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 = IBLND_MAX_RDMA_FRAGS;
 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 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 732 {
 733         /*
 734          * One WR for the LNet message
 735          * And ibc_max_frags for the transfer WRs
 736          */
 737         unsigned int ret = 1 + conn->ibc_max_frags;
 738
 739         /* account for a maximum of ibc_queue_depth in-flight transfers */
 740         ret *= conn->ibc_queue_depth;
 741         return ret;
 742 }
 743
 744 kib_conn_t *
 745 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 746                    int state, int version)
 747 {
 748         /* CAVEAT EMPTOR:
 749          * If the new conn is created successfully it takes over the caller's
 750          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 751          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 752          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 753          * to destroy 'cmid' here since I'm called from the CM which still has
 754          * its ref on 'cmid'). */
 755         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 756         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 757         kib_dev_t              *dev;
 758         struct ib_qp_init_attr *init_qp_attr;
 759         struct kib_sched_info   *sched;
 760 #ifdef HAVE_IB_CQ_INIT_ATTR
 761         struct ib_cq_init_attr  cq_attr = {};
 762 #endif
 763         kib_conn_t              *conn;
 764         struct ib_cq            *cq;
 765         unsigned long           flags;
 766         int                     cpt;
 767         int                     rc;
 768         int                     i;
 769
 770         LASSERT(net != NULL);
 771         LASSERT(!in_interrupt());
 772
 773         dev = net->ibn_dev;
 774
 775         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 776         sched = kiblnd_get_scheduler(cpt);
 777
 778         if (sched == NULL) {
 779                 CERROR("no schedulers available. node is unhealthy\n");
 780                 goto failed_0;
 781         }
 782
 783         /*
 784          * The cpt might have changed if we ended up selecting a non cpt
 785          * native scheduler. So use the scheduler's cpt instead.
 786          */
 787         cpt = sched->ibs_cpt;
 788
 789         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 790                          sizeof(*init_qp_attr));
 791         if (init_qp_attr == NULL) {
 792                 CERROR("Can't allocate qp_attr for %s\n",
 793                        libcfs_nid2str(peer_ni->ibp_nid));
 794                 goto failed_0;
 795         }
 796
 797         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 798         if (conn == NULL) {
 799                 CERROR("Can't allocate connection for %s\n",
 800                        libcfs_nid2str(peer_ni->ibp_nid));
 801                 goto failed_1;
 802         }
 803
 804         conn->ibc_state = IBLND_CONN_INIT;
 805         conn->ibc_version = version;
 806         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 807         cmid->context = conn;                   /* for future CM callbacks */
 808         conn->ibc_cmid = cmid;
 809         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 810         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 811
 812         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 813         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 814         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 815         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 816         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 817         INIT_LIST_HEAD(&conn->ibc_active_txs);
 818         spin_lock_init(&conn->ibc_lock);
 819
 820         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 821                          sizeof(*conn->ibc_connvars));
 822         if (conn->ibc_connvars == NULL) {
 823                 CERROR("Can't allocate in-progress connection state\n");
 824                 goto failed_2;
 825         }
 826
 827         write_lock_irqsave(glock, flags);
 828         if (dev->ibd_failover) {
 829                 write_unlock_irqrestore(glock, flags);
 830                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 831                 goto failed_2;
 832         }
 833
 834         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 835                 /* wakeup failover thread and teardown connection */
 836                 if (kiblnd_dev_can_failover(dev)) {
 837                         list_add_tail(&dev->ibd_fail_list,
 838                                       &kiblnd_data.kib_failed_devs);
 839                         wake_up(&kiblnd_data.kib_failover_waitq);
 840                 }
 841
 842                 write_unlock_irqrestore(glock, flags);
 843                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 844                        cmid->device->name, dev->ibd_ifname);
 845                 goto failed_2;
 846         }
 847
 848         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 849         conn->ibc_hdev = dev->ibd_hdev;
 850
 851         kiblnd_setup_mtu_locked(cmid);
 852
 853         write_unlock_irqrestore(glock, flags);
 854
 855         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 856                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 857         if (conn->ibc_rxs == NULL) {
 858                 CERROR("Cannot allocate RX buffers\n");
 859                 goto failed_2;
 860         }
 861
 862         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 863                                 IBLND_RX_MSG_PAGES(conn));
 864         if (rc != 0)
 865                 goto failed_2;
 866
 867         kiblnd_map_rx_descs(conn);
 868
 869 #ifdef HAVE_IB_CQ_INIT_ATTR
 870         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 871         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 872         cq = ib_create_cq(cmid->device,
 873                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 874                           &cq_attr);
 875 #else
 876         cq = ib_create_cq(cmid->device,
 877                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 878                           IBLND_CQ_ENTRIES(conn),
 879                           kiblnd_get_completion_vector(conn, cpt));
 880 #endif
 881         if (IS_ERR(cq)) {
 882                 /*
 883                  * on MLX-5 (possibly MLX-4 as well) this error could be
 884                  * hit if the concurrent_sends and/or peer_tx_credits is set
 885                  * too high. Or due to an MLX-5 bug which tries to
 886                  * allocate 256kb via kmalloc for WR cookie array
 887                  */
 888                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 889                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 890                 goto failed_2;
 891         }
 892
 893         conn->ibc_cq = cq;
 894
 895         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 896         if (rc != 0) {
 897                 CERROR("Can't request completion notification: %d\n", rc);
 898                 goto failed_2;
 899         }
 900
 901         init_qp_attr->event_handler = kiblnd_qp_event;
 902         init_qp_attr->qp_context = conn;
 903         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 904         init_qp_attr->cap.max_recv_sge = 1;
 905         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 906         init_qp_attr->qp_type = IB_QPT_RC;
 907         init_qp_attr->send_cq = cq;
 908         init_qp_attr->recv_cq = cq;
 909
 910         conn->ibc_sched = sched;
 911
 912         do {
 913                 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 914                 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 915
 916                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 917                 if (!rc || conn->ibc_queue_depth < 2)
 918                         break;
 919
 920                 conn->ibc_queue_depth--;
 921         } while (rc);
 922
 923         if (rc) {
 924                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 925                        "send_sge: %d, recv_sge: %d\n",
 926                        rc, init_qp_attr->cap.max_send_wr,
 927                        init_qp_attr->cap.max_recv_wr,
 928                        init_qp_attr->cap.max_send_sge,
 929                        init_qp_attr->cap.max_recv_sge);
 930                 goto failed_2;
 931         }
 932
 933         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 934                 CWARN("peer %s - queue depth reduced from %u to %u"
 935                       "  to allow for qp creation\n",
 936                       libcfs_nid2str(peer_ni->ibp_nid),
 937                       peer_ni->ibp_queue_depth,
 938                       conn->ibc_queue_depth);
 939
 940         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 941
 942         /* 1 ref for caller and each rxmsg */
 943         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 944         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 945
 946         /* post receives */
 947         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 948                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 949                 if (rc != 0) {
 950                         CERROR("Can't post rxmsg: %d\n", rc);
 951
 952                         /* Make posted receives complete */
 953                         kiblnd_abort_receives(conn);
 954
 955                         /* correct # of posted buffers
 956                          * NB locking needed now I'm racing with completion */
 957                         spin_lock_irqsave(&sched->ibs_lock, flags);
 958                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 959                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 960
 961                         /* cmid will be destroyed by CM(ofed) after cm_callback
 962                          * returned, so we can't refer it anymore
 963                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 964                         rdma_destroy_qp(conn->ibc_cmid);
 965                         conn->ibc_cmid = NULL;
 966
 967                         /* Drop my own and unused rxbuffer refcounts */
 968                         while (i++ <= IBLND_RX_MSGS(conn))
 969                                 kiblnd_conn_decref(conn);
 970
 971                         return NULL;
 972                 }
 973         }
 974
 975         /* Init successful! */
 976         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 977                  state == IBLND_CONN_PASSIVE_WAIT);
 978         conn->ibc_state = state;
 979
 980         /* 1 more conn */
 981         atomic_inc(&net->ibn_nconns);
 982         return conn;
 983
 984  failed_2:
 985         kiblnd_destroy_conn(conn, true);
 986  failed_1:
 987         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 988  failed_0:
 989         return NULL;
 990 }
 991
 992 void
 993 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 994 {
 995         struct rdma_cm_id *cmid = conn->ibc_cmid;
 996         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
 997         int                rc;
 998
 999         LASSERT (!in_interrupt());
1000         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1001         LASSERT(list_empty(&conn->ibc_early_rxs));
1002         LASSERT(list_empty(&conn->ibc_tx_noops));
1003         LASSERT(list_empty(&conn->ibc_tx_queue));
1004         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1005         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1006         LASSERT(list_empty(&conn->ibc_active_txs));
1007         LASSERT (conn->ibc_noops_posted == 0);
1008         LASSERT (conn->ibc_nsends_posted == 0);
1009
1010         switch (conn->ibc_state) {
1011         default:
1012                 /* conn must be completely disengaged from the network */
1013                 LBUG();
1014
1015         case IBLND_CONN_DISCONNECTED:
1016                 /* connvars should have been freed already */
1017                 LASSERT (conn->ibc_connvars == NULL);
1018                 break;
1019
1020         case IBLND_CONN_INIT:
1021                 break;
1022         }
1023
1024         /* conn->ibc_cmid might be destroyed by CM already */
1025         if (cmid != NULL && cmid->qp != NULL)
1026                 rdma_destroy_qp(cmid);
1027
1028         if (conn->ibc_cq != NULL) {
1029                 rc = ib_destroy_cq(conn->ibc_cq);
1030                 if (rc != 0)
1031                         CWARN("Error destroying CQ: %d\n", rc);
1032         }
1033
1034         if (conn->ibc_rx_pages != NULL)
1035                 kiblnd_unmap_rx_descs(conn);
1036
1037         if (conn->ibc_rxs != NULL) {
1038                 LIBCFS_FREE(conn->ibc_rxs,
1039                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1040         }
1041
1042         if (conn->ibc_connvars != NULL)
1043                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1044
1045         if (conn->ibc_hdev != NULL)
1046                 kiblnd_hdev_decref(conn->ibc_hdev);
1047
1048         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1049         if (conn->ibc_state != IBLND_CONN_INIT) {
1050                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1051
1052                 kiblnd_peer_decref(peer_ni);
1053                 rdma_destroy_id(cmid);
1054                 atomic_dec(&net->ibn_nconns);
1055         }
1056
1057         if (free_conn)
1058                 LIBCFS_FREE(conn, sizeof(*conn));
1059 }
1060
1061 int
1062 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1063 {
1064         kib_conn_t              *conn;
1065         struct list_head        *ctmp;
1066         struct list_head        *cnxt;
1067         int                     count = 0;
1068
1069         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1070                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1071
1072                 CDEBUG(D_NET, "Closing conn -> %s, "
1073                               "version: %x, reason: %d\n",
1074                        libcfs_nid2str(peer_ni->ibp_nid),
1075                        conn->ibc_version, why);
1076
1077                 kiblnd_close_conn_locked(conn, why);
1078                 count++;
1079         }
1080
1081         return count;
1082 }
1083
1084 int
1085 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1086                                 int version, __u64 incarnation)
1087 {
1088         kib_conn_t              *conn;
1089         struct list_head        *ctmp;
1090         struct list_head        *cnxt;
1091         int                     count = 0;
1092
1093         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1094                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1095
1096                 if (conn->ibc_version     == version &&
1097                     conn->ibc_incarnation == incarnation)
1098                         continue;
1099
1100                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1101                               "incarnation:%#llx(%x, %#llx)\n",
1102                        libcfs_nid2str(peer_ni->ibp_nid),
1103                        conn->ibc_version, conn->ibc_incarnation,
1104                        version, incarnation);
1105
1106                 kiblnd_close_conn_locked(conn, -ESTALE);
1107                 count++;
1108         }
1109
1110         return count;
1111 }
1112
1113 static int
1114 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1115 {
1116         kib_peer_ni_t           *peer_ni;
1117         struct list_head        *ptmp;
1118         struct list_head        *pnxt;
1119         int                     lo;
1120         int                     hi;
1121         int                     i;
1122         unsigned long           flags;
1123         int                     count = 0;
1124
1125         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1126
1127         if (nid != LNET_NID_ANY)
1128                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1129         else {
1130                 lo = 0;
1131                 hi = kiblnd_data.kib_peer_hash_size - 1;
1132         }
1133
1134         for (i = lo; i <= hi; i++) {
1135                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1136
1137                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1138                         LASSERT(!kiblnd_peer_idle(peer_ni));
1139
1140                         if (peer_ni->ibp_ni != ni)
1141                                 continue;
1142
1143                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1144                                 continue;
1145
1146                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1147                 }
1148         }
1149
1150         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1151
1152         /* wildcards always succeed */
1153         if (nid == LNET_NID_ANY)
1154                 return 0;
1155
1156         return (count == 0) ? -ENOENT : 0;
1157 }
1158
1159 static int
1160 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1161 {
1162         struct libcfs_ioctl_data *data = arg;
1163         int                       rc = -EINVAL;
1164
1165         switch(cmd) {
1166         case IOC_LIBCFS_GET_PEER: {
1167                 lnet_nid_t   nid = 0;
1168                 int          count = 0;
1169
1170                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1171                                           &nid, &count);
1172                 data->ioc_nid    = nid;
1173                 data->ioc_count  = count;
1174                 break;
1175         }
1176
1177         case IOC_LIBCFS_DEL_PEER: {
1178                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1179                 break;
1180         }
1181         case IOC_LIBCFS_GET_CONN: {
1182                 kib_conn_t *conn;
1183
1184                 rc = 0;
1185                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1186                 if (conn == NULL) {
1187                         rc = -ENOENT;
1188                         break;
1189                 }
1190
1191                 LASSERT(conn->ibc_cmid != NULL);
1192                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1193                 if (conn->ibc_cmid->route.path_rec == NULL)
1194                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1195                 else
1196                         data->ioc_u32[0] =
1197                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1198                 kiblnd_conn_decref(conn);
1199                 break;
1200         }
1201         case IOC_LIBCFS_CLOSE_CONNECTION: {
1202                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1203                 break;
1204         }
1205
1206         default:
1207                 break;
1208         }
1209
1210         return rc;
1211 }
1212
1213 static void
1214 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1215 {
1216         cfs_time_t      last_alive = 0;
1217         cfs_time_t      now = cfs_time_current();
1218         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1219         kib_peer_ni_t   *peer_ni;
1220         unsigned long   flags;
1221
1222         read_lock_irqsave(glock, flags);
1223
1224         peer_ni = kiblnd_find_peer_locked(ni, nid);
1225         if (peer_ni != NULL)
1226                 last_alive = peer_ni->ibp_last_alive;
1227
1228         read_unlock_irqrestore(glock, flags);
1229
1230         if (last_alive != 0)
1231                 *when = last_alive;
1232
1233         /* peer_ni is not persistent in hash, trigger peer_ni creation
1234          * and connection establishment with a NULL tx */
1235         if (peer_ni == NULL)
1236                 kiblnd_launch_tx(ni, NULL, nid);
1237
1238         CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1239                libcfs_nid2str(nid), peer_ni,
1240                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1241         return;
1242 }
1243
1244 static void
1245 kiblnd_free_pages(kib_pages_t *p)
1246 {
1247         int     npages = p->ibp_npages;
1248         int     i;
1249
1250         for (i = 0; i < npages; i++) {
1251                 if (p->ibp_pages[i] != NULL)
1252                         __free_page(p->ibp_pages[i]);
1253         }
1254
1255         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1256 }
1257
1258 int
1259 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1260 {
1261         kib_pages_t     *p;
1262         int             i;
1263
1264         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1265                          offsetof(kib_pages_t, ibp_pages[npages]));
1266         if (p == NULL) {
1267                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1268                 return -ENOMEM;
1269         }
1270
1271         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1272         p->ibp_npages = npages;
1273
1274         for (i = 0; i < npages; i++) {
1275                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1276                                                      GFP_NOFS);
1277                 if (p->ibp_pages[i] == NULL) {
1278                         CERROR("Can't allocate page %d of %d\n", i, npages);
1279                         kiblnd_free_pages(p);
1280                         return -ENOMEM;
1281                 }
1282         }
1283
1284         *pp = p;
1285         return 0;
1286 }
1287
1288 void
1289 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1290 {
1291         kib_rx_t *rx;
1292         int       i;
1293
1294         LASSERT (conn->ibc_rxs != NULL);
1295         LASSERT (conn->ibc_hdev != NULL);
1296
1297         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1298                 rx = &conn->ibc_rxs[i];
1299
1300                 LASSERT(rx->rx_nob >= 0); /* not posted */
1301
1302                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1303                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1304                                                           rx->rx_msgaddr),
1305                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1306         }
1307
1308         kiblnd_free_pages(conn->ibc_rx_pages);
1309
1310         conn->ibc_rx_pages = NULL;
1311 }
1312
1313 void
1314 kiblnd_map_rx_descs(kib_conn_t *conn)
1315 {
1316         kib_rx_t       *rx;
1317         struct page    *pg;
1318         int             pg_off;
1319         int             ipg;
1320         int             i;
1321
1322         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1323                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1324                 rx = &conn->ibc_rxs[i];
1325
1326                 rx->rx_conn = conn;
1327                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1328
1329                 rx->rx_msgaddr =
1330                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1331                                               rx->rx_msg, IBLND_MSG_SIZE,
1332                                               DMA_FROM_DEVICE);
1333                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1334                                                   rx->rx_msgaddr));
1335                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1336
1337                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1338                        i, rx->rx_msg, rx->rx_msgaddr,
1339                        (__u64)(page_to_phys(pg) + pg_off));
1340
1341                 pg_off += IBLND_MSG_SIZE;
1342                 LASSERT(pg_off <= PAGE_SIZE);
1343
1344                 if (pg_off == PAGE_SIZE) {
1345                         pg_off = 0;
1346                         ipg++;
1347                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1348                 }
1349         }
1350 }
1351
1352 static void
1353 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1354 {
1355         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1356         kib_tx_t       *tx;
1357         int             i;
1358
1359         LASSERT (tpo->tpo_pool.po_allocated == 0);
1360
1361         if (hdev == NULL)
1362                 return;
1363
1364         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1365                 tx = &tpo->tpo_tx_descs[i];
1366                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1367                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1368                                                           tx->tx_msgaddr),
1369                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1370         }
1371
1372         kiblnd_hdev_decref(hdev);
1373         tpo->tpo_hdev = NULL;
1374 }
1375
1376 static kib_hca_dev_t *
1377 kiblnd_current_hdev(kib_dev_t *dev)
1378 {
1379         kib_hca_dev_t *hdev;
1380         unsigned long  flags;
1381         int            i = 0;
1382
1383         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1384         while (dev->ibd_failover) {
1385                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1386                 if (i++ % 50 == 0)
1387                         CDEBUG(D_NET, "%s: Wait for failover\n",
1388                                dev->ibd_ifname);
1389                 set_current_state(TASK_INTERRUPTIBLE);
1390                 schedule_timeout(cfs_time_seconds(1) / 100);
1391
1392                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1393         }
1394
1395         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1396         hdev = dev->ibd_hdev;
1397
1398         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1399
1400         return hdev;
1401 }
1402
1403 static void
1404 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1405 {
1406         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1407         kib_pool_t     *pool  = &tpo->tpo_pool;
1408         kib_net_t      *net   = pool->po_owner->ps_net;
1409         kib_dev_t      *dev;
1410         struct page    *page;
1411         kib_tx_t       *tx;
1412         int             page_offset;
1413         int             ipage;
1414         int             i;
1415
1416         LASSERT (net != NULL);
1417
1418         dev = net->ibn_dev;
1419
1420         /* pre-mapped messages are not bigger than 1 page */
1421         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1422
1423         /* No fancy arithmetic when we do the buffer calculations */
1424         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1425
1426         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1427
1428         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1429                 page = txpgs->ibp_pages[ipage];
1430                 tx = &tpo->tpo_tx_descs[i];
1431
1432                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1433                                            page_offset);
1434
1435                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1436                                                        tx->tx_msg,
1437                                                        IBLND_MSG_SIZE,
1438                                                        DMA_TO_DEVICE);
1439                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1440                                                   tx->tx_msgaddr));
1441                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1442
1443                 list_add(&tx->tx_list, &pool->po_free_list);
1444
1445                 page_offset += IBLND_MSG_SIZE;
1446                 LASSERT(page_offset <= PAGE_SIZE);
1447
1448                 if (page_offset == PAGE_SIZE) {
1449                         page_offset = 0;
1450                         ipage++;
1451                         LASSERT(ipage <= txpgs->ibp_npages);
1452                 }
1453         }
1454 }
1455
1456 static void
1457 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1458 {
1459         LASSERT(fpo->fpo_map_count == 0);
1460
1461         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1462                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1463         } else {
1464                 struct kib_fast_reg_descriptor *frd, *tmp;
1465                 int i = 0;
1466
1467                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1468                                          frd_list) {
1469                         list_del(&frd->frd_list);
1470 #ifndef HAVE_IB_MAP_MR_SG
1471                         ib_free_fast_reg_page_list(frd->frd_frpl);
1472 #endif
1473                         ib_dereg_mr(frd->frd_mr);
1474                         LIBCFS_FREE(frd, sizeof(*frd));
1475                         i++;
1476                 }
1477                 if (i < fpo->fast_reg.fpo_pool_size)
1478                         CERROR("FastReg pool still has %d regions registered\n",
1479                                 fpo->fast_reg.fpo_pool_size - i);
1480         }
1481
1482         if (fpo->fpo_hdev)
1483                 kiblnd_hdev_decref(fpo->fpo_hdev);
1484
1485         LIBCFS_FREE(fpo, sizeof(*fpo));
1486 }
1487
1488 static void
1489 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1490 {
1491         kib_fmr_pool_t *fpo, *tmp;
1492
1493         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1494                 list_del(&fpo->fpo_list);
1495                 kiblnd_destroy_fmr_pool(fpo);
1496         }
1497 }
1498
1499 static int
1500 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1501                      int ncpts)
1502 {
1503         int size = tunables->lnd_fmr_pool_size / ncpts;
1504
1505         return max(IBLND_FMR_POOL, size);
1506 }
1507
1508 static int
1509 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1510                          int ncpts)
1511 {
1512         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1513
1514         return max(IBLND_FMR_POOL_FLUSH, size);
1515 }
1516
1517 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1518 {
1519         struct ib_fmr_pool_param param = {
1520                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1521                 .page_shift        = PAGE_SHIFT,
1522                 .access            = (IB_ACCESS_LOCAL_WRITE |
1523                                       IB_ACCESS_REMOTE_WRITE),
1524                 .pool_size         = fps->fps_pool_size,
1525                 .dirty_watermark   = fps->fps_flush_trigger,
1526                 .flush_function    = NULL,
1527                 .flush_arg         = NULL,
1528                 .cache             = !!fps->fps_cache };
1529         int rc = 0;
1530
1531         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1532                                                    &param);
1533         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1534                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1535                 if (rc != -ENOSYS)
1536                         CERROR("Failed to create FMR pool: %d\n", rc);
1537                 else
1538                         CERROR("FMRs are not supported\n");
1539         }
1540         fpo->fpo_is_fmr = true;
1541
1542         return rc;
1543 }
1544
1545 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1546                                   __u32 dev_caps)
1547 {
1548         struct kib_fast_reg_descriptor *frd, *tmp;
1549         int i, rc;
1550
1551         fpo->fpo_is_fmr = false;
1552
1553         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1554         fpo->fast_reg.fpo_pool_size = 0;
1555         for (i = 0; i < fps->fps_pool_size; i++) {
1556                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1557                                  sizeof(*frd));
1558                 if (!frd) {
1559                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1560                         rc = -ENOMEM;
1561                         goto out;
1562                 }
1563                 frd->frd_mr = NULL;
1564
1565 #ifndef HAVE_IB_MAP_MR_SG
1566                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1567                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1568                 if (IS_ERR(frd->frd_frpl)) {
1569                         rc = PTR_ERR(frd->frd_frpl);
1570                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1571                                 rc);
1572                         frd->frd_frpl = NULL;
1573                         goto out_middle;
1574                 }
1575 #endif
1576
1577 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1578                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1579                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1580 #else
1581                 /*
1582                  * it is expected to get here if this is an MLX-5 card.
1583                  * MLX-4 cards will always use FMR and MLX-5 cards will
1584                  * always use fast_reg. It turns out that some MLX-5 cards
1585                  * (possibly due to older FW versions) do not natively support
1586                  * gaps. So we will need to track them here.
1587                  */
1588                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1589                                           (dev_caps &
1590                                                 IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) ?
1591                                                 IB_MR_TYPE_SG_GAPS :
1592                                                 IB_MR_TYPE_MEM_REG,
1593                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1594 #endif
1595                 if (IS_ERR(frd->frd_mr)) {
1596                         rc = PTR_ERR(frd->frd_mr);
1597                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1598                         frd->frd_mr = NULL;
1599                         goto out_middle;
1600                 }
1601
1602                 /* There appears to be a bug in MLX5 code where you must
1603                  * invalidate the rkey of a new FastReg pool before first
1604                  * using it. Thus, I am marking the FRD invalid here. */
1605                 frd->frd_valid = false;
1606
1607                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1608                 fpo->fast_reg.fpo_pool_size++;
1609         }
1610
1611         return 0;
1612
1613 out_middle:
1614         if (frd->frd_mr)
1615                 ib_dereg_mr(frd->frd_mr);
1616 #ifndef HAVE_IB_MAP_MR_SG
1617         if (frd->frd_frpl)
1618                 ib_free_fast_reg_page_list(frd->frd_frpl);
1619 #endif
1620         LIBCFS_FREE(frd, sizeof(*frd));
1621
1622 out:
1623         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1624                                  frd_list) {
1625                 list_del(&frd->frd_list);
1626 #ifndef HAVE_IB_MAP_MR_SG
1627                 ib_free_fast_reg_page_list(frd->frd_frpl);
1628 #endif
1629                 ib_dereg_mr(frd->frd_mr);
1630                 LIBCFS_FREE(frd, sizeof(*frd));
1631         }
1632
1633         return rc;
1634 }
1635
1636 static int
1637 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1638 {
1639         kib_dev_t *dev = fps->fps_net->ibn_dev;
1640         kib_fmr_pool_t *fpo;
1641         int rc;
1642
1643         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1644         if (!fpo) {
1645                 return -ENOMEM;
1646         }
1647         memset(fpo, 0, sizeof(*fpo));
1648
1649         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1650
1651         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1652                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1653         else
1654                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1655         if (rc)
1656                 goto out_fpo;
1657
1658         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1659         fpo->fpo_owner    = fps;
1660         *pp_fpo = fpo;
1661
1662         return 0;
1663
1664 out_fpo:
1665         kiblnd_hdev_decref(fpo->fpo_hdev);
1666         LIBCFS_FREE(fpo, sizeof(*fpo));
1667         return rc;
1668 }
1669
1670 static void
1671 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1672 {
1673         if (fps->fps_net == NULL) /* intialized? */
1674                 return;
1675
1676         spin_lock(&fps->fps_lock);
1677
1678         while (!list_empty(&fps->fps_pool_list)) {
1679                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1680                                                  kib_fmr_pool_t, fpo_list);
1681                 fpo->fpo_failed = 1;
1682                 list_del(&fpo->fpo_list);
1683                 if (fpo->fpo_map_count == 0)
1684                         list_add(&fpo->fpo_list, zombies);
1685                 else
1686                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1687         }
1688
1689         spin_unlock(&fps->fps_lock);
1690 }
1691
1692 static void
1693 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1694 {
1695         if (fps->fps_net != NULL) { /* initialized? */
1696                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1697                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1698         }
1699 }
1700
1701 static int
1702 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1703                         kib_net_t *net,
1704                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1705 {
1706         kib_fmr_pool_t *fpo;
1707         int             rc;
1708
1709         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1710
1711         fps->fps_net = net;
1712         fps->fps_cpt = cpt;
1713
1714         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1715         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1716         fps->fps_cache = tunables->lnd_fmr_cache;
1717
1718         spin_lock_init(&fps->fps_lock);
1719         INIT_LIST_HEAD(&fps->fps_pool_list);
1720         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1721
1722         rc = kiblnd_create_fmr_pool(fps, &fpo);
1723         if (rc == 0)
1724                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1725
1726         return rc;
1727 }
1728
1729 static int
1730 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1731 {
1732         if (fpo->fpo_map_count != 0) /* still in use */
1733                 return 0;
1734         if (fpo->fpo_failed)
1735                 return 1;
1736         return cfs_time_aftereq(now, fpo->fpo_deadline);
1737 }
1738
1739 static int
1740 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1741 {
1742         kib_hca_dev_t   *hdev;
1743         __u64           *pages = tx->tx_pages;
1744         int             npages;
1745         int             size;
1746         int             i;
1747
1748         hdev = tx->tx_pool->tpo_hdev;
1749
1750         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1751                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1752                         size += hdev->ibh_page_size) {
1753                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1754                                            hdev->ibh_page_mask) + size;
1755                 }
1756         }
1757
1758         return npages;
1759 }
1760
1761 void
1762 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1763 {
1764         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1765         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1766         kib_fmr_poolset_t *fps;
1767         cfs_time_t         now = cfs_time_current();
1768         kib_fmr_pool_t    *tmp;
1769         int                rc;
1770
1771         if (!fpo)
1772                 return;
1773
1774         fps = fpo->fpo_owner;
1775         if (fpo->fpo_is_fmr) {
1776                 if (fmr->fmr_pfmr) {
1777                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1778                         LASSERT(!rc);
1779                         fmr->fmr_pfmr = NULL;
1780                 }
1781
1782                 if (status) {
1783                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1784                         LASSERT(!rc);
1785                 }
1786         } else {
1787                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1788
1789                 if (frd) {
1790                         frd->frd_valid = false;
1791                         spin_lock(&fps->fps_lock);
1792                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1793                         spin_unlock(&fps->fps_lock);
1794                         fmr->fmr_frd = NULL;
1795                 }
1796         }
1797         fmr->fmr_pool = NULL;
1798
1799         spin_lock(&fps->fps_lock);
1800         fpo->fpo_map_count--;   /* decref the pool */
1801
1802         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1803                 /* the first pool is persistent */
1804                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1805                         continue;
1806
1807                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1808                         list_move(&fpo->fpo_list, &zombies);
1809                         fps->fps_version++;
1810                 }
1811         }
1812         spin_unlock(&fps->fps_lock);
1813
1814         if (!list_empty(&zombies))
1815                 kiblnd_destroy_fmr_pool_list(&zombies);
1816 }
1817
1818 int
1819 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1820                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1821 {
1822         kib_fmr_pool_t *fpo;
1823         __u64 *pages = tx->tx_pages;
1824         __u64 version;
1825         bool is_rx = (rd != tx->tx_rd);
1826         bool tx_pages_mapped = 0;
1827         int npages = 0;
1828         int rc;
1829
1830 again:
1831         spin_lock(&fps->fps_lock);
1832         version = fps->fps_version;
1833         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1834                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1835                 fpo->fpo_map_count++;
1836
1837                 if (fpo->fpo_is_fmr) {
1838                         struct ib_pool_fmr *pfmr;
1839
1840                         spin_unlock(&fps->fps_lock);
1841
1842                         if (!tx_pages_mapped) {
1843                                 npages = kiblnd_map_tx_pages(tx, rd);
1844                                 tx_pages_mapped = 1;
1845                         }
1846
1847                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1848                                                     pages, npages, iov);
1849                         if (likely(!IS_ERR(pfmr))) {
1850                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1851                                                       : pfmr->fmr->lkey;
1852                                 fmr->fmr_frd  = NULL;
1853                                 fmr->fmr_pfmr = pfmr;
1854                                 fmr->fmr_pool = fpo;
1855                                 return 0;
1856                         }
1857                         rc = PTR_ERR(pfmr);
1858                 } else {
1859                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1860                                 struct kib_fast_reg_descriptor *frd;
1861 #ifdef HAVE_IB_MAP_MR_SG
1862                                 struct ib_reg_wr *wr;
1863                                 int n;
1864 #else
1865                                 struct ib_rdma_wr *wr;
1866                                 struct ib_fast_reg_page_list *frpl;
1867 #endif
1868                                 struct ib_mr *mr;
1869
1870                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1871                                                         struct kib_fast_reg_descriptor,
1872                                                         frd_list);
1873                                 list_del(&frd->frd_list);
1874                                 spin_unlock(&fps->fps_lock);
1875
1876 #ifndef HAVE_IB_MAP_MR_SG
1877                                 frpl = frd->frd_frpl;
1878 #endif
1879                                 mr   = frd->frd_mr;
1880
1881                                 if (!frd->frd_valid) {
1882                                         struct ib_rdma_wr *inv_wr;
1883                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1884
1885                                         inv_wr = &frd->frd_inv_wr;
1886                                         memset(inv_wr, 0, sizeof(*inv_wr));
1887
1888                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1889                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1890                                         inv_wr->wr.ex.invalidate_rkey = key;
1891
1892                                         /* Bump the key */
1893                                         key = ib_inc_rkey(key);
1894                                         ib_update_fast_reg_key(mr, key);
1895                                 }
1896
1897 #ifdef HAVE_IB_MAP_MR_SG
1898 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1899                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1900                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1901 #else
1902                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1903                                                  tx->tx_nfrags, PAGE_SIZE);
1904 #endif
1905                                 if (unlikely(n != tx->tx_nfrags)) {
1906                                         CERROR("Failed to map mr %d/%d "
1907                                                "elements\n", n, tx->tx_nfrags);
1908                                         return n < 0 ? n : -EINVAL;
1909                                 }
1910
1911                                 wr = &frd->frd_fastreg_wr;
1912                                 memset(wr, 0, sizeof(*wr));
1913
1914                                 wr->wr.opcode = IB_WR_REG_MR;
1915                                 wr->wr.wr_id  = IBLND_WID_MR;
1916                                 wr->wr.num_sge = 0;
1917                                 wr->wr.send_flags = 0;
1918                                 wr->mr = mr;
1919                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1920                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1921                                               IB_ACCESS_REMOTE_WRITE);
1922 #else
1923                                 if (!tx_pages_mapped) {
1924                                         npages = kiblnd_map_tx_pages(tx, rd);
1925                                         tx_pages_mapped = 1;
1926                                 }
1927
1928                                 LASSERT(npages <= frpl->max_page_list_len);
1929                                 memcpy(frpl->page_list, pages,
1930                                         sizeof(*pages) * npages);
1931
1932                                 /* Prepare FastReg WR */
1933                                 wr = &frd->frd_fastreg_wr;
1934                                 memset(wr, 0, sizeof(*wr));
1935
1936                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1937                                 wr->wr.wr_id  = IBLND_WID_MR;
1938
1939                                 wr->wr.wr.fast_reg.iova_start = iov;
1940                                 wr->wr.wr.fast_reg.page_list  = frpl;
1941                                 wr->wr.wr.fast_reg.page_list_len = npages;
1942                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1943                                 wr->wr.wr.fast_reg.length = nob;
1944                                 wr->wr.wr.fast_reg.rkey =
1945                                                 is_rx ? mr->rkey : mr->lkey;
1946                                 wr->wr.wr.fast_reg.access_flags =
1947                                                 (IB_ACCESS_LOCAL_WRITE |
1948                                                  IB_ACCESS_REMOTE_WRITE);
1949 #endif
1950
1951                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1952                                 fmr->fmr_frd  = frd;
1953                                 fmr->fmr_pfmr = NULL;
1954                                 fmr->fmr_pool = fpo;
1955                                 return 0;
1956                         }
1957                         spin_unlock(&fps->fps_lock);
1958                         rc = -EAGAIN;
1959                 }
1960
1961                 spin_lock(&fps->fps_lock);
1962                 fpo->fpo_map_count--;
1963                 if (rc != -EAGAIN) {
1964                         spin_unlock(&fps->fps_lock);
1965                         return rc;
1966                 }
1967
1968                 /* EAGAIN and ... */
1969                 if (version != fps->fps_version) {
1970                         spin_unlock(&fps->fps_lock);
1971                         goto again;
1972                 }
1973         }
1974
1975         if (fps->fps_increasing) {
1976                 spin_unlock(&fps->fps_lock);
1977                 CDEBUG(D_NET, "Another thread is allocating new "
1978                        "FMR pool, waiting for her to complete\n");
1979                 schedule();
1980                 goto again;
1981
1982         }
1983
1984         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1985                 /* someone failed recently */
1986                 spin_unlock(&fps->fps_lock);
1987                 return -EAGAIN;
1988         }
1989
1990         fps->fps_increasing = 1;
1991         spin_unlock(&fps->fps_lock);
1992
1993         CDEBUG(D_NET, "Allocate new FMR pool\n");
1994         rc = kiblnd_create_fmr_pool(fps, &fpo);
1995         spin_lock(&fps->fps_lock);
1996         fps->fps_increasing = 0;
1997         if (rc == 0) {
1998                 fps->fps_version++;
1999                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2000         } else {
2001                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2002         }
2003         spin_unlock(&fps->fps_lock);
2004
2005         goto again;
2006 }
2007
2008 static void
2009 kiblnd_fini_pool(kib_pool_t *pool)
2010 {
2011         LASSERT(list_empty(&pool->po_free_list));
2012         LASSERT(pool->po_allocated == 0);
2013
2014         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2015 }
2016
2017 static void
2018 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2019 {
2020         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2021
2022         memset(pool, 0, sizeof(kib_pool_t));
2023         INIT_LIST_HEAD(&pool->po_free_list);
2024         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2025         pool->po_owner    = ps;
2026         pool->po_size     = size;
2027 }
2028
2029 static void
2030 kiblnd_destroy_pool_list(struct list_head *head)
2031 {
2032         kib_pool_t *pool;
2033
2034         while (!list_empty(head)) {
2035                 pool = list_entry(head->next, kib_pool_t, po_list);
2036                 list_del(&pool->po_list);
2037
2038                 LASSERT(pool->po_owner != NULL);
2039                 pool->po_owner->ps_pool_destroy(pool);
2040         }
2041 }
2042
2043 static void
2044 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2045 {
2046         if (ps->ps_net == NULL) /* intialized? */
2047                 return;
2048
2049         spin_lock(&ps->ps_lock);
2050         while (!list_empty(&ps->ps_pool_list)) {
2051                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2052                                             kib_pool_t, po_list);
2053                 po->po_failed = 1;
2054                 list_del(&po->po_list);
2055                 if (po->po_allocated == 0)
2056                         list_add(&po->po_list, zombies);
2057                 else
2058                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2059         }
2060         spin_unlock(&ps->ps_lock);
2061 }
2062
2063 static void
2064 kiblnd_fini_poolset(kib_poolset_t *ps)
2065 {
2066         if (ps->ps_net != NULL) { /* initialized? */
2067                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2068                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2069         }
2070 }
2071
2072 static int
2073 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2074                     kib_net_t *net, char *name, int size,
2075                     kib_ps_pool_create_t po_create,
2076                     kib_ps_pool_destroy_t po_destroy,
2077                     kib_ps_node_init_t nd_init,
2078                     kib_ps_node_fini_t nd_fini)
2079 {
2080         kib_pool_t      *pool;
2081         int             rc;
2082
2083         memset(ps, 0, sizeof(kib_poolset_t));
2084
2085         ps->ps_cpt          = cpt;
2086         ps->ps_net          = net;
2087         ps->ps_pool_create  = po_create;
2088         ps->ps_pool_destroy = po_destroy;
2089         ps->ps_node_init    = nd_init;
2090         ps->ps_node_fini    = nd_fini;
2091         ps->ps_pool_size    = size;
2092         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2093             >= sizeof(ps->ps_name))
2094                 return -E2BIG;
2095         spin_lock_init(&ps->ps_lock);
2096         INIT_LIST_HEAD(&ps->ps_pool_list);
2097         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2098
2099         rc = ps->ps_pool_create(ps, size, &pool);
2100         if (rc == 0)
2101                 list_add(&pool->po_list, &ps->ps_pool_list);
2102         else
2103                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2104
2105         return rc;
2106 }
2107
2108 static int
2109 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2110 {
2111         if (pool->po_allocated != 0) /* still in use */
2112                 return 0;
2113         if (pool->po_failed)
2114                 return 1;
2115         return cfs_time_aftereq(now, pool->po_deadline);
2116 }
2117
2118 void
2119 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2120 {
2121         struct list_head zombies = LIST_HEAD_INIT(zombies);
2122         kib_poolset_t   *ps = pool->po_owner;
2123         kib_pool_t      *tmp;
2124         cfs_time_t       now = cfs_time_current();
2125
2126         spin_lock(&ps->ps_lock);
2127
2128         if (ps->ps_node_fini != NULL)
2129                 ps->ps_node_fini(pool, node);
2130
2131         LASSERT(pool->po_allocated > 0);
2132         list_add(node, &pool->po_free_list);
2133         pool->po_allocated--;
2134
2135         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2136                 /* the first pool is persistent */
2137                 if (ps->ps_pool_list.next == &pool->po_list)
2138                         continue;
2139
2140                 if (kiblnd_pool_is_idle(pool, now))
2141                         list_move(&pool->po_list, &zombies);
2142         }
2143         spin_unlock(&ps->ps_lock);
2144
2145         if (!list_empty(&zombies))
2146                 kiblnd_destroy_pool_list(&zombies);
2147 }
2148
2149 struct list_head *
2150 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2151 {
2152         struct list_head        *node;
2153         kib_pool_t              *pool;
2154         int                     rc;
2155         unsigned int            interval = 1;
2156         cfs_time_t              time_before;
2157         unsigned int            trips = 0;
2158
2159 again:
2160         spin_lock(&ps->ps_lock);
2161         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2162                 if (list_empty(&pool->po_free_list))
2163                         continue;
2164
2165                 pool->po_allocated++;
2166                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2167                 node = pool->po_free_list.next;
2168                 list_del(node);
2169
2170                 if (ps->ps_node_init != NULL) {
2171                         /* still hold the lock */
2172                         ps->ps_node_init(pool, node);
2173                 }
2174                 spin_unlock(&ps->ps_lock);
2175                 return node;
2176         }
2177
2178         /* no available tx pool and ... */
2179         if (ps->ps_increasing) {
2180                 /* another thread is allocating a new pool */
2181                 spin_unlock(&ps->ps_lock);
2182                 trips++;
2183                 CDEBUG(D_NET, "Another thread is allocating new "
2184                        "%s pool, waiting %d HZs for her to complete."
2185                        "trips = %d\n",
2186                        ps->ps_name, interval, trips);
2187
2188                 set_current_state(TASK_INTERRUPTIBLE);
2189                 schedule_timeout(interval);
2190                 if (interval < cfs_time_seconds(1))
2191                         interval *= 2;
2192
2193                 goto again;
2194         }
2195
2196         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2197                 /* someone failed recently */
2198                 spin_unlock(&ps->ps_lock);
2199                 return NULL;
2200         }
2201
2202         ps->ps_increasing = 1;
2203         spin_unlock(&ps->ps_lock);
2204
2205         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2206         time_before = cfs_time_current();
2207         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2208         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2209                cfs_time_current() - time_before);
2210
2211         spin_lock(&ps->ps_lock);
2212         ps->ps_increasing = 0;
2213         if (rc == 0) {
2214                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2215         } else {
2216                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2217                 CERROR("Can't allocate new %s pool because out of memory\n",
2218                        ps->ps_name);
2219         }
2220         spin_unlock(&ps->ps_lock);
2221
2222         goto again;
2223 }
2224
2225 static void
2226 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2227 {
2228         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2229         int             i;
2230
2231         LASSERT (pool->po_allocated == 0);
2232
2233         if (tpo->tpo_tx_pages != NULL) {
2234                 kiblnd_unmap_tx_pool(tpo);
2235                 kiblnd_free_pages(tpo->tpo_tx_pages);
2236         }
2237
2238         if (tpo->tpo_tx_descs == NULL)
2239                 goto out;
2240
2241         for (i = 0; i < pool->po_size; i++) {
2242                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2243                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2244
2245                 list_del(&tx->tx_list);
2246                 if (tx->tx_pages != NULL)
2247                         LIBCFS_FREE(tx->tx_pages,
2248                                     LNET_MAX_IOV *
2249                                     sizeof(*tx->tx_pages));
2250                 if (tx->tx_frags != NULL)
2251                         LIBCFS_FREE(tx->tx_frags,
2252                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2253                                     sizeof(*tx->tx_frags));
2254                 if (tx->tx_wrq != NULL)
2255                         LIBCFS_FREE(tx->tx_wrq,
2256                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2257                                     sizeof(*tx->tx_wrq));
2258                 if (tx->tx_sge != NULL)
2259                         LIBCFS_FREE(tx->tx_sge,
2260                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2261                                     sizeof(*tx->tx_sge));
2262                 if (tx->tx_rd != NULL)
2263                         LIBCFS_FREE(tx->tx_rd,
2264                                     offsetof(kib_rdma_desc_t,
2265                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2266         }
2267
2268         LIBCFS_FREE(tpo->tpo_tx_descs,
2269                     pool->po_size * sizeof(kib_tx_t));
2270 out:
2271         kiblnd_fini_pool(pool);
2272         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2273 }
2274
2275 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2276 {
2277         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2278         int ntx;
2279
2280         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2281         ntx = tunables->lnd_ntx / ncpts;
2282
2283         return max(IBLND_TX_POOL, ntx);
2284 }
2285
2286 static int
2287 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2288 {
2289         int            i;
2290         int            npg;
2291         kib_pool_t    *pool;
2292         kib_tx_pool_t *tpo;
2293
2294         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2295         if (tpo == NULL) {
2296                 CERROR("Failed to allocate TX pool\n");
2297                 return -ENOMEM;
2298         }
2299
2300         pool = &tpo->tpo_pool;
2301         kiblnd_init_pool(ps, pool, size);
2302         tpo->tpo_tx_descs = NULL;
2303         tpo->tpo_tx_pages = NULL;
2304
2305         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2306         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2307                 CERROR("Can't allocate tx pages: %d\n", npg);
2308                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2309                 return -ENOMEM;
2310         }
2311
2312         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2313                          size * sizeof(kib_tx_t));
2314         if (tpo->tpo_tx_descs == NULL) {
2315                 CERROR("Can't allocate %d tx descriptors\n", size);
2316                 ps->ps_pool_destroy(pool);
2317                 return -ENOMEM;
2318         }
2319
2320         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2321
2322         for (i = 0; i < size; i++) {
2323                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2324                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2325
2326                 tx->tx_pool = tpo;
2327                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2328                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2329                                          lnet_cpt_table(), ps->ps_cpt,
2330                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2331                         if (tx->tx_pages == NULL)
2332                                 break;
2333                 }
2334
2335                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2336                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2337                                  sizeof(*tx->tx_frags));
2338                 if (tx->tx_frags == NULL)
2339                         break;
2340
2341                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2342
2343                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2344                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2345                                  sizeof(*tx->tx_wrq));
2346                 if (tx->tx_wrq == NULL)
2347                         break;
2348
2349                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2350                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2351                                  sizeof(*tx->tx_sge));
2352                 if (tx->tx_sge == NULL)
2353                         break;
2354
2355                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2356                                  offsetof(kib_rdma_desc_t,
2357                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2358                 if (tx->tx_rd == NULL)
2359                         break;
2360         }
2361
2362         if (i == size) {
2363                 kiblnd_map_tx_pool(tpo);
2364                 *pp_po = pool;
2365                 return 0;
2366         }
2367
2368         ps->ps_pool_destroy(pool);
2369         return -ENOMEM;
2370 }
2371
2372 static void
2373 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2374 {
2375         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2376                                              tps_poolset);
2377         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2378
2379         tx->tx_cookie = tps->tps_next_tx_cookie++;
2380 }
2381
2382 static void
2383 kiblnd_net_fini_pools(kib_net_t *net)
2384 {
2385         int     i;
2386
2387         cfs_cpt_for_each(i, lnet_cpt_table()) {
2388                 kib_tx_poolset_t        *tps;
2389                 kib_fmr_poolset_t       *fps;
2390
2391                 if (net->ibn_tx_ps != NULL) {
2392                         tps = net->ibn_tx_ps[i];
2393                         kiblnd_fini_poolset(&tps->tps_poolset);
2394                 }
2395
2396                 if (net->ibn_fmr_ps != NULL) {
2397                         fps = net->ibn_fmr_ps[i];
2398                         kiblnd_fini_fmr_poolset(fps);
2399                 }
2400         }
2401
2402         if (net->ibn_tx_ps != NULL) {
2403                 cfs_percpt_free(net->ibn_tx_ps);
2404                 net->ibn_tx_ps = NULL;
2405         }
2406
2407         if (net->ibn_fmr_ps != NULL) {
2408                 cfs_percpt_free(net->ibn_fmr_ps);
2409                 net->ibn_fmr_ps = NULL;
2410         }
2411 }
2412
2413 static int
2414 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2415                       int ncpts)
2416 {
2417         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2418 #ifdef HAVE_IB_GET_DMA_MR
2419         unsigned long   flags;
2420 #endif
2421         int             cpt;
2422         int             rc;
2423         int             i;
2424
2425         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2426
2427 #ifdef HAVE_IB_GET_DMA_MR
2428         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2429         /*
2430          * if lnd_map_on_demand is zero then we have effectively disabled
2431          * FMR or FastReg and we're using global memory regions
2432          * exclusively.
2433          */
2434         if (!tunables->lnd_map_on_demand) {
2435                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2436                                            flags);
2437                 goto create_tx_pool;
2438         }
2439
2440         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2441 #endif
2442
2443         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2444                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2445                        tunables->lnd_fmr_pool_size,
2446                        tunables->lnd_ntx / 4);
2447                 rc = -EINVAL;
2448                 goto failed;
2449         }
2450
2451         /* TX pool must be created later than FMR, see LU-2268
2452          * for details */
2453         LASSERT(net->ibn_tx_ps == NULL);
2454
2455         /* premapping can fail if ibd_nmr > 1, so we always create
2456          * FMR pool and map-on-demand if premapping failed */
2457
2458         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2459                                            sizeof(kib_fmr_poolset_t));
2460         if (net->ibn_fmr_ps == NULL) {
2461                 CERROR("Failed to allocate FMR pool array\n");
2462                 rc = -ENOMEM;
2463                 goto failed;
2464         }
2465
2466         for (i = 0; i < ncpts; i++) {
2467                 cpt = (cpts == NULL) ? i : cpts[i];
2468                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2469                                              net, tunables);
2470                 if (rc != 0) {
2471                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2472                                cpt, rc);
2473                         goto failed;
2474                 }
2475         }
2476
2477         if (i > 0)
2478                 LASSERT(i == ncpts);
2479
2480 #ifdef HAVE_IB_GET_DMA_MR
2481  create_tx_pool:
2482 #endif
2483         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2484                                           sizeof(kib_tx_poolset_t));
2485         if (net->ibn_tx_ps == NULL) {
2486                 CERROR("Failed to allocate tx pool array\n");
2487                 rc = -ENOMEM;
2488                 goto failed;
2489         }
2490
2491         for (i = 0; i < ncpts; i++) {
2492                 cpt = (cpts == NULL) ? i : cpts[i];
2493                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2494                                          cpt, net, "TX",
2495                                          kiblnd_tx_pool_size(ni, ncpts),
2496                                          kiblnd_create_tx_pool,
2497                                          kiblnd_destroy_tx_pool,
2498                                          kiblnd_tx_init, NULL);
2499                 if (rc != 0) {
2500                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2501                                cpt, rc);
2502                         goto failed;
2503                 }
2504         }
2505
2506         return 0;
2507  failed:
2508         kiblnd_net_fini_pools(net);
2509         LASSERT(rc != 0);
2510         return rc;
2511 }
2512
2513 static int
2514 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2515 {
2516         struct ib_device_attr *dev_attr;
2517         int rc = 0;
2518
2519         /* It's safe to assume a HCA can handle a page size
2520          * matching that of the native system */
2521         hdev->ibh_page_shift = PAGE_SHIFT;
2522         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2523         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2524
2525 #ifndef HAVE_IB_DEVICE_ATTRS
2526         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2527         if (dev_attr == NULL) {
2528                 CERROR("Out of memory\n");
2529                 return -ENOMEM;
2530         }
2531
2532         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2533         if (rc != 0) {
2534                 CERROR("Failed to query IB device: %d\n", rc);
2535                 goto out_clean_attr;
2536         }
2537 #else
2538         dev_attr = &hdev->ibh_ibdev->attrs;
2539 #endif
2540
2541         hdev->ibh_mr_size = dev_attr->max_mr_size;
2542
2543         /* Setup device Memory Registration capabilities */
2544         if (hdev->ibh_ibdev->alloc_fmr &&
2545             hdev->ibh_ibdev->dealloc_fmr &&
2546             hdev->ibh_ibdev->map_phys_fmr &&
2547             hdev->ibh_ibdev->unmap_fmr) {
2548                 LCONSOLE_INFO("Using FMR for registration\n");
2549                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2550         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2551                 LCONSOLE_INFO("Using FastReg for registration\n");
2552                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2553 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2554                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2555                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2556 #endif
2557         } else {
2558                 rc = -ENOSYS;
2559         }
2560
2561         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2562                 hdev->ibh_mr_shift = 64;
2563         else if (rc != 0)
2564                 rc = -EINVAL;
2565
2566 #ifndef HAVE_IB_DEVICE_ATTRS
2567 out_clean_attr:
2568         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2569 #endif
2570
2571         if (rc == -ENOSYS)
2572                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2573                        "register memory: %d\n", rc);
2574         else if (rc == -EINVAL)
2575                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2576         return rc;
2577 }
2578
2579 #ifdef HAVE_IB_GET_DMA_MR
2580 static void
2581 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2582 {
2583         if (hdev->ibh_mrs == NULL)
2584                 return;
2585
2586         ib_dereg_mr(hdev->ibh_mrs);
2587
2588         hdev->ibh_mrs = NULL;
2589 }
2590 #endif
2591
2592 void
2593 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2594 {
2595 #ifdef HAVE_IB_GET_DMA_MR
2596         kiblnd_hdev_cleanup_mrs(hdev);
2597 #endif
2598
2599         if (hdev->ibh_pd != NULL)
2600                 ib_dealloc_pd(hdev->ibh_pd);
2601
2602         if (hdev->ibh_cmid != NULL)
2603                 rdma_destroy_id(hdev->ibh_cmid);
2604
2605         LIBCFS_FREE(hdev, sizeof(*hdev));
2606 }
2607
2608 #ifdef HAVE_IB_GET_DMA_MR
2609 static int
2610 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2611 {
2612         struct ib_mr *mr;
2613         int           acflags = IB_ACCESS_LOCAL_WRITE |
2614                                 IB_ACCESS_REMOTE_WRITE;
2615
2616         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2617         if (IS_ERR(mr)) {
2618                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2619                 kiblnd_hdev_cleanup_mrs(hdev);
2620                 return PTR_ERR(mr);
2621         }
2622
2623         hdev->ibh_mrs = mr;
2624
2625         return 0;
2626 }
2627 #endif
2628
2629 static int
2630 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2631 {       /* DUMMY */
2632         return 0;
2633 }
2634
2635 static int
2636 kiblnd_dev_need_failover(kib_dev_t *dev)
2637 {
2638         struct rdma_cm_id  *cmid;
2639         struct sockaddr_in  srcaddr;
2640         struct sockaddr_in  dstaddr;
2641         int                 rc;
2642
2643         if (dev->ibd_hdev == NULL || /* initializing */
2644             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2645             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2646                 return 1;
2647
2648         /* XXX: it's UGLY, but I don't have better way to find
2649          * ib-bonding HCA failover because:
2650          *
2651          * a. no reliable CM event for HCA failover...
2652          * b. no OFED API to get ib_device for current net_device...
2653          *
2654          * We have only two choices at this point:
2655          *
2656          * a. rdma_bind_addr(), it will conflict with listener cmid
2657          * b. rdma_resolve_addr() to zero addr */
2658         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2659                                      IB_QPT_RC);
2660         if (IS_ERR(cmid)) {
2661                 rc = PTR_ERR(cmid);
2662                 CERROR("Failed to create cmid for failover: %d\n", rc);
2663                 return rc;
2664         }
2665
2666         memset(&srcaddr, 0, sizeof(srcaddr));
2667         srcaddr.sin_family      = AF_INET;
2668         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2669
2670         memset(&dstaddr, 0, sizeof(dstaddr));
2671         dstaddr.sin_family = AF_INET;
2672         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2673                                (struct sockaddr *)&dstaddr, 1);
2674         if (rc != 0 || cmid->device == NULL) {
2675                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2676                        dev->ibd_ifname, &dev->ibd_ifip,
2677                        cmid->device, rc);
2678                 rdma_destroy_id(cmid);
2679                 return rc;
2680         }
2681
2682         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2683         rdma_destroy_id(cmid);
2684         return rc;
2685 }
2686
2687 int
2688 kiblnd_dev_failover(kib_dev_t *dev)
2689 {
2690         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2691         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2692         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2693         struct rdma_cm_id  *cmid  = NULL;
2694         kib_hca_dev_t      *hdev  = NULL;
2695         kib_hca_dev_t      *old;
2696         struct ib_pd       *pd;
2697         kib_net_t          *net;
2698         struct sockaddr_in  addr;
2699         unsigned long       flags;
2700         int                 rc = 0;
2701         int                 i;
2702
2703         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2704                  dev->ibd_can_failover ||
2705                  dev->ibd_hdev == NULL);
2706
2707         rc = kiblnd_dev_need_failover(dev);
2708         if (rc <= 0)
2709                 goto out;
2710
2711         if (dev->ibd_hdev != NULL &&
2712             dev->ibd_hdev->ibh_cmid != NULL) {
2713                 /* XXX it's not good to close old listener at here,
2714                  * because we can fail to create new listener.
2715                  * But we have to close it now, otherwise rdma_bind_addr
2716                  * will return EADDRINUSE... How crap! */
2717                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2718
2719                 cmid = dev->ibd_hdev->ibh_cmid;
2720                 /* make next schedule of kiblnd_dev_need_failover()
2721                  * return 1 for me */
2722                 dev->ibd_hdev->ibh_cmid  = NULL;
2723                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2724
2725                 rdma_destroy_id(cmid);
2726         }
2727
2728         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2729                                      IB_QPT_RC);
2730         if (IS_ERR(cmid)) {
2731                 rc = PTR_ERR(cmid);
2732                 CERROR("Failed to create cmid for failover: %d\n", rc);
2733                 goto out;
2734         }
2735
2736         memset(&addr, 0, sizeof(addr));
2737         addr.sin_family      = AF_INET;
2738         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2739         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2740
2741         /* Bind to failover device or port */
2742         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2743         if (rc != 0 || cmid->device == NULL) {
2744                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2745                        dev->ibd_ifname, &dev->ibd_ifip,
2746                        cmid->device, rc);
2747                 rdma_destroy_id(cmid);
2748                 goto out;
2749         }
2750
2751         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2752         if (hdev == NULL) {
2753                 CERROR("Failed to allocate kib_hca_dev\n");
2754                 rdma_destroy_id(cmid);
2755                 rc = -ENOMEM;
2756                 goto out;
2757         }
2758
2759         atomic_set(&hdev->ibh_ref, 1);
2760         hdev->ibh_dev   = dev;
2761         hdev->ibh_cmid  = cmid;
2762         hdev->ibh_ibdev = cmid->device;
2763
2764 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2765         pd = ib_alloc_pd(cmid->device, 0);
2766 #else
2767         pd = ib_alloc_pd(cmid->device);
2768 #endif
2769         if (IS_ERR(pd)) {
2770                 rc = PTR_ERR(pd);
2771                 CERROR("Can't allocate PD: %d\n", rc);
2772                 goto out;
2773         }
2774
2775         hdev->ibh_pd = pd;
2776
2777         rc = rdma_listen(cmid, 0);
2778         if (rc != 0) {
2779                 CERROR("Can't start new listener: %d\n", rc);
2780                 goto out;
2781         }
2782
2783         rc = kiblnd_hdev_get_attr(hdev);
2784         if (rc != 0) {
2785                 CERROR("Can't get device attributes: %d\n", rc);
2786                 goto out;
2787         }
2788
2789 #ifdef HAVE_IB_GET_DMA_MR
2790         rc = kiblnd_hdev_setup_mrs(hdev);
2791         if (rc != 0) {
2792                 CERROR("Can't setup device: %d\n", rc);
2793                 goto out;
2794         }
2795 #endif
2796
2797         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2798
2799         old = dev->ibd_hdev;
2800         dev->ibd_hdev = hdev;   /* take over the refcount */
2801         hdev = old;
2802
2803         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2804                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2805                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2806                                             &zombie_tpo);
2807
2808                         if (net->ibn_fmr_ps != NULL)
2809                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2810                                                         &zombie_fpo);
2811                 }
2812         }
2813
2814         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2815  out:
2816         if (!list_empty(&zombie_tpo))
2817                 kiblnd_destroy_pool_list(&zombie_tpo);
2818         if (!list_empty(&zombie_ppo))
2819                 kiblnd_destroy_pool_list(&zombie_ppo);
2820         if (!list_empty(&zombie_fpo))
2821                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2822         if (hdev != NULL)
2823                 kiblnd_hdev_decref(hdev);
2824
2825         if (rc != 0)
2826                 dev->ibd_failed_failover++;
2827         else
2828                 dev->ibd_failed_failover = 0;
2829
2830         return rc;
2831 }
2832
2833 void
2834 kiblnd_destroy_dev (kib_dev_t *dev)
2835 {
2836         LASSERT (dev->ibd_nnets == 0);
2837         LASSERT(list_empty(&dev->ibd_nets));
2838
2839         list_del(&dev->ibd_fail_list);
2840         list_del(&dev->ibd_list);
2841
2842         if (dev->ibd_hdev != NULL)
2843                 kiblnd_hdev_decref(dev->ibd_hdev);
2844
2845         LIBCFS_FREE(dev, sizeof(*dev));
2846 }
2847
2848 static kib_dev_t *
2849 kiblnd_create_dev(char *ifname)
2850 {
2851         struct net_device *netdev;
2852         kib_dev_t         *dev;
2853         __u32              netmask;
2854         __u32              ip;
2855         int                up;
2856         int                rc;
2857
2858         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2859         if (rc != 0) {
2860                 CERROR("Can't query IPoIB interface %s: %d\n",
2861                        ifname, rc);
2862                 return NULL;
2863         }
2864
2865         if (!up) {
2866                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2867                 return NULL;
2868         }
2869
2870         LIBCFS_ALLOC(dev, sizeof(*dev));
2871         if (dev == NULL)
2872                 return NULL;
2873
2874         netdev = dev_get_by_name(&init_net, ifname);
2875         if (netdev == NULL) {
2876                 dev->ibd_can_failover = 0;
2877         } else {
2878                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2879                 dev_put(netdev);
2880         }
2881
2882         INIT_LIST_HEAD(&dev->ibd_nets);
2883         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2884         INIT_LIST_HEAD(&dev->ibd_fail_list);
2885         dev->ibd_ifip = ip;
2886         strcpy(&dev->ibd_ifname[0], ifname);
2887
2888         /* initialize the device */
2889         rc = kiblnd_dev_failover(dev);
2890         if (rc != 0) {
2891                 CERROR("Can't initialize device: %d\n", rc);
2892                 LIBCFS_FREE(dev, sizeof(*dev));
2893                 return NULL;
2894         }
2895
2896         list_add_tail(&dev->ibd_list,
2897                           &kiblnd_data.kib_devs);
2898         return dev;
2899 }
2900
2901 static void
2902 kiblnd_base_shutdown(void)
2903 {
2904         struct kib_sched_info   *sched;
2905         int                     i;
2906
2907         LASSERT(list_empty(&kiblnd_data.kib_devs));
2908
2909         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2910                atomic_read(&libcfs_kmemory));
2911
2912         switch (kiblnd_data.kib_init) {
2913         default:
2914                 LBUG();
2915
2916         case IBLND_INIT_ALL:
2917         case IBLND_INIT_DATA:
2918                 LASSERT (kiblnd_data.kib_peers != NULL);
2919                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2920                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2921                 }
2922                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2923                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2924                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2925                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2926
2927                 /* flag threads to terminate; wake and wait for them to die */
2928                 kiblnd_data.kib_shutdown = 1;
2929
2930                 /* NB: we really want to stop scheduler threads net by net
2931                  * instead of the whole module, this should be improved
2932                  * with dynamic configuration LNet */
2933                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2934                         wake_up_all(&sched->ibs_waitq);
2935
2936                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2937                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2938
2939                 i = 2;
2940                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2941                         i++;
2942                         /* power of 2? */
2943                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2944                                "Waiting for %d threads to terminate\n",
2945                                atomic_read(&kiblnd_data.kib_nthreads));
2946                         set_current_state(TASK_UNINTERRUPTIBLE);
2947                         schedule_timeout(cfs_time_seconds(1));
2948                 }
2949
2950                 /* fall through */
2951
2952         case IBLND_INIT_NOTHING:
2953                 break;
2954         }
2955
2956         if (kiblnd_data.kib_peers != NULL) {
2957                 LIBCFS_FREE(kiblnd_data.kib_peers,
2958                             sizeof(struct list_head) *
2959                             kiblnd_data.kib_peer_hash_size);
2960         }
2961
2962         if (kiblnd_data.kib_scheds != NULL)
2963                 cfs_percpt_free(kiblnd_data.kib_scheds);
2964
2965         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2966                atomic_read(&libcfs_kmemory));
2967
2968         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2969         module_put(THIS_MODULE);
2970 }
2971
2972 static void
2973 kiblnd_shutdown(struct lnet_ni *ni)
2974 {
2975         kib_net_t        *net = ni->ni_data;
2976         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2977         int               i;
2978         unsigned long     flags;
2979
2980         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2981
2982         if (net == NULL)
2983                 goto out;
2984
2985         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2986                atomic_read(&libcfs_kmemory));
2987
2988         write_lock_irqsave(g_lock, flags);
2989         net->ibn_shutdown = 1;
2990         write_unlock_irqrestore(g_lock, flags);
2991
2992         switch (net->ibn_init) {
2993         default:
2994                 LBUG();
2995
2996         case IBLND_INIT_ALL:
2997                 /* nuke all existing peers within this net */
2998                 kiblnd_del_peer(ni, LNET_NID_ANY);
2999
3000                 /* Wait for all peer_ni state to clean up */
3001                 i = 2;
3002                 while (atomic_read(&net->ibn_npeers) != 0) {
3003                         i++;
3004                         /* power of 2? */
3005                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3006                                "%s: waiting for %d peers to disconnect\n",
3007                                libcfs_nid2str(ni->ni_nid),
3008                                atomic_read(&net->ibn_npeers));
3009                         set_current_state(TASK_UNINTERRUPTIBLE);
3010                         schedule_timeout(cfs_time_seconds(1));
3011                 }
3012
3013                 kiblnd_net_fini_pools(net);
3014
3015                 write_lock_irqsave(g_lock, flags);
3016                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3017                 net->ibn_dev->ibd_nnets--;
3018                 list_del(&net->ibn_list);
3019                 write_unlock_irqrestore(g_lock, flags);
3020
3021                 /* fall through */
3022
3023         case IBLND_INIT_NOTHING:
3024                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3025
3026                 if (net->ibn_dev != NULL &&
3027                     net->ibn_dev->ibd_nnets == 0)
3028                         kiblnd_destroy_dev(net->ibn_dev);
3029
3030                 break;
3031         }
3032
3033         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3034                atomic_read(&libcfs_kmemory));
3035
3036         net->ibn_init = IBLND_INIT_NOTHING;
3037         ni->ni_data = NULL;
3038
3039         LIBCFS_FREE(net, sizeof(*net));
3040
3041 out:
3042         if (list_empty(&kiblnd_data.kib_devs))
3043                 kiblnd_base_shutdown();
3044         return;
3045 }
3046
3047 static int
3048 kiblnd_base_startup(void)
3049 {
3050         struct kib_sched_info   *sched;
3051         int                     rc;
3052         int                     i;
3053
3054         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3055
3056         try_module_get(THIS_MODULE);
3057         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3058
3059         rwlock_init(&kiblnd_data.kib_global_lock);
3060
3061         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3062         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3063
3064         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3065         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3066                      sizeof(struct list_head) *
3067                      kiblnd_data.kib_peer_hash_size);
3068         if (kiblnd_data.kib_peers == NULL)
3069                 goto failed;
3070
3071         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3072                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3073
3074         spin_lock_init(&kiblnd_data.kib_connd_lock);
3075         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3076         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3077         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3078         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3079
3080         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3081         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3082
3083         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3084                                                   sizeof(*sched));
3085         if (kiblnd_data.kib_scheds == NULL)
3086                 goto failed;
3087
3088         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3089                 int     nthrs;
3090
3091                 spin_lock_init(&sched->ibs_lock);
3092                 INIT_LIST_HEAD(&sched->ibs_conns);
3093                 init_waitqueue_head(&sched->ibs_waitq);
3094
3095                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3096                 if (*kiblnd_tunables.kib_nscheds > 0) {
3097                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3098                 } else {
3099                         /* max to half of CPUs, another half is reserved for
3100                          * upper layer modules */
3101                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3102                 }
3103
3104                 sched->ibs_nthreads_max = nthrs;
3105                 sched->ibs_cpt = i;
3106         }
3107
3108         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3109
3110         /* lists/ptrs/locks initialised */
3111         kiblnd_data.kib_init = IBLND_INIT_DATA;
3112         /*****************************************************/
3113
3114         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3115         if (rc != 0) {
3116                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3117                 goto failed;
3118         }
3119
3120         if (*kiblnd_tunables.kib_dev_failover != 0)
3121                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3122                                          "kiblnd_failover");
3123
3124         if (rc != 0) {
3125                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3126                 goto failed;
3127         }
3128
3129         /* flag everything initialised */
3130         kiblnd_data.kib_init = IBLND_INIT_ALL;
3131         /*****************************************************/
3132
3133         return 0;
3134
3135  failed:
3136         kiblnd_base_shutdown();
3137         return -ENETDOWN;
3138 }
3139
3140 static int
3141 kiblnd_start_schedulers(struct kib_sched_info *sched)
3142 {
3143         int     rc = 0;
3144         int     nthrs;
3145         int     i;
3146
3147         if (sched->ibs_nthreads == 0) {
3148                 if (*kiblnd_tunables.kib_nscheds > 0) {
3149                         nthrs = sched->ibs_nthreads_max;
3150                 } else {
3151                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3152                                                sched->ibs_cpt);
3153                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3154                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3155                 }
3156         } else {
3157                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3158                 /* increase one thread if there is new interface */
3159                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3160         }
3161
3162         for (i = 0; i < nthrs; i++) {
3163                 long    id;
3164                 char    name[20];
3165                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3166                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3167                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3168                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3169                 if (rc == 0)
3170                         continue;
3171
3172                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3173                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3174                 break;
3175         }
3176
3177         sched->ibs_nthreads += i;
3178         return rc;
3179 }
3180
3181 static int
3182 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3183 {
3184         int     cpt;
3185         int     rc;
3186         int     i;
3187
3188         for (i = 0; i < ncpts; i++) {
3189                 struct kib_sched_info *sched;
3190
3191                 cpt = (cpts == NULL) ? i : cpts[i];
3192                 sched = kiblnd_data.kib_scheds[cpt];
3193
3194                 if (!newdev && sched->ibs_nthreads > 0)
3195                         continue;
3196
3197                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3198                 if (rc != 0) {
3199                         CERROR("Failed to start scheduler threads for %s\n",
3200                                dev->ibd_ifname);
3201                         return rc;
3202                 }
3203         }
3204         return 0;
3205 }
3206
3207 static kib_dev_t *
3208 kiblnd_dev_search(char *ifname)
3209 {
3210         kib_dev_t       *alias = NULL;
3211         kib_dev_t       *dev;
3212         char            *colon;
3213         char            *colon2;
3214
3215         colon = strchr(ifname, ':');
3216         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3217                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3218                         return dev;
3219
3220                 if (alias != NULL)
3221                         continue;
3222
3223                 colon2 = strchr(dev->ibd_ifname, ':');
3224                 if (colon != NULL)
3225                         *colon = 0;
3226                 if (colon2 != NULL)
3227                         *colon2 = 0;
3228
3229                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3230                         alias = dev;
3231
3232                 if (colon != NULL)
3233                         *colon = ':';
3234                 if (colon2 != NULL)
3235                         *colon2 = ':';
3236         }
3237         return alias;
3238 }
3239
3240 static int
3241 kiblnd_startup(struct lnet_ni *ni)
3242 {
3243         char                     *ifname;
3244         kib_dev_t                *ibdev = NULL;
3245         kib_net_t                *net;
3246         unsigned long             flags;
3247         int                       rc;
3248         int                       newdev;
3249         int                       node_id;
3250
3251         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3252
3253         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3254                 rc = kiblnd_base_startup();
3255                 if (rc != 0)
3256                         return rc;
3257         }
3258
3259         LIBCFS_ALLOC(net, sizeof(*net));
3260         ni->ni_data = net;
3261         if (net == NULL)
3262                 goto failed;
3263
3264         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3265
3266         kiblnd_tunables_setup(ni);
3267
3268         if (ni->ni_interfaces[0] != NULL) {
3269                 /* Use the IPoIB interface specified in 'networks=' */
3270
3271                 CLASSERT(LNET_INTERFACES_NUM > 1);
3272                 if (ni->ni_interfaces[1] != NULL) {
3273                         CERROR("Multiple interfaces not supported\n");
3274                         goto failed;
3275                 }
3276
3277                 ifname = ni->ni_interfaces[0];
3278         } else {
3279                 ifname = *kiblnd_tunables.kib_default_ipif;
3280         }
3281
3282         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3283                 CERROR("IPoIB interface name too long: %s\n", ifname);
3284                 goto failed;
3285         }
3286
3287         ibdev = kiblnd_dev_search(ifname);
3288
3289         newdev = ibdev == NULL;
3290         /* hmm...create kib_dev even for alias */
3291         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3292                 ibdev = kiblnd_create_dev(ifname);
3293
3294         if (ibdev == NULL)
3295                 goto failed;
3296
3297         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3298         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3299
3300         net->ibn_dev = ibdev;
3301         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3302
3303         rc = kiblnd_dev_start_threads(ibdev, newdev,
3304                                       ni->ni_cpts, ni->ni_ncpts);
3305         if (rc != 0)
3306                 goto failed;
3307
3308         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3309         if (rc != 0) {
3310                 CERROR("Failed to initialize NI pools: %d\n", rc);
3311                 goto failed;
3312         }
3313
3314         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3315         ibdev->ibd_nnets++;
3316         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3317         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3318
3319         net->ibn_init = IBLND_INIT_ALL;
3320
3321         return 0;
3322
3323 failed:
3324         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3325                 kiblnd_destroy_dev(ibdev);
3326
3327         kiblnd_shutdown(ni);
3328
3329         CDEBUG(D_NET, "kiblnd_startup failed\n");
3330         return -ENETDOWN;
3331 }
3332
3333 static struct lnet_lnd the_o2iblnd = {
3334         .lnd_type       = O2IBLND,
3335         .lnd_startup    = kiblnd_startup,
3336         .lnd_shutdown   = kiblnd_shutdown,
3337         .lnd_ctl        = kiblnd_ctl,
3338         .lnd_query      = kiblnd_query,
3339         .lnd_send       = kiblnd_send,
3340         .lnd_recv       = kiblnd_recv,
3341 };
3342
3343 static void __exit ko2iblnd_exit(void)
3344 {
3345         lnet_unregister_lnd(&the_o2iblnd);
3346 }
3347
3348 static int __init ko2iblnd_init(void)
3349 {
3350         int rc;
3351
3352         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3353         CLASSERT(offsetof(kib_msg_t,
3354                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3355                  IBLND_MSG_SIZE);
3356         CLASSERT(offsetof(kib_msg_t,
3357                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3358                  <= IBLND_MSG_SIZE);
3359
3360         rc = kiblnd_tunables_init();
3361         if (rc != 0)
3362                 return rc;
3363
3364         lnet_register_lnd(&the_o2iblnd);
3365
3366         return 0;
3367 }
3368
3369 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3370 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3371 MODULE_VERSION("2.8.0");
3372 MODULE_LICENSE("GPL");
3373
3374 module_init(ko2iblnd_init);
3375 module_exit(ko2iblnd_exit);