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 %lld "
 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, ktime_to_ns(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, "   tx_noops:\n");
 608         list_for_each(tmp, &conn->ibc_tx_noops)
 609                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 610
 611         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 612         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 613                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 614
 615         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 616         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 617                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 620         list_for_each(tmp, &conn->ibc_tx_queue)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   active_txs:\n");
 624         list_for_each(tmp, &conn->ibc_active_txs)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   rxs:\n");
 628         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 629                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 630
 631         spin_unlock(&conn->ibc_lock);
 632 }
 633
 634 int
 635 kiblnd_translate_mtu(int value)
 636 {
 637         switch (value) {
 638         default:
 639                 return -1;
 640         case 0:
 641                 return 0;
 642         case 256:
 643                 return IB_MTU_256;
 644         case 512:
 645                 return IB_MTU_512;
 646         case 1024:
 647                 return IB_MTU_1024;
 648         case 2048:
 649                 return IB_MTU_2048;
 650         case 4096:
 651                 return IB_MTU_4096;
 652         }
 653 }
 654
 655 static void
 656 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 657 {
 658         int           mtu;
 659
 660         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 661         if (cmid->route.path_rec == NULL)
 662                 return;
 663
 664         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 665         LASSERT (mtu >= 0);
 666         if (mtu != 0)
 667                 cmid->route.path_rec->mtu = mtu;
 668 }
 669
 670 static int
 671 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 672 {
 673         cpumask_t       *mask;
 674         int             vectors;
 675         int             off;
 676         int             i;
 677         lnet_nid_t      ibp_nid;
 678
 679         vectors = conn->ibc_cmid->device->num_comp_vectors;
 680         if (vectors <= 1)
 681                 return 0;
 682
 683         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 684
 685         /* hash NID to CPU id in this partition... */
 686         ibp_nid = conn->ibc_peer->ibp_nid;
 687         off = do_div(ibp_nid, cpumask_weight(mask));
 688         for_each_cpu(i, mask) {
 689                 if (off-- == 0)
 690                         return i % vectors;
 691         }
 692
 693         LBUG();
 694         return 1;
 695 }
 696
 697 /*
 698  * Get the scheduler bound to this CPT. If the scheduler has no
 699  * threads, which means that the CPT has no CPUs, then grab the
 700  * next scheduler that we can use.
 701  *
 702  * This case would be triggered if a NUMA node is configured with
 703  * no associated CPUs.
 704  */
 705 static struct kib_sched_info *
 706 kiblnd_get_scheduler(int cpt)
 707 {
 708         struct kib_sched_info *sched;
 709         int i;
 710
 711         sched = kiblnd_data.kib_scheds[cpt];
 712
 713         if (sched->ibs_nthreads > 0)
 714                 return sched;
 715
 716         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 717                 if (sched->ibs_nthreads > 0) {
 718                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 719                                         cpt, sched->ibs_cpt);
 720                         return sched;
 721                 }
 722         }
 723
 724         return NULL;
 725 }
 726
 727 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 728 {
 729         /*
 730          * One WR for the LNet message
 731          * And ibc_max_frags for the transfer WRs
 732          */
 733         unsigned int ret = 1 + conn->ibc_max_frags;
 734         __u32 dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
 735
 736         /* FastReg needs two extra WRs for map and invalidate */
 737         if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
 738                 ret += 2;
 739
 740         /* account for a maximum of ibc_queue_depth in-flight transfers */
 741         ret *= conn->ibc_queue_depth;
 742         return ret;
 743 }
 744
 745 kib_conn_t *
 746 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 747                    int state, int version)
 748 {
 749         /* CAVEAT EMPTOR:
 750          * If the new conn is created successfully it takes over the caller's
 751          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 752          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 753          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 754          * to destroy 'cmid' here since I'm called from the CM which still has
 755          * its ref on 'cmid'). */
 756         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 757         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 758         kib_dev_t              *dev;
 759         struct ib_qp_init_attr *init_qp_attr;
 760         struct kib_sched_info   *sched;
 761 #ifdef HAVE_IB_CQ_INIT_ATTR
 762         struct ib_cq_init_attr  cq_attr = {};
 763 #endif
 764         kib_conn_t              *conn;
 765         struct ib_cq            *cq;
 766         unsigned long           flags;
 767         int                     cpt;
 768         int                     rc;
 769         int                     i;
 770
 771         LASSERT(net != NULL);
 772         LASSERT(!in_interrupt());
 773
 774         dev = net->ibn_dev;
 775
 776         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 777         sched = kiblnd_get_scheduler(cpt);
 778
 779         if (sched == NULL) {
 780                 CERROR("no schedulers available. node is unhealthy\n");
 781                 goto failed_0;
 782         }
 783
 784         /*
 785          * The cpt might have changed if we ended up selecting a non cpt
 786          * native scheduler. So use the scheduler's cpt instead.
 787          */
 788         cpt = sched->ibs_cpt;
 789
 790         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 791                          sizeof(*init_qp_attr));
 792         if (init_qp_attr == NULL) {
 793                 CERROR("Can't allocate qp_attr for %s\n",
 794                        libcfs_nid2str(peer_ni->ibp_nid));
 795                 goto failed_0;
 796         }
 797
 798         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 799         if (conn == NULL) {
 800                 CERROR("Can't allocate connection for %s\n",
 801                        libcfs_nid2str(peer_ni->ibp_nid));
 802                 goto failed_1;
 803         }
 804
 805         conn->ibc_state = IBLND_CONN_INIT;
 806         conn->ibc_version = version;
 807         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 808         cmid->context = conn;                   /* for future CM callbacks */
 809         conn->ibc_cmid = cmid;
 810         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 811         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 812
 813         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 814         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 815         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 816         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 817         INIT_LIST_HEAD(&conn->ibc_active_txs);
 818         spin_lock_init(&conn->ibc_lock);
 819
 820         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 821                          sizeof(*conn->ibc_connvars));
 822         if (conn->ibc_connvars == NULL) {
 823                 CERROR("Can't allocate in-progress connection state\n");
 824                 goto failed_2;
 825         }
 826
 827         write_lock_irqsave(glock, flags);
 828         if (dev->ibd_failover) {
 829                 write_unlock_irqrestore(glock, flags);
 830                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 831                 goto failed_2;
 832         }
 833
 834         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 835                 /* wakeup failover thread and teardown connection */
 836                 if (kiblnd_dev_can_failover(dev)) {
 837                         list_add_tail(&dev->ibd_fail_list,
 838                                       &kiblnd_data.kib_failed_devs);
 839                         wake_up(&kiblnd_data.kib_failover_waitq);
 840                 }
 841
 842                 write_unlock_irqrestore(glock, flags);
 843                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 844                        cmid->device->name, dev->ibd_ifname);
 845                 goto failed_2;
 846         }
 847
 848         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 849         conn->ibc_hdev = dev->ibd_hdev;
 850
 851         kiblnd_setup_mtu_locked(cmid);
 852
 853         write_unlock_irqrestore(glock, flags);
 854
 855         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 856                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 857         if (conn->ibc_rxs == NULL) {
 858                 CERROR("Cannot allocate RX buffers\n");
 859                 goto failed_2;
 860         }
 861
 862         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 863                                 IBLND_RX_MSG_PAGES(conn));
 864         if (rc != 0)
 865                 goto failed_2;
 866
 867         kiblnd_map_rx_descs(conn);
 868
 869 #ifdef HAVE_IB_CQ_INIT_ATTR
 870         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 871         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 872         cq = ib_create_cq(cmid->device,
 873                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 874                           &cq_attr);
 875 #else
 876         cq = ib_create_cq(cmid->device,
 877                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 878                           IBLND_CQ_ENTRIES(conn),
 879                           kiblnd_get_completion_vector(conn, cpt));
 880 #endif
 881         if (IS_ERR(cq)) {
 882                 /*
 883                  * on MLX-5 (possibly MLX-4 as well) this error could be
 884                  * hit if the concurrent_sends and/or peer_tx_credits is set
 885                  * too high. Or due to an MLX-5 bug which tries to
 886                  * allocate 256kb via kmalloc for WR cookie array
 887                  */
 888                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 889                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 890                 goto failed_2;
 891         }
 892
 893         conn->ibc_cq = cq;
 894
 895         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 896         if (rc != 0) {
 897                 CERROR("Can't request completion notification: %d\n", rc);
 898                 goto failed_2;
 899         }
 900
 901         init_qp_attr->event_handler = kiblnd_qp_event;
 902         init_qp_attr->qp_context = conn;
 903         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 904         init_qp_attr->cap.max_recv_sge = 1;
 905         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 906         init_qp_attr->qp_type = IB_QPT_RC;
 907         init_qp_attr->send_cq = cq;
 908         init_qp_attr->recv_cq = cq;
 909
 910         conn->ibc_sched = sched;
 911
 912         do {
 913                 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 914                 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 915
 916                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 917                 if (!rc || conn->ibc_queue_depth < 2)
 918                         break;
 919
 920                 conn->ibc_queue_depth--;
 921         } while (rc);
 922
 923         if (rc) {
 924                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 925                        "send_sge: %d, recv_sge: %d\n",
 926                        rc, init_qp_attr->cap.max_send_wr,
 927                        init_qp_attr->cap.max_recv_wr,
 928                        init_qp_attr->cap.max_send_sge,
 929                        init_qp_attr->cap.max_recv_sge);
 930                 goto failed_2;
 931         }
 932
 933         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 934                 CWARN("peer %s - queue depth reduced from %u to %u"
 935                       "  to allow for qp creation\n",
 936                       libcfs_nid2str(peer_ni->ibp_nid),
 937                       peer_ni->ibp_queue_depth,
 938                       conn->ibc_queue_depth);
 939
 940         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 941
 942         /* 1 ref for caller and each rxmsg */
 943         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 944         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 945
 946         /* post receives */
 947         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 948                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 949                 if (rc != 0) {
 950                         CERROR("Can't post rxmsg: %d\n", rc);
 951
 952                         /* Make posted receives complete */
 953                         kiblnd_abort_receives(conn);
 954
 955                         /* correct # of posted buffers
 956                          * NB locking needed now I'm racing with completion */
 957                         spin_lock_irqsave(&sched->ibs_lock, flags);
 958                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 959                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 960
 961                         /* cmid will be destroyed by CM(ofed) after cm_callback
 962                          * returned, so we can't refer it anymore
 963                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 964                         rdma_destroy_qp(conn->ibc_cmid);
 965                         conn->ibc_cmid = NULL;
 966
 967                         /* Drop my own and unused rxbuffer refcounts */
 968                         while (i++ <= IBLND_RX_MSGS(conn))
 969                                 kiblnd_conn_decref(conn);
 970
 971                         return NULL;
 972                 }
 973         }
 974
 975         /* Init successful! */
 976         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 977                  state == IBLND_CONN_PASSIVE_WAIT);
 978         conn->ibc_state = state;
 979
 980         /* 1 more conn */
 981         atomic_inc(&net->ibn_nconns);
 982         return conn;
 983
 984  failed_2:
 985         kiblnd_destroy_conn(conn);
 986         LIBCFS_FREE(conn, sizeof(*conn));
 987  failed_1:
 988         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 989  failed_0:
 990         return NULL;
 991 }
 992
 993 void
 994 kiblnd_destroy_conn(kib_conn_t *conn)
 995 {
 996         struct rdma_cm_id *cmid = conn->ibc_cmid;
 997         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
 998         int                rc;
 999
1000         LASSERT (!in_interrupt());
1001         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1002         LASSERT(list_empty(&conn->ibc_tx_noops));
1003         LASSERT(list_empty(&conn->ibc_tx_queue));
1004         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1005         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1006         LASSERT(list_empty(&conn->ibc_active_txs));
1007         LASSERT (conn->ibc_noops_posted == 0);
1008         LASSERT (conn->ibc_nsends_posted == 0);
1009
1010         switch (conn->ibc_state) {
1011         default:
1012                 /* conn must be completely disengaged from the network */
1013                 LBUG();
1014
1015         case IBLND_CONN_DISCONNECTED:
1016                 /* connvars should have been freed already */
1017                 LASSERT (conn->ibc_connvars == NULL);
1018                 break;
1019
1020         case IBLND_CONN_INIT:
1021                 break;
1022         }
1023
1024         /* conn->ibc_cmid might be destroyed by CM already */
1025         if (cmid != NULL && cmid->qp != NULL)
1026                 rdma_destroy_qp(cmid);
1027
1028         if (conn->ibc_cq != NULL) {
1029                 rc = ib_destroy_cq(conn->ibc_cq);
1030                 if (rc != 0)
1031                         CWARN("Error destroying CQ: %d\n", rc);
1032         }
1033
1034         if (conn->ibc_rx_pages != NULL)
1035                 kiblnd_unmap_rx_descs(conn);
1036
1037         if (conn->ibc_rxs != NULL) {
1038                 LIBCFS_FREE(conn->ibc_rxs,
1039                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1040         }
1041
1042         if (conn->ibc_connvars != NULL)
1043                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1044
1045         if (conn->ibc_hdev != NULL)
1046                 kiblnd_hdev_decref(conn->ibc_hdev);
1047
1048         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1049         if (conn->ibc_state != IBLND_CONN_INIT) {
1050                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1051
1052                 kiblnd_peer_decref(peer_ni);
1053                 rdma_destroy_id(cmid);
1054                 atomic_dec(&net->ibn_nconns);
1055         }
1056 }
1057
1058 int
1059 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1060 {
1061         kib_conn_t              *conn;
1062         struct list_head        *ctmp;
1063         struct list_head        *cnxt;
1064         int                     count = 0;
1065
1066         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1067                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1068
1069                 CDEBUG(D_NET, "Closing conn -> %s, "
1070                               "version: %x, reason: %d\n",
1071                        libcfs_nid2str(peer_ni->ibp_nid),
1072                        conn->ibc_version, why);
1073
1074                 kiblnd_close_conn_locked(conn, why);
1075                 count++;
1076         }
1077
1078         return count;
1079 }
1080
1081 int
1082 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1083                                 int version, __u64 incarnation)
1084 {
1085         kib_conn_t              *conn;
1086         struct list_head        *ctmp;
1087         struct list_head        *cnxt;
1088         int                     count = 0;
1089
1090         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1091                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1092
1093                 if (conn->ibc_version     == version &&
1094                     conn->ibc_incarnation == incarnation)
1095                         continue;
1096
1097                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1098                               "incarnation:%#llx(%x, %#llx)\n",
1099                        libcfs_nid2str(peer_ni->ibp_nid),
1100                        conn->ibc_version, conn->ibc_incarnation,
1101                        version, incarnation);
1102
1103                 kiblnd_close_conn_locked(conn, -ESTALE);
1104                 count++;
1105         }
1106
1107         return count;
1108 }
1109
1110 static int
1111 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1112 {
1113         kib_peer_ni_t           *peer_ni;
1114         struct list_head        *ptmp;
1115         struct list_head        *pnxt;
1116         int                     lo;
1117         int                     hi;
1118         int                     i;
1119         unsigned long           flags;
1120         int                     count = 0;
1121
1122         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1123
1124         if (nid != LNET_NID_ANY)
1125                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1126         else {
1127                 lo = 0;
1128                 hi = kiblnd_data.kib_peer_hash_size - 1;
1129         }
1130
1131         for (i = lo; i <= hi; i++) {
1132                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1133
1134                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1135                         LASSERT(!kiblnd_peer_idle(peer_ni));
1136
1137                         if (peer_ni->ibp_ni != ni)
1138                                 continue;
1139
1140                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1141                                 continue;
1142
1143                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1144                 }
1145         }
1146
1147         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1148
1149         /* wildcards always succeed */
1150         if (nid == LNET_NID_ANY)
1151                 return 0;
1152
1153         return (count == 0) ? -ENOENT : 0;
1154 }
1155
1156 static int
1157 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1158 {
1159         struct libcfs_ioctl_data *data = arg;
1160         int                       rc = -EINVAL;
1161
1162         switch(cmd) {
1163         case IOC_LIBCFS_GET_PEER: {
1164                 lnet_nid_t   nid = 0;
1165                 int          count = 0;
1166
1167                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1168                                           &nid, &count);
1169                 data->ioc_nid    = nid;
1170                 data->ioc_count  = count;
1171                 break;
1172         }
1173
1174         case IOC_LIBCFS_DEL_PEER: {
1175                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1176                 break;
1177         }
1178         case IOC_LIBCFS_GET_CONN: {
1179                 kib_conn_t *conn;
1180
1181                 rc = 0;
1182                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1183                 if (conn == NULL) {
1184                         rc = -ENOENT;
1185                         break;
1186                 }
1187
1188                 LASSERT(conn->ibc_cmid != NULL);
1189                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1190                 if (conn->ibc_cmid->route.path_rec == NULL)
1191                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1192                 else
1193                         data->ioc_u32[0] =
1194                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1195                 kiblnd_conn_decref(conn);
1196                 break;
1197         }
1198         case IOC_LIBCFS_CLOSE_CONNECTION: {
1199                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1200                 break;
1201         }
1202
1203         default:
1204                 break;
1205         }
1206
1207         return rc;
1208 }
1209
1210 static void
1211 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1212 {
1213         time64_t last_alive = 0;
1214         time64_t now = ktime_get_seconds();
1215         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1216         kib_peer_ni_t *peer_ni;
1217         unsigned long flags;
1218
1219         read_lock_irqsave(glock, flags);
1220
1221         peer_ni = kiblnd_find_peer_locked(ni, nid);
1222         if (peer_ni != NULL)
1223                 last_alive = peer_ni->ibp_last_alive;
1224
1225         read_unlock_irqrestore(glock, flags);
1226
1227         if (last_alive != 0)
1228                 *when = last_alive;
1229
1230         /* peer_ni is not persistent in hash, trigger peer_ni creation
1231          * and connection establishment with a NULL tx */
1232         if (peer_ni == NULL)
1233                 kiblnd_launch_tx(ni, NULL, nid);
1234
1235         CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1236                libcfs_nid2str(nid), peer_ni,
1237                last_alive ? now - last_alive : -1);
1238         return;
1239 }
1240
1241 static void
1242 kiblnd_free_pages(kib_pages_t *p)
1243 {
1244         int     npages = p->ibp_npages;
1245         int     i;
1246
1247         for (i = 0; i < npages; i++) {
1248                 if (p->ibp_pages[i] != NULL)
1249                         __free_page(p->ibp_pages[i]);
1250         }
1251
1252         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1253 }
1254
1255 int
1256 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1257 {
1258         kib_pages_t     *p;
1259         int             i;
1260
1261         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1262                          offsetof(kib_pages_t, ibp_pages[npages]));
1263         if (p == NULL) {
1264                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1265                 return -ENOMEM;
1266         }
1267
1268         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1269         p->ibp_npages = npages;
1270
1271         for (i = 0; i < npages; i++) {
1272                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1273                                                      GFP_NOFS);
1274                 if (p->ibp_pages[i] == NULL) {
1275                         CERROR("Can't allocate page %d of %d\n", i, npages);
1276                         kiblnd_free_pages(p);
1277                         return -ENOMEM;
1278                 }
1279         }
1280
1281         *pp = p;
1282         return 0;
1283 }
1284
1285 void
1286 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1287 {
1288         kib_rx_t *rx;
1289         int       i;
1290
1291         LASSERT (conn->ibc_rxs != NULL);
1292         LASSERT (conn->ibc_hdev != NULL);
1293
1294         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1295                 rx = &conn->ibc_rxs[i];
1296
1297                 LASSERT(rx->rx_nob >= 0); /* not posted */
1298
1299                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1300                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1301                                                           rx->rx_msgaddr),
1302                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1303         }
1304
1305         kiblnd_free_pages(conn->ibc_rx_pages);
1306
1307         conn->ibc_rx_pages = NULL;
1308 }
1309
1310 void
1311 kiblnd_map_rx_descs(kib_conn_t *conn)
1312 {
1313         kib_rx_t       *rx;
1314         struct page    *pg;
1315         int             pg_off;
1316         int             ipg;
1317         int             i;
1318
1319         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1320                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1321                 rx = &conn->ibc_rxs[i];
1322
1323                 rx->rx_conn = conn;
1324                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1325
1326                 rx->rx_msgaddr =
1327                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1328                                               rx->rx_msg, IBLND_MSG_SIZE,
1329                                               DMA_FROM_DEVICE);
1330                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1331                                                   rx->rx_msgaddr));
1332                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1333
1334                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1335                        i, rx->rx_msg, rx->rx_msgaddr,
1336                        (__u64)(page_to_phys(pg) + pg_off));
1337
1338                 pg_off += IBLND_MSG_SIZE;
1339                 LASSERT(pg_off <= PAGE_SIZE);
1340
1341                 if (pg_off == PAGE_SIZE) {
1342                         pg_off = 0;
1343                         ipg++;
1344                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1345                 }
1346         }
1347 }
1348
1349 static void
1350 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1351 {
1352         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1353         kib_tx_t       *tx;
1354         int             i;
1355
1356         LASSERT (tpo->tpo_pool.po_allocated == 0);
1357
1358         if (hdev == NULL)
1359                 return;
1360
1361         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1362                 tx = &tpo->tpo_tx_descs[i];
1363                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1364                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1365                                                           tx->tx_msgaddr),
1366                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1367         }
1368
1369         kiblnd_hdev_decref(hdev);
1370         tpo->tpo_hdev = NULL;
1371 }
1372
1373 static kib_hca_dev_t *
1374 kiblnd_current_hdev(kib_dev_t *dev)
1375 {
1376         kib_hca_dev_t *hdev;
1377         unsigned long  flags;
1378         int            i = 0;
1379
1380         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1381         while (dev->ibd_failover) {
1382                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1383                 if (i++ % 50 == 0)
1384                         CDEBUG(D_NET, "%s: Wait for failover\n",
1385                                dev->ibd_ifname);
1386                 set_current_state(TASK_INTERRUPTIBLE);
1387                 schedule_timeout(cfs_time_seconds(1) / 100);
1388
1389                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1390         }
1391
1392         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1393         hdev = dev->ibd_hdev;
1394
1395         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1396
1397         return hdev;
1398 }
1399
1400 static void
1401 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1402 {
1403         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1404         kib_pool_t     *pool  = &tpo->tpo_pool;
1405         kib_net_t      *net   = pool->po_owner->ps_net;
1406         kib_dev_t      *dev;
1407         struct page    *page;
1408         kib_tx_t       *tx;
1409         int             page_offset;
1410         int             ipage;
1411         int             i;
1412
1413         LASSERT (net != NULL);
1414
1415         dev = net->ibn_dev;
1416
1417         /* pre-mapped messages are not bigger than 1 page */
1418         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1419
1420         /* No fancy arithmetic when we do the buffer calculations */
1421         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1422
1423         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1424
1425         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1426                 page = txpgs->ibp_pages[ipage];
1427                 tx = &tpo->tpo_tx_descs[i];
1428
1429                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1430                                            page_offset);
1431
1432                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1433                                                        tx->tx_msg,
1434                                                        IBLND_MSG_SIZE,
1435                                                        DMA_TO_DEVICE);
1436                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1437                                                   tx->tx_msgaddr));
1438                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1439
1440                 list_add(&tx->tx_list, &pool->po_free_list);
1441
1442                 page_offset += IBLND_MSG_SIZE;
1443                 LASSERT(page_offset <= PAGE_SIZE);
1444
1445                 if (page_offset == PAGE_SIZE) {
1446                         page_offset = 0;
1447                         ipage++;
1448                         LASSERT(ipage <= txpgs->ibp_npages);
1449                 }
1450         }
1451 }
1452
1453 static void
1454 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1455 {
1456         LASSERT(fpo->fpo_map_count == 0);
1457
1458         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1459                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1460         } else {
1461                 struct kib_fast_reg_descriptor *frd, *tmp;
1462                 int i = 0;
1463
1464                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1465                                          frd_list) {
1466                         list_del(&frd->frd_list);
1467 #ifndef HAVE_IB_MAP_MR_SG
1468                         ib_free_fast_reg_page_list(frd->frd_frpl);
1469 #endif
1470                         ib_dereg_mr(frd->frd_mr);
1471                         LIBCFS_FREE(frd, sizeof(*frd));
1472                         i++;
1473                 }
1474                 if (i < fpo->fast_reg.fpo_pool_size)
1475                         CERROR("FastReg pool still has %d regions registered\n",
1476                                 fpo->fast_reg.fpo_pool_size - i);
1477         }
1478
1479         if (fpo->fpo_hdev)
1480                 kiblnd_hdev_decref(fpo->fpo_hdev);
1481
1482         LIBCFS_FREE(fpo, sizeof(*fpo));
1483 }
1484
1485 static void
1486 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1487 {
1488         kib_fmr_pool_t *fpo, *tmp;
1489
1490         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1491                 list_del(&fpo->fpo_list);
1492                 kiblnd_destroy_fmr_pool(fpo);
1493         }
1494 }
1495
1496 static int
1497 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1498                      int ncpts)
1499 {
1500         int size = tunables->lnd_fmr_pool_size / ncpts;
1501
1502         return max(IBLND_FMR_POOL, size);
1503 }
1504
1505 static int
1506 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1507                          int ncpts)
1508 {
1509         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1510
1511         return max(IBLND_FMR_POOL_FLUSH, size);
1512 }
1513
1514 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1515 {
1516         struct ib_fmr_pool_param param = {
1517                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1518                 .page_shift        = PAGE_SHIFT,
1519                 .access            = (IB_ACCESS_LOCAL_WRITE |
1520                                       IB_ACCESS_REMOTE_WRITE),
1521                 .pool_size         = fps->fps_pool_size,
1522                 .dirty_watermark   = fps->fps_flush_trigger,
1523                 .flush_function    = NULL,
1524                 .flush_arg         = NULL,
1525                 .cache             = !!fps->fps_cache };
1526         int rc = 0;
1527
1528         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1529                                                    &param);
1530         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1531                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1532                 if (rc != -ENOSYS)
1533                         CERROR("Failed to create FMR pool: %d\n", rc);
1534                 else
1535                         CERROR("FMRs are not supported\n");
1536         }
1537         fpo->fpo_is_fmr = true;
1538
1539         return rc;
1540 }
1541
1542 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1543                                   __u32 dev_caps)
1544 {
1545         struct kib_fast_reg_descriptor *frd, *tmp;
1546         int i, rc;
1547
1548         fpo->fpo_is_fmr = false;
1549
1550         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1551         fpo->fast_reg.fpo_pool_size = 0;
1552         for (i = 0; i < fps->fps_pool_size; i++) {
1553                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1554                                  sizeof(*frd));
1555                 if (!frd) {
1556                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1557                         rc = -ENOMEM;
1558                         goto out;
1559                 }
1560                 frd->frd_mr = NULL;
1561
1562 #ifndef HAVE_IB_MAP_MR_SG
1563                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1564                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1565                 if (IS_ERR(frd->frd_frpl)) {
1566                         rc = PTR_ERR(frd->frd_frpl);
1567                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1568                                 rc);
1569                         frd->frd_frpl = NULL;
1570                         goto out_middle;
1571                 }
1572 #endif
1573
1574 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1575                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1576                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1577 #else
1578                 /*
1579                  * it is expected to get here if this is an MLX-5 card.
1580                  * MLX-4 cards will always use FMR and MLX-5 cards will
1581                  * always use fast_reg. It turns out that some MLX-5 cards
1582                  * (possibly due to older FW versions) do not natively support
1583                  * gaps. So we will need to track them here.
1584                  */
1585                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1586 #ifdef IB_MR_TYPE_SG_GAPS
1587                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1588                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1589                                                 IB_MR_TYPE_SG_GAPS :
1590                                                 IB_MR_TYPE_MEM_REG,
1591 #else
1592                                                 IB_MR_TYPE_MEM_REG,
1593 #endif
1594                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1595                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1596                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1597                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1598 #endif
1599                 if (IS_ERR(frd->frd_mr)) {
1600                         rc = PTR_ERR(frd->frd_mr);
1601                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1602                         frd->frd_mr = NULL;
1603                         goto out_middle;
1604                 }
1605
1606                 /* There appears to be a bug in MLX5 code where you must
1607                  * invalidate the rkey of a new FastReg pool before first
1608                  * using it. Thus, I am marking the FRD invalid here. */
1609                 frd->frd_valid = false;
1610
1611                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1612                 fpo->fast_reg.fpo_pool_size++;
1613         }
1614
1615         return 0;
1616
1617 out_middle:
1618         if (frd->frd_mr)
1619                 ib_dereg_mr(frd->frd_mr);
1620 #ifndef HAVE_IB_MAP_MR_SG
1621         if (frd->frd_frpl)
1622                 ib_free_fast_reg_page_list(frd->frd_frpl);
1623 #endif
1624         LIBCFS_FREE(frd, sizeof(*frd));
1625
1626 out:
1627         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1628                                  frd_list) {
1629                 list_del(&frd->frd_list);
1630 #ifndef HAVE_IB_MAP_MR_SG
1631                 ib_free_fast_reg_page_list(frd->frd_frpl);
1632 #endif
1633                 ib_dereg_mr(frd->frd_mr);
1634                 LIBCFS_FREE(frd, sizeof(*frd));
1635         }
1636
1637         return rc;
1638 }
1639
1640 static int
1641 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1642 {
1643         kib_dev_t *dev = fps->fps_net->ibn_dev;
1644         kib_fmr_pool_t *fpo;
1645         int rc;
1646
1647         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1648         if (!fpo) {
1649                 return -ENOMEM;
1650         }
1651         memset(fpo, 0, sizeof(*fpo));
1652
1653         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1654
1655         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1656                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1657         else
1658                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1659         if (rc)
1660                 goto out_fpo;
1661
1662         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1663         fpo->fpo_owner = fps;
1664         *pp_fpo = fpo;
1665
1666         return 0;
1667
1668 out_fpo:
1669         kiblnd_hdev_decref(fpo->fpo_hdev);
1670         LIBCFS_FREE(fpo, sizeof(*fpo));
1671         return rc;
1672 }
1673
1674 static void
1675 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1676 {
1677         if (fps->fps_net == NULL) /* intialized? */
1678                 return;
1679
1680         spin_lock(&fps->fps_lock);
1681
1682         while (!list_empty(&fps->fps_pool_list)) {
1683                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1684                                                  kib_fmr_pool_t, fpo_list);
1685                 fpo->fpo_failed = 1;
1686                 list_del(&fpo->fpo_list);
1687                 if (fpo->fpo_map_count == 0)
1688                         list_add(&fpo->fpo_list, zombies);
1689                 else
1690                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1691         }
1692
1693         spin_unlock(&fps->fps_lock);
1694 }
1695
1696 static void
1697 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1698 {
1699         if (fps->fps_net != NULL) { /* initialized? */
1700                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1701                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1702         }
1703 }
1704
1705 static int
1706 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1707                         kib_net_t *net,
1708                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1709 {
1710         kib_fmr_pool_t *fpo;
1711         int             rc;
1712
1713         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1714
1715         fps->fps_net = net;
1716         fps->fps_cpt = cpt;
1717
1718         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1719         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1720         fps->fps_cache = tunables->lnd_fmr_cache;
1721
1722         spin_lock_init(&fps->fps_lock);
1723         INIT_LIST_HEAD(&fps->fps_pool_list);
1724         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1725
1726         rc = kiblnd_create_fmr_pool(fps, &fpo);
1727         if (rc == 0)
1728                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1729
1730         return rc;
1731 }
1732
1733 static int
1734 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, time64_t now)
1735 {
1736         if (fpo->fpo_map_count != 0) /* still in use */
1737                 return 0;
1738         if (fpo->fpo_failed)
1739                 return 1;
1740         return now >= fpo->fpo_deadline;
1741 }
1742
1743 static int
1744 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1745 {
1746         kib_hca_dev_t   *hdev;
1747         __u64           *pages = tx->tx_pages;
1748         int             npages;
1749         int             size;
1750         int             i;
1751
1752         hdev = tx->tx_pool->tpo_hdev;
1753
1754         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1755                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1756                         size += hdev->ibh_page_size) {
1757                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1758                                            hdev->ibh_page_mask) + size;
1759                 }
1760         }
1761
1762         return npages;
1763 }
1764
1765 void
1766 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1767 {
1768         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1769         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1770         kib_fmr_poolset_t *fps;
1771         time64_t now = ktime_get_seconds();
1772         kib_fmr_pool_t *tmp;
1773         int rc;
1774
1775         if (!fpo)
1776                 return;
1777
1778         fps = fpo->fpo_owner;
1779         if (fpo->fpo_is_fmr) {
1780                 if (fmr->fmr_pfmr) {
1781                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1782                         LASSERT(!rc);
1783                         fmr->fmr_pfmr = NULL;
1784                 }
1785
1786                 if (status) {
1787                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1788                         LASSERT(!rc);
1789                 }
1790         } else {
1791                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1792
1793                 if (frd) {
1794                         frd->frd_valid = false;
1795                         spin_lock(&fps->fps_lock);
1796                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1797                         spin_unlock(&fps->fps_lock);
1798                         fmr->fmr_frd = NULL;
1799                 }
1800         }
1801         fmr->fmr_pool = NULL;
1802
1803         spin_lock(&fps->fps_lock);
1804         fpo->fpo_map_count--;   /* decref the pool */
1805
1806         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1807                 /* the first pool is persistent */
1808                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1809                         continue;
1810
1811                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1812                         list_move(&fpo->fpo_list, &zombies);
1813                         fps->fps_version++;
1814                 }
1815         }
1816         spin_unlock(&fps->fps_lock);
1817
1818         if (!list_empty(&zombies))
1819                 kiblnd_destroy_fmr_pool_list(&zombies);
1820 }
1821
1822 int
1823 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1824                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1825 {
1826         kib_fmr_pool_t *fpo;
1827         __u64 *pages = tx->tx_pages;
1828         __u64 version;
1829         bool is_rx = (rd != tx->tx_rd);
1830         bool tx_pages_mapped = 0;
1831         int npages = 0;
1832         int rc;
1833
1834 again:
1835         spin_lock(&fps->fps_lock);
1836         version = fps->fps_version;
1837         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1838                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1839                 fpo->fpo_map_count++;
1840
1841                 if (fpo->fpo_is_fmr) {
1842                         struct ib_pool_fmr *pfmr;
1843
1844                         spin_unlock(&fps->fps_lock);
1845
1846                         if (!tx_pages_mapped) {
1847                                 npages = kiblnd_map_tx_pages(tx, rd);
1848                                 tx_pages_mapped = 1;
1849                         }
1850
1851                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1852                                                     pages, npages, iov);
1853                         if (likely(!IS_ERR(pfmr))) {
1854                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1855                                                       : pfmr->fmr->lkey;
1856                                 fmr->fmr_frd  = NULL;
1857                                 fmr->fmr_pfmr = pfmr;
1858                                 fmr->fmr_pool = fpo;
1859                                 return 0;
1860                         }
1861                         rc = PTR_ERR(pfmr);
1862                 } else {
1863                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1864                                 struct kib_fast_reg_descriptor *frd;
1865 #ifdef HAVE_IB_MAP_MR_SG
1866                                 struct ib_reg_wr *wr;
1867                                 int n;
1868 #else
1869                                 struct ib_rdma_wr *wr;
1870                                 struct ib_fast_reg_page_list *frpl;
1871 #endif
1872                                 struct ib_mr *mr;
1873
1874                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1875                                                         struct kib_fast_reg_descriptor,
1876                                                         frd_list);
1877                                 list_del(&frd->frd_list);
1878                                 spin_unlock(&fps->fps_lock);
1879
1880 #ifndef HAVE_IB_MAP_MR_SG
1881                                 frpl = frd->frd_frpl;
1882 #endif
1883                                 mr   = frd->frd_mr;
1884
1885                                 if (!frd->frd_valid) {
1886                                         struct ib_rdma_wr *inv_wr;
1887                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1888
1889                                         inv_wr = &frd->frd_inv_wr;
1890                                         memset(inv_wr, 0, sizeof(*inv_wr));
1891
1892                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1893                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1894                                         inv_wr->wr.ex.invalidate_rkey = key;
1895
1896                                         /* Bump the key */
1897                                         key = ib_inc_rkey(key);
1898                                         ib_update_fast_reg_key(mr, key);
1899                                 }
1900
1901 #ifdef HAVE_IB_MAP_MR_SG
1902 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1903                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1904                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1905 #else
1906                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1907                                                  tx->tx_nfrags, PAGE_SIZE);
1908 #endif
1909                                 if (unlikely(n != tx->tx_nfrags)) {
1910                                         CERROR("Failed to map mr %d/%d "
1911                                                "elements\n", n, tx->tx_nfrags);
1912                                         return n < 0 ? n : -EINVAL;
1913                                 }
1914
1915                                 wr = &frd->frd_fastreg_wr;
1916                                 memset(wr, 0, sizeof(*wr));
1917
1918                                 wr->wr.opcode = IB_WR_REG_MR;
1919                                 wr->wr.wr_id  = IBLND_WID_MR;
1920                                 wr->wr.num_sge = 0;
1921                                 wr->wr.send_flags = 0;
1922                                 wr->mr = mr;
1923                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1924                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1925                                               IB_ACCESS_REMOTE_WRITE);
1926 #else
1927                                 if (!tx_pages_mapped) {
1928                                         npages = kiblnd_map_tx_pages(tx, rd);
1929                                         tx_pages_mapped = 1;
1930                                 }
1931
1932                                 LASSERT(npages <= frpl->max_page_list_len);
1933                                 memcpy(frpl->page_list, pages,
1934                                         sizeof(*pages) * npages);
1935
1936                                 /* Prepare FastReg WR */
1937                                 wr = &frd->frd_fastreg_wr;
1938                                 memset(wr, 0, sizeof(*wr));
1939
1940                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1941                                 wr->wr.wr_id  = IBLND_WID_MR;
1942
1943                                 wr->wr.wr.fast_reg.iova_start = iov;
1944                                 wr->wr.wr.fast_reg.page_list  = frpl;
1945                                 wr->wr.wr.fast_reg.page_list_len = npages;
1946                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1947                                 wr->wr.wr.fast_reg.length = nob;
1948                                 wr->wr.wr.fast_reg.rkey =
1949                                                 is_rx ? mr->rkey : mr->lkey;
1950                                 wr->wr.wr.fast_reg.access_flags =
1951                                                 (IB_ACCESS_LOCAL_WRITE |
1952                                                  IB_ACCESS_REMOTE_WRITE);
1953 #endif
1954
1955                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1956                                 fmr->fmr_frd  = frd;
1957                                 fmr->fmr_pfmr = NULL;
1958                                 fmr->fmr_pool = fpo;
1959                                 return 0;
1960                         }
1961                         spin_unlock(&fps->fps_lock);
1962                         rc = -EAGAIN;
1963                 }
1964
1965                 spin_lock(&fps->fps_lock);
1966                 fpo->fpo_map_count--;
1967                 if (rc != -EAGAIN) {
1968                         spin_unlock(&fps->fps_lock);
1969                         return rc;
1970                 }
1971
1972                 /* EAGAIN and ... */
1973                 if (version != fps->fps_version) {
1974                         spin_unlock(&fps->fps_lock);
1975                         goto again;
1976                 }
1977         }
1978
1979         if (fps->fps_increasing) {
1980                 spin_unlock(&fps->fps_lock);
1981                 CDEBUG(D_NET, "Another thread is allocating new "
1982                        "FMR pool, waiting for her to complete\n");
1983                 schedule();
1984                 goto again;
1985
1986         }
1987
1988         if (ktime_get_seconds() < fps->fps_next_retry) {
1989                 /* someone failed recently */
1990                 spin_unlock(&fps->fps_lock);
1991                 return -EAGAIN;
1992         }
1993
1994         fps->fps_increasing = 1;
1995         spin_unlock(&fps->fps_lock);
1996
1997         CDEBUG(D_NET, "Allocate new FMR pool\n");
1998         rc = kiblnd_create_fmr_pool(fps, &fpo);
1999         spin_lock(&fps->fps_lock);
2000         fps->fps_increasing = 0;
2001         if (rc == 0) {
2002                 fps->fps_version++;
2003                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2004         } else {
2005                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2006         }
2007         spin_unlock(&fps->fps_lock);
2008
2009         goto again;
2010 }
2011
2012 static void
2013 kiblnd_fini_pool(kib_pool_t *pool)
2014 {
2015         LASSERT(list_empty(&pool->po_free_list));
2016         LASSERT(pool->po_allocated == 0);
2017
2018         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2019 }
2020
2021 static void
2022 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2023 {
2024         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2025
2026         memset(pool, 0, sizeof(kib_pool_t));
2027         INIT_LIST_HEAD(&pool->po_free_list);
2028         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2029         pool->po_owner = ps;
2030         pool->po_size = size;
2031 }
2032
2033 static void
2034 kiblnd_destroy_pool_list(struct list_head *head)
2035 {
2036         kib_pool_t *pool;
2037
2038         while (!list_empty(head)) {
2039                 pool = list_entry(head->next, kib_pool_t, po_list);
2040                 list_del(&pool->po_list);
2041
2042                 LASSERT(pool->po_owner != NULL);
2043                 pool->po_owner->ps_pool_destroy(pool);
2044         }
2045 }
2046
2047 static void
2048 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2049 {
2050         if (ps->ps_net == NULL) /* intialized? */
2051                 return;
2052
2053         spin_lock(&ps->ps_lock);
2054         while (!list_empty(&ps->ps_pool_list)) {
2055                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2056                                             kib_pool_t, po_list);
2057                 po->po_failed = 1;
2058                 list_del(&po->po_list);
2059                 if (po->po_allocated == 0)
2060                         list_add(&po->po_list, zombies);
2061                 else
2062                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2063         }
2064         spin_unlock(&ps->ps_lock);
2065 }
2066
2067 static void
2068 kiblnd_fini_poolset(kib_poolset_t *ps)
2069 {
2070         if (ps->ps_net != NULL) { /* initialized? */
2071                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2072                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2073         }
2074 }
2075
2076 static int
2077 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2078                     kib_net_t *net, char *name, int size,
2079                     kib_ps_pool_create_t po_create,
2080                     kib_ps_pool_destroy_t po_destroy,
2081                     kib_ps_node_init_t nd_init,
2082                     kib_ps_node_fini_t nd_fini)
2083 {
2084         kib_pool_t      *pool;
2085         int             rc;
2086
2087         memset(ps, 0, sizeof(kib_poolset_t));
2088
2089         ps->ps_cpt          = cpt;
2090         ps->ps_net          = net;
2091         ps->ps_pool_create  = po_create;
2092         ps->ps_pool_destroy = po_destroy;
2093         ps->ps_node_init    = nd_init;
2094         ps->ps_node_fini    = nd_fini;
2095         ps->ps_pool_size    = size;
2096         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2097             >= sizeof(ps->ps_name))
2098                 return -E2BIG;
2099         spin_lock_init(&ps->ps_lock);
2100         INIT_LIST_HEAD(&ps->ps_pool_list);
2101         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2102
2103         rc = ps->ps_pool_create(ps, size, &pool);
2104         if (rc == 0)
2105                 list_add(&pool->po_list, &ps->ps_pool_list);
2106         else
2107                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2108
2109         return rc;
2110 }
2111
2112 static int
2113 kiblnd_pool_is_idle(kib_pool_t *pool, time64_t now)
2114 {
2115         if (pool->po_allocated != 0) /* still in use */
2116                 return 0;
2117         if (pool->po_failed)
2118                 return 1;
2119         return now >= pool->po_deadline;
2120 }
2121
2122 void
2123 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2124 {
2125         struct list_head zombies = LIST_HEAD_INIT(zombies);
2126         kib_poolset_t *ps = pool->po_owner;
2127         kib_pool_t *tmp;
2128         time64_t now = ktime_get_seconds();
2129
2130         spin_lock(&ps->ps_lock);
2131
2132         if (ps->ps_node_fini != NULL)
2133                 ps->ps_node_fini(pool, node);
2134
2135         LASSERT(pool->po_allocated > 0);
2136         list_add(node, &pool->po_free_list);
2137         pool->po_allocated--;
2138
2139         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2140                 /* the first pool is persistent */
2141                 if (ps->ps_pool_list.next == &pool->po_list)
2142                         continue;
2143
2144                 if (kiblnd_pool_is_idle(pool, now))
2145                         list_move(&pool->po_list, &zombies);
2146         }
2147         spin_unlock(&ps->ps_lock);
2148
2149         if (!list_empty(&zombies))
2150                 kiblnd_destroy_pool_list(&zombies);
2151 }
2152
2153 struct list_head *
2154 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2155 {
2156         struct list_head        *node;
2157         kib_pool_t              *pool;
2158         int                     rc;
2159         unsigned int            interval = 1;
2160         ktime_t time_before;
2161         unsigned int trips = 0;
2162
2163 again:
2164         spin_lock(&ps->ps_lock);
2165         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2166                 if (list_empty(&pool->po_free_list))
2167                         continue;
2168
2169                 pool->po_allocated++;
2170                 pool->po_deadline = ktime_get_seconds() +
2171                                     IBLND_POOL_DEADLINE;
2172                 node = pool->po_free_list.next;
2173                 list_del(node);
2174
2175                 if (ps->ps_node_init != NULL) {
2176                         /* still hold the lock */
2177                         ps->ps_node_init(pool, node);
2178                 }
2179                 spin_unlock(&ps->ps_lock);
2180                 return node;
2181         }
2182
2183         /* no available tx pool and ... */
2184         if (ps->ps_increasing) {
2185                 /* another thread is allocating a new pool */
2186                 spin_unlock(&ps->ps_lock);
2187                 trips++;
2188                 CDEBUG(D_NET, "Another thread is allocating new "
2189                        "%s pool, waiting %d HZs for her to complete."
2190                        "trips = %d\n",
2191                        ps->ps_name, interval, trips);
2192
2193                 set_current_state(TASK_INTERRUPTIBLE);
2194                 schedule_timeout(interval);
2195                 if (interval < cfs_time_seconds(1))
2196                         interval *= 2;
2197
2198                 goto again;
2199         }
2200
2201         if (ktime_get_seconds() < ps->ps_next_retry) {
2202                 /* someone failed recently */
2203                 spin_unlock(&ps->ps_lock);
2204                 return NULL;
2205         }
2206
2207         ps->ps_increasing = 1;
2208         spin_unlock(&ps->ps_lock);
2209
2210         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2211         time_before = ktime_get();
2212         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2213         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2214                ktime_ms_delta(ktime_get(), time_before));
2215
2216         spin_lock(&ps->ps_lock);
2217         ps->ps_increasing = 0;
2218         if (rc == 0) {
2219                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2220         } else {
2221                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2222                 CERROR("Can't allocate new %s pool because out of memory\n",
2223                        ps->ps_name);
2224         }
2225         spin_unlock(&ps->ps_lock);
2226
2227         goto again;
2228 }
2229
2230 static void
2231 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2232 {
2233         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2234         int             i;
2235
2236         LASSERT (pool->po_allocated == 0);
2237
2238         if (tpo->tpo_tx_pages != NULL) {
2239                 kiblnd_unmap_tx_pool(tpo);
2240                 kiblnd_free_pages(tpo->tpo_tx_pages);
2241         }
2242
2243         if (tpo->tpo_tx_descs == NULL)
2244                 goto out;
2245
2246         for (i = 0; i < pool->po_size; i++) {
2247                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2248                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2249
2250                 list_del(&tx->tx_list);
2251                 if (tx->tx_pages != NULL)
2252                         LIBCFS_FREE(tx->tx_pages,
2253                                     LNET_MAX_IOV *
2254                                     sizeof(*tx->tx_pages));
2255                 if (tx->tx_frags != NULL)
2256                         LIBCFS_FREE(tx->tx_frags,
2257                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2258                                     sizeof(*tx->tx_frags));
2259                 if (tx->tx_wrq != NULL)
2260                         LIBCFS_FREE(tx->tx_wrq,
2261                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2262                                     sizeof(*tx->tx_wrq));
2263                 if (tx->tx_sge != NULL)
2264                         LIBCFS_FREE(tx->tx_sge,
2265                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2266                                     sizeof(*tx->tx_sge));
2267                 if (tx->tx_rd != NULL)
2268                         LIBCFS_FREE(tx->tx_rd,
2269                                     offsetof(kib_rdma_desc_t,
2270                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2271         }
2272
2273         LIBCFS_FREE(tpo->tpo_tx_descs,
2274                     pool->po_size * sizeof(kib_tx_t));
2275 out:
2276         kiblnd_fini_pool(pool);
2277         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2278 }
2279
2280 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2281 {
2282         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2283         int ntx;
2284
2285         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2286         ntx = tunables->lnd_ntx / ncpts;
2287
2288         return max(IBLND_TX_POOL, ntx);
2289 }
2290
2291 static int
2292 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2293 {
2294         int            i;
2295         int            npg;
2296         kib_pool_t    *pool;
2297         kib_tx_pool_t *tpo;
2298
2299         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2300         if (tpo == NULL) {
2301                 CERROR("Failed to allocate TX pool\n");
2302                 return -ENOMEM;
2303         }
2304
2305         pool = &tpo->tpo_pool;
2306         kiblnd_init_pool(ps, pool, size);
2307         tpo->tpo_tx_descs = NULL;
2308         tpo->tpo_tx_pages = NULL;
2309
2310         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2311         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2312                 CERROR("Can't allocate tx pages: %d\n", npg);
2313                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2314                 return -ENOMEM;
2315         }
2316
2317         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2318                          size * sizeof(kib_tx_t));
2319         if (tpo->tpo_tx_descs == NULL) {
2320                 CERROR("Can't allocate %d tx descriptors\n", size);
2321                 ps->ps_pool_destroy(pool);
2322                 return -ENOMEM;
2323         }
2324
2325         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2326
2327         for (i = 0; i < size; i++) {
2328                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2329                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2330
2331                 tx->tx_pool = tpo;
2332                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2333                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2334                                          lnet_cpt_table(), ps->ps_cpt,
2335                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2336                         if (tx->tx_pages == NULL)
2337                                 break;
2338                 }
2339
2340                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2341                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2342                                  sizeof(*tx->tx_frags));
2343                 if (tx->tx_frags == NULL)
2344                         break;
2345
2346                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2347
2348                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2349                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2350                                  sizeof(*tx->tx_wrq));
2351                 if (tx->tx_wrq == NULL)
2352                         break;
2353
2354                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2355                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2356                                  sizeof(*tx->tx_sge));
2357                 if (tx->tx_sge == NULL)
2358                         break;
2359
2360                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2361                                  offsetof(kib_rdma_desc_t,
2362                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2363                 if (tx->tx_rd == NULL)
2364                         break;
2365         }
2366
2367         if (i == size) {
2368                 kiblnd_map_tx_pool(tpo);
2369                 *pp_po = pool;
2370                 return 0;
2371         }
2372
2373         ps->ps_pool_destroy(pool);
2374         return -ENOMEM;
2375 }
2376
2377 static void
2378 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2379 {
2380         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2381                                              tps_poolset);
2382         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2383
2384         tx->tx_cookie = tps->tps_next_tx_cookie++;
2385 }
2386
2387 static void
2388 kiblnd_net_fini_pools(kib_net_t *net)
2389 {
2390         int     i;
2391
2392         cfs_cpt_for_each(i, lnet_cpt_table()) {
2393                 kib_tx_poolset_t        *tps;
2394                 kib_fmr_poolset_t       *fps;
2395
2396                 if (net->ibn_tx_ps != NULL) {
2397                         tps = net->ibn_tx_ps[i];
2398                         kiblnd_fini_poolset(&tps->tps_poolset);
2399                 }
2400
2401                 if (net->ibn_fmr_ps != NULL) {
2402                         fps = net->ibn_fmr_ps[i];
2403                         kiblnd_fini_fmr_poolset(fps);
2404                 }
2405         }
2406
2407         if (net->ibn_tx_ps != NULL) {
2408                 cfs_percpt_free(net->ibn_tx_ps);
2409                 net->ibn_tx_ps = NULL;
2410         }
2411
2412         if (net->ibn_fmr_ps != NULL) {
2413                 cfs_percpt_free(net->ibn_fmr_ps);
2414                 net->ibn_fmr_ps = NULL;
2415         }
2416 }
2417
2418 static int
2419 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2420                       int ncpts)
2421 {
2422         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2423 #ifdef HAVE_IB_GET_DMA_MR
2424         unsigned long   flags;
2425 #endif
2426         int             cpt;
2427         int             rc;
2428         int             i;
2429
2430         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2431
2432 #ifdef HAVE_IB_GET_DMA_MR
2433         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2434         /*
2435          * if lnd_map_on_demand is zero then we have effectively disabled
2436          * FMR or FastReg and we're using global memory regions
2437          * exclusively.
2438          */
2439         if (!tunables->lnd_map_on_demand) {
2440                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2441                                            flags);
2442                 goto create_tx_pool;
2443         }
2444
2445         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2446 #endif
2447
2448         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2449                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2450                        tunables->lnd_fmr_pool_size,
2451                        tunables->lnd_ntx / 4);
2452                 rc = -EINVAL;
2453                 goto failed;
2454         }
2455
2456         /* TX pool must be created later than FMR, see LU-2268
2457          * for details */
2458         LASSERT(net->ibn_tx_ps == NULL);
2459
2460         /* premapping can fail if ibd_nmr > 1, so we always create
2461          * FMR pool and map-on-demand if premapping failed */
2462
2463         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2464                                            sizeof(kib_fmr_poolset_t));
2465         if (net->ibn_fmr_ps == NULL) {
2466                 CERROR("Failed to allocate FMR pool array\n");
2467                 rc = -ENOMEM;
2468                 goto failed;
2469         }
2470
2471         for (i = 0; i < ncpts; i++) {
2472                 cpt = (cpts == NULL) ? i : cpts[i];
2473                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2474                                              net, tunables);
2475                 if (rc != 0) {
2476                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2477                                cpt, rc);
2478                         goto failed;
2479                 }
2480         }
2481
2482         if (i > 0)
2483                 LASSERT(i == ncpts);
2484
2485 #ifdef HAVE_IB_GET_DMA_MR
2486  create_tx_pool:
2487 #endif
2488         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2489                                           sizeof(kib_tx_poolset_t));
2490         if (net->ibn_tx_ps == NULL) {
2491                 CERROR("Failed to allocate tx pool array\n");
2492                 rc = -ENOMEM;
2493                 goto failed;
2494         }
2495
2496         for (i = 0; i < ncpts; i++) {
2497                 cpt = (cpts == NULL) ? i : cpts[i];
2498                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2499                                          cpt, net, "TX",
2500                                          kiblnd_tx_pool_size(ni, ncpts),
2501                                          kiblnd_create_tx_pool,
2502                                          kiblnd_destroy_tx_pool,
2503                                          kiblnd_tx_init, NULL);
2504                 if (rc != 0) {
2505                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2506                                cpt, rc);
2507                         goto failed;
2508                 }
2509         }
2510
2511         return 0;
2512  failed:
2513         kiblnd_net_fini_pools(net);
2514         LASSERT(rc != 0);
2515         return rc;
2516 }
2517
2518 static int
2519 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2520 {
2521         struct ib_device_attr *dev_attr;
2522         int rc = 0;
2523
2524         /* It's safe to assume a HCA can handle a page size
2525          * matching that of the native system */
2526         hdev->ibh_page_shift = PAGE_SHIFT;
2527         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2528         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2529
2530 #ifndef HAVE_IB_DEVICE_ATTRS
2531         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2532         if (dev_attr == NULL) {
2533                 CERROR("Out of memory\n");
2534                 return -ENOMEM;
2535         }
2536
2537         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2538         if (rc != 0) {
2539                 CERROR("Failed to query IB device: %d\n", rc);
2540                 goto out_clean_attr;
2541         }
2542 #else
2543         dev_attr = &hdev->ibh_ibdev->attrs;
2544 #endif
2545
2546         hdev->ibh_mr_size = dev_attr->max_mr_size;
2547
2548         /* Setup device Memory Registration capabilities */
2549         if (hdev->ibh_ibdev->alloc_fmr &&
2550             hdev->ibh_ibdev->dealloc_fmr &&
2551             hdev->ibh_ibdev->map_phys_fmr &&
2552             hdev->ibh_ibdev->unmap_fmr) {
2553                 LCONSOLE_INFO("Using FMR for registration\n");
2554                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2555         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2556                 LCONSOLE_INFO("Using FastReg for registration\n");
2557                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2558 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2559 #ifdef IB_DEVICE_SG_GAPS_REG
2560                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2561                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2562 #endif
2563 #endif
2564         } else {
2565                 rc = -ENOSYS;
2566         }
2567
2568         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2569                 hdev->ibh_mr_shift = 64;
2570         else if (rc != 0)
2571                 rc = -EINVAL;
2572
2573 #ifndef HAVE_IB_DEVICE_ATTRS
2574 out_clean_attr:
2575         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2576 #endif
2577
2578         if (rc == -ENOSYS)
2579                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2580                        "register memory: %d\n", rc);
2581         else if (rc == -EINVAL)
2582                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2583         return rc;
2584 }
2585
2586 #ifdef HAVE_IB_GET_DMA_MR
2587 static void
2588 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2589 {
2590         if (hdev->ibh_mrs == NULL)
2591                 return;
2592
2593         ib_dereg_mr(hdev->ibh_mrs);
2594
2595         hdev->ibh_mrs = NULL;
2596 }
2597 #endif
2598
2599 void
2600 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2601 {
2602 #ifdef HAVE_IB_GET_DMA_MR
2603         kiblnd_hdev_cleanup_mrs(hdev);
2604 #endif
2605
2606         if (hdev->ibh_pd != NULL)
2607                 ib_dealloc_pd(hdev->ibh_pd);
2608
2609         if (hdev->ibh_cmid != NULL)
2610                 rdma_destroy_id(hdev->ibh_cmid);
2611
2612         LIBCFS_FREE(hdev, sizeof(*hdev));
2613 }
2614
2615 #ifdef HAVE_IB_GET_DMA_MR
2616 static int
2617 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2618 {
2619         struct ib_mr *mr;
2620         int           acflags = IB_ACCESS_LOCAL_WRITE |
2621                                 IB_ACCESS_REMOTE_WRITE;
2622
2623         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2624         if (IS_ERR(mr)) {
2625                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2626                 kiblnd_hdev_cleanup_mrs(hdev);
2627                 return PTR_ERR(mr);
2628         }
2629
2630         hdev->ibh_mrs = mr;
2631
2632         return 0;
2633 }
2634 #endif
2635
2636 static int
2637 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2638 {       /* DUMMY */
2639         return 0;
2640 }
2641
2642 static int
2643 kiblnd_dev_need_failover(kib_dev_t *dev)
2644 {
2645         struct rdma_cm_id  *cmid;
2646         struct sockaddr_in  srcaddr;
2647         struct sockaddr_in  dstaddr;
2648         int                 rc;
2649
2650         if (dev->ibd_hdev == NULL || /* initializing */
2651             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2652             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2653                 return 1;
2654
2655         /* XXX: it's UGLY, but I don't have better way to find
2656          * ib-bonding HCA failover because:
2657          *
2658          * a. no reliable CM event for HCA failover...
2659          * b. no OFED API to get ib_device for current net_device...
2660          *
2661          * We have only two choices at this point:
2662          *
2663          * a. rdma_bind_addr(), it will conflict with listener cmid
2664          * b. rdma_resolve_addr() to zero addr */
2665         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2666                                      IB_QPT_RC);
2667         if (IS_ERR(cmid)) {
2668                 rc = PTR_ERR(cmid);
2669                 CERROR("Failed to create cmid for failover: %d\n", rc);
2670                 return rc;
2671         }
2672
2673         memset(&srcaddr, 0, sizeof(srcaddr));
2674         srcaddr.sin_family      = AF_INET;
2675         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2676
2677         memset(&dstaddr, 0, sizeof(dstaddr));
2678         dstaddr.sin_family = AF_INET;
2679         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2680                                (struct sockaddr *)&dstaddr, 1);
2681         if (rc != 0 || cmid->device == NULL) {
2682                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2683                        dev->ibd_ifname, &dev->ibd_ifip,
2684                        cmid->device, rc);
2685                 rdma_destroy_id(cmid);
2686                 return rc;
2687         }
2688
2689         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2690         rdma_destroy_id(cmid);
2691         return rc;
2692 }
2693
2694 int
2695 kiblnd_dev_failover(kib_dev_t *dev)
2696 {
2697         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2698         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2699         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2700         struct rdma_cm_id  *cmid  = NULL;
2701         kib_hca_dev_t      *hdev  = NULL;
2702         kib_hca_dev_t      *old;
2703         struct ib_pd       *pd;
2704         kib_net_t          *net;
2705         struct sockaddr_in  addr;
2706         unsigned long       flags;
2707         int                 rc = 0;
2708         int                 i;
2709
2710         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2711                  dev->ibd_can_failover ||
2712                  dev->ibd_hdev == NULL);
2713
2714         rc = kiblnd_dev_need_failover(dev);
2715         if (rc <= 0)
2716                 goto out;
2717
2718         if (dev->ibd_hdev != NULL &&
2719             dev->ibd_hdev->ibh_cmid != NULL) {
2720                 /* XXX it's not good to close old listener at here,
2721                  * because we can fail to create new listener.
2722                  * But we have to close it now, otherwise rdma_bind_addr
2723                  * will return EADDRINUSE... How crap! */
2724                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2725
2726                 cmid = dev->ibd_hdev->ibh_cmid;
2727                 /* make next schedule of kiblnd_dev_need_failover()
2728                  * return 1 for me */
2729                 dev->ibd_hdev->ibh_cmid  = NULL;
2730                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2731
2732                 rdma_destroy_id(cmid);
2733         }
2734
2735         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2736                                      IB_QPT_RC);
2737         if (IS_ERR(cmid)) {
2738                 rc = PTR_ERR(cmid);
2739                 CERROR("Failed to create cmid for failover: %d\n", rc);
2740                 goto out;
2741         }
2742
2743         memset(&addr, 0, sizeof(addr));
2744         addr.sin_family      = AF_INET;
2745         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2746         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2747
2748         /* Bind to failover device or port */
2749         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2750         if (rc != 0 || cmid->device == NULL) {
2751                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2752                        dev->ibd_ifname, &dev->ibd_ifip,
2753                        cmid->device, rc);
2754                 rdma_destroy_id(cmid);
2755                 goto out;
2756         }
2757
2758         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2759         if (hdev == NULL) {
2760                 CERROR("Failed to allocate kib_hca_dev\n");
2761                 rdma_destroy_id(cmid);
2762                 rc = -ENOMEM;
2763                 goto out;
2764         }
2765
2766         atomic_set(&hdev->ibh_ref, 1);
2767         hdev->ibh_dev   = dev;
2768         hdev->ibh_cmid  = cmid;
2769         hdev->ibh_ibdev = cmid->device;
2770
2771 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2772         pd = ib_alloc_pd(cmid->device, 0);
2773 #else
2774         pd = ib_alloc_pd(cmid->device);
2775 #endif
2776         if (IS_ERR(pd)) {
2777                 rc = PTR_ERR(pd);
2778                 CERROR("Can't allocate PD: %d\n", rc);
2779                 goto out;
2780         }
2781
2782         hdev->ibh_pd = pd;
2783
2784         rc = rdma_listen(cmid, 0);
2785         if (rc != 0) {
2786                 CERROR("Can't start new listener: %d\n", rc);
2787                 goto out;
2788         }
2789
2790         rc = kiblnd_hdev_get_attr(hdev);
2791         if (rc != 0) {
2792                 CERROR("Can't get device attributes: %d\n", rc);
2793                 goto out;
2794         }
2795
2796 #ifdef HAVE_IB_GET_DMA_MR
2797         rc = kiblnd_hdev_setup_mrs(hdev);
2798         if (rc != 0) {
2799                 CERROR("Can't setup device: %d\n", rc);
2800                 goto out;
2801         }
2802 #endif
2803
2804         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2805
2806         old = dev->ibd_hdev;
2807         dev->ibd_hdev = hdev;   /* take over the refcount */
2808         hdev = old;
2809
2810         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2811                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2812                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2813                                             &zombie_tpo);
2814
2815                         if (net->ibn_fmr_ps != NULL)
2816                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2817                                                         &zombie_fpo);
2818                 }
2819         }
2820
2821         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2822  out:
2823         if (!list_empty(&zombie_tpo))
2824                 kiblnd_destroy_pool_list(&zombie_tpo);
2825         if (!list_empty(&zombie_ppo))
2826                 kiblnd_destroy_pool_list(&zombie_ppo);
2827         if (!list_empty(&zombie_fpo))
2828                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2829         if (hdev != NULL)
2830                 kiblnd_hdev_decref(hdev);
2831
2832         if (rc != 0)
2833                 dev->ibd_failed_failover++;
2834         else
2835                 dev->ibd_failed_failover = 0;
2836
2837         return rc;
2838 }
2839
2840 void
2841 kiblnd_destroy_dev (kib_dev_t *dev)
2842 {
2843         LASSERT (dev->ibd_nnets == 0);
2844         LASSERT(list_empty(&dev->ibd_nets));
2845
2846         list_del(&dev->ibd_fail_list);
2847         list_del(&dev->ibd_list);
2848
2849         if (dev->ibd_hdev != NULL)
2850                 kiblnd_hdev_decref(dev->ibd_hdev);
2851
2852         LIBCFS_FREE(dev, sizeof(*dev));
2853 }
2854
2855 static kib_dev_t *
2856 kiblnd_create_dev(char *ifname)
2857 {
2858         struct net_device *netdev;
2859         kib_dev_t         *dev;
2860         __u32              netmask;
2861         __u32              ip;
2862         int                up;
2863         int                rc;
2864
2865         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2866         if (rc != 0) {
2867                 CERROR("Can't query IPoIB interface %s: %d\n",
2868                        ifname, rc);
2869                 return NULL;
2870         }
2871
2872         if (!up) {
2873                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2874                 return NULL;
2875         }
2876
2877         LIBCFS_ALLOC(dev, sizeof(*dev));
2878         if (dev == NULL)
2879                 return NULL;
2880
2881         netdev = dev_get_by_name(&init_net, ifname);
2882         if (netdev == NULL) {
2883                 dev->ibd_can_failover = 0;
2884         } else {
2885                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2886                 dev_put(netdev);
2887         }
2888
2889         INIT_LIST_HEAD(&dev->ibd_nets);
2890         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2891         INIT_LIST_HEAD(&dev->ibd_fail_list);
2892         dev->ibd_ifip = ip;
2893         strcpy(&dev->ibd_ifname[0], ifname);
2894
2895         /* initialize the device */
2896         rc = kiblnd_dev_failover(dev);
2897         if (rc != 0) {
2898                 CERROR("Can't initialize device: %d\n", rc);
2899                 LIBCFS_FREE(dev, sizeof(*dev));
2900                 return NULL;
2901         }
2902
2903         list_add_tail(&dev->ibd_list,
2904                           &kiblnd_data.kib_devs);
2905         return dev;
2906 }
2907
2908 static void
2909 kiblnd_base_shutdown(void)
2910 {
2911         struct kib_sched_info   *sched;
2912         int                     i;
2913
2914         LASSERT(list_empty(&kiblnd_data.kib_devs));
2915
2916         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2917                atomic_read(&libcfs_kmemory));
2918
2919         switch (kiblnd_data.kib_init) {
2920         default:
2921                 LBUG();
2922
2923         case IBLND_INIT_ALL:
2924         case IBLND_INIT_DATA:
2925                 LASSERT (kiblnd_data.kib_peers != NULL);
2926                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2927                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2928                 }
2929                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2930                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2931                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2932                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2933
2934                 /* flag threads to terminate; wake and wait for them to die */
2935                 kiblnd_data.kib_shutdown = 1;
2936
2937                 /* NB: we really want to stop scheduler threads net by net
2938                  * instead of the whole module, this should be improved
2939                  * with dynamic configuration LNet */
2940                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2941                         wake_up_all(&sched->ibs_waitq);
2942
2943                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2944                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2945
2946                 i = 2;
2947                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2948                         i++;
2949                         /* power of 2? */
2950                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2951                                "Waiting for %d threads to terminate\n",
2952                                atomic_read(&kiblnd_data.kib_nthreads));
2953                         set_current_state(TASK_UNINTERRUPTIBLE);
2954                         schedule_timeout(cfs_time_seconds(1));
2955                 }
2956
2957                 /* fall through */
2958
2959         case IBLND_INIT_NOTHING:
2960                 break;
2961         }
2962
2963         if (kiblnd_data.kib_peers != NULL) {
2964                 LIBCFS_FREE(kiblnd_data.kib_peers,
2965                             sizeof(struct list_head) *
2966                             kiblnd_data.kib_peer_hash_size);
2967         }
2968
2969         if (kiblnd_data.kib_scheds != NULL)
2970                 cfs_percpt_free(kiblnd_data.kib_scheds);
2971
2972         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2973                atomic_read(&libcfs_kmemory));
2974
2975         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2976         module_put(THIS_MODULE);
2977 }
2978
2979 static void
2980 kiblnd_shutdown(struct lnet_ni *ni)
2981 {
2982         kib_net_t        *net = ni->ni_data;
2983         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2984         int               i;
2985         unsigned long     flags;
2986
2987         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2988
2989         if (net == NULL)
2990                 goto out;
2991
2992         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2993                atomic_read(&libcfs_kmemory));
2994
2995         write_lock_irqsave(g_lock, flags);
2996         net->ibn_shutdown = 1;
2997         write_unlock_irqrestore(g_lock, flags);
2998
2999         switch (net->ibn_init) {
3000         default:
3001                 LBUG();
3002
3003         case IBLND_INIT_ALL:
3004                 /* nuke all existing peers within this net */
3005                 kiblnd_del_peer(ni, LNET_NID_ANY);
3006
3007                 /* Wait for all peer_ni state to clean up */
3008                 i = 2;
3009                 while (atomic_read(&net->ibn_npeers) != 0) {
3010                         i++;
3011                         /* power of 2? */
3012                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3013                                "%s: waiting for %d peers to disconnect\n",
3014                                libcfs_nid2str(ni->ni_nid),
3015                                atomic_read(&net->ibn_npeers));
3016                         set_current_state(TASK_UNINTERRUPTIBLE);
3017                         schedule_timeout(cfs_time_seconds(1));
3018                 }
3019
3020                 kiblnd_net_fini_pools(net);
3021
3022                 write_lock_irqsave(g_lock, flags);
3023                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3024                 net->ibn_dev->ibd_nnets--;
3025                 list_del(&net->ibn_list);
3026                 write_unlock_irqrestore(g_lock, flags);
3027
3028                 /* fall through */
3029
3030         case IBLND_INIT_NOTHING:
3031                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3032
3033                 if (net->ibn_dev != NULL &&
3034                     net->ibn_dev->ibd_nnets == 0)
3035                         kiblnd_destroy_dev(net->ibn_dev);
3036
3037                 break;
3038         }
3039
3040         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3041                atomic_read(&libcfs_kmemory));
3042
3043         net->ibn_init = IBLND_INIT_NOTHING;
3044         ni->ni_data = NULL;
3045
3046         LIBCFS_FREE(net, sizeof(*net));
3047
3048 out:
3049         if (list_empty(&kiblnd_data.kib_devs))
3050                 kiblnd_base_shutdown();
3051         return;
3052 }
3053
3054 static int
3055 kiblnd_base_startup(void)
3056 {
3057         struct kib_sched_info   *sched;
3058         int                     rc;
3059         int                     i;
3060
3061         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3062
3063         try_module_get(THIS_MODULE);
3064         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3065
3066         rwlock_init(&kiblnd_data.kib_global_lock);
3067
3068         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3069         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3070
3071         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3072         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3073                      sizeof(struct list_head) *
3074                      kiblnd_data.kib_peer_hash_size);
3075         if (kiblnd_data.kib_peers == NULL)
3076                 goto failed;
3077
3078         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3079                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3080
3081         spin_lock_init(&kiblnd_data.kib_connd_lock);
3082         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3083         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3084         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3085         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3086
3087         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3088         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3089
3090         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3091                                                   sizeof(*sched));
3092         if (kiblnd_data.kib_scheds == NULL)
3093                 goto failed;
3094
3095         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3096                 int     nthrs;
3097
3098                 spin_lock_init(&sched->ibs_lock);
3099                 INIT_LIST_HEAD(&sched->ibs_conns);
3100                 init_waitqueue_head(&sched->ibs_waitq);
3101
3102                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3103                 if (*kiblnd_tunables.kib_nscheds > 0) {
3104                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3105                 } else {
3106                         /* max to half of CPUs, another half is reserved for
3107                          * upper layer modules */
3108                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3109                 }
3110
3111                 sched->ibs_nthreads_max = nthrs;
3112                 sched->ibs_cpt = i;
3113         }
3114
3115         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3116
3117         /* lists/ptrs/locks initialised */
3118         kiblnd_data.kib_init = IBLND_INIT_DATA;
3119         /*****************************************************/
3120
3121         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3122         if (rc != 0) {
3123                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3124                 goto failed;
3125         }
3126
3127         if (*kiblnd_tunables.kib_dev_failover != 0)
3128                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3129                                          "kiblnd_failover");
3130
3131         if (rc != 0) {
3132                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3133                 goto failed;
3134         }
3135
3136         /* flag everything initialised */
3137         kiblnd_data.kib_init = IBLND_INIT_ALL;
3138         /*****************************************************/
3139
3140         return 0;
3141
3142  failed:
3143         kiblnd_base_shutdown();
3144         return -ENETDOWN;
3145 }
3146
3147 static int
3148 kiblnd_start_schedulers(struct kib_sched_info *sched)
3149 {
3150         int     rc = 0;
3151         int     nthrs;
3152         int     i;
3153
3154         if (sched->ibs_nthreads == 0) {
3155                 if (*kiblnd_tunables.kib_nscheds > 0) {
3156                         nthrs = sched->ibs_nthreads_max;
3157                 } else {
3158                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3159                                                sched->ibs_cpt);
3160                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3161                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3162                 }
3163         } else {
3164                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3165                 /* increase one thread if there is new interface */
3166                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3167         }
3168
3169         for (i = 0; i < nthrs; i++) {
3170                 long    id;
3171                 char    name[20];
3172                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3173                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3174                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3175                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3176                 if (rc == 0)
3177                         continue;
3178
3179                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3180                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3181                 break;
3182         }
3183
3184         sched->ibs_nthreads += i;
3185         return rc;
3186 }
3187
3188 static int
3189 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3190 {
3191         int     cpt;
3192         int     rc;
3193         int     i;
3194
3195         for (i = 0; i < ncpts; i++) {
3196                 struct kib_sched_info *sched;
3197
3198                 cpt = (cpts == NULL) ? i : cpts[i];
3199                 sched = kiblnd_data.kib_scheds[cpt];
3200
3201                 if (!newdev && sched->ibs_nthreads > 0)
3202                         continue;
3203
3204                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3205                 if (rc != 0) {
3206                         CERROR("Failed to start scheduler threads for %s\n",
3207                                dev->ibd_ifname);
3208                         return rc;
3209                 }
3210         }
3211         return 0;
3212 }
3213
3214 static kib_dev_t *
3215 kiblnd_dev_search(char *ifname)
3216 {
3217         kib_dev_t       *alias = NULL;
3218         kib_dev_t       *dev;
3219         char            *colon;
3220         char            *colon2;
3221
3222         colon = strchr(ifname, ':');
3223         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3224                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3225                         return dev;
3226
3227                 if (alias != NULL)
3228                         continue;
3229
3230                 colon2 = strchr(dev->ibd_ifname, ':');
3231                 if (colon != NULL)
3232                         *colon = 0;
3233                 if (colon2 != NULL)
3234                         *colon2 = 0;
3235
3236                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3237                         alias = dev;
3238
3239                 if (colon != NULL)
3240                         *colon = ':';
3241                 if (colon2 != NULL)
3242                         *colon2 = ':';
3243         }
3244         return alias;
3245 }
3246
3247 static int
3248 kiblnd_startup(struct lnet_ni *ni)
3249 {
3250         char                     *ifname;
3251         kib_dev_t                *ibdev = NULL;
3252         kib_net_t                *net;
3253         unsigned long             flags;
3254         int                       rc;
3255         int                       newdev;
3256         int                       node_id;
3257
3258         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3259
3260         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3261                 rc = kiblnd_base_startup();
3262                 if (rc != 0)
3263                         return rc;
3264         }
3265
3266         LIBCFS_ALLOC(net, sizeof(*net));
3267         ni->ni_data = net;
3268         if (net == NULL)
3269                 goto failed;
3270
3271         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3272
3273         kiblnd_tunables_setup(ni);
3274
3275         if (ni->ni_interfaces[0] != NULL) {
3276                 /* Use the IPoIB interface specified in 'networks=' */
3277
3278                 CLASSERT(LNET_INTERFACES_NUM > 1);
3279                 if (ni->ni_interfaces[1] != NULL) {
3280                         CERROR("Multiple interfaces not supported\n");
3281                         goto failed;
3282                 }
3283
3284                 ifname = ni->ni_interfaces[0];
3285         } else {
3286                 ifname = *kiblnd_tunables.kib_default_ipif;
3287         }
3288
3289         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3290                 CERROR("IPoIB interface name too long: %s\n", ifname);
3291                 goto failed;
3292         }
3293
3294         ibdev = kiblnd_dev_search(ifname);
3295
3296         newdev = ibdev == NULL;
3297         /* hmm...create kib_dev even for alias */
3298         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3299                 ibdev = kiblnd_create_dev(ifname);
3300
3301         if (ibdev == NULL)
3302                 goto failed;
3303
3304         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3305         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3306
3307         net->ibn_dev = ibdev;
3308         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3309
3310         rc = kiblnd_dev_start_threads(ibdev, newdev,
3311                                       ni->ni_cpts, ni->ni_ncpts);
3312         if (rc != 0)
3313                 goto failed;
3314
3315         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3316         if (rc != 0) {
3317                 CERROR("Failed to initialize NI pools: %d\n", rc);
3318                 goto failed;
3319         }
3320
3321         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3322         ibdev->ibd_nnets++;
3323         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3324         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3325
3326         net->ibn_init = IBLND_INIT_ALL;
3327
3328         return 0;
3329
3330 failed:
3331         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3332                 kiblnd_destroy_dev(ibdev);
3333
3334         kiblnd_shutdown(ni);
3335
3336         CDEBUG(D_NET, "kiblnd_startup failed\n");
3337         return -ENETDOWN;
3338 }
3339
3340 static struct lnet_lnd the_o2iblnd = {
3341         .lnd_type       = O2IBLND,
3342         .lnd_startup    = kiblnd_startup,
3343         .lnd_shutdown   = kiblnd_shutdown,
3344         .lnd_ctl        = kiblnd_ctl,
3345         .lnd_query      = kiblnd_query,
3346         .lnd_send       = kiblnd_send,
3347         .lnd_recv       = kiblnd_recv,
3348 };
3349
3350 static void __exit ko2iblnd_exit(void)
3351 {
3352         lnet_unregister_lnd(&the_o2iblnd);
3353 }
3354
3355 static int __init ko2iblnd_init(void)
3356 {
3357         int rc;
3358
3359         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3360         CLASSERT(offsetof(kib_msg_t,
3361                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3362                  IBLND_MSG_SIZE);
3363         CLASSERT(offsetof(kib_msg_t,
3364                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3365                  <= IBLND_MSG_SIZE);
3366
3367         rc = kiblnd_tunables_init();
3368         if (rc != 0)
3369                 return rc;
3370
3371         lnet_register_lnd(&the_o2iblnd);
3372
3373         return 0;
3374 }
3375
3376 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3377 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3378 MODULE_VERSION("2.8.0");
3379 MODULE_LICENSE("GPL");
3380
3381 module_init(ko2iblnd_init);
3382 module_exit(ko2iblnd_exit);