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 /*
 702  * Get the scheduler bound to this CPT. If the scheduler has no
 703  * threads, which means that the CPT has no CPUs, then grab the
 704  * next scheduler that we can use.
 705  *
 706  * This case would be triggered if a NUMA node is configured with
 707  * no associated CPUs.
 708  */
 709 static struct kib_sched_info *
 710 kiblnd_get_scheduler(int cpt)
 711 {
 712         struct kib_sched_info *sched;
 713         int i;
 714
 715         sched = kiblnd_data.kib_scheds[cpt];
 716
 717         if (sched->ibs_nthreads > 0)
 718                 return sched;
 719
 720         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 721                 if (sched->ibs_nthreads > 0) {
 722                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 723                                         cpt, sched->ibs_cpt);
 724                         return sched;
 725                 }
 726         }
 727
 728         return NULL;
 729 }
 730
 731 kib_conn_t *
 732 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 733                    int state, int version)
 734 {
 735         /* CAVEAT EMPTOR:
 736          * If the new conn is created successfully it takes over the caller's
 737          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 738          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 739          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 740          * to destroy 'cmid' here since I'm called from the CM which still has
 741          * its ref on 'cmid'). */
 742         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 743         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 744         kib_dev_t              *dev;
 745         struct ib_qp_init_attr *init_qp_attr;
 746         struct kib_sched_info   *sched;
 747 #ifdef HAVE_IB_CQ_INIT_ATTR
 748         struct ib_cq_init_attr  cq_attr = {};
 749 #endif
 750         kib_conn_t              *conn;
 751         struct ib_cq            *cq;
 752         unsigned long           flags;
 753         int                     cpt;
 754         int                     rc;
 755         int                     i;
 756
 757         LASSERT(net != NULL);
 758         LASSERT(!in_interrupt());
 759
 760         dev = net->ibn_dev;
 761
 762         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 763         sched = kiblnd_get_scheduler(cpt);
 764
 765         if (sched == NULL) {
 766                 CERROR("no schedulers available. node is unhealthy\n");
 767                 goto failed_0;
 768         }
 769
 770         /*
 771          * The cpt might have changed if we ended up selecting a non cpt
 772          * native scheduler. So use the scheduler's cpt instead.
 773          */
 774         cpt = sched->ibs_cpt;
 775
 776         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 777                          sizeof(*init_qp_attr));
 778         if (init_qp_attr == NULL) {
 779                 CERROR("Can't allocate qp_attr for %s\n",
 780                        libcfs_nid2str(peer_ni->ibp_nid));
 781                 goto failed_0;
 782         }
 783
 784         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 785         if (conn == NULL) {
 786                 CERROR("Can't allocate connection for %s\n",
 787                        libcfs_nid2str(peer_ni->ibp_nid));
 788                 goto failed_1;
 789         }
 790
 791         conn->ibc_state = IBLND_CONN_INIT;
 792         conn->ibc_version = version;
 793         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 794         cmid->context = conn;                   /* for future CM callbacks */
 795         conn->ibc_cmid = cmid;
 796         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 797         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 798
 799         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 800         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 801         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 802         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 803         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 804         INIT_LIST_HEAD(&conn->ibc_active_txs);
 805         spin_lock_init(&conn->ibc_lock);
 806
 807         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 808                          sizeof(*conn->ibc_connvars));
 809         if (conn->ibc_connvars == NULL) {
 810                 CERROR("Can't allocate in-progress connection state\n");
 811                 goto failed_2;
 812         }
 813
 814         write_lock_irqsave(glock, flags);
 815         if (dev->ibd_failover) {
 816                 write_unlock_irqrestore(glock, flags);
 817                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 818                 goto failed_2;
 819         }
 820
 821         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 822                 /* wakeup failover thread and teardown connection */
 823                 if (kiblnd_dev_can_failover(dev)) {
 824                         list_add_tail(&dev->ibd_fail_list,
 825                                       &kiblnd_data.kib_failed_devs);
 826                         wake_up(&kiblnd_data.kib_failover_waitq);
 827                 }
 828
 829                 write_unlock_irqrestore(glock, flags);
 830                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 831                        cmid->device->name, dev->ibd_ifname);
 832                 goto failed_2;
 833         }
 834
 835         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 836         conn->ibc_hdev = dev->ibd_hdev;
 837
 838         kiblnd_setup_mtu_locked(cmid);
 839
 840         write_unlock_irqrestore(glock, flags);
 841
 842         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 843                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 844         if (conn->ibc_rxs == NULL) {
 845                 CERROR("Cannot allocate RX buffers\n");
 846                 goto failed_2;
 847         }
 848
 849         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 850                                 IBLND_RX_MSG_PAGES(conn));
 851         if (rc != 0)
 852                 goto failed_2;
 853
 854         kiblnd_map_rx_descs(conn);
 855
 856 #ifdef HAVE_IB_CQ_INIT_ATTR
 857         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 858         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 859         cq = ib_create_cq(cmid->device,
 860                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 861                           &cq_attr);
 862 #else
 863         cq = ib_create_cq(cmid->device,
 864                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 865                           IBLND_CQ_ENTRIES(conn),
 866                           kiblnd_get_completion_vector(conn, cpt));
 867 #endif
 868         if (IS_ERR(cq)) {
 869                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 870                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 871                 goto failed_2;
 872         }
 873
 874         conn->ibc_cq = cq;
 875
 876         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 877         if (rc != 0) {
 878                 CERROR("Can't request completion notification: %d\n", rc);
 879                 goto failed_2;
 880         }
 881
 882         init_qp_attr->event_handler = kiblnd_qp_event;
 883         init_qp_attr->qp_context = conn;
 884         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 885         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 886         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 887         init_qp_attr->cap.max_recv_sge = 1;
 888         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 889         init_qp_attr->qp_type = IB_QPT_RC;
 890         init_qp_attr->send_cq = cq;
 891         init_qp_attr->recv_cq = cq;
 892
 893         conn->ibc_sched = sched;
 894
 895         do {
 896                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 897                 if (!rc || init_qp_attr->cap.max_send_wr < 16)
 898                         break;
 899
 900                 init_qp_attr->cap.max_send_wr -= init_qp_attr->cap.max_send_wr / 4;
 901         } while (rc);
 902
 903         if (rc) {
 904                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 905                        "send_sge: %d, recv_sge: %d\n",
 906                        rc, init_qp_attr->cap.max_send_wr,
 907                        init_qp_attr->cap.max_recv_wr,
 908                        init_qp_attr->cap.max_send_sge,
 909                        init_qp_attr->cap.max_recv_sge);
 910                 goto failed_2;
 911         }
 912
 913         if (init_qp_attr->cap.max_send_wr != IBLND_SEND_WRS(conn))
 914                 CDEBUG(D_NET, "original send wr %d, created with %d\n",
 915                         IBLND_SEND_WRS(conn), init_qp_attr->cap.max_send_wr);
 916
 917         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 918
 919         /* 1 ref for caller and each rxmsg */
 920         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 921         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 922
 923         /* post receives */
 924         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 925                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 926                 if (rc != 0) {
 927                         CERROR("Can't post rxmsg: %d\n", rc);
 928
 929                         /* Make posted receives complete */
 930                         kiblnd_abort_receives(conn);
 931
 932                         /* correct # of posted buffers
 933                          * NB locking needed now I'm racing with completion */
 934                         spin_lock_irqsave(&sched->ibs_lock, flags);
 935                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 936                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 937
 938                         /* cmid will be destroyed by CM(ofed) after cm_callback
 939                          * returned, so we can't refer it anymore
 940                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 941                         rdma_destroy_qp(conn->ibc_cmid);
 942                         conn->ibc_cmid = NULL;
 943
 944                         /* Drop my own and unused rxbuffer refcounts */
 945                         while (i++ <= IBLND_RX_MSGS(conn))
 946                                 kiblnd_conn_decref(conn);
 947
 948                         return NULL;
 949                 }
 950         }
 951
 952         /* Init successful! */
 953         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 954                  state == IBLND_CONN_PASSIVE_WAIT);
 955         conn->ibc_state = state;
 956
 957         /* 1 more conn */
 958         atomic_inc(&net->ibn_nconns);
 959         return conn;
 960
 961  failed_2:
 962         kiblnd_destroy_conn(conn, true);
 963  failed_1:
 964         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 965  failed_0:
 966         return NULL;
 967 }
 968
 969 void
 970 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 971 {
 972         struct rdma_cm_id *cmid = conn->ibc_cmid;
 973         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
 974         int                rc;
 975
 976         LASSERT (!in_interrupt());
 977         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 978         LASSERT(list_empty(&conn->ibc_early_rxs));
 979         LASSERT(list_empty(&conn->ibc_tx_noops));
 980         LASSERT(list_empty(&conn->ibc_tx_queue));
 981         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 982         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 983         LASSERT(list_empty(&conn->ibc_active_txs));
 984         LASSERT (conn->ibc_noops_posted == 0);
 985         LASSERT (conn->ibc_nsends_posted == 0);
 986
 987         switch (conn->ibc_state) {
 988         default:
 989                 /* conn must be completely disengaged from the network */
 990                 LBUG();
 991
 992         case IBLND_CONN_DISCONNECTED:
 993                 /* connvars should have been freed already */
 994                 LASSERT (conn->ibc_connvars == NULL);
 995                 break;
 996
 997         case IBLND_CONN_INIT:
 998                 break;
 999         }
1000
1001         /* conn->ibc_cmid might be destroyed by CM already */
1002         if (cmid != NULL && cmid->qp != NULL)
1003                 rdma_destroy_qp(cmid);
1004
1005         if (conn->ibc_cq != NULL) {
1006                 rc = ib_destroy_cq(conn->ibc_cq);
1007                 if (rc != 0)
1008                         CWARN("Error destroying CQ: %d\n", rc);
1009         }
1010
1011         if (conn->ibc_rx_pages != NULL)
1012                 kiblnd_unmap_rx_descs(conn);
1013
1014         if (conn->ibc_rxs != NULL) {
1015                 LIBCFS_FREE(conn->ibc_rxs,
1016                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1017         }
1018
1019         if (conn->ibc_connvars != NULL)
1020                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1021
1022         if (conn->ibc_hdev != NULL)
1023                 kiblnd_hdev_decref(conn->ibc_hdev);
1024
1025         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1026         if (conn->ibc_state != IBLND_CONN_INIT) {
1027                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1028
1029                 kiblnd_peer_decref(peer_ni);
1030                 rdma_destroy_id(cmid);
1031                 atomic_dec(&net->ibn_nconns);
1032         }
1033
1034         if (free_conn)
1035                 LIBCFS_FREE(conn, sizeof(*conn));
1036 }
1037
1038 int
1039 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1040 {
1041         kib_conn_t              *conn;
1042         struct list_head        *ctmp;
1043         struct list_head        *cnxt;
1044         int                     count = 0;
1045
1046         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1047                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1048
1049                 CDEBUG(D_NET, "Closing conn -> %s, "
1050                               "version: %x, reason: %d\n",
1051                        libcfs_nid2str(peer_ni->ibp_nid),
1052                        conn->ibc_version, why);
1053
1054                 kiblnd_close_conn_locked(conn, why);
1055                 count++;
1056         }
1057
1058         return count;
1059 }
1060
1061 int
1062 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1063                                 int version, __u64 incarnation)
1064 {
1065         kib_conn_t              *conn;
1066         struct list_head        *ctmp;
1067         struct list_head        *cnxt;
1068         int                     count = 0;
1069
1070         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1071                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1072
1073                 if (conn->ibc_version     == version &&
1074                     conn->ibc_incarnation == incarnation)
1075                         continue;
1076
1077                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1078                               "incarnation:%#llx(%x, %#llx)\n",
1079                        libcfs_nid2str(peer_ni->ibp_nid),
1080                        conn->ibc_version, conn->ibc_incarnation,
1081                        version, incarnation);
1082
1083                 kiblnd_close_conn_locked(conn, -ESTALE);
1084                 count++;
1085         }
1086
1087         return count;
1088 }
1089
1090 static int
1091 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1092 {
1093         kib_peer_ni_t           *peer_ni;
1094         struct list_head        *ptmp;
1095         struct list_head        *pnxt;
1096         int                     lo;
1097         int                     hi;
1098         int                     i;
1099         unsigned long           flags;
1100         int                     count = 0;
1101
1102         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1103
1104         if (nid != LNET_NID_ANY)
1105                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1106         else {
1107                 lo = 0;
1108                 hi = kiblnd_data.kib_peer_hash_size - 1;
1109         }
1110
1111         for (i = lo; i <= hi; i++) {
1112                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1113
1114                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1115                         LASSERT(!kiblnd_peer_idle(peer_ni));
1116
1117                         if (peer_ni->ibp_ni != ni)
1118                                 continue;
1119
1120                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1121                                 continue;
1122
1123                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1124                 }
1125         }
1126
1127         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1128
1129         /* wildcards always succeed */
1130         if (nid == LNET_NID_ANY)
1131                 return 0;
1132
1133         return (count == 0) ? -ENOENT : 0;
1134 }
1135
1136 static int
1137 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1138 {
1139         struct libcfs_ioctl_data *data = arg;
1140         int                       rc = -EINVAL;
1141
1142         switch(cmd) {
1143         case IOC_LIBCFS_GET_PEER: {
1144                 lnet_nid_t   nid = 0;
1145                 int          count = 0;
1146
1147                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1148                                           &nid, &count);
1149                 data->ioc_nid    = nid;
1150                 data->ioc_count  = count;
1151                 break;
1152         }
1153
1154         case IOC_LIBCFS_DEL_PEER: {
1155                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1156                 break;
1157         }
1158         case IOC_LIBCFS_GET_CONN: {
1159                 kib_conn_t *conn;
1160
1161                 rc = 0;
1162                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1163                 if (conn == NULL) {
1164                         rc = -ENOENT;
1165                         break;
1166                 }
1167
1168                 LASSERT(conn->ibc_cmid != NULL);
1169                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1170                 if (conn->ibc_cmid->route.path_rec == NULL)
1171                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1172                 else
1173                         data->ioc_u32[0] =
1174                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1175                 kiblnd_conn_decref(conn);
1176                 break;
1177         }
1178         case IOC_LIBCFS_CLOSE_CONNECTION: {
1179                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1180                 break;
1181         }
1182
1183         default:
1184                 break;
1185         }
1186
1187         return rc;
1188 }
1189
1190 static void
1191 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1192 {
1193         cfs_time_t      last_alive = 0;
1194         cfs_time_t      now = cfs_time_current();
1195         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1196         kib_peer_ni_t   *peer_ni;
1197         unsigned long   flags;
1198
1199         read_lock_irqsave(glock, flags);
1200
1201         peer_ni = kiblnd_find_peer_locked(ni, nid);
1202         if (peer_ni != NULL)
1203                 last_alive = peer_ni->ibp_last_alive;
1204
1205         read_unlock_irqrestore(glock, flags);
1206
1207         if (last_alive != 0)
1208                 *when = last_alive;
1209
1210         /* peer_ni is not persistent in hash, trigger peer_ni creation
1211          * and connection establishment with a NULL tx */
1212         if (peer_ni == NULL)
1213                 kiblnd_launch_tx(ni, NULL, nid);
1214
1215         CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1216                libcfs_nid2str(nid), peer_ni,
1217                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1218         return;
1219 }
1220
1221 static void
1222 kiblnd_free_pages(kib_pages_t *p)
1223 {
1224         int     npages = p->ibp_npages;
1225         int     i;
1226
1227         for (i = 0; i < npages; i++) {
1228                 if (p->ibp_pages[i] != NULL)
1229                         __free_page(p->ibp_pages[i]);
1230         }
1231
1232         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1233 }
1234
1235 int
1236 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1237 {
1238         kib_pages_t     *p;
1239         int             i;
1240
1241         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1242                          offsetof(kib_pages_t, ibp_pages[npages]));
1243         if (p == NULL) {
1244                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1245                 return -ENOMEM;
1246         }
1247
1248         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1249         p->ibp_npages = npages;
1250
1251         for (i = 0; i < npages; i++) {
1252                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1253                                                      GFP_NOFS);
1254                 if (p->ibp_pages[i] == NULL) {
1255                         CERROR("Can't allocate page %d of %d\n", i, npages);
1256                         kiblnd_free_pages(p);
1257                         return -ENOMEM;
1258                 }
1259         }
1260
1261         *pp = p;
1262         return 0;
1263 }
1264
1265 void
1266 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1267 {
1268         kib_rx_t *rx;
1269         int       i;
1270
1271         LASSERT (conn->ibc_rxs != NULL);
1272         LASSERT (conn->ibc_hdev != NULL);
1273
1274         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1275                 rx = &conn->ibc_rxs[i];
1276
1277                 LASSERT(rx->rx_nob >= 0); /* not posted */
1278
1279                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1280                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1281                                                           rx->rx_msgaddr),
1282                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1283         }
1284
1285         kiblnd_free_pages(conn->ibc_rx_pages);
1286
1287         conn->ibc_rx_pages = NULL;
1288 }
1289
1290 void
1291 kiblnd_map_rx_descs(kib_conn_t *conn)
1292 {
1293         kib_rx_t       *rx;
1294         struct page    *pg;
1295         int             pg_off;
1296         int             ipg;
1297         int             i;
1298
1299         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1300                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1301                 rx = &conn->ibc_rxs[i];
1302
1303                 rx->rx_conn = conn;
1304                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1305
1306                 rx->rx_msgaddr =
1307                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1308                                               rx->rx_msg, IBLND_MSG_SIZE,
1309                                               DMA_FROM_DEVICE);
1310                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1311                                                   rx->rx_msgaddr));
1312                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1313
1314                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1315                        i, rx->rx_msg, rx->rx_msgaddr,
1316                        (__u64)(page_to_phys(pg) + pg_off));
1317
1318                 pg_off += IBLND_MSG_SIZE;
1319                 LASSERT(pg_off <= PAGE_SIZE);
1320
1321                 if (pg_off == PAGE_SIZE) {
1322                         pg_off = 0;
1323                         ipg++;
1324                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1325                 }
1326         }
1327 }
1328
1329 static void
1330 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1331 {
1332         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1333         kib_tx_t       *tx;
1334         int             i;
1335
1336         LASSERT (tpo->tpo_pool.po_allocated == 0);
1337
1338         if (hdev == NULL)
1339                 return;
1340
1341         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1342                 tx = &tpo->tpo_tx_descs[i];
1343                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1344                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1345                                                           tx->tx_msgaddr),
1346                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1347         }
1348
1349         kiblnd_hdev_decref(hdev);
1350         tpo->tpo_hdev = NULL;
1351 }
1352
1353 static kib_hca_dev_t *
1354 kiblnd_current_hdev(kib_dev_t *dev)
1355 {
1356         kib_hca_dev_t *hdev;
1357         unsigned long  flags;
1358         int            i = 0;
1359
1360         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1361         while (dev->ibd_failover) {
1362                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1363                 if (i++ % 50 == 0)
1364                         CDEBUG(D_NET, "%s: Wait for failover\n",
1365                                dev->ibd_ifname);
1366                 set_current_state(TASK_INTERRUPTIBLE);
1367                 schedule_timeout(cfs_time_seconds(1) / 100);
1368
1369                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1370         }
1371
1372         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1373         hdev = dev->ibd_hdev;
1374
1375         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1376
1377         return hdev;
1378 }
1379
1380 static void
1381 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1382 {
1383         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1384         kib_pool_t     *pool  = &tpo->tpo_pool;
1385         kib_net_t      *net   = pool->po_owner->ps_net;
1386         kib_dev_t      *dev;
1387         struct page    *page;
1388         kib_tx_t       *tx;
1389         int             page_offset;
1390         int             ipage;
1391         int             i;
1392
1393         LASSERT (net != NULL);
1394
1395         dev = net->ibn_dev;
1396
1397         /* pre-mapped messages are not bigger than 1 page */
1398         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1399
1400         /* No fancy arithmetic when we do the buffer calculations */
1401         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1402
1403         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1404
1405         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1406                 page = txpgs->ibp_pages[ipage];
1407                 tx = &tpo->tpo_tx_descs[i];
1408
1409                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1410                                            page_offset);
1411
1412                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1413                                                        tx->tx_msg,
1414                                                        IBLND_MSG_SIZE,
1415                                                        DMA_TO_DEVICE);
1416                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1417                                                   tx->tx_msgaddr));
1418                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1419
1420                 list_add(&tx->tx_list, &pool->po_free_list);
1421
1422                 page_offset += IBLND_MSG_SIZE;
1423                 LASSERT(page_offset <= PAGE_SIZE);
1424
1425                 if (page_offset == PAGE_SIZE) {
1426                         page_offset = 0;
1427                         ipage++;
1428                         LASSERT(ipage <= txpgs->ibp_npages);
1429                 }
1430         }
1431 }
1432
1433 #ifdef HAVE_IB_GET_DMA_MR
1434 struct ib_mr *
1435 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1436                       int negotiated_nfrags)
1437 {
1438         kib_net_t     *net   = ni->ni_data;
1439         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
1440         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1441         int     mod;
1442         __u16   nfrags;
1443
1444         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1445         mod = tunables->lnd_map_on_demand;
1446         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1447
1448         LASSERT(hdev->ibh_mrs != NULL);
1449
1450         if (mod > 0 && nfrags <= rd->rd_nfrags)
1451                 return NULL;
1452
1453         return hdev->ibh_mrs;
1454 }
1455 #endif
1456
1457 static void
1458 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1459 {
1460         LASSERT(fpo->fpo_map_count == 0);
1461
1462         if (fpo->fpo_is_fmr) {
1463                 if (fpo->fmr.fpo_fmr_pool)
1464                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1465         } else {
1466                 struct kib_fast_reg_descriptor *frd, *tmp;
1467                 int i = 0;
1468
1469                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1470                                          frd_list) {
1471                         list_del(&frd->frd_list);
1472 #ifndef HAVE_IB_MAP_MR_SG
1473                         ib_free_fast_reg_page_list(frd->frd_frpl);
1474 #endif
1475                         ib_dereg_mr(frd->frd_mr);
1476                         LIBCFS_FREE(frd, sizeof(*frd));
1477                         i++;
1478                 }
1479                 if (i < fpo->fast_reg.fpo_pool_size)
1480                         CERROR("FastReg pool still has %d regions registered\n",
1481                                 fpo->fast_reg.fpo_pool_size - i);
1482         }
1483
1484         if (fpo->fpo_hdev)
1485                 kiblnd_hdev_decref(fpo->fpo_hdev);
1486
1487         LIBCFS_FREE(fpo, sizeof(*fpo));
1488 }
1489
1490 static void
1491 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1492 {
1493         kib_fmr_pool_t *fpo, *tmp;
1494
1495         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1496                 list_del(&fpo->fpo_list);
1497                 kiblnd_destroy_fmr_pool(fpo);
1498         }
1499 }
1500
1501 static int
1502 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1503                      int ncpts)
1504 {
1505         int size = tunables->lnd_fmr_pool_size / ncpts;
1506
1507         return max(IBLND_FMR_POOL, size);
1508 }
1509
1510 static int
1511 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1512                          int ncpts)
1513 {
1514         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1515
1516         return max(IBLND_FMR_POOL_FLUSH, size);
1517 }
1518
1519 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1520 {
1521         struct ib_fmr_pool_param param = {
1522                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1523                 .page_shift        = PAGE_SHIFT,
1524                 .access            = (IB_ACCESS_LOCAL_WRITE |
1525                                       IB_ACCESS_REMOTE_WRITE),
1526                 .pool_size         = fps->fps_pool_size,
1527                 .dirty_watermark   = fps->fps_flush_trigger,
1528                 .flush_function    = NULL,
1529                 .flush_arg         = NULL,
1530                 .cache             = !!fps->fps_cache };
1531         int rc = 0;
1532
1533         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1534                                                    &param);
1535         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1536                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1537                 if (rc != -ENOSYS)
1538                         CERROR("Failed to create FMR pool: %d\n", rc);
1539                 else
1540                         CERROR("FMRs are not supported\n");
1541         }
1542         fpo->fpo_is_fmr = true;
1543
1544         return rc;
1545 }
1546
1547 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1548                                   __u32 dev_caps)
1549 {
1550         struct kib_fast_reg_descriptor *frd, *tmp;
1551         int i, rc;
1552
1553         fpo->fpo_is_fmr = false;
1554
1555         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1556         fpo->fast_reg.fpo_pool_size = 0;
1557         for (i = 0; i < fps->fps_pool_size; i++) {
1558                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1559                                  sizeof(*frd));
1560                 if (!frd) {
1561                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1562                         rc = -ENOMEM;
1563                         goto out;
1564                 }
1565                 frd->frd_mr = NULL;
1566
1567 #ifndef HAVE_IB_MAP_MR_SG
1568                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1569                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1570                 if (IS_ERR(frd->frd_frpl)) {
1571                         rc = PTR_ERR(frd->frd_frpl);
1572                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1573                                 rc);
1574                         frd->frd_frpl = NULL;
1575                         goto out_middle;
1576                 }
1577 #endif
1578
1579 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1580                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1581                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1582 #else
1583                 /*
1584                  * it is expected to get here if this is an MLX-5 card.
1585                  * MLX-4 cards will always use FMR and MLX-5 cards will
1586                  * always use fast_reg. It turns out that some MLX-5 cards
1587                  * (possibly due to older FW versions) do not natively support
1588                  * gaps. So we will need to track them here.
1589                  */
1590                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1591                                           (dev_caps &
1592                                                 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) ?
1593                                                 IB_MR_TYPE_SG_GAPS :
1594                                                 IB_MR_TYPE_MEM_REG,
1595                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1596 #endif
1597                 if (IS_ERR(frd->frd_mr)) {
1598                         rc = PTR_ERR(frd->frd_mr);
1599                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1600                         frd->frd_mr = NULL;
1601                         goto out_middle;
1602                 }
1603
1604                 /* There appears to be a bug in MLX5 code where you must
1605                  * invalidate the rkey of a new FastReg pool before first
1606                  * using it. Thus, I am marking the FRD invalid here. */
1607                 frd->frd_valid = false;
1608
1609                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1610                 fpo->fast_reg.fpo_pool_size++;
1611         }
1612
1613         return 0;
1614
1615 out_middle:
1616         if (frd->frd_mr)
1617                 ib_dereg_mr(frd->frd_mr);
1618 #ifndef HAVE_IB_MAP_MR_SG
1619         if (frd->frd_frpl)
1620                 ib_free_fast_reg_page_list(frd->frd_frpl);
1621 #endif
1622         LIBCFS_FREE(frd, sizeof(*frd));
1623
1624 out:
1625         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1626                                  frd_list) {
1627                 list_del(&frd->frd_list);
1628 #ifndef HAVE_IB_MAP_MR_SG
1629                 ib_free_fast_reg_page_list(frd->frd_frpl);
1630 #endif
1631                 ib_dereg_mr(frd->frd_mr);
1632                 LIBCFS_FREE(frd, sizeof(*frd));
1633         }
1634
1635         return rc;
1636 }
1637
1638 static int
1639 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1640 {
1641         kib_dev_t *dev = fps->fps_net->ibn_dev;
1642         kib_fmr_pool_t *fpo;
1643         int rc;
1644
1645         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1646         if (!fpo) {
1647                 return -ENOMEM;
1648         }
1649         memset(fpo, 0, sizeof(*fpo));
1650
1651         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1652
1653         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1654                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1655         else
1656                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1657         if (rc)
1658                 goto out_fpo;
1659
1660         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1661         fpo->fpo_owner    = fps;
1662         *pp_fpo = fpo;
1663
1664         return 0;
1665
1666 out_fpo:
1667         kiblnd_hdev_decref(fpo->fpo_hdev);
1668         LIBCFS_FREE(fpo, sizeof(*fpo));
1669         return rc;
1670 }
1671
1672 static void
1673 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1674 {
1675         if (fps->fps_net == NULL) /* intialized? */
1676                 return;
1677
1678         spin_lock(&fps->fps_lock);
1679
1680         while (!list_empty(&fps->fps_pool_list)) {
1681                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1682                                                  kib_fmr_pool_t, fpo_list);
1683                 fpo->fpo_failed = 1;
1684                 list_del(&fpo->fpo_list);
1685                 if (fpo->fpo_map_count == 0)
1686                         list_add(&fpo->fpo_list, zombies);
1687                 else
1688                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1689         }
1690
1691         spin_unlock(&fps->fps_lock);
1692 }
1693
1694 static void
1695 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1696 {
1697         if (fps->fps_net != NULL) { /* initialized? */
1698                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1699                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1700         }
1701 }
1702
1703 static int
1704 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1705                         kib_net_t *net,
1706                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1707 {
1708         kib_fmr_pool_t *fpo;
1709         int             rc;
1710
1711         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1712
1713         fps->fps_net = net;
1714         fps->fps_cpt = cpt;
1715
1716         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1717         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1718         fps->fps_cache = tunables->lnd_fmr_cache;
1719
1720         spin_lock_init(&fps->fps_lock);
1721         INIT_LIST_HEAD(&fps->fps_pool_list);
1722         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1723
1724         rc = kiblnd_create_fmr_pool(fps, &fpo);
1725         if (rc == 0)
1726                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1727
1728         return rc;
1729 }
1730
1731 static int
1732 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1733 {
1734         if (fpo->fpo_map_count != 0) /* still in use */
1735                 return 0;
1736         if (fpo->fpo_failed)
1737                 return 1;
1738         return cfs_time_aftereq(now, fpo->fpo_deadline);
1739 }
1740
1741 static int
1742 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1743 {
1744         kib_hca_dev_t   *hdev;
1745         __u64           *pages = tx->tx_pages;
1746         int             npages;
1747         int             size;
1748         int             i;
1749
1750         hdev = tx->tx_pool->tpo_hdev;
1751
1752         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1753                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1754                         size += hdev->ibh_page_size) {
1755                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1756                                            hdev->ibh_page_mask) + size;
1757                 }
1758         }
1759
1760         return npages;
1761 }
1762
1763 void
1764 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1765 {
1766         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1767         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1768         kib_fmr_poolset_t *fps;
1769         cfs_time_t         now = cfs_time_current();
1770         kib_fmr_pool_t    *tmp;
1771         int                rc;
1772
1773         if (!fpo)
1774                 return;
1775
1776         fps = fpo->fpo_owner;
1777         if (fpo->fpo_is_fmr) {
1778                 if (fmr->fmr_pfmr) {
1779                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1780                         LASSERT(!rc);
1781                         fmr->fmr_pfmr = NULL;
1782                 }
1783
1784                 if (status) {
1785                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1786                         LASSERT(!rc);
1787                 }
1788         } else {
1789                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1790
1791                 if (frd) {
1792                         frd->frd_valid = false;
1793                         spin_lock(&fps->fps_lock);
1794                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1795                         spin_unlock(&fps->fps_lock);
1796                         fmr->fmr_frd = NULL;
1797                 }
1798         }
1799         fmr->fmr_pool = NULL;
1800
1801         spin_lock(&fps->fps_lock);
1802         fpo->fpo_map_count--;   /* decref the pool */
1803
1804         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1805                 /* the first pool is persistent */
1806                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1807                         continue;
1808
1809                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1810                         list_move(&fpo->fpo_list, &zombies);
1811                         fps->fps_version++;
1812                 }
1813         }
1814         spin_unlock(&fps->fps_lock);
1815
1816         if (!list_empty(&zombies))
1817                 kiblnd_destroy_fmr_pool_list(&zombies);
1818 }
1819
1820 int
1821 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1822                     __u32 nob, __u64 iov, kib_fmr_t *fmr, bool *is_fastreg)
1823 {
1824         kib_fmr_pool_t *fpo;
1825         __u64 *pages = tx->tx_pages;
1826         __u64 version;
1827         bool is_rx = (rd != tx->tx_rd);
1828         bool tx_pages_mapped = 0;
1829         int npages = 0;
1830         int rc;
1831
1832 again:
1833         spin_lock(&fps->fps_lock);
1834         version = fps->fps_version;
1835         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1836                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1837                 fpo->fpo_map_count++;
1838
1839                 if (fpo->fpo_is_fmr) {
1840                         struct ib_pool_fmr *pfmr;
1841
1842                         *is_fastreg = 0;
1843                         spin_unlock(&fps->fps_lock);
1844
1845                         if (!tx_pages_mapped) {
1846                                 npages = kiblnd_map_tx_pages(tx, rd);
1847                                 tx_pages_mapped = 1;
1848                         }
1849
1850                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1851                                                     pages, npages, iov);
1852                         if (likely(!IS_ERR(pfmr))) {
1853                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1854                                                       : pfmr->fmr->lkey;
1855                                 fmr->fmr_frd  = NULL;
1856                                 fmr->fmr_pfmr = pfmr;
1857                                 fmr->fmr_pool = fpo;
1858                                 return 0;
1859                         }
1860                         rc = PTR_ERR(pfmr);
1861                 } else {
1862                         *is_fastreg = 1;
1863                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1864                                 struct kib_fast_reg_descriptor *frd;
1865 #ifdef HAVE_IB_MAP_MR_SG
1866                                 struct ib_reg_wr *wr;
1867                                 int n;
1868 #else
1869                                 struct ib_rdma_wr *wr;
1870                                 struct ib_fast_reg_page_list *frpl;
1871 #endif
1872                                 struct ib_mr *mr;
1873
1874                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1875                                                         struct kib_fast_reg_descriptor,
1876                                                         frd_list);
1877                                 list_del(&frd->frd_list);
1878                                 spin_unlock(&fps->fps_lock);
1879
1880 #ifndef HAVE_IB_MAP_MR_SG
1881                                 frpl = frd->frd_frpl;
1882 #endif
1883                                 mr   = frd->frd_mr;
1884
1885                                 if (!frd->frd_valid) {
1886                                         struct ib_rdma_wr *inv_wr;
1887                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1888
1889                                         inv_wr = &frd->frd_inv_wr;
1890                                         memset(inv_wr, 0, sizeof(*inv_wr));
1891
1892                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1893                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1894                                         inv_wr->wr.ex.invalidate_rkey = key;
1895
1896                                         /* Bump the key */
1897                                         key = ib_inc_rkey(key);
1898                                         ib_update_fast_reg_key(mr, key);
1899                                 }
1900
1901 #ifdef HAVE_IB_MAP_MR_SG
1902 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1903                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1904                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1905 #else
1906                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1907                                                  tx->tx_nfrags, PAGE_SIZE);
1908 #endif
1909                                 if (unlikely(n != tx->tx_nfrags)) {
1910                                         CERROR("Failed to map mr %d/%d "
1911                                                "elements\n", n, tx->tx_nfrags);
1912                                         return n < 0 ? n : -EINVAL;
1913                                 }
1914
1915                                 wr = &frd->frd_fastreg_wr;
1916                                 memset(wr, 0, sizeof(*wr));
1917
1918                                 wr->wr.opcode = IB_WR_REG_MR;
1919                                 wr->wr.wr_id  = IBLND_WID_MR;
1920                                 wr->wr.num_sge = 0;
1921                                 wr->wr.send_flags = 0;
1922                                 wr->mr = mr;
1923                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1924                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1925                                               IB_ACCESS_REMOTE_WRITE);
1926 #else
1927                                 if (!tx_pages_mapped) {
1928                                         npages = kiblnd_map_tx_pages(tx, rd);
1929                                         tx_pages_mapped = 1;
1930                                 }
1931
1932                                 LASSERT(npages <= frpl->max_page_list_len);
1933                                 memcpy(frpl->page_list, pages,
1934                                         sizeof(*pages) * npages);
1935
1936                                 /* Prepare FastReg WR */
1937                                 wr = &frd->frd_fastreg_wr;
1938                                 memset(wr, 0, sizeof(*wr));
1939
1940                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1941                                 wr->wr.wr_id  = IBLND_WID_MR;
1942
1943                                 wr->wr.wr.fast_reg.iova_start = iov;
1944                                 wr->wr.wr.fast_reg.page_list  = frpl;
1945                                 wr->wr.wr.fast_reg.page_list_len = npages;
1946                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1947                                 wr->wr.wr.fast_reg.length = nob;
1948                                 wr->wr.wr.fast_reg.rkey =
1949                                                 is_rx ? mr->rkey : mr->lkey;
1950                                 wr->wr.wr.fast_reg.access_flags =
1951                                                 (IB_ACCESS_LOCAL_WRITE |
1952                                                  IB_ACCESS_REMOTE_WRITE);
1953 #endif
1954
1955                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1956                                 fmr->fmr_frd  = frd;
1957                                 fmr->fmr_pfmr = NULL;
1958                                 fmr->fmr_pool = fpo;
1959                                 return 0;
1960                         }
1961                         spin_unlock(&fps->fps_lock);
1962                         rc = -EAGAIN;
1963                 }
1964
1965                 spin_lock(&fps->fps_lock);
1966                 fpo->fpo_map_count--;
1967                 if (rc != -EAGAIN) {
1968                         spin_unlock(&fps->fps_lock);
1969                         return rc;
1970                 }
1971
1972                 /* EAGAIN and ... */
1973                 if (version != fps->fps_version) {
1974                         spin_unlock(&fps->fps_lock);
1975                         goto again;
1976                 }
1977         }
1978
1979         if (fps->fps_increasing) {
1980                 spin_unlock(&fps->fps_lock);
1981                 CDEBUG(D_NET, "Another thread is allocating new "
1982                        "FMR pool, waiting for her to complete\n");
1983                 schedule();
1984                 goto again;
1985
1986         }
1987
1988         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1989                 /* someone failed recently */
1990                 spin_unlock(&fps->fps_lock);
1991                 return -EAGAIN;
1992         }
1993
1994         fps->fps_increasing = 1;
1995         spin_unlock(&fps->fps_lock);
1996
1997         CDEBUG(D_NET, "Allocate new FMR pool\n");
1998         rc = kiblnd_create_fmr_pool(fps, &fpo);
1999         spin_lock(&fps->fps_lock);
2000         fps->fps_increasing = 0;
2001         if (rc == 0) {
2002                 fps->fps_version++;
2003                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2004         } else {
2005                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2006         }
2007         spin_unlock(&fps->fps_lock);
2008
2009         goto again;
2010 }
2011
2012 static void
2013 kiblnd_fini_pool(kib_pool_t *pool)
2014 {
2015         LASSERT(list_empty(&pool->po_free_list));
2016         LASSERT(pool->po_allocated == 0);
2017
2018         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2019 }
2020
2021 static void
2022 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2023 {
2024         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2025
2026         memset(pool, 0, sizeof(kib_pool_t));
2027         INIT_LIST_HEAD(&pool->po_free_list);
2028         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2029         pool->po_owner    = ps;
2030         pool->po_size     = size;
2031 }
2032
2033 static void
2034 kiblnd_destroy_pool_list(struct list_head *head)
2035 {
2036         kib_pool_t *pool;
2037
2038         while (!list_empty(head)) {
2039                 pool = list_entry(head->next, kib_pool_t, po_list);
2040                 list_del(&pool->po_list);
2041
2042                 LASSERT(pool->po_owner != NULL);
2043                 pool->po_owner->ps_pool_destroy(pool);
2044         }
2045 }
2046
2047 static void
2048 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2049 {
2050         if (ps->ps_net == NULL) /* intialized? */
2051                 return;
2052
2053         spin_lock(&ps->ps_lock);
2054         while (!list_empty(&ps->ps_pool_list)) {
2055                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2056                                             kib_pool_t, po_list);
2057                 po->po_failed = 1;
2058                 list_del(&po->po_list);
2059                 if (po->po_allocated == 0)
2060                         list_add(&po->po_list, zombies);
2061                 else
2062                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2063         }
2064         spin_unlock(&ps->ps_lock);
2065 }
2066
2067 static void
2068 kiblnd_fini_poolset(kib_poolset_t *ps)
2069 {
2070         if (ps->ps_net != NULL) { /* initialized? */
2071                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2072                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2073         }
2074 }
2075
2076 static int
2077 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2078                     kib_net_t *net, char *name, int size,
2079                     kib_ps_pool_create_t po_create,
2080                     kib_ps_pool_destroy_t po_destroy,
2081                     kib_ps_node_init_t nd_init,
2082                     kib_ps_node_fini_t nd_fini)
2083 {
2084         kib_pool_t      *pool;
2085         int             rc;
2086
2087         memset(ps, 0, sizeof(kib_poolset_t));
2088
2089         ps->ps_cpt          = cpt;
2090         ps->ps_net          = net;
2091         ps->ps_pool_create  = po_create;
2092         ps->ps_pool_destroy = po_destroy;
2093         ps->ps_node_init    = nd_init;
2094         ps->ps_node_fini    = nd_fini;
2095         ps->ps_pool_size    = size;
2096         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2097             >= sizeof(ps->ps_name))
2098                 return -E2BIG;
2099         spin_lock_init(&ps->ps_lock);
2100         INIT_LIST_HEAD(&ps->ps_pool_list);
2101         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2102
2103         rc = ps->ps_pool_create(ps, size, &pool);
2104         if (rc == 0)
2105                 list_add(&pool->po_list, &ps->ps_pool_list);
2106         else
2107                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2108
2109         return rc;
2110 }
2111
2112 static int
2113 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2114 {
2115         if (pool->po_allocated != 0) /* still in use */
2116                 return 0;
2117         if (pool->po_failed)
2118                 return 1;
2119         return cfs_time_aftereq(now, pool->po_deadline);
2120 }
2121
2122 void
2123 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2124 {
2125         struct list_head zombies = LIST_HEAD_INIT(zombies);
2126         kib_poolset_t   *ps = pool->po_owner;
2127         kib_pool_t      *tmp;
2128         cfs_time_t       now = cfs_time_current();
2129
2130         spin_lock(&ps->ps_lock);
2131
2132         if (ps->ps_node_fini != NULL)
2133                 ps->ps_node_fini(pool, node);
2134
2135         LASSERT(pool->po_allocated > 0);
2136         list_add(node, &pool->po_free_list);
2137         pool->po_allocated--;
2138
2139         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2140                 /* the first pool is persistent */
2141                 if (ps->ps_pool_list.next == &pool->po_list)
2142                         continue;
2143
2144                 if (kiblnd_pool_is_idle(pool, now))
2145                         list_move(&pool->po_list, &zombies);
2146         }
2147         spin_unlock(&ps->ps_lock);
2148
2149         if (!list_empty(&zombies))
2150                 kiblnd_destroy_pool_list(&zombies);
2151 }
2152
2153 struct list_head *
2154 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2155 {
2156         struct list_head        *node;
2157         kib_pool_t              *pool;
2158         int                     rc;
2159         unsigned int            interval = 1;
2160         cfs_time_t              time_before;
2161         unsigned int            trips = 0;
2162
2163 again:
2164         spin_lock(&ps->ps_lock);
2165         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2166                 if (list_empty(&pool->po_free_list))
2167                         continue;
2168
2169                 pool->po_allocated++;
2170                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2171                 node = pool->po_free_list.next;
2172                 list_del(node);
2173
2174                 if (ps->ps_node_init != NULL) {
2175                         /* still hold the lock */
2176                         ps->ps_node_init(pool, node);
2177                 }
2178                 spin_unlock(&ps->ps_lock);
2179                 return node;
2180         }
2181
2182         /* no available tx pool and ... */
2183         if (ps->ps_increasing) {
2184                 /* another thread is allocating a new pool */
2185                 spin_unlock(&ps->ps_lock);
2186                 trips++;
2187                 CDEBUG(D_NET, "Another thread is allocating new "
2188                        "%s pool, waiting %d HZs for her to complete."
2189                        "trips = %d\n",
2190                        ps->ps_name, interval, trips);
2191
2192                 set_current_state(TASK_INTERRUPTIBLE);
2193                 schedule_timeout(interval);
2194                 if (interval < cfs_time_seconds(1))
2195                         interval *= 2;
2196
2197                 goto again;
2198         }
2199
2200         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2201                 /* someone failed recently */
2202                 spin_unlock(&ps->ps_lock);
2203                 return NULL;
2204         }
2205
2206         ps->ps_increasing = 1;
2207         spin_unlock(&ps->ps_lock);
2208
2209         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2210         time_before = cfs_time_current();
2211         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2212         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2213                cfs_time_current() - time_before);
2214
2215         spin_lock(&ps->ps_lock);
2216         ps->ps_increasing = 0;
2217         if (rc == 0) {
2218                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2219         } else {
2220                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2221                 CERROR("Can't allocate new %s pool because out of memory\n",
2222                        ps->ps_name);
2223         }
2224         spin_unlock(&ps->ps_lock);
2225
2226         goto again;
2227 }
2228
2229 static void
2230 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2231 {
2232         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2233         int             i;
2234
2235         LASSERT (pool->po_allocated == 0);
2236
2237         if (tpo->tpo_tx_pages != NULL) {
2238                 kiblnd_unmap_tx_pool(tpo);
2239                 kiblnd_free_pages(tpo->tpo_tx_pages);
2240         }
2241
2242         if (tpo->tpo_tx_descs == NULL)
2243                 goto out;
2244
2245         for (i = 0; i < pool->po_size; i++) {
2246                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2247                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2248
2249                 list_del(&tx->tx_list);
2250                 if (tx->tx_pages != NULL)
2251                         LIBCFS_FREE(tx->tx_pages,
2252                                     LNET_MAX_IOV *
2253                                     sizeof(*tx->tx_pages));
2254                 if (tx->tx_frags != NULL)
2255                         LIBCFS_FREE(tx->tx_frags,
2256                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2257                                     sizeof(*tx->tx_frags));
2258                 if (tx->tx_wrq != NULL)
2259                         LIBCFS_FREE(tx->tx_wrq,
2260                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2261                                     sizeof(*tx->tx_wrq));
2262                 if (tx->tx_sge != NULL)
2263                         LIBCFS_FREE(tx->tx_sge,
2264                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2265                                     sizeof(*tx->tx_sge));
2266                 if (tx->tx_rd != NULL)
2267                         LIBCFS_FREE(tx->tx_rd,
2268                                     offsetof(kib_rdma_desc_t,
2269                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2270         }
2271
2272         LIBCFS_FREE(tpo->tpo_tx_descs,
2273                     pool->po_size * sizeof(kib_tx_t));
2274 out:
2275         kiblnd_fini_pool(pool);
2276         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2277 }
2278
2279 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2280 {
2281         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2282         int ntx;
2283
2284         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2285         ntx = tunables->lnd_ntx / ncpts;
2286
2287         return max(IBLND_TX_POOL, ntx);
2288 }
2289
2290 static int
2291 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2292 {
2293         int            i;
2294         int            npg;
2295         kib_pool_t    *pool;
2296         kib_tx_pool_t *tpo;
2297
2298         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2299         if (tpo == NULL) {
2300                 CERROR("Failed to allocate TX pool\n");
2301                 return -ENOMEM;
2302         }
2303
2304         pool = &tpo->tpo_pool;
2305         kiblnd_init_pool(ps, pool, size);
2306         tpo->tpo_tx_descs = NULL;
2307         tpo->tpo_tx_pages = NULL;
2308
2309         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2310         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2311                 CERROR("Can't allocate tx pages: %d\n", npg);
2312                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2313                 return -ENOMEM;
2314         }
2315
2316         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2317                          size * sizeof(kib_tx_t));
2318         if (tpo->tpo_tx_descs == NULL) {
2319                 CERROR("Can't allocate %d tx descriptors\n", size);
2320                 ps->ps_pool_destroy(pool);
2321                 return -ENOMEM;
2322         }
2323
2324         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2325
2326         for (i = 0; i < size; i++) {
2327                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2328                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2329
2330                 tx->tx_pool = tpo;
2331                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2332                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2333                                          lnet_cpt_table(), ps->ps_cpt,
2334                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2335                         if (tx->tx_pages == NULL)
2336                                 break;
2337                 }
2338
2339                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2340                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2341                                  sizeof(*tx->tx_frags));
2342                 if (tx->tx_frags == NULL)
2343                         break;
2344
2345                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2346
2347                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2348                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2349                                  sizeof(*tx->tx_wrq));
2350                 if (tx->tx_wrq == NULL)
2351                         break;
2352
2353                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2354                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2355                                  sizeof(*tx->tx_sge));
2356                 if (tx->tx_sge == NULL)
2357                         break;
2358
2359                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2360                                  offsetof(kib_rdma_desc_t,
2361                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2362                 if (tx->tx_rd == NULL)
2363                         break;
2364         }
2365
2366         if (i == size) {
2367                 kiblnd_map_tx_pool(tpo);
2368                 *pp_po = pool;
2369                 return 0;
2370         }
2371
2372         ps->ps_pool_destroy(pool);
2373         return -ENOMEM;
2374 }
2375
2376 static void
2377 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2378 {
2379         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2380                                              tps_poolset);
2381         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2382
2383         tx->tx_cookie = tps->tps_next_tx_cookie++;
2384 }
2385
2386 static void
2387 kiblnd_net_fini_pools(kib_net_t *net)
2388 {
2389         int     i;
2390
2391         cfs_cpt_for_each(i, lnet_cpt_table()) {
2392                 kib_tx_poolset_t        *tps;
2393                 kib_fmr_poolset_t       *fps;
2394
2395                 if (net->ibn_tx_ps != NULL) {
2396                         tps = net->ibn_tx_ps[i];
2397                         kiblnd_fini_poolset(&tps->tps_poolset);
2398                 }
2399
2400                 if (net->ibn_fmr_ps != NULL) {
2401                         fps = net->ibn_fmr_ps[i];
2402                         kiblnd_fini_fmr_poolset(fps);
2403                 }
2404         }
2405
2406         if (net->ibn_tx_ps != NULL) {
2407                 cfs_percpt_free(net->ibn_tx_ps);
2408                 net->ibn_tx_ps = NULL;
2409         }
2410
2411         if (net->ibn_fmr_ps != NULL) {
2412                 cfs_percpt_free(net->ibn_fmr_ps);
2413                 net->ibn_fmr_ps = NULL;
2414         }
2415 }
2416
2417 static int
2418 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2419                       int ncpts)
2420 {
2421         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2422 #ifdef HAVE_IB_GET_DMA_MR
2423         unsigned long   flags;
2424 #endif
2425         int             cpt;
2426         int             rc;
2427         int             i;
2428
2429         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2430
2431 #ifdef HAVE_IB_GET_DMA_MR
2432         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2433         if (tunables->lnd_map_on_demand == 0) {
2434                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2435                                            flags);
2436                 goto create_tx_pool;
2437         }
2438
2439         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2440 #endif
2441
2442         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2443                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2444                        tunables->lnd_fmr_pool_size,
2445                        tunables->lnd_ntx / 4);
2446                 rc = -EINVAL;
2447                 goto failed;
2448         }
2449
2450         /* TX pool must be created later than FMR, see LU-2268
2451          * for details */
2452         LASSERT(net->ibn_tx_ps == NULL);
2453
2454         /* premapping can fail if ibd_nmr > 1, so we always create
2455          * FMR pool and map-on-demand if premapping failed */
2456
2457         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2458                                            sizeof(kib_fmr_poolset_t));
2459         if (net->ibn_fmr_ps == NULL) {
2460                 CERROR("Failed to allocate FMR pool array\n");
2461                 rc = -ENOMEM;
2462                 goto failed;
2463         }
2464
2465         for (i = 0; i < ncpts; i++) {
2466                 cpt = (cpts == NULL) ? i : cpts[i];
2467                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2468                                              net, tunables);
2469                 if (rc != 0) {
2470                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2471                                cpt, rc);
2472                         goto failed;
2473                 }
2474         }
2475
2476         if (i > 0)
2477                 LASSERT(i == ncpts);
2478
2479 #ifdef HAVE_IB_GET_DMA_MR
2480  create_tx_pool:
2481 #endif
2482         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2483                                           sizeof(kib_tx_poolset_t));
2484         if (net->ibn_tx_ps == NULL) {
2485                 CERROR("Failed to allocate tx pool array\n");
2486                 rc = -ENOMEM;
2487                 goto failed;
2488         }
2489
2490         for (i = 0; i < ncpts; i++) {
2491                 cpt = (cpts == NULL) ? i : cpts[i];
2492                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2493                                          cpt, net, "TX",
2494                                          kiblnd_tx_pool_size(ni, ncpts),
2495                                          kiblnd_create_tx_pool,
2496                                          kiblnd_destroy_tx_pool,
2497                                          kiblnd_tx_init, NULL);
2498                 if (rc != 0) {
2499                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2500                                cpt, rc);
2501                         goto failed;
2502                 }
2503         }
2504
2505         return 0;
2506  failed:
2507         kiblnd_net_fini_pools(net);
2508         LASSERT(rc != 0);
2509         return rc;
2510 }
2511
2512 static int
2513 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2514 {
2515         struct ib_device_attr *dev_attr;
2516         int rc = 0;
2517
2518         /* It's safe to assume a HCA can handle a page size
2519          * matching that of the native system */
2520         hdev->ibh_page_shift = PAGE_SHIFT;
2521         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2522         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2523
2524 #ifndef HAVE_IB_DEVICE_ATTRS
2525         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2526         if (dev_attr == NULL) {
2527                 CERROR("Out of memory\n");
2528                 return -ENOMEM;
2529         }
2530
2531         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2532         if (rc != 0) {
2533                 CERROR("Failed to query IB device: %d\n", rc);
2534                 goto out_clean_attr;
2535         }
2536 #else
2537         dev_attr = &hdev->ibh_ibdev->attrs;
2538 #endif
2539
2540         hdev->ibh_mr_size = dev_attr->max_mr_size;
2541
2542         /* Setup device Memory Registration capabilities */
2543         if (hdev->ibh_ibdev->alloc_fmr &&
2544             hdev->ibh_ibdev->dealloc_fmr &&
2545             hdev->ibh_ibdev->map_phys_fmr &&
2546             hdev->ibh_ibdev->unmap_fmr) {
2547                 LCONSOLE_INFO("Using FMR for registration\n");
2548                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2549         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2550                 LCONSOLE_INFO("Using FastReg for registration\n");
2551                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2552 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2553                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2554                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2555 #endif
2556         } else {
2557                 rc = -ENOSYS;
2558         }
2559
2560         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2561                 hdev->ibh_mr_shift = 64;
2562         else if (rc != 0)
2563                 rc = -EINVAL;
2564
2565 #ifndef HAVE_IB_DEVICE_ATTRS
2566 out_clean_attr:
2567         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2568 #endif
2569
2570         if (rc == -ENOSYS)
2571                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2572                        "register memory: %d\n", rc);
2573         else if (rc == -EINVAL)
2574                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2575         return rc;
2576 }
2577
2578 #ifdef HAVE_IB_GET_DMA_MR
2579 static void
2580 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2581 {
2582         if (hdev->ibh_mrs == NULL)
2583                 return;
2584
2585         ib_dereg_mr(hdev->ibh_mrs);
2586
2587         hdev->ibh_mrs = NULL;
2588 }
2589 #endif
2590
2591 void
2592 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2593 {
2594 #ifdef HAVE_IB_GET_DMA_MR
2595         kiblnd_hdev_cleanup_mrs(hdev);
2596 #endif
2597
2598         if (hdev->ibh_pd != NULL)
2599                 ib_dealloc_pd(hdev->ibh_pd);
2600
2601         if (hdev->ibh_cmid != NULL)
2602                 rdma_destroy_id(hdev->ibh_cmid);
2603
2604         LIBCFS_FREE(hdev, sizeof(*hdev));
2605 }
2606
2607 #ifdef HAVE_IB_GET_DMA_MR
2608 static int
2609 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2610 {
2611         struct ib_mr *mr;
2612         int           acflags = IB_ACCESS_LOCAL_WRITE |
2613                                 IB_ACCESS_REMOTE_WRITE;
2614
2615         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2616         if (IS_ERR(mr)) {
2617                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2618                 kiblnd_hdev_cleanup_mrs(hdev);
2619                 return PTR_ERR(mr);
2620         }
2621
2622         hdev->ibh_mrs = mr;
2623
2624         return 0;
2625 }
2626 #endif
2627
2628 static int
2629 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2630 {       /* DUMMY */
2631         return 0;
2632 }
2633
2634 static int
2635 kiblnd_dev_need_failover(kib_dev_t *dev)
2636 {
2637         struct rdma_cm_id  *cmid;
2638         struct sockaddr_in  srcaddr;
2639         struct sockaddr_in  dstaddr;
2640         int                 rc;
2641
2642         if (dev->ibd_hdev == NULL || /* initializing */
2643             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2644             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2645                 return 1;
2646
2647         /* XXX: it's UGLY, but I don't have better way to find
2648          * ib-bonding HCA failover because:
2649          *
2650          * a. no reliable CM event for HCA failover...
2651          * b. no OFED API to get ib_device for current net_device...
2652          *
2653          * We have only two choices at this point:
2654          *
2655          * a. rdma_bind_addr(), it will conflict with listener cmid
2656          * b. rdma_resolve_addr() to zero addr */
2657         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2658                                      IB_QPT_RC);
2659         if (IS_ERR(cmid)) {
2660                 rc = PTR_ERR(cmid);
2661                 CERROR("Failed to create cmid for failover: %d\n", rc);
2662                 return rc;
2663         }
2664
2665         memset(&srcaddr, 0, sizeof(srcaddr));
2666         srcaddr.sin_family      = AF_INET;
2667         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2668
2669         memset(&dstaddr, 0, sizeof(dstaddr));
2670         dstaddr.sin_family = AF_INET;
2671         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2672                                (struct sockaddr *)&dstaddr, 1);
2673         if (rc != 0 || cmid->device == NULL) {
2674                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2675                        dev->ibd_ifname, &dev->ibd_ifip,
2676                        cmid->device, rc);
2677                 rdma_destroy_id(cmid);
2678                 return rc;
2679         }
2680
2681         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2682         rdma_destroy_id(cmid);
2683         return rc;
2684 }
2685
2686 int
2687 kiblnd_dev_failover(kib_dev_t *dev)
2688 {
2689         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2690         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2691         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2692         struct rdma_cm_id  *cmid  = NULL;
2693         kib_hca_dev_t      *hdev  = NULL;
2694         kib_hca_dev_t      *old;
2695         struct ib_pd       *pd;
2696         kib_net_t          *net;
2697         struct sockaddr_in  addr;
2698         unsigned long       flags;
2699         int                 rc = 0;
2700         int                 i;
2701
2702         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2703                  dev->ibd_can_failover ||
2704                  dev->ibd_hdev == NULL);
2705
2706         rc = kiblnd_dev_need_failover(dev);
2707         if (rc <= 0)
2708                 goto out;
2709
2710         if (dev->ibd_hdev != NULL &&
2711             dev->ibd_hdev->ibh_cmid != NULL) {
2712                 /* XXX it's not good to close old listener at here,
2713                  * because we can fail to create new listener.
2714                  * But we have to close it now, otherwise rdma_bind_addr
2715                  * will return EADDRINUSE... How crap! */
2716                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2717
2718                 cmid = dev->ibd_hdev->ibh_cmid;
2719                 /* make next schedule of kiblnd_dev_need_failover()
2720                  * return 1 for me */
2721                 dev->ibd_hdev->ibh_cmid  = NULL;
2722                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2723
2724                 rdma_destroy_id(cmid);
2725         }
2726
2727         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2728                                      IB_QPT_RC);
2729         if (IS_ERR(cmid)) {
2730                 rc = PTR_ERR(cmid);
2731                 CERROR("Failed to create cmid for failover: %d\n", rc);
2732                 goto out;
2733         }
2734
2735         memset(&addr, 0, sizeof(addr));
2736         addr.sin_family      = AF_INET;
2737         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2738         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2739
2740         /* Bind to failover device or port */
2741         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2742         if (rc != 0 || cmid->device == NULL) {
2743                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2744                        dev->ibd_ifname, &dev->ibd_ifip,
2745                        cmid->device, rc);
2746                 rdma_destroy_id(cmid);
2747                 goto out;
2748         }
2749
2750         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2751         if (hdev == NULL) {
2752                 CERROR("Failed to allocate kib_hca_dev\n");
2753                 rdma_destroy_id(cmid);
2754                 rc = -ENOMEM;
2755                 goto out;
2756         }
2757
2758         atomic_set(&hdev->ibh_ref, 1);
2759         hdev->ibh_dev   = dev;
2760         hdev->ibh_cmid  = cmid;
2761         hdev->ibh_ibdev = cmid->device;
2762
2763 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2764         pd = ib_alloc_pd(cmid->device, 0);
2765 #else
2766         pd = ib_alloc_pd(cmid->device);
2767 #endif
2768         if (IS_ERR(pd)) {
2769                 rc = PTR_ERR(pd);
2770                 CERROR("Can't allocate PD: %d\n", rc);
2771                 goto out;
2772         }
2773
2774         hdev->ibh_pd = pd;
2775
2776         rc = rdma_listen(cmid, 0);
2777         if (rc != 0) {
2778                 CERROR("Can't start new listener: %d\n", rc);
2779                 goto out;
2780         }
2781
2782         rc = kiblnd_hdev_get_attr(hdev);
2783         if (rc != 0) {
2784                 CERROR("Can't get device attributes: %d\n", rc);
2785                 goto out;
2786         }
2787
2788 #ifdef HAVE_IB_GET_DMA_MR
2789         rc = kiblnd_hdev_setup_mrs(hdev);
2790         if (rc != 0) {
2791                 CERROR("Can't setup device: %d\n", rc);
2792                 goto out;
2793         }
2794 #endif
2795
2796         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2797
2798         old = dev->ibd_hdev;
2799         dev->ibd_hdev = hdev;   /* take over the refcount */
2800         hdev = old;
2801
2802         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2803                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2804                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2805                                             &zombie_tpo);
2806
2807                         if (net->ibn_fmr_ps != NULL)
2808                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2809                                                         &zombie_fpo);
2810                 }
2811         }
2812
2813         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2814  out:
2815         if (!list_empty(&zombie_tpo))
2816                 kiblnd_destroy_pool_list(&zombie_tpo);
2817         if (!list_empty(&zombie_ppo))
2818                 kiblnd_destroy_pool_list(&zombie_ppo);
2819         if (!list_empty(&zombie_fpo))
2820                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2821         if (hdev != NULL)
2822                 kiblnd_hdev_decref(hdev);
2823
2824         if (rc != 0)
2825                 dev->ibd_failed_failover++;
2826         else
2827                 dev->ibd_failed_failover = 0;
2828
2829         return rc;
2830 }
2831
2832 void
2833 kiblnd_destroy_dev (kib_dev_t *dev)
2834 {
2835         LASSERT (dev->ibd_nnets == 0);
2836         LASSERT(list_empty(&dev->ibd_nets));
2837
2838         list_del(&dev->ibd_fail_list);
2839         list_del(&dev->ibd_list);
2840
2841         if (dev->ibd_hdev != NULL)
2842                 kiblnd_hdev_decref(dev->ibd_hdev);
2843
2844         LIBCFS_FREE(dev, sizeof(*dev));
2845 }
2846
2847 static kib_dev_t *
2848 kiblnd_create_dev(char *ifname)
2849 {
2850         struct net_device *netdev;
2851         kib_dev_t         *dev;
2852         __u32              netmask;
2853         __u32              ip;
2854         int                up;
2855         int                rc;
2856
2857         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2858         if (rc != 0) {
2859                 CERROR("Can't query IPoIB interface %s: %d\n",
2860                        ifname, rc);
2861                 return NULL;
2862         }
2863
2864         if (!up) {
2865                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2866                 return NULL;
2867         }
2868
2869         LIBCFS_ALLOC(dev, sizeof(*dev));
2870         if (dev == NULL)
2871                 return NULL;
2872
2873         netdev = dev_get_by_name(&init_net, ifname);
2874         if (netdev == NULL) {
2875                 dev->ibd_can_failover = 0;
2876         } else {
2877                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2878                 dev_put(netdev);
2879         }
2880
2881         INIT_LIST_HEAD(&dev->ibd_nets);
2882         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2883         INIT_LIST_HEAD(&dev->ibd_fail_list);
2884         dev->ibd_ifip = ip;
2885         strcpy(&dev->ibd_ifname[0], ifname);
2886
2887         /* initialize the device */
2888         rc = kiblnd_dev_failover(dev);
2889         if (rc != 0) {
2890                 CERROR("Can't initialize device: %d\n", rc);
2891                 LIBCFS_FREE(dev, sizeof(*dev));
2892                 return NULL;
2893         }
2894
2895         list_add_tail(&dev->ibd_list,
2896                           &kiblnd_data.kib_devs);
2897         return dev;
2898 }
2899
2900 static void
2901 kiblnd_base_shutdown(void)
2902 {
2903         struct kib_sched_info   *sched;
2904         int                     i;
2905
2906         LASSERT(list_empty(&kiblnd_data.kib_devs));
2907
2908         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2909                atomic_read(&libcfs_kmemory));
2910
2911         switch (kiblnd_data.kib_init) {
2912         default:
2913                 LBUG();
2914
2915         case IBLND_INIT_ALL:
2916         case IBLND_INIT_DATA:
2917                 LASSERT (kiblnd_data.kib_peers != NULL);
2918                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2919                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2920                 }
2921                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2922                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2923                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2924                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2925
2926                 /* flag threads to terminate; wake and wait for them to die */
2927                 kiblnd_data.kib_shutdown = 1;
2928
2929                 /* NB: we really want to stop scheduler threads net by net
2930                  * instead of the whole module, this should be improved
2931                  * with dynamic configuration LNet */
2932                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2933                         wake_up_all(&sched->ibs_waitq);
2934
2935                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2936                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2937
2938                 i = 2;
2939                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2940                         i++;
2941                         /* power of 2? */
2942                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2943                                "Waiting for %d threads to terminate\n",
2944                                atomic_read(&kiblnd_data.kib_nthreads));
2945                         set_current_state(TASK_UNINTERRUPTIBLE);
2946                         schedule_timeout(cfs_time_seconds(1));
2947                 }
2948
2949                 /* fall through */
2950
2951         case IBLND_INIT_NOTHING:
2952                 break;
2953         }
2954
2955         if (kiblnd_data.kib_peers != NULL) {
2956                 LIBCFS_FREE(kiblnd_data.kib_peers,
2957                             sizeof(struct list_head) *
2958                             kiblnd_data.kib_peer_hash_size);
2959         }
2960
2961         if (kiblnd_data.kib_scheds != NULL)
2962                 cfs_percpt_free(kiblnd_data.kib_scheds);
2963
2964         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2965                atomic_read(&libcfs_kmemory));
2966
2967         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2968         module_put(THIS_MODULE);
2969 }
2970
2971 static void
2972 kiblnd_shutdown(struct lnet_ni *ni)
2973 {
2974         kib_net_t        *net = ni->ni_data;
2975         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2976         int               i;
2977         unsigned long     flags;
2978
2979         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2980
2981         if (net == NULL)
2982                 goto out;
2983
2984         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2985                atomic_read(&libcfs_kmemory));
2986
2987         write_lock_irqsave(g_lock, flags);
2988         net->ibn_shutdown = 1;
2989         write_unlock_irqrestore(g_lock, flags);
2990
2991         switch (net->ibn_init) {
2992         default:
2993                 LBUG();
2994
2995         case IBLND_INIT_ALL:
2996                 /* nuke all existing peers within this net */
2997                 kiblnd_del_peer(ni, LNET_NID_ANY);
2998
2999                 /* Wait for all peer_ni state to clean up */
3000                 i = 2;
3001                 while (atomic_read(&net->ibn_npeers) != 0) {
3002                         i++;
3003                         /* power of 2? */
3004                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3005                                "%s: waiting for %d peers to disconnect\n",
3006                                libcfs_nid2str(ni->ni_nid),
3007                                atomic_read(&net->ibn_npeers));
3008                         set_current_state(TASK_UNINTERRUPTIBLE);
3009                         schedule_timeout(cfs_time_seconds(1));
3010                 }
3011
3012                 kiblnd_net_fini_pools(net);
3013
3014                 write_lock_irqsave(g_lock, flags);
3015                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3016                 net->ibn_dev->ibd_nnets--;
3017                 list_del(&net->ibn_list);
3018                 write_unlock_irqrestore(g_lock, flags);
3019
3020                 /* fall through */
3021
3022         case IBLND_INIT_NOTHING:
3023                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3024
3025                 if (net->ibn_dev != NULL &&
3026                     net->ibn_dev->ibd_nnets == 0)
3027                         kiblnd_destroy_dev(net->ibn_dev);
3028
3029                 break;
3030         }
3031
3032         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3033                atomic_read(&libcfs_kmemory));
3034
3035         net->ibn_init = IBLND_INIT_NOTHING;
3036         ni->ni_data = NULL;
3037
3038         LIBCFS_FREE(net, sizeof(*net));
3039
3040 out:
3041         if (list_empty(&kiblnd_data.kib_devs))
3042                 kiblnd_base_shutdown();
3043         return;
3044 }
3045
3046 static int
3047 kiblnd_base_startup(void)
3048 {
3049         struct kib_sched_info   *sched;
3050         int                     rc;
3051         int                     i;
3052
3053         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3054
3055         try_module_get(THIS_MODULE);
3056         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3057
3058         rwlock_init(&kiblnd_data.kib_global_lock);
3059
3060         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3061         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3062
3063         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3064         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3065                      sizeof(struct list_head) *
3066                      kiblnd_data.kib_peer_hash_size);
3067         if (kiblnd_data.kib_peers == NULL)
3068                 goto failed;
3069
3070         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3071                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3072
3073         spin_lock_init(&kiblnd_data.kib_connd_lock);
3074         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3075         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3076         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3077         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3078
3079         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3080         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3081
3082         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3083                                                   sizeof(*sched));
3084         if (kiblnd_data.kib_scheds == NULL)
3085                 goto failed;
3086
3087         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3088                 int     nthrs;
3089
3090                 spin_lock_init(&sched->ibs_lock);
3091                 INIT_LIST_HEAD(&sched->ibs_conns);
3092                 init_waitqueue_head(&sched->ibs_waitq);
3093
3094                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3095                 if (*kiblnd_tunables.kib_nscheds > 0) {
3096                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3097                 } else {
3098                         /* max to half of CPUs, another half is reserved for
3099                          * upper layer modules */
3100                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3101                 }
3102
3103                 sched->ibs_nthreads_max = nthrs;
3104                 sched->ibs_cpt = i;
3105         }
3106
3107         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3108
3109         /* lists/ptrs/locks initialised */
3110         kiblnd_data.kib_init = IBLND_INIT_DATA;
3111         /*****************************************************/
3112
3113         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3114         if (rc != 0) {
3115                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3116                 goto failed;
3117         }
3118
3119         if (*kiblnd_tunables.kib_dev_failover != 0)
3120                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3121                                          "kiblnd_failover");
3122
3123         if (rc != 0) {
3124                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3125                 goto failed;
3126         }
3127
3128         /* flag everything initialised */
3129         kiblnd_data.kib_init = IBLND_INIT_ALL;
3130         /*****************************************************/
3131
3132         return 0;
3133
3134  failed:
3135         kiblnd_base_shutdown();
3136         return -ENETDOWN;
3137 }
3138
3139 static int
3140 kiblnd_start_schedulers(struct kib_sched_info *sched)
3141 {
3142         int     rc = 0;
3143         int     nthrs;
3144         int     i;
3145
3146         if (sched->ibs_nthreads == 0) {
3147                 if (*kiblnd_tunables.kib_nscheds > 0) {
3148                         nthrs = sched->ibs_nthreads_max;
3149                 } else {
3150                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3151                                                sched->ibs_cpt);
3152                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3153                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3154                 }
3155         } else {
3156                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3157                 /* increase one thread if there is new interface */
3158                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3159         }
3160
3161         for (i = 0; i < nthrs; i++) {
3162                 long    id;
3163                 char    name[20];
3164                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3165                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3166                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3167                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3168                 if (rc == 0)
3169                         continue;
3170
3171                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3172                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3173                 break;
3174         }
3175
3176         sched->ibs_nthreads += i;
3177         return rc;
3178 }
3179
3180 static int
3181 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3182 {
3183         int     cpt;
3184         int     rc;
3185         int     i;
3186
3187         for (i = 0; i < ncpts; i++) {
3188                 struct kib_sched_info *sched;
3189
3190                 cpt = (cpts == NULL) ? i : cpts[i];
3191                 sched = kiblnd_data.kib_scheds[cpt];
3192
3193                 if (!newdev && sched->ibs_nthreads > 0)
3194                         continue;
3195
3196                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3197                 if (rc != 0) {
3198                         CERROR("Failed to start scheduler threads for %s\n",
3199                                dev->ibd_ifname);
3200                         return rc;
3201                 }
3202         }
3203         return 0;
3204 }
3205
3206 static kib_dev_t *
3207 kiblnd_dev_search(char *ifname)
3208 {
3209         kib_dev_t       *alias = NULL;
3210         kib_dev_t       *dev;
3211         char            *colon;
3212         char            *colon2;
3213
3214         colon = strchr(ifname, ':');
3215         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3216                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3217                         return dev;
3218
3219                 if (alias != NULL)
3220                         continue;
3221
3222                 colon2 = strchr(dev->ibd_ifname, ':');
3223                 if (colon != NULL)
3224                         *colon = 0;
3225                 if (colon2 != NULL)
3226                         *colon2 = 0;
3227
3228                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3229                         alias = dev;
3230
3231                 if (colon != NULL)
3232                         *colon = ':';
3233                 if (colon2 != NULL)
3234                         *colon2 = ':';
3235         }
3236         return alias;
3237 }
3238
3239 static int
3240 kiblnd_startup(struct lnet_ni *ni)
3241 {
3242         char                     *ifname;
3243         kib_dev_t                *ibdev = NULL;
3244         kib_net_t                *net;
3245         unsigned long             flags;
3246         int                       rc;
3247         int                       newdev;
3248         int                       node_id;
3249
3250         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3251
3252         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3253                 rc = kiblnd_base_startup();
3254                 if (rc != 0)
3255                         return rc;
3256         }
3257
3258         LIBCFS_ALLOC(net, sizeof(*net));
3259         ni->ni_data = net;
3260         if (net == NULL)
3261                 goto failed;
3262
3263         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3264
3265         kiblnd_tunables_setup(ni);
3266
3267         if (ni->ni_interfaces[0] != NULL) {
3268                 /* Use the IPoIB interface specified in 'networks=' */
3269
3270                 CLASSERT(LNET_INTERFACES_NUM > 1);
3271                 if (ni->ni_interfaces[1] != NULL) {
3272                         CERROR("Multiple interfaces not supported\n");
3273                         goto failed;
3274                 }
3275
3276                 ifname = ni->ni_interfaces[0];
3277         } else {
3278                 ifname = *kiblnd_tunables.kib_default_ipif;
3279         }
3280
3281         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3282                 CERROR("IPoIB interface name too long: %s\n", ifname);
3283                 goto failed;
3284         }
3285
3286         ibdev = kiblnd_dev_search(ifname);
3287
3288         newdev = ibdev == NULL;
3289         /* hmm...create kib_dev even for alias */
3290         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3291                 ibdev = kiblnd_create_dev(ifname);
3292
3293         if (ibdev == NULL)
3294                 goto failed;
3295
3296         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3297         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3298
3299         net->ibn_dev = ibdev;
3300         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3301
3302         rc = kiblnd_dev_start_threads(ibdev, newdev,
3303                                       ni->ni_cpts, ni->ni_ncpts);
3304         if (rc != 0)
3305                 goto failed;
3306
3307         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3308         if (rc != 0) {
3309                 CERROR("Failed to initialize NI pools: %d\n", rc);
3310                 goto failed;
3311         }
3312
3313         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3314         ibdev->ibd_nnets++;
3315         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3316         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3317
3318         net->ibn_init = IBLND_INIT_ALL;
3319
3320         return 0;
3321
3322 failed:
3323         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3324                 kiblnd_destroy_dev(ibdev);
3325
3326         kiblnd_shutdown(ni);
3327
3328         CDEBUG(D_NET, "kiblnd_startup failed\n");
3329         return -ENETDOWN;
3330 }
3331
3332 static struct lnet_lnd the_o2iblnd = {
3333         .lnd_type       = O2IBLND,
3334         .lnd_startup    = kiblnd_startup,
3335         .lnd_shutdown   = kiblnd_shutdown,
3336         .lnd_ctl        = kiblnd_ctl,
3337         .lnd_query      = kiblnd_query,
3338         .lnd_send       = kiblnd_send,
3339         .lnd_recv       = kiblnd_recv,
3340 };
3341
3342 static void __exit ko2iblnd_exit(void)
3343 {
3344         lnet_unregister_lnd(&the_o2iblnd);
3345 }
3346
3347 static int __init ko2iblnd_init(void)
3348 {
3349         int rc;
3350
3351         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3352         CLASSERT(offsetof(kib_msg_t,
3353                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3354                  IBLND_MSG_SIZE);
3355         CLASSERT(offsetof(kib_msg_t,
3356                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3357                  <= IBLND_MSG_SIZE);
3358
3359         rc = kiblnd_tunables_init();
3360         if (rc != 0)
3361                 return rc;
3362
3363         lnet_register_lnd(&the_o2iblnd);
3364
3365         return 0;
3366 }
3367
3368 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3369 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3370 MODULE_VERSION("2.8.0");
3371 MODULE_LICENSE("GPL");
3372
3373 module_init(ko2iblnd_init);
3374 module_exit(ko2iblnd_exit);