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, 2017, 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         __u32 dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
 739
 740         /* FastReg needs two extra WRs for map and invalidate */
 741         if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
 742                 ret += 2;
 743
 744         /* account for a maximum of ibc_queue_depth in-flight transfers */
 745         ret *= conn->ibc_queue_depth;
 746         return ret;
 747 }
 748
 749 kib_conn_t *
 750 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 751                    int state, int version)
 752 {
 753         /* CAVEAT EMPTOR:
 754          * If the new conn is created successfully it takes over the caller's
 755          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 756          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 757          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 758          * to destroy 'cmid' here since I'm called from the CM which still has
 759          * its ref on 'cmid'). */
 760         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 761         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 762         kib_dev_t              *dev;
 763         struct ib_qp_init_attr *init_qp_attr;
 764         struct kib_sched_info   *sched;
 765 #ifdef HAVE_IB_CQ_INIT_ATTR
 766         struct ib_cq_init_attr  cq_attr = {};
 767 #endif
 768         kib_conn_t              *conn;
 769         struct ib_cq            *cq;
 770         unsigned long           flags;
 771         int                     cpt;
 772         int                     rc;
 773         int                     i;
 774
 775         LASSERT(net != NULL);
 776         LASSERT(!in_interrupt());
 777
 778         dev = net->ibn_dev;
 779
 780         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 781         sched = kiblnd_get_scheduler(cpt);
 782
 783         if (sched == NULL) {
 784                 CERROR("no schedulers available. node is unhealthy\n");
 785                 goto failed_0;
 786         }
 787
 788         /*
 789          * The cpt might have changed if we ended up selecting a non cpt
 790          * native scheduler. So use the scheduler's cpt instead.
 791          */
 792         cpt = sched->ibs_cpt;
 793
 794         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 795                          sizeof(*init_qp_attr));
 796         if (init_qp_attr == NULL) {
 797                 CERROR("Can't allocate qp_attr for %s\n",
 798                        libcfs_nid2str(peer_ni->ibp_nid));
 799                 goto failed_0;
 800         }
 801
 802         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 803         if (conn == NULL) {
 804                 CERROR("Can't allocate connection for %s\n",
 805                        libcfs_nid2str(peer_ni->ibp_nid));
 806                 goto failed_1;
 807         }
 808
 809         conn->ibc_state = IBLND_CONN_INIT;
 810         conn->ibc_version = version;
 811         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 812         cmid->context = conn;                   /* for future CM callbacks */
 813         conn->ibc_cmid = cmid;
 814         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 815         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 816
 817         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 818         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 819         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 820         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 821         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 822         INIT_LIST_HEAD(&conn->ibc_active_txs);
 823         spin_lock_init(&conn->ibc_lock);
 824
 825         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 826                          sizeof(*conn->ibc_connvars));
 827         if (conn->ibc_connvars == NULL) {
 828                 CERROR("Can't allocate in-progress connection state\n");
 829                 goto failed_2;
 830         }
 831
 832         write_lock_irqsave(glock, flags);
 833         if (dev->ibd_failover) {
 834                 write_unlock_irqrestore(glock, flags);
 835                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 836                 goto failed_2;
 837         }
 838
 839         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 840                 /* wakeup failover thread and teardown connection */
 841                 if (kiblnd_dev_can_failover(dev)) {
 842                         list_add_tail(&dev->ibd_fail_list,
 843                                       &kiblnd_data.kib_failed_devs);
 844                         wake_up(&kiblnd_data.kib_failover_waitq);
 845                 }
 846
 847                 write_unlock_irqrestore(glock, flags);
 848                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 849                        cmid->device->name, dev->ibd_ifname);
 850                 goto failed_2;
 851         }
 852
 853         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 854         conn->ibc_hdev = dev->ibd_hdev;
 855
 856         kiblnd_setup_mtu_locked(cmid);
 857
 858         write_unlock_irqrestore(glock, flags);
 859
 860         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 861                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 862         if (conn->ibc_rxs == NULL) {
 863                 CERROR("Cannot allocate RX buffers\n");
 864                 goto failed_2;
 865         }
 866
 867         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 868                                 IBLND_RX_MSG_PAGES(conn));
 869         if (rc != 0)
 870                 goto failed_2;
 871
 872         kiblnd_map_rx_descs(conn);
 873
 874 #ifdef HAVE_IB_CQ_INIT_ATTR
 875         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 876         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 877         cq = ib_create_cq(cmid->device,
 878                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 879                           &cq_attr);
 880 #else
 881         cq = ib_create_cq(cmid->device,
 882                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 883                           IBLND_CQ_ENTRIES(conn),
 884                           kiblnd_get_completion_vector(conn, cpt));
 885 #endif
 886         if (IS_ERR(cq)) {
 887                 /*
 888                  * on MLX-5 (possibly MLX-4 as well) this error could be
 889                  * hit if the concurrent_sends and/or peer_tx_credits is set
 890                  * too high. Or due to an MLX-5 bug which tries to
 891                  * allocate 256kb via kmalloc for WR cookie array
 892                  */
 893                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 894                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 895                 goto failed_2;
 896         }
 897
 898         conn->ibc_cq = cq;
 899
 900         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 901         if (rc != 0) {
 902                 CERROR("Can't request completion notification: %d\n", rc);
 903                 goto failed_2;
 904         }
 905
 906         init_qp_attr->event_handler = kiblnd_qp_event;
 907         init_qp_attr->qp_context = conn;
 908         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 909         init_qp_attr->cap.max_recv_sge = 1;
 910         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 911         init_qp_attr->qp_type = IB_QPT_RC;
 912         init_qp_attr->send_cq = cq;
 913         init_qp_attr->recv_cq = cq;
 914
 915         conn->ibc_sched = sched;
 916
 917         do {
 918                 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 919                 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 920
 921                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 922                 if (!rc || conn->ibc_queue_depth < 2)
 923                         break;
 924
 925                 conn->ibc_queue_depth--;
 926         } while (rc);
 927
 928         if (rc) {
 929                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 930                        "send_sge: %d, recv_sge: %d\n",
 931                        rc, init_qp_attr->cap.max_send_wr,
 932                        init_qp_attr->cap.max_recv_wr,
 933                        init_qp_attr->cap.max_send_sge,
 934                        init_qp_attr->cap.max_recv_sge);
 935                 goto failed_2;
 936         }
 937
 938         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 939                 CWARN("peer %s - queue depth reduced from %u to %u"
 940                       "  to allow for qp creation\n",
 941                       libcfs_nid2str(peer_ni->ibp_nid),
 942                       peer_ni->ibp_queue_depth,
 943                       conn->ibc_queue_depth);
 944
 945         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 946
 947         /* 1 ref for caller and each rxmsg */
 948         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 949         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 950
 951         /* post receives */
 952         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 953                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 954                 if (rc != 0) {
 955                         CERROR("Can't post rxmsg: %d\n", rc);
 956
 957                         /* Make posted receives complete */
 958                         kiblnd_abort_receives(conn);
 959
 960                         /* correct # of posted buffers
 961                          * NB locking needed now I'm racing with completion */
 962                         spin_lock_irqsave(&sched->ibs_lock, flags);
 963                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 964                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 965
 966                         /* cmid will be destroyed by CM(ofed) after cm_callback
 967                          * returned, so we can't refer it anymore
 968                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 969                         rdma_destroy_qp(conn->ibc_cmid);
 970                         conn->ibc_cmid = NULL;
 971
 972                         /* Drop my own and unused rxbuffer refcounts */
 973                         while (i++ <= IBLND_RX_MSGS(conn))
 974                                 kiblnd_conn_decref(conn);
 975
 976                         return NULL;
 977                 }
 978         }
 979
 980         /* Init successful! */
 981         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 982                  state == IBLND_CONN_PASSIVE_WAIT);
 983         conn->ibc_state = state;
 984
 985         /* 1 more conn */
 986         atomic_inc(&net->ibn_nconns);
 987         return conn;
 988
 989  failed_2:
 990         kiblnd_destroy_conn(conn, true);
 991  failed_1:
 992         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 993  failed_0:
 994         return NULL;
 995 }
 996
 997 void
 998 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 999 {
1000         struct rdma_cm_id *cmid = conn->ibc_cmid;
1001         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
1002         int                rc;
1003
1004         LASSERT (!in_interrupt());
1005         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1006         LASSERT(list_empty(&conn->ibc_early_rxs));
1007         LASSERT(list_empty(&conn->ibc_tx_noops));
1008         LASSERT(list_empty(&conn->ibc_tx_queue));
1009         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1010         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1011         LASSERT(list_empty(&conn->ibc_active_txs));
1012         LASSERT (conn->ibc_noops_posted == 0);
1013         LASSERT (conn->ibc_nsends_posted == 0);
1014
1015         switch (conn->ibc_state) {
1016         default:
1017                 /* conn must be completely disengaged from the network */
1018                 LBUG();
1019
1020         case IBLND_CONN_DISCONNECTED:
1021                 /* connvars should have been freed already */
1022                 LASSERT (conn->ibc_connvars == NULL);
1023                 break;
1024
1025         case IBLND_CONN_INIT:
1026                 break;
1027         }
1028
1029         /* conn->ibc_cmid might be destroyed by CM already */
1030         if (cmid != NULL && cmid->qp != NULL)
1031                 rdma_destroy_qp(cmid);
1032
1033         if (conn->ibc_cq != NULL) {
1034                 rc = ib_destroy_cq(conn->ibc_cq);
1035                 if (rc != 0)
1036                         CWARN("Error destroying CQ: %d\n", rc);
1037         }
1038
1039         if (conn->ibc_rx_pages != NULL)
1040                 kiblnd_unmap_rx_descs(conn);
1041
1042         if (conn->ibc_rxs != NULL) {
1043                 LIBCFS_FREE(conn->ibc_rxs,
1044                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1045         }
1046
1047         if (conn->ibc_connvars != NULL)
1048                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1049
1050         if (conn->ibc_hdev != NULL)
1051                 kiblnd_hdev_decref(conn->ibc_hdev);
1052
1053         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1054         if (conn->ibc_state != IBLND_CONN_INIT) {
1055                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1056
1057                 kiblnd_peer_decref(peer_ni);
1058                 rdma_destroy_id(cmid);
1059                 atomic_dec(&net->ibn_nconns);
1060         }
1061
1062         if (free_conn)
1063                 LIBCFS_FREE(conn, sizeof(*conn));
1064 }
1065
1066 int
1067 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1068 {
1069         kib_conn_t              *conn;
1070         struct list_head        *ctmp;
1071         struct list_head        *cnxt;
1072         int                     count = 0;
1073
1074         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1075                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1076
1077                 CDEBUG(D_NET, "Closing conn -> %s, "
1078                               "version: %x, reason: %d\n",
1079                        libcfs_nid2str(peer_ni->ibp_nid),
1080                        conn->ibc_version, why);
1081
1082                 kiblnd_close_conn_locked(conn, why);
1083                 count++;
1084         }
1085
1086         return count;
1087 }
1088
1089 int
1090 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1091                                 int version, __u64 incarnation)
1092 {
1093         kib_conn_t              *conn;
1094         struct list_head        *ctmp;
1095         struct list_head        *cnxt;
1096         int                     count = 0;
1097
1098         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1099                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1100
1101                 if (conn->ibc_version     == version &&
1102                     conn->ibc_incarnation == incarnation)
1103                         continue;
1104
1105                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1106                               "incarnation:%#llx(%x, %#llx)\n",
1107                        libcfs_nid2str(peer_ni->ibp_nid),
1108                        conn->ibc_version, conn->ibc_incarnation,
1109                        version, incarnation);
1110
1111                 kiblnd_close_conn_locked(conn, -ESTALE);
1112                 count++;
1113         }
1114
1115         return count;
1116 }
1117
1118 static int
1119 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1120 {
1121         kib_peer_ni_t           *peer_ni;
1122         struct list_head        *ptmp;
1123         struct list_head        *pnxt;
1124         int                     lo;
1125         int                     hi;
1126         int                     i;
1127         unsigned long           flags;
1128         int                     count = 0;
1129
1130         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1131
1132         if (nid != LNET_NID_ANY)
1133                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1134         else {
1135                 lo = 0;
1136                 hi = kiblnd_data.kib_peer_hash_size - 1;
1137         }
1138
1139         for (i = lo; i <= hi; i++) {
1140                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1141
1142                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1143                         LASSERT(!kiblnd_peer_idle(peer_ni));
1144
1145                         if (peer_ni->ibp_ni != ni)
1146                                 continue;
1147
1148                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1149                                 continue;
1150
1151                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1152                 }
1153         }
1154
1155         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1156
1157         /* wildcards always succeed */
1158         if (nid == LNET_NID_ANY)
1159                 return 0;
1160
1161         return (count == 0) ? -ENOENT : 0;
1162 }
1163
1164 static int
1165 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1166 {
1167         struct libcfs_ioctl_data *data = arg;
1168         int                       rc = -EINVAL;
1169
1170         switch(cmd) {
1171         case IOC_LIBCFS_GET_PEER: {
1172                 lnet_nid_t   nid = 0;
1173                 int          count = 0;
1174
1175                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1176                                           &nid, &count);
1177                 data->ioc_nid    = nid;
1178                 data->ioc_count  = count;
1179                 break;
1180         }
1181
1182         case IOC_LIBCFS_DEL_PEER: {
1183                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1184                 break;
1185         }
1186         case IOC_LIBCFS_GET_CONN: {
1187                 kib_conn_t *conn;
1188
1189                 rc = 0;
1190                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1191                 if (conn == NULL) {
1192                         rc = -ENOENT;
1193                         break;
1194                 }
1195
1196                 LASSERT(conn->ibc_cmid != NULL);
1197                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1198                 if (conn->ibc_cmid->route.path_rec == NULL)
1199                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1200                 else
1201                         data->ioc_u32[0] =
1202                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1203                 kiblnd_conn_decref(conn);
1204                 break;
1205         }
1206         case IOC_LIBCFS_CLOSE_CONNECTION: {
1207                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1208                 break;
1209         }
1210
1211         default:
1212                 break;
1213         }
1214
1215         return rc;
1216 }
1217
1218 static void
1219 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1220 {
1221         time64_t last_alive = 0;
1222         time64_t now = ktime_get_seconds();
1223         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1224         kib_peer_ni_t *peer_ni;
1225         unsigned long flags;
1226
1227         read_lock_irqsave(glock, flags);
1228
1229         peer_ni = kiblnd_find_peer_locked(ni, nid);
1230         if (peer_ni != NULL)
1231                 last_alive = peer_ni->ibp_last_alive;
1232
1233         read_unlock_irqrestore(glock, flags);
1234
1235         if (last_alive != 0)
1236                 *when = last_alive;
1237
1238         /* peer_ni is not persistent in hash, trigger peer_ni creation
1239          * and connection establishment with a NULL tx */
1240         if (peer_ni == NULL)
1241                 kiblnd_launch_tx(ni, NULL, nid);
1242
1243         CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1244                libcfs_nid2str(nid), peer_ni,
1245                last_alive ? now - last_alive : -1);
1246         return;
1247 }
1248
1249 static void
1250 kiblnd_free_pages(kib_pages_t *p)
1251 {
1252         int     npages = p->ibp_npages;
1253         int     i;
1254
1255         for (i = 0; i < npages; i++) {
1256                 if (p->ibp_pages[i] != NULL)
1257                         __free_page(p->ibp_pages[i]);
1258         }
1259
1260         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1261 }
1262
1263 int
1264 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1265 {
1266         kib_pages_t     *p;
1267         int             i;
1268
1269         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1270                          offsetof(kib_pages_t, ibp_pages[npages]));
1271         if (p == NULL) {
1272                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1273                 return -ENOMEM;
1274         }
1275
1276         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1277         p->ibp_npages = npages;
1278
1279         for (i = 0; i < npages; i++) {
1280                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1281                                                      GFP_NOFS);
1282                 if (p->ibp_pages[i] == NULL) {
1283                         CERROR("Can't allocate page %d of %d\n", i, npages);
1284                         kiblnd_free_pages(p);
1285                         return -ENOMEM;
1286                 }
1287         }
1288
1289         *pp = p;
1290         return 0;
1291 }
1292
1293 void
1294 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1295 {
1296         kib_rx_t *rx;
1297         int       i;
1298
1299         LASSERT (conn->ibc_rxs != NULL);
1300         LASSERT (conn->ibc_hdev != NULL);
1301
1302         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1303                 rx = &conn->ibc_rxs[i];
1304
1305                 LASSERT(rx->rx_nob >= 0); /* not posted */
1306
1307                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1308                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1309                                                           rx->rx_msgaddr),
1310                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1311         }
1312
1313         kiblnd_free_pages(conn->ibc_rx_pages);
1314
1315         conn->ibc_rx_pages = NULL;
1316 }
1317
1318 void
1319 kiblnd_map_rx_descs(kib_conn_t *conn)
1320 {
1321         kib_rx_t       *rx;
1322         struct page    *pg;
1323         int             pg_off;
1324         int             ipg;
1325         int             i;
1326
1327         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1328                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1329                 rx = &conn->ibc_rxs[i];
1330
1331                 rx->rx_conn = conn;
1332                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1333
1334                 rx->rx_msgaddr =
1335                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1336                                               rx->rx_msg, IBLND_MSG_SIZE,
1337                                               DMA_FROM_DEVICE);
1338                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1339                                                   rx->rx_msgaddr));
1340                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1341
1342                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1343                        i, rx->rx_msg, rx->rx_msgaddr,
1344                        (__u64)(page_to_phys(pg) + pg_off));
1345
1346                 pg_off += IBLND_MSG_SIZE;
1347                 LASSERT(pg_off <= PAGE_SIZE);
1348
1349                 if (pg_off == PAGE_SIZE) {
1350                         pg_off = 0;
1351                         ipg++;
1352                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1353                 }
1354         }
1355 }
1356
1357 static void
1358 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1359 {
1360         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1361         kib_tx_t       *tx;
1362         int             i;
1363
1364         LASSERT (tpo->tpo_pool.po_allocated == 0);
1365
1366         if (hdev == NULL)
1367                 return;
1368
1369         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1370                 tx = &tpo->tpo_tx_descs[i];
1371                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1372                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1373                                                           tx->tx_msgaddr),
1374                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1375         }
1376
1377         kiblnd_hdev_decref(hdev);
1378         tpo->tpo_hdev = NULL;
1379 }
1380
1381 static kib_hca_dev_t *
1382 kiblnd_current_hdev(kib_dev_t *dev)
1383 {
1384         kib_hca_dev_t *hdev;
1385         unsigned long  flags;
1386         int            i = 0;
1387
1388         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1389         while (dev->ibd_failover) {
1390                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1391                 if (i++ % 50 == 0)
1392                         CDEBUG(D_NET, "%s: Wait for failover\n",
1393                                dev->ibd_ifname);
1394                 set_current_state(TASK_INTERRUPTIBLE);
1395                 schedule_timeout(cfs_time_seconds(1) / 100);
1396
1397                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1398         }
1399
1400         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1401         hdev = dev->ibd_hdev;
1402
1403         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1404
1405         return hdev;
1406 }
1407
1408 static void
1409 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1410 {
1411         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1412         kib_pool_t     *pool  = &tpo->tpo_pool;
1413         kib_net_t      *net   = pool->po_owner->ps_net;
1414         kib_dev_t      *dev;
1415         struct page    *page;
1416         kib_tx_t       *tx;
1417         int             page_offset;
1418         int             ipage;
1419         int             i;
1420
1421         LASSERT (net != NULL);
1422
1423         dev = net->ibn_dev;
1424
1425         /* pre-mapped messages are not bigger than 1 page */
1426         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1427
1428         /* No fancy arithmetic when we do the buffer calculations */
1429         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1430
1431         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1432
1433         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1434                 page = txpgs->ibp_pages[ipage];
1435                 tx = &tpo->tpo_tx_descs[i];
1436
1437                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1438                                            page_offset);
1439
1440                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1441                                                        tx->tx_msg,
1442                                                        IBLND_MSG_SIZE,
1443                                                        DMA_TO_DEVICE);
1444                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1445                                                   tx->tx_msgaddr));
1446                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1447
1448                 list_add(&tx->tx_list, &pool->po_free_list);
1449
1450                 page_offset += IBLND_MSG_SIZE;
1451                 LASSERT(page_offset <= PAGE_SIZE);
1452
1453                 if (page_offset == PAGE_SIZE) {
1454                         page_offset = 0;
1455                         ipage++;
1456                         LASSERT(ipage <= txpgs->ibp_npages);
1457                 }
1458         }
1459 }
1460
1461 static void
1462 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1463 {
1464         LASSERT(fpo->fpo_map_count == 0);
1465
1466         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1467                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1468         } else {
1469                 struct kib_fast_reg_descriptor *frd, *tmp;
1470                 int i = 0;
1471
1472                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1473                                          frd_list) {
1474                         list_del(&frd->frd_list);
1475 #ifndef HAVE_IB_MAP_MR_SG
1476                         ib_free_fast_reg_page_list(frd->frd_frpl);
1477 #endif
1478                         ib_dereg_mr(frd->frd_mr);
1479                         LIBCFS_FREE(frd, sizeof(*frd));
1480                         i++;
1481                 }
1482                 if (i < fpo->fast_reg.fpo_pool_size)
1483                         CERROR("FastReg pool still has %d regions registered\n",
1484                                 fpo->fast_reg.fpo_pool_size - i);
1485         }
1486
1487         if (fpo->fpo_hdev)
1488                 kiblnd_hdev_decref(fpo->fpo_hdev);
1489
1490         LIBCFS_FREE(fpo, sizeof(*fpo));
1491 }
1492
1493 static void
1494 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1495 {
1496         kib_fmr_pool_t *fpo, *tmp;
1497
1498         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1499                 list_del(&fpo->fpo_list);
1500                 kiblnd_destroy_fmr_pool(fpo);
1501         }
1502 }
1503
1504 static int
1505 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1506                      int ncpts)
1507 {
1508         int size = tunables->lnd_fmr_pool_size / ncpts;
1509
1510         return max(IBLND_FMR_POOL, size);
1511 }
1512
1513 static int
1514 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1515                          int ncpts)
1516 {
1517         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1518
1519         return max(IBLND_FMR_POOL_FLUSH, size);
1520 }
1521
1522 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1523 {
1524         struct ib_fmr_pool_param param = {
1525                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1526                 .page_shift        = PAGE_SHIFT,
1527                 .access            = (IB_ACCESS_LOCAL_WRITE |
1528                                       IB_ACCESS_REMOTE_WRITE),
1529                 .pool_size         = fps->fps_pool_size,
1530                 .dirty_watermark   = fps->fps_flush_trigger,
1531                 .flush_function    = NULL,
1532                 .flush_arg         = NULL,
1533                 .cache             = !!fps->fps_cache };
1534         int rc = 0;
1535
1536         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1537                                                    &param);
1538         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1539                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1540                 if (rc != -ENOSYS)
1541                         CERROR("Failed to create FMR pool: %d\n", rc);
1542                 else
1543                         CERROR("FMRs are not supported\n");
1544         }
1545         fpo->fpo_is_fmr = true;
1546
1547         return rc;
1548 }
1549
1550 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1551                                   __u32 dev_caps)
1552 {
1553         struct kib_fast_reg_descriptor *frd, *tmp;
1554         int i, rc;
1555
1556         fpo->fpo_is_fmr = false;
1557
1558         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1559         fpo->fast_reg.fpo_pool_size = 0;
1560         for (i = 0; i < fps->fps_pool_size; i++) {
1561                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1562                                  sizeof(*frd));
1563                 if (!frd) {
1564                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1565                         rc = -ENOMEM;
1566                         goto out;
1567                 }
1568                 frd->frd_mr = NULL;
1569
1570 #ifndef HAVE_IB_MAP_MR_SG
1571                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1572                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1573                 if (IS_ERR(frd->frd_frpl)) {
1574                         rc = PTR_ERR(frd->frd_frpl);
1575                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1576                                 rc);
1577                         frd->frd_frpl = NULL;
1578                         goto out_middle;
1579                 }
1580 #endif
1581
1582 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1583                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1584                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1585 #else
1586                 /*
1587                  * it is expected to get here if this is an MLX-5 card.
1588                  * MLX-4 cards will always use FMR and MLX-5 cards will
1589                  * always use fast_reg. It turns out that some MLX-5 cards
1590                  * (possibly due to older FW versions) do not natively support
1591                  * gaps. So we will need to track them here.
1592                  */
1593                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1594 #ifdef IB_MR_TYPE_SG_GAPS
1595                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1596                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1597                                                 IB_MR_TYPE_SG_GAPS :
1598                                                 IB_MR_TYPE_MEM_REG,
1599 #else
1600                                                 IB_MR_TYPE_MEM_REG,
1601 #endif
1602                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1603                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1604                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1605                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1606 #endif
1607                 if (IS_ERR(frd->frd_mr)) {
1608                         rc = PTR_ERR(frd->frd_mr);
1609                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1610                         frd->frd_mr = NULL;
1611                         goto out_middle;
1612                 }
1613
1614                 /* There appears to be a bug in MLX5 code where you must
1615                  * invalidate the rkey of a new FastReg pool before first
1616                  * using it. Thus, I am marking the FRD invalid here. */
1617                 frd->frd_valid = false;
1618
1619                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1620                 fpo->fast_reg.fpo_pool_size++;
1621         }
1622
1623         return 0;
1624
1625 out_middle:
1626         if (frd->frd_mr)
1627                 ib_dereg_mr(frd->frd_mr);
1628 #ifndef HAVE_IB_MAP_MR_SG
1629         if (frd->frd_frpl)
1630                 ib_free_fast_reg_page_list(frd->frd_frpl);
1631 #endif
1632         LIBCFS_FREE(frd, sizeof(*frd));
1633
1634 out:
1635         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1636                                  frd_list) {
1637                 list_del(&frd->frd_list);
1638 #ifndef HAVE_IB_MAP_MR_SG
1639                 ib_free_fast_reg_page_list(frd->frd_frpl);
1640 #endif
1641                 ib_dereg_mr(frd->frd_mr);
1642                 LIBCFS_FREE(frd, sizeof(*frd));
1643         }
1644
1645         return rc;
1646 }
1647
1648 static int
1649 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1650 {
1651         kib_dev_t *dev = fps->fps_net->ibn_dev;
1652         kib_fmr_pool_t *fpo;
1653         int rc;
1654
1655         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1656         if (!fpo) {
1657                 return -ENOMEM;
1658         }
1659         memset(fpo, 0, sizeof(*fpo));
1660
1661         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1662
1663         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1664                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1665         else
1666                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1667         if (rc)
1668                 goto out_fpo;
1669
1670         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1671         fpo->fpo_owner    = fps;
1672         *pp_fpo = fpo;
1673
1674         return 0;
1675
1676 out_fpo:
1677         kiblnd_hdev_decref(fpo->fpo_hdev);
1678         LIBCFS_FREE(fpo, sizeof(*fpo));
1679         return rc;
1680 }
1681
1682 static void
1683 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1684 {
1685         if (fps->fps_net == NULL) /* intialized? */
1686                 return;
1687
1688         spin_lock(&fps->fps_lock);
1689
1690         while (!list_empty(&fps->fps_pool_list)) {
1691                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1692                                                  kib_fmr_pool_t, fpo_list);
1693                 fpo->fpo_failed = 1;
1694                 list_del(&fpo->fpo_list);
1695                 if (fpo->fpo_map_count == 0)
1696                         list_add(&fpo->fpo_list, zombies);
1697                 else
1698                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1699         }
1700
1701         spin_unlock(&fps->fps_lock);
1702 }
1703
1704 static void
1705 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1706 {
1707         if (fps->fps_net != NULL) { /* initialized? */
1708                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1709                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1710         }
1711 }
1712
1713 static int
1714 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1715                         kib_net_t *net,
1716                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1717 {
1718         kib_fmr_pool_t *fpo;
1719         int             rc;
1720
1721         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1722
1723         fps->fps_net = net;
1724         fps->fps_cpt = cpt;
1725
1726         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1727         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1728         fps->fps_cache = tunables->lnd_fmr_cache;
1729
1730         spin_lock_init(&fps->fps_lock);
1731         INIT_LIST_HEAD(&fps->fps_pool_list);
1732         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1733
1734         rc = kiblnd_create_fmr_pool(fps, &fpo);
1735         if (rc == 0)
1736                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1737
1738         return rc;
1739 }
1740
1741 static int
1742 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1743 {
1744         if (fpo->fpo_map_count != 0) /* still in use */
1745                 return 0;
1746         if (fpo->fpo_failed)
1747                 return 1;
1748         return cfs_time_aftereq(now, fpo->fpo_deadline);
1749 }
1750
1751 static int
1752 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1753 {
1754         kib_hca_dev_t   *hdev;
1755         __u64           *pages = tx->tx_pages;
1756         int             npages;
1757         int             size;
1758         int             i;
1759
1760         hdev = tx->tx_pool->tpo_hdev;
1761
1762         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1763                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1764                         size += hdev->ibh_page_size) {
1765                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1766                                            hdev->ibh_page_mask) + size;
1767                 }
1768         }
1769
1770         return npages;
1771 }
1772
1773 void
1774 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1775 {
1776         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1777         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1778         kib_fmr_poolset_t *fps;
1779         cfs_time_t         now = cfs_time_current();
1780         kib_fmr_pool_t    *tmp;
1781         int                rc;
1782
1783         if (!fpo)
1784                 return;
1785
1786         fps = fpo->fpo_owner;
1787         if (fpo->fpo_is_fmr) {
1788                 if (fmr->fmr_pfmr) {
1789                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1790                         LASSERT(!rc);
1791                         fmr->fmr_pfmr = NULL;
1792                 }
1793
1794                 if (status) {
1795                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1796                         LASSERT(!rc);
1797                 }
1798         } else {
1799                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1800
1801                 if (frd) {
1802                         frd->frd_valid = false;
1803                         spin_lock(&fps->fps_lock);
1804                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1805                         spin_unlock(&fps->fps_lock);
1806                         fmr->fmr_frd = NULL;
1807                 }
1808         }
1809         fmr->fmr_pool = NULL;
1810
1811         spin_lock(&fps->fps_lock);
1812         fpo->fpo_map_count--;   /* decref the pool */
1813
1814         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1815                 /* the first pool is persistent */
1816                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1817                         continue;
1818
1819                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1820                         list_move(&fpo->fpo_list, &zombies);
1821                         fps->fps_version++;
1822                 }
1823         }
1824         spin_unlock(&fps->fps_lock);
1825
1826         if (!list_empty(&zombies))
1827                 kiblnd_destroy_fmr_pool_list(&zombies);
1828 }
1829
1830 int
1831 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1832                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1833 {
1834         kib_fmr_pool_t *fpo;
1835         __u64 *pages = tx->tx_pages;
1836         __u64 version;
1837         bool is_rx = (rd != tx->tx_rd);
1838         bool tx_pages_mapped = 0;
1839         int npages = 0;
1840         int rc;
1841
1842 again:
1843         spin_lock(&fps->fps_lock);
1844         version = fps->fps_version;
1845         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1846                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1847                 fpo->fpo_map_count++;
1848
1849                 if (fpo->fpo_is_fmr) {
1850                         struct ib_pool_fmr *pfmr;
1851
1852                         spin_unlock(&fps->fps_lock);
1853
1854                         if (!tx_pages_mapped) {
1855                                 npages = kiblnd_map_tx_pages(tx, rd);
1856                                 tx_pages_mapped = 1;
1857                         }
1858
1859                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1860                                                     pages, npages, iov);
1861                         if (likely(!IS_ERR(pfmr))) {
1862                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1863                                                       : pfmr->fmr->lkey;
1864                                 fmr->fmr_frd  = NULL;
1865                                 fmr->fmr_pfmr = pfmr;
1866                                 fmr->fmr_pool = fpo;
1867                                 return 0;
1868                         }
1869                         rc = PTR_ERR(pfmr);
1870                 } else {
1871                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1872                                 struct kib_fast_reg_descriptor *frd;
1873 #ifdef HAVE_IB_MAP_MR_SG
1874                                 struct ib_reg_wr *wr;
1875                                 int n;
1876 #else
1877                                 struct ib_rdma_wr *wr;
1878                                 struct ib_fast_reg_page_list *frpl;
1879 #endif
1880                                 struct ib_mr *mr;
1881
1882                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1883                                                         struct kib_fast_reg_descriptor,
1884                                                         frd_list);
1885                                 list_del(&frd->frd_list);
1886                                 spin_unlock(&fps->fps_lock);
1887
1888 #ifndef HAVE_IB_MAP_MR_SG
1889                                 frpl = frd->frd_frpl;
1890 #endif
1891                                 mr   = frd->frd_mr;
1892
1893                                 if (!frd->frd_valid) {
1894                                         struct ib_rdma_wr *inv_wr;
1895                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1896
1897                                         inv_wr = &frd->frd_inv_wr;
1898                                         memset(inv_wr, 0, sizeof(*inv_wr));
1899
1900                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1901                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1902                                         inv_wr->wr.ex.invalidate_rkey = key;
1903
1904                                         /* Bump the key */
1905                                         key = ib_inc_rkey(key);
1906                                         ib_update_fast_reg_key(mr, key);
1907                                 }
1908
1909 #ifdef HAVE_IB_MAP_MR_SG
1910 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1911                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1912                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1913 #else
1914                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1915                                                  tx->tx_nfrags, PAGE_SIZE);
1916 #endif
1917                                 if (unlikely(n != tx->tx_nfrags)) {
1918                                         CERROR("Failed to map mr %d/%d "
1919                                                "elements\n", n, tx->tx_nfrags);
1920                                         return n < 0 ? n : -EINVAL;
1921                                 }
1922
1923                                 wr = &frd->frd_fastreg_wr;
1924                                 memset(wr, 0, sizeof(*wr));
1925
1926                                 wr->wr.opcode = IB_WR_REG_MR;
1927                                 wr->wr.wr_id  = IBLND_WID_MR;
1928                                 wr->wr.num_sge = 0;
1929                                 wr->wr.send_flags = 0;
1930                                 wr->mr = mr;
1931                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1932                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1933                                               IB_ACCESS_REMOTE_WRITE);
1934 #else
1935                                 if (!tx_pages_mapped) {
1936                                         npages = kiblnd_map_tx_pages(tx, rd);
1937                                         tx_pages_mapped = 1;
1938                                 }
1939
1940                                 LASSERT(npages <= frpl->max_page_list_len);
1941                                 memcpy(frpl->page_list, pages,
1942                                         sizeof(*pages) * npages);
1943
1944                                 /* Prepare FastReg WR */
1945                                 wr = &frd->frd_fastreg_wr;
1946                                 memset(wr, 0, sizeof(*wr));
1947
1948                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1949                                 wr->wr.wr_id  = IBLND_WID_MR;
1950
1951                                 wr->wr.wr.fast_reg.iova_start = iov;
1952                                 wr->wr.wr.fast_reg.page_list  = frpl;
1953                                 wr->wr.wr.fast_reg.page_list_len = npages;
1954                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1955                                 wr->wr.wr.fast_reg.length = nob;
1956                                 wr->wr.wr.fast_reg.rkey =
1957                                                 is_rx ? mr->rkey : mr->lkey;
1958                                 wr->wr.wr.fast_reg.access_flags =
1959                                                 (IB_ACCESS_LOCAL_WRITE |
1960                                                  IB_ACCESS_REMOTE_WRITE);
1961 #endif
1962
1963                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1964                                 fmr->fmr_frd  = frd;
1965                                 fmr->fmr_pfmr = NULL;
1966                                 fmr->fmr_pool = fpo;
1967                                 return 0;
1968                         }
1969                         spin_unlock(&fps->fps_lock);
1970                         rc = -EAGAIN;
1971                 }
1972
1973                 spin_lock(&fps->fps_lock);
1974                 fpo->fpo_map_count--;
1975                 if (rc != -EAGAIN) {
1976                         spin_unlock(&fps->fps_lock);
1977                         return rc;
1978                 }
1979
1980                 /* EAGAIN and ... */
1981                 if (version != fps->fps_version) {
1982                         spin_unlock(&fps->fps_lock);
1983                         goto again;
1984                 }
1985         }
1986
1987         if (fps->fps_increasing) {
1988                 spin_unlock(&fps->fps_lock);
1989                 CDEBUG(D_NET, "Another thread is allocating new "
1990                        "FMR pool, waiting for her to complete\n");
1991                 schedule();
1992                 goto again;
1993
1994         }
1995
1996         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
1997                 /* someone failed recently */
1998                 spin_unlock(&fps->fps_lock);
1999                 return -EAGAIN;
2000         }
2001
2002         fps->fps_increasing = 1;
2003         spin_unlock(&fps->fps_lock);
2004
2005         CDEBUG(D_NET, "Allocate new FMR pool\n");
2006         rc = kiblnd_create_fmr_pool(fps, &fpo);
2007         spin_lock(&fps->fps_lock);
2008         fps->fps_increasing = 0;
2009         if (rc == 0) {
2010                 fps->fps_version++;
2011                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2012         } else {
2013                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2014         }
2015         spin_unlock(&fps->fps_lock);
2016
2017         goto again;
2018 }
2019
2020 static void
2021 kiblnd_fini_pool(kib_pool_t *pool)
2022 {
2023         LASSERT(list_empty(&pool->po_free_list));
2024         LASSERT(pool->po_allocated == 0);
2025
2026         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2027 }
2028
2029 static void
2030 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2031 {
2032         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2033
2034         memset(pool, 0, sizeof(kib_pool_t));
2035         INIT_LIST_HEAD(&pool->po_free_list);
2036         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2037         pool->po_owner    = ps;
2038         pool->po_size     = size;
2039 }
2040
2041 static void
2042 kiblnd_destroy_pool_list(struct list_head *head)
2043 {
2044         kib_pool_t *pool;
2045
2046         while (!list_empty(head)) {
2047                 pool = list_entry(head->next, kib_pool_t, po_list);
2048                 list_del(&pool->po_list);
2049
2050                 LASSERT(pool->po_owner != NULL);
2051                 pool->po_owner->ps_pool_destroy(pool);
2052         }
2053 }
2054
2055 static void
2056 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2057 {
2058         if (ps->ps_net == NULL) /* intialized? */
2059                 return;
2060
2061         spin_lock(&ps->ps_lock);
2062         while (!list_empty(&ps->ps_pool_list)) {
2063                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2064                                             kib_pool_t, po_list);
2065                 po->po_failed = 1;
2066                 list_del(&po->po_list);
2067                 if (po->po_allocated == 0)
2068                         list_add(&po->po_list, zombies);
2069                 else
2070                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2071         }
2072         spin_unlock(&ps->ps_lock);
2073 }
2074
2075 static void
2076 kiblnd_fini_poolset(kib_poolset_t *ps)
2077 {
2078         if (ps->ps_net != NULL) { /* initialized? */
2079                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2080                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2081         }
2082 }
2083
2084 static int
2085 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2086                     kib_net_t *net, char *name, int size,
2087                     kib_ps_pool_create_t po_create,
2088                     kib_ps_pool_destroy_t po_destroy,
2089                     kib_ps_node_init_t nd_init,
2090                     kib_ps_node_fini_t nd_fini)
2091 {
2092         kib_pool_t      *pool;
2093         int             rc;
2094
2095         memset(ps, 0, sizeof(kib_poolset_t));
2096
2097         ps->ps_cpt          = cpt;
2098         ps->ps_net          = net;
2099         ps->ps_pool_create  = po_create;
2100         ps->ps_pool_destroy = po_destroy;
2101         ps->ps_node_init    = nd_init;
2102         ps->ps_node_fini    = nd_fini;
2103         ps->ps_pool_size    = size;
2104         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2105             >= sizeof(ps->ps_name))
2106                 return -E2BIG;
2107         spin_lock_init(&ps->ps_lock);
2108         INIT_LIST_HEAD(&ps->ps_pool_list);
2109         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2110
2111         rc = ps->ps_pool_create(ps, size, &pool);
2112         if (rc == 0)
2113                 list_add(&pool->po_list, &ps->ps_pool_list);
2114         else
2115                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2116
2117         return rc;
2118 }
2119
2120 static int
2121 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2122 {
2123         if (pool->po_allocated != 0) /* still in use */
2124                 return 0;
2125         if (pool->po_failed)
2126                 return 1;
2127         return cfs_time_aftereq(now, pool->po_deadline);
2128 }
2129
2130 void
2131 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2132 {
2133         struct list_head zombies = LIST_HEAD_INIT(zombies);
2134         kib_poolset_t   *ps = pool->po_owner;
2135         kib_pool_t      *tmp;
2136         cfs_time_t       now = cfs_time_current();
2137
2138         spin_lock(&ps->ps_lock);
2139
2140         if (ps->ps_node_fini != NULL)
2141                 ps->ps_node_fini(pool, node);
2142
2143         LASSERT(pool->po_allocated > 0);
2144         list_add(node, &pool->po_free_list);
2145         pool->po_allocated--;
2146
2147         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2148                 /* the first pool is persistent */
2149                 if (ps->ps_pool_list.next == &pool->po_list)
2150                         continue;
2151
2152                 if (kiblnd_pool_is_idle(pool, now))
2153                         list_move(&pool->po_list, &zombies);
2154         }
2155         spin_unlock(&ps->ps_lock);
2156
2157         if (!list_empty(&zombies))
2158                 kiblnd_destroy_pool_list(&zombies);
2159 }
2160
2161 struct list_head *
2162 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2163 {
2164         struct list_head        *node;
2165         kib_pool_t              *pool;
2166         int                     rc;
2167         unsigned int            interval = 1;
2168         cfs_time_t              time_before;
2169         unsigned int            trips = 0;
2170
2171 again:
2172         spin_lock(&ps->ps_lock);
2173         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2174                 if (list_empty(&pool->po_free_list))
2175                         continue;
2176
2177                 pool->po_allocated++;
2178                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2179                 node = pool->po_free_list.next;
2180                 list_del(node);
2181
2182                 if (ps->ps_node_init != NULL) {
2183                         /* still hold the lock */
2184                         ps->ps_node_init(pool, node);
2185                 }
2186                 spin_unlock(&ps->ps_lock);
2187                 return node;
2188         }
2189
2190         /* no available tx pool and ... */
2191         if (ps->ps_increasing) {
2192                 /* another thread is allocating a new pool */
2193                 spin_unlock(&ps->ps_lock);
2194                 trips++;
2195                 CDEBUG(D_NET, "Another thread is allocating new "
2196                        "%s pool, waiting %d HZs for her to complete."
2197                        "trips = %d\n",
2198                        ps->ps_name, interval, trips);
2199
2200                 set_current_state(TASK_INTERRUPTIBLE);
2201                 schedule_timeout(interval);
2202                 if (interval < cfs_time_seconds(1))
2203                         interval *= 2;
2204
2205                 goto again;
2206         }
2207
2208         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2209                 /* someone failed recently */
2210                 spin_unlock(&ps->ps_lock);
2211                 return NULL;
2212         }
2213
2214         ps->ps_increasing = 1;
2215         spin_unlock(&ps->ps_lock);
2216
2217         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2218         time_before = cfs_time_current();
2219         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2220         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2221                cfs_time_current() - time_before);
2222
2223         spin_lock(&ps->ps_lock);
2224         ps->ps_increasing = 0;
2225         if (rc == 0) {
2226                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2227         } else {
2228                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2229                 CERROR("Can't allocate new %s pool because out of memory\n",
2230                        ps->ps_name);
2231         }
2232         spin_unlock(&ps->ps_lock);
2233
2234         goto again;
2235 }
2236
2237 static void
2238 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2239 {
2240         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2241         int             i;
2242
2243         LASSERT (pool->po_allocated == 0);
2244
2245         if (tpo->tpo_tx_pages != NULL) {
2246                 kiblnd_unmap_tx_pool(tpo);
2247                 kiblnd_free_pages(tpo->tpo_tx_pages);
2248         }
2249
2250         if (tpo->tpo_tx_descs == NULL)
2251                 goto out;
2252
2253         for (i = 0; i < pool->po_size; i++) {
2254                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2255                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2256
2257                 list_del(&tx->tx_list);
2258                 if (tx->tx_pages != NULL)
2259                         LIBCFS_FREE(tx->tx_pages,
2260                                     LNET_MAX_IOV *
2261                                     sizeof(*tx->tx_pages));
2262                 if (tx->tx_frags != NULL)
2263                         LIBCFS_FREE(tx->tx_frags,
2264                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2265                                     sizeof(*tx->tx_frags));
2266                 if (tx->tx_wrq != NULL)
2267                         LIBCFS_FREE(tx->tx_wrq,
2268                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2269                                     sizeof(*tx->tx_wrq));
2270                 if (tx->tx_sge != NULL)
2271                         LIBCFS_FREE(tx->tx_sge,
2272                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2273                                     sizeof(*tx->tx_sge));
2274                 if (tx->tx_rd != NULL)
2275                         LIBCFS_FREE(tx->tx_rd,
2276                                     offsetof(kib_rdma_desc_t,
2277                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2278         }
2279
2280         LIBCFS_FREE(tpo->tpo_tx_descs,
2281                     pool->po_size * sizeof(kib_tx_t));
2282 out:
2283         kiblnd_fini_pool(pool);
2284         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2285 }
2286
2287 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2288 {
2289         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2290         int ntx;
2291
2292         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2293         ntx = tunables->lnd_ntx / ncpts;
2294
2295         return max(IBLND_TX_POOL, ntx);
2296 }
2297
2298 static int
2299 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2300 {
2301         int            i;
2302         int            npg;
2303         kib_pool_t    *pool;
2304         kib_tx_pool_t *tpo;
2305
2306         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2307         if (tpo == NULL) {
2308                 CERROR("Failed to allocate TX pool\n");
2309                 return -ENOMEM;
2310         }
2311
2312         pool = &tpo->tpo_pool;
2313         kiblnd_init_pool(ps, pool, size);
2314         tpo->tpo_tx_descs = NULL;
2315         tpo->tpo_tx_pages = NULL;
2316
2317         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2318         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2319                 CERROR("Can't allocate tx pages: %d\n", npg);
2320                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2321                 return -ENOMEM;
2322         }
2323
2324         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2325                          size * sizeof(kib_tx_t));
2326         if (tpo->tpo_tx_descs == NULL) {
2327                 CERROR("Can't allocate %d tx descriptors\n", size);
2328                 ps->ps_pool_destroy(pool);
2329                 return -ENOMEM;
2330         }
2331
2332         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2333
2334         for (i = 0; i < size; i++) {
2335                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2336                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2337
2338                 tx->tx_pool = tpo;
2339                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2340                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2341                                          lnet_cpt_table(), ps->ps_cpt,
2342                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2343                         if (tx->tx_pages == NULL)
2344                                 break;
2345                 }
2346
2347                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2348                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2349                                  sizeof(*tx->tx_frags));
2350                 if (tx->tx_frags == NULL)
2351                         break;
2352
2353                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2354
2355                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2356                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2357                                  sizeof(*tx->tx_wrq));
2358                 if (tx->tx_wrq == NULL)
2359                         break;
2360
2361                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2362                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2363                                  sizeof(*tx->tx_sge));
2364                 if (tx->tx_sge == NULL)
2365                         break;
2366
2367                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2368                                  offsetof(kib_rdma_desc_t,
2369                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2370                 if (tx->tx_rd == NULL)
2371                         break;
2372         }
2373
2374         if (i == size) {
2375                 kiblnd_map_tx_pool(tpo);
2376                 *pp_po = pool;
2377                 return 0;
2378         }
2379
2380         ps->ps_pool_destroy(pool);
2381         return -ENOMEM;
2382 }
2383
2384 static void
2385 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2386 {
2387         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2388                                              tps_poolset);
2389         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2390
2391         tx->tx_cookie = tps->tps_next_tx_cookie++;
2392 }
2393
2394 static void
2395 kiblnd_net_fini_pools(kib_net_t *net)
2396 {
2397         int     i;
2398
2399         cfs_cpt_for_each(i, lnet_cpt_table()) {
2400                 kib_tx_poolset_t        *tps;
2401                 kib_fmr_poolset_t       *fps;
2402
2403                 if (net->ibn_tx_ps != NULL) {
2404                         tps = net->ibn_tx_ps[i];
2405                         kiblnd_fini_poolset(&tps->tps_poolset);
2406                 }
2407
2408                 if (net->ibn_fmr_ps != NULL) {
2409                         fps = net->ibn_fmr_ps[i];
2410                         kiblnd_fini_fmr_poolset(fps);
2411                 }
2412         }
2413
2414         if (net->ibn_tx_ps != NULL) {
2415                 cfs_percpt_free(net->ibn_tx_ps);
2416                 net->ibn_tx_ps = NULL;
2417         }
2418
2419         if (net->ibn_fmr_ps != NULL) {
2420                 cfs_percpt_free(net->ibn_fmr_ps);
2421                 net->ibn_fmr_ps = NULL;
2422         }
2423 }
2424
2425 static int
2426 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2427                       int ncpts)
2428 {
2429         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2430 #ifdef HAVE_IB_GET_DMA_MR
2431         unsigned long   flags;
2432 #endif
2433         int             cpt;
2434         int             rc;
2435         int             i;
2436
2437         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2438
2439 #ifdef HAVE_IB_GET_DMA_MR
2440         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2441         /*
2442          * if lnd_map_on_demand is zero then we have effectively disabled
2443          * FMR or FastReg and we're using global memory regions
2444          * exclusively.
2445          */
2446         if (!tunables->lnd_map_on_demand) {
2447                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2448                                            flags);
2449                 goto create_tx_pool;
2450         }
2451
2452         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2453 #endif
2454
2455         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2456                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2457                        tunables->lnd_fmr_pool_size,
2458                        tunables->lnd_ntx / 4);
2459                 rc = -EINVAL;
2460                 goto failed;
2461         }
2462
2463         /* TX pool must be created later than FMR, see LU-2268
2464          * for details */
2465         LASSERT(net->ibn_tx_ps == NULL);
2466
2467         /* premapping can fail if ibd_nmr > 1, so we always create
2468          * FMR pool and map-on-demand if premapping failed */
2469
2470         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2471                                            sizeof(kib_fmr_poolset_t));
2472         if (net->ibn_fmr_ps == NULL) {
2473                 CERROR("Failed to allocate FMR pool array\n");
2474                 rc = -ENOMEM;
2475                 goto failed;
2476         }
2477
2478         for (i = 0; i < ncpts; i++) {
2479                 cpt = (cpts == NULL) ? i : cpts[i];
2480                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2481                                              net, tunables);
2482                 if (rc != 0) {
2483                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2484                                cpt, rc);
2485                         goto failed;
2486                 }
2487         }
2488
2489         if (i > 0)
2490                 LASSERT(i == ncpts);
2491
2492 #ifdef HAVE_IB_GET_DMA_MR
2493  create_tx_pool:
2494 #endif
2495         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2496                                           sizeof(kib_tx_poolset_t));
2497         if (net->ibn_tx_ps == NULL) {
2498                 CERROR("Failed to allocate tx pool array\n");
2499                 rc = -ENOMEM;
2500                 goto failed;
2501         }
2502
2503         for (i = 0; i < ncpts; i++) {
2504                 cpt = (cpts == NULL) ? i : cpts[i];
2505                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2506                                          cpt, net, "TX",
2507                                          kiblnd_tx_pool_size(ni, ncpts),
2508                                          kiblnd_create_tx_pool,
2509                                          kiblnd_destroy_tx_pool,
2510                                          kiblnd_tx_init, NULL);
2511                 if (rc != 0) {
2512                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2513                                cpt, rc);
2514                         goto failed;
2515                 }
2516         }
2517
2518         return 0;
2519  failed:
2520         kiblnd_net_fini_pools(net);
2521         LASSERT(rc != 0);
2522         return rc;
2523 }
2524
2525 static int
2526 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2527 {
2528         struct ib_device_attr *dev_attr;
2529         int rc = 0;
2530
2531         /* It's safe to assume a HCA can handle a page size
2532          * matching that of the native system */
2533         hdev->ibh_page_shift = PAGE_SHIFT;
2534         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2535         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2536
2537 #ifndef HAVE_IB_DEVICE_ATTRS
2538         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2539         if (dev_attr == NULL) {
2540                 CERROR("Out of memory\n");
2541                 return -ENOMEM;
2542         }
2543
2544         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2545         if (rc != 0) {
2546                 CERROR("Failed to query IB device: %d\n", rc);
2547                 goto out_clean_attr;
2548         }
2549 #else
2550         dev_attr = &hdev->ibh_ibdev->attrs;
2551 #endif
2552
2553         hdev->ibh_mr_size = dev_attr->max_mr_size;
2554
2555         /* Setup device Memory Registration capabilities */
2556         if (hdev->ibh_ibdev->alloc_fmr &&
2557             hdev->ibh_ibdev->dealloc_fmr &&
2558             hdev->ibh_ibdev->map_phys_fmr &&
2559             hdev->ibh_ibdev->unmap_fmr) {
2560                 LCONSOLE_INFO("Using FMR for registration\n");
2561                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2562         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2563                 LCONSOLE_INFO("Using FastReg for registration\n");
2564                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2565 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2566 #ifdef IB_DEVICE_SG_GAPS_REG
2567                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2568                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2569 #endif
2570 #endif
2571         } else {
2572                 rc = -ENOSYS;
2573         }
2574
2575         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2576                 hdev->ibh_mr_shift = 64;
2577         else if (rc != 0)
2578                 rc = -EINVAL;
2579
2580 #ifndef HAVE_IB_DEVICE_ATTRS
2581 out_clean_attr:
2582         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2583 #endif
2584
2585         if (rc == -ENOSYS)
2586                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2587                        "register memory: %d\n", rc);
2588         else if (rc == -EINVAL)
2589                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2590         return rc;
2591 }
2592
2593 #ifdef HAVE_IB_GET_DMA_MR
2594 static void
2595 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2596 {
2597         if (hdev->ibh_mrs == NULL)
2598                 return;
2599
2600         ib_dereg_mr(hdev->ibh_mrs);
2601
2602         hdev->ibh_mrs = NULL;
2603 }
2604 #endif
2605
2606 void
2607 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2608 {
2609 #ifdef HAVE_IB_GET_DMA_MR
2610         kiblnd_hdev_cleanup_mrs(hdev);
2611 #endif
2612
2613         if (hdev->ibh_pd != NULL)
2614                 ib_dealloc_pd(hdev->ibh_pd);
2615
2616         if (hdev->ibh_cmid != NULL)
2617                 rdma_destroy_id(hdev->ibh_cmid);
2618
2619         LIBCFS_FREE(hdev, sizeof(*hdev));
2620 }
2621
2622 #ifdef HAVE_IB_GET_DMA_MR
2623 static int
2624 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2625 {
2626         struct ib_mr *mr;
2627         int           acflags = IB_ACCESS_LOCAL_WRITE |
2628                                 IB_ACCESS_REMOTE_WRITE;
2629
2630         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2631         if (IS_ERR(mr)) {
2632                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2633                 kiblnd_hdev_cleanup_mrs(hdev);
2634                 return PTR_ERR(mr);
2635         }
2636
2637         hdev->ibh_mrs = mr;
2638
2639         return 0;
2640 }
2641 #endif
2642
2643 static int
2644 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2645 {       /* DUMMY */
2646         return 0;
2647 }
2648
2649 static int
2650 kiblnd_dev_need_failover(kib_dev_t *dev)
2651 {
2652         struct rdma_cm_id  *cmid;
2653         struct sockaddr_in  srcaddr;
2654         struct sockaddr_in  dstaddr;
2655         int                 rc;
2656
2657         if (dev->ibd_hdev == NULL || /* initializing */
2658             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2659             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2660                 return 1;
2661
2662         /* XXX: it's UGLY, but I don't have better way to find
2663          * ib-bonding HCA failover because:
2664          *
2665          * a. no reliable CM event for HCA failover...
2666          * b. no OFED API to get ib_device for current net_device...
2667          *
2668          * We have only two choices at this point:
2669          *
2670          * a. rdma_bind_addr(), it will conflict with listener cmid
2671          * b. rdma_resolve_addr() to zero addr */
2672         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2673                                      IB_QPT_RC);
2674         if (IS_ERR(cmid)) {
2675                 rc = PTR_ERR(cmid);
2676                 CERROR("Failed to create cmid for failover: %d\n", rc);
2677                 return rc;
2678         }
2679
2680         memset(&srcaddr, 0, sizeof(srcaddr));
2681         srcaddr.sin_family      = AF_INET;
2682         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2683
2684         memset(&dstaddr, 0, sizeof(dstaddr));
2685         dstaddr.sin_family = AF_INET;
2686         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2687                                (struct sockaddr *)&dstaddr, 1);
2688         if (rc != 0 || cmid->device == NULL) {
2689                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2690                        dev->ibd_ifname, &dev->ibd_ifip,
2691                        cmid->device, rc);
2692                 rdma_destroy_id(cmid);
2693                 return rc;
2694         }
2695
2696         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2697         rdma_destroy_id(cmid);
2698         return rc;
2699 }
2700
2701 int
2702 kiblnd_dev_failover(kib_dev_t *dev)
2703 {
2704         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2705         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2706         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2707         struct rdma_cm_id  *cmid  = NULL;
2708         kib_hca_dev_t      *hdev  = NULL;
2709         kib_hca_dev_t      *old;
2710         struct ib_pd       *pd;
2711         kib_net_t          *net;
2712         struct sockaddr_in  addr;
2713         unsigned long       flags;
2714         int                 rc = 0;
2715         int                 i;
2716
2717         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2718                  dev->ibd_can_failover ||
2719                  dev->ibd_hdev == NULL);
2720
2721         rc = kiblnd_dev_need_failover(dev);
2722         if (rc <= 0)
2723                 goto out;
2724
2725         if (dev->ibd_hdev != NULL &&
2726             dev->ibd_hdev->ibh_cmid != NULL) {
2727                 /* XXX it's not good to close old listener at here,
2728                  * because we can fail to create new listener.
2729                  * But we have to close it now, otherwise rdma_bind_addr
2730                  * will return EADDRINUSE... How crap! */
2731                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2732
2733                 cmid = dev->ibd_hdev->ibh_cmid;
2734                 /* make next schedule of kiblnd_dev_need_failover()
2735                  * return 1 for me */
2736                 dev->ibd_hdev->ibh_cmid  = NULL;
2737                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2738
2739                 rdma_destroy_id(cmid);
2740         }
2741
2742         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2743                                      IB_QPT_RC);
2744         if (IS_ERR(cmid)) {
2745                 rc = PTR_ERR(cmid);
2746                 CERROR("Failed to create cmid for failover: %d\n", rc);
2747                 goto out;
2748         }
2749
2750         memset(&addr, 0, sizeof(addr));
2751         addr.sin_family      = AF_INET;
2752         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2753         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2754
2755         /* Bind to failover device or port */
2756         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2757         if (rc != 0 || cmid->device == NULL) {
2758                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2759                        dev->ibd_ifname, &dev->ibd_ifip,
2760                        cmid->device, rc);
2761                 rdma_destroy_id(cmid);
2762                 goto out;
2763         }
2764
2765         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2766         if (hdev == NULL) {
2767                 CERROR("Failed to allocate kib_hca_dev\n");
2768                 rdma_destroy_id(cmid);
2769                 rc = -ENOMEM;
2770                 goto out;
2771         }
2772
2773         atomic_set(&hdev->ibh_ref, 1);
2774         hdev->ibh_dev   = dev;
2775         hdev->ibh_cmid  = cmid;
2776         hdev->ibh_ibdev = cmid->device;
2777
2778 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2779         pd = ib_alloc_pd(cmid->device, 0);
2780 #else
2781         pd = ib_alloc_pd(cmid->device);
2782 #endif
2783         if (IS_ERR(pd)) {
2784                 rc = PTR_ERR(pd);
2785                 CERROR("Can't allocate PD: %d\n", rc);
2786                 goto out;
2787         }
2788
2789         hdev->ibh_pd = pd;
2790
2791         rc = rdma_listen(cmid, 0);
2792         if (rc != 0) {
2793                 CERROR("Can't start new listener: %d\n", rc);
2794                 goto out;
2795         }
2796
2797         rc = kiblnd_hdev_get_attr(hdev);
2798         if (rc != 0) {
2799                 CERROR("Can't get device attributes: %d\n", rc);
2800                 goto out;
2801         }
2802
2803 #ifdef HAVE_IB_GET_DMA_MR
2804         rc = kiblnd_hdev_setup_mrs(hdev);
2805         if (rc != 0) {
2806                 CERROR("Can't setup device: %d\n", rc);
2807                 goto out;
2808         }
2809 #endif
2810
2811         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2812
2813         old = dev->ibd_hdev;
2814         dev->ibd_hdev = hdev;   /* take over the refcount */
2815         hdev = old;
2816
2817         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2818                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2819                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2820                                             &zombie_tpo);
2821
2822                         if (net->ibn_fmr_ps != NULL)
2823                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2824                                                         &zombie_fpo);
2825                 }
2826         }
2827
2828         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2829  out:
2830         if (!list_empty(&zombie_tpo))
2831                 kiblnd_destroy_pool_list(&zombie_tpo);
2832         if (!list_empty(&zombie_ppo))
2833                 kiblnd_destroy_pool_list(&zombie_ppo);
2834         if (!list_empty(&zombie_fpo))
2835                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2836         if (hdev != NULL)
2837                 kiblnd_hdev_decref(hdev);
2838
2839         if (rc != 0)
2840                 dev->ibd_failed_failover++;
2841         else
2842                 dev->ibd_failed_failover = 0;
2843
2844         return rc;
2845 }
2846
2847 void
2848 kiblnd_destroy_dev (kib_dev_t *dev)
2849 {
2850         LASSERT (dev->ibd_nnets == 0);
2851         LASSERT(list_empty(&dev->ibd_nets));
2852
2853         list_del(&dev->ibd_fail_list);
2854         list_del(&dev->ibd_list);
2855
2856         if (dev->ibd_hdev != NULL)
2857                 kiblnd_hdev_decref(dev->ibd_hdev);
2858
2859         LIBCFS_FREE(dev, sizeof(*dev));
2860 }
2861
2862 static kib_dev_t *
2863 kiblnd_create_dev(char *ifname)
2864 {
2865         struct net_device *netdev;
2866         kib_dev_t         *dev;
2867         __u32              netmask;
2868         __u32              ip;
2869         int                up;
2870         int                rc;
2871
2872         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2873         if (rc != 0) {
2874                 CERROR("Can't query IPoIB interface %s: %d\n",
2875                        ifname, rc);
2876                 return NULL;
2877         }
2878
2879         if (!up) {
2880                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2881                 return NULL;
2882         }
2883
2884         LIBCFS_ALLOC(dev, sizeof(*dev));
2885         if (dev == NULL)
2886                 return NULL;
2887
2888         netdev = dev_get_by_name(&init_net, ifname);
2889         if (netdev == NULL) {
2890                 dev->ibd_can_failover = 0;
2891         } else {
2892                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2893                 dev_put(netdev);
2894         }
2895
2896         INIT_LIST_HEAD(&dev->ibd_nets);
2897         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2898         INIT_LIST_HEAD(&dev->ibd_fail_list);
2899         dev->ibd_ifip = ip;
2900         strcpy(&dev->ibd_ifname[0], ifname);
2901
2902         /* initialize the device */
2903         rc = kiblnd_dev_failover(dev);
2904         if (rc != 0) {
2905                 CERROR("Can't initialize device: %d\n", rc);
2906                 LIBCFS_FREE(dev, sizeof(*dev));
2907                 return NULL;
2908         }
2909
2910         list_add_tail(&dev->ibd_list,
2911                           &kiblnd_data.kib_devs);
2912         return dev;
2913 }
2914
2915 static void
2916 kiblnd_base_shutdown(void)
2917 {
2918         struct kib_sched_info   *sched;
2919         int                     i;
2920
2921         LASSERT(list_empty(&kiblnd_data.kib_devs));
2922
2923         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2924                atomic_read(&libcfs_kmemory));
2925
2926         switch (kiblnd_data.kib_init) {
2927         default:
2928                 LBUG();
2929
2930         case IBLND_INIT_ALL:
2931         case IBLND_INIT_DATA:
2932                 LASSERT (kiblnd_data.kib_peers != NULL);
2933                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2934                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2935                 }
2936                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2937                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2938                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2939                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2940
2941                 /* flag threads to terminate; wake and wait for them to die */
2942                 kiblnd_data.kib_shutdown = 1;
2943
2944                 /* NB: we really want to stop scheduler threads net by net
2945                  * instead of the whole module, this should be improved
2946                  * with dynamic configuration LNet */
2947                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2948                         wake_up_all(&sched->ibs_waitq);
2949
2950                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2951                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2952
2953                 i = 2;
2954                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2955                         i++;
2956                         /* power of 2? */
2957                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2958                                "Waiting for %d threads to terminate\n",
2959                                atomic_read(&kiblnd_data.kib_nthreads));
2960                         set_current_state(TASK_UNINTERRUPTIBLE);
2961                         schedule_timeout(cfs_time_seconds(1));
2962                 }
2963
2964                 /* fall through */
2965
2966         case IBLND_INIT_NOTHING:
2967                 break;
2968         }
2969
2970         if (kiblnd_data.kib_peers != NULL) {
2971                 LIBCFS_FREE(kiblnd_data.kib_peers,
2972                             sizeof(struct list_head) *
2973                             kiblnd_data.kib_peer_hash_size);
2974         }
2975
2976         if (kiblnd_data.kib_scheds != NULL)
2977                 cfs_percpt_free(kiblnd_data.kib_scheds);
2978
2979         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2980                atomic_read(&libcfs_kmemory));
2981
2982         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2983         module_put(THIS_MODULE);
2984 }
2985
2986 static void
2987 kiblnd_shutdown(struct lnet_ni *ni)
2988 {
2989         kib_net_t        *net = ni->ni_data;
2990         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2991         int               i;
2992         unsigned long     flags;
2993
2994         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2995
2996         if (net == NULL)
2997                 goto out;
2998
2999         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3000                atomic_read(&libcfs_kmemory));
3001
3002         write_lock_irqsave(g_lock, flags);
3003         net->ibn_shutdown = 1;
3004         write_unlock_irqrestore(g_lock, flags);
3005
3006         switch (net->ibn_init) {
3007         default:
3008                 LBUG();
3009
3010         case IBLND_INIT_ALL:
3011                 /* nuke all existing peers within this net */
3012                 kiblnd_del_peer(ni, LNET_NID_ANY);
3013
3014                 /* Wait for all peer_ni state to clean up */
3015                 i = 2;
3016                 while (atomic_read(&net->ibn_npeers) != 0) {
3017                         i++;
3018                         /* power of 2? */
3019                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3020                                "%s: waiting for %d peers to disconnect\n",
3021                                libcfs_nid2str(ni->ni_nid),
3022                                atomic_read(&net->ibn_npeers));
3023                         set_current_state(TASK_UNINTERRUPTIBLE);
3024                         schedule_timeout(cfs_time_seconds(1));
3025                 }
3026
3027                 kiblnd_net_fini_pools(net);
3028
3029                 write_lock_irqsave(g_lock, flags);
3030                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3031                 net->ibn_dev->ibd_nnets--;
3032                 list_del(&net->ibn_list);
3033                 write_unlock_irqrestore(g_lock, flags);
3034
3035                 /* fall through */
3036
3037         case IBLND_INIT_NOTHING:
3038                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3039
3040                 if (net->ibn_dev != NULL &&
3041                     net->ibn_dev->ibd_nnets == 0)
3042                         kiblnd_destroy_dev(net->ibn_dev);
3043
3044                 break;
3045         }
3046
3047         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3048                atomic_read(&libcfs_kmemory));
3049
3050         net->ibn_init = IBLND_INIT_NOTHING;
3051         ni->ni_data = NULL;
3052
3053         LIBCFS_FREE(net, sizeof(*net));
3054
3055 out:
3056         if (list_empty(&kiblnd_data.kib_devs))
3057                 kiblnd_base_shutdown();
3058         return;
3059 }
3060
3061 static int
3062 kiblnd_base_startup(void)
3063 {
3064         struct kib_sched_info   *sched;
3065         int                     rc;
3066         int                     i;
3067
3068         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3069
3070         try_module_get(THIS_MODULE);
3071         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3072
3073         rwlock_init(&kiblnd_data.kib_global_lock);
3074
3075         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3076         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3077
3078         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3079         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3080                      sizeof(struct list_head) *
3081                      kiblnd_data.kib_peer_hash_size);
3082         if (kiblnd_data.kib_peers == NULL)
3083                 goto failed;
3084
3085         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3086                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3087
3088         spin_lock_init(&kiblnd_data.kib_connd_lock);
3089         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3090         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3091         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3092         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3093
3094         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3095         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3096
3097         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3098                                                   sizeof(*sched));
3099         if (kiblnd_data.kib_scheds == NULL)
3100                 goto failed;
3101
3102         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3103                 int     nthrs;
3104
3105                 spin_lock_init(&sched->ibs_lock);
3106                 INIT_LIST_HEAD(&sched->ibs_conns);
3107                 init_waitqueue_head(&sched->ibs_waitq);
3108
3109                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3110                 if (*kiblnd_tunables.kib_nscheds > 0) {
3111                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3112                 } else {
3113                         /* max to half of CPUs, another half is reserved for
3114                          * upper layer modules */
3115                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3116                 }
3117
3118                 sched->ibs_nthreads_max = nthrs;
3119                 sched->ibs_cpt = i;
3120         }
3121
3122         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3123
3124         /* lists/ptrs/locks initialised */
3125         kiblnd_data.kib_init = IBLND_INIT_DATA;
3126         /*****************************************************/
3127
3128         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3129         if (rc != 0) {
3130                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3131                 goto failed;
3132         }
3133
3134         if (*kiblnd_tunables.kib_dev_failover != 0)
3135                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3136                                          "kiblnd_failover");
3137
3138         if (rc != 0) {
3139                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3140                 goto failed;
3141         }
3142
3143         /* flag everything initialised */
3144         kiblnd_data.kib_init = IBLND_INIT_ALL;
3145         /*****************************************************/
3146
3147         return 0;
3148
3149  failed:
3150         kiblnd_base_shutdown();
3151         return -ENETDOWN;
3152 }
3153
3154 static int
3155 kiblnd_start_schedulers(struct kib_sched_info *sched)
3156 {
3157         int     rc = 0;
3158         int     nthrs;
3159         int     i;
3160
3161         if (sched->ibs_nthreads == 0) {
3162                 if (*kiblnd_tunables.kib_nscheds > 0) {
3163                         nthrs = sched->ibs_nthreads_max;
3164                 } else {
3165                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3166                                                sched->ibs_cpt);
3167                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3168                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3169                 }
3170         } else {
3171                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3172                 /* increase one thread if there is new interface */
3173                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3174         }
3175
3176         for (i = 0; i < nthrs; i++) {
3177                 long    id;
3178                 char    name[20];
3179                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3180                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3181                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3182                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3183                 if (rc == 0)
3184                         continue;
3185
3186                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3187                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3188                 break;
3189         }
3190
3191         sched->ibs_nthreads += i;
3192         return rc;
3193 }
3194
3195 static int
3196 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3197 {
3198         int     cpt;
3199         int     rc;
3200         int     i;
3201
3202         for (i = 0; i < ncpts; i++) {
3203                 struct kib_sched_info *sched;
3204
3205                 cpt = (cpts == NULL) ? i : cpts[i];
3206                 sched = kiblnd_data.kib_scheds[cpt];
3207
3208                 if (!newdev && sched->ibs_nthreads > 0)
3209                         continue;
3210
3211                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3212                 if (rc != 0) {
3213                         CERROR("Failed to start scheduler threads for %s\n",
3214                                dev->ibd_ifname);
3215                         return rc;
3216                 }
3217         }
3218         return 0;
3219 }
3220
3221 static kib_dev_t *
3222 kiblnd_dev_search(char *ifname)
3223 {
3224         kib_dev_t       *alias = NULL;
3225         kib_dev_t       *dev;
3226         char            *colon;
3227         char            *colon2;
3228
3229         colon = strchr(ifname, ':');
3230         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3231                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3232                         return dev;
3233
3234                 if (alias != NULL)
3235                         continue;
3236
3237                 colon2 = strchr(dev->ibd_ifname, ':');
3238                 if (colon != NULL)
3239                         *colon = 0;
3240                 if (colon2 != NULL)
3241                         *colon2 = 0;
3242
3243                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3244                         alias = dev;
3245
3246                 if (colon != NULL)
3247                         *colon = ':';
3248                 if (colon2 != NULL)
3249                         *colon2 = ':';
3250         }
3251         return alias;
3252 }
3253
3254 static int
3255 kiblnd_startup(struct lnet_ni *ni)
3256 {
3257         char                     *ifname;
3258         kib_dev_t                *ibdev = NULL;
3259         kib_net_t                *net;
3260         unsigned long             flags;
3261         int                       rc;
3262         int                       newdev;
3263         int                       node_id;
3264
3265         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3266
3267         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3268                 rc = kiblnd_base_startup();
3269                 if (rc != 0)
3270                         return rc;
3271         }
3272
3273         LIBCFS_ALLOC(net, sizeof(*net));
3274         ni->ni_data = net;
3275         if (net == NULL)
3276                 goto failed;
3277
3278         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3279
3280         kiblnd_tunables_setup(ni);
3281
3282         if (ni->ni_interfaces[0] != NULL) {
3283                 /* Use the IPoIB interface specified in 'networks=' */
3284
3285                 CLASSERT(LNET_INTERFACES_NUM > 1);
3286                 if (ni->ni_interfaces[1] != NULL) {
3287                         CERROR("Multiple interfaces not supported\n");
3288                         goto failed;
3289                 }
3290
3291                 ifname = ni->ni_interfaces[0];
3292         } else {
3293                 ifname = *kiblnd_tunables.kib_default_ipif;
3294         }
3295
3296         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3297                 CERROR("IPoIB interface name too long: %s\n", ifname);
3298                 goto failed;
3299         }
3300
3301         ibdev = kiblnd_dev_search(ifname);
3302
3303         newdev = ibdev == NULL;
3304         /* hmm...create kib_dev even for alias */
3305         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3306                 ibdev = kiblnd_create_dev(ifname);
3307
3308         if (ibdev == NULL)
3309                 goto failed;
3310
3311         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3312         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3313
3314         net->ibn_dev = ibdev;
3315         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3316
3317         rc = kiblnd_dev_start_threads(ibdev, newdev,
3318                                       ni->ni_cpts, ni->ni_ncpts);
3319         if (rc != 0)
3320                 goto failed;
3321
3322         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3323         if (rc != 0) {
3324                 CERROR("Failed to initialize NI pools: %d\n", rc);
3325                 goto failed;
3326         }
3327
3328         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3329         ibdev->ibd_nnets++;
3330         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3331         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3332
3333         net->ibn_init = IBLND_INIT_ALL;
3334
3335         return 0;
3336
3337 failed:
3338         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3339                 kiblnd_destroy_dev(ibdev);
3340
3341         kiblnd_shutdown(ni);
3342
3343         CDEBUG(D_NET, "kiblnd_startup failed\n");
3344         return -ENETDOWN;
3345 }
3346
3347 static struct lnet_lnd the_o2iblnd = {
3348         .lnd_type       = O2IBLND,
3349         .lnd_startup    = kiblnd_startup,
3350         .lnd_shutdown   = kiblnd_shutdown,
3351         .lnd_ctl        = kiblnd_ctl,
3352         .lnd_query      = kiblnd_query,
3353         .lnd_send       = kiblnd_send,
3354         .lnd_recv       = kiblnd_recv,
3355 };
3356
3357 static void __exit ko2iblnd_exit(void)
3358 {
3359         lnet_unregister_lnd(&the_o2iblnd);
3360 }
3361
3362 static int __init ko2iblnd_init(void)
3363 {
3364         int rc;
3365
3366         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3367         CLASSERT(offsetof(kib_msg_t,
3368                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3369                  IBLND_MSG_SIZE);
3370         CLASSERT(offsetof(kib_msg_t,
3371                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3372                  <= IBLND_MSG_SIZE);
3373
3374         rc = kiblnd_tunables_init();
3375         if (rc != 0)
3376                 return rc;
3377
3378         lnet_register_lnd(&the_o2iblnd);
3379
3380         return 0;
3381 }
3382
3383 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3384 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3385 MODULE_VERSION("2.8.0");
3386 MODULE_LICENSE("GPL");
3387
3388 module_init(ko2iblnd_init);
3389 module_exit(ko2iblnd_exit);