3102
3103 static int
3104 kiblnd_start_schedulers(struct kib_sched_info *sched)
3105 {
3106         int     rc = 0;
3107         int     nthrs;
3108         int     i;
3109
3110         if (sched->ibs_nthreads == 0) {
3111                 if (*kiblnd_tunables.kib_nscheds > 0) {
3112                         nthrs = sched->ibs_nthreads_max;
3113                 } else {
3114                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3115                                                sched->ibs_cpt);
3116                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3117                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3118                 }
3119         } else {
3120                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3121                 /* increase one thread if there is new interface */
3122                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3123         }
3124
3125         for (i = 0; i < nthrs; i++) {
3126                 long    id;
3127                 char    name[20];
3128                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3129                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3130                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3131                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3132                 if (rc == 0)
3133                         continue;
3134
3135                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3136                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3137                 break;
3138         }
3139
3140         sched->ibs_nthreads += i;
3141         return rc;
3142 }
3143
3144 static int
3145 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3146 {
3147         int     cpt;
3148         int     rc;
3149         int     i;
3150
3151         for (i = 0; i < ncpts; i++) {
3152                 struct kib_sched_info *sched;
3153
3154                 cpt = (cpts == NULL) ? i : cpts[i];
3155                 sched = kiblnd_data.kib_scheds[cpt];
3156
3157                 if (!newdev && sched->ibs_nthreads > 0)
3158                         continue;
3159
3160                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3161                 if (rc != 0) {
3162                         CERROR("Failed to start scheduler threads for %s\n",
3163                                dev->ibd_ifname);
3164                         return rc;
3165                 }
3166         }
3167         return 0;
3168 }
3169
3170 static kib_dev_t *
3171 kiblnd_dev_search(char *ifname)
3172 {
3173         kib_dev_t       *alias = NULL;
3174         kib_dev_t       *dev;
3175         char            *colon;
3176         char            *colon2;
3177
3178         colon = strchr(ifname, ':');
3179         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3180                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3181                         return dev;
3182
3183                 if (alias != NULL)
3184                         continue;
3185
3186                 colon2 = strchr(dev->ibd_ifname, ':');
3187                 if (colon != NULL)
3188                         *colon = 0;
3189                 if (colon2 != NULL)
3190                         *colon2 = 0;
3191
3192                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3193                         alias = dev;
3194
3195                 if (colon != NULL)
3196                         *colon = ':';
3197                 if (colon2 != NULL)
3198                         *colon2 = ':';
3199         }
3200         return alias;
3201 }
3202
3203 static int
3204 kiblnd_startup(struct lnet_ni *ni)
3205 {
3206         char                     *ifname;
3207         kib_dev_t                *ibdev = NULL;
3208         kib_net_t                *net;
3209         unsigned long             flags;
3210         int                       rc;
3211         int                       newdev;
3212         int                       node_id;
3213
3214         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3215
3216         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3217                 rc = kiblnd_base_startup();
3218                 if (rc != 0)
3219                         return rc;
3220         }
3221
3222         LIBCFS_ALLOC(net, sizeof(*net));
3223         ni->ni_data = net;
3224         if (net == NULL)
3225                 goto failed;
3226
3227         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3228
3229         kiblnd_tunables_setup(ni);
3230
3231         if (ni->ni_interfaces[0] != NULL) {
3232                 /* Use the IPoIB interface specified in 'networks=' */
3233
3234                 CLASSERT (LNET_MAX_INTERFACES > 1);
3235                 if (ni->ni_interfaces[1] != NULL) {
3236                         CERROR("Multiple interfaces not supported\n");
3237                         goto failed;
3238                 }
3239
3240                 ifname = ni->ni_interfaces[0];
3241         } else {
3242                 ifname = *kiblnd_tunables.kib_default_ipif;
3243         }
3244
3245         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3246                 CERROR("IPoIB interface name too long: %s\n", ifname);
3247                 goto failed;
3248         }
3249
3250         ibdev = kiblnd_dev_search(ifname);
3251
3252         newdev = ibdev == NULL;
3253         /* hmm...create kib_dev even for alias */
3254         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3255                 ibdev = kiblnd_create_dev(ifname);
3256
3257         if (ibdev == NULL)
3258                 goto failed;
3259
3260         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3261         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3262
3263         net->ibn_dev = ibdev;
3264         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3265
3266         rc = kiblnd_dev_start_threads(ibdev, newdev,
3267                                       ni->ni_cpts, ni->ni_ncpts);
3268         if (rc != 0)
3269                 goto failed;
3270
3271         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3272         if (rc != 0) {
3273                 CERROR("Failed to initialize NI pools: %d\n", rc);
3274                 goto failed;
3275         }
3276
3277         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3278         ibdev->ibd_nnets++;
3279         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3280         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3281
3282         net->ibn_init = IBLND_INIT_ALL;
3283
3284         return 0;
3285
3286 failed:
3287         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3288                 kiblnd_destroy_dev(ibdev);
3289
3290         kiblnd_shutdown(ni);
3291
3292         CDEBUG(D_NET, "kiblnd_startup failed\n");
3293         return -ENETDOWN;
3294 }
3295
3296 static struct lnet_lnd the_o2iblnd = {
3297         .lnd_type       = O2IBLND,
3298         .lnd_startup    = kiblnd_startup,
3299         .lnd_shutdown   = kiblnd_shutdown,
3300         .lnd_ctl        = kiblnd_ctl,
3301         .lnd_query      = kiblnd_query,
3302         .lnd_send       = kiblnd_send,
3303         .lnd_recv       = kiblnd_recv,
3304 };
3305
3306 static void __exit ko2iblnd_exit(void)
3307 {
3308         lnet_unregister_lnd(&the_o2iblnd);
3309 }
3310
3311 static int __init ko2iblnd_init(void)
3312 {
3313         int rc;
3314
3315         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3316         CLASSERT(offsetof(kib_msg_t,
3317                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3318                  IBLND_MSG_SIZE);
3319         CLASSERT(offsetof(kib_msg_t,
3320                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3321                  <= IBLND_MSG_SIZE);
3322
3323         rc = kiblnd_tunables_init();
3324         if (rc != 0)
3325                 return rc;
3326
3327         lnet_register_lnd(&the_o2iblnd);
3328
3329         return 0;
3330 }
3331
3332 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3333 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3334 MODULE_VERSION("2.8.0");
3335 MODULE_LICENSE("GPL");
3336
3337 module_init(ko2iblnd_init);
3338 module_exit(ko2iblnd_exit);