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 <linux/inetdevice.h>
  39
  40 #include "o2iblnd.h"
  41
  42 static const struct lnet_lnd the_o2iblnd;
  43
  44 struct kib_data kiblnd_data;
  45
  46 static __u32
  47 kiblnd_cksum (void *ptr, int nob)
  48 {
  49         char  *c  = ptr;
  50         __u32  sum = 0;
  51
  52         while (nob-- > 0)
  53                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  54
  55         /* ensure I don't return 0 (== no checksum) */
  56         return (sum == 0) ? 1 : sum;
  57 }
  58
  59 static char *
  60 kiblnd_msgtype2str(int type)
  61 {
  62         switch (type) {
  63         case IBLND_MSG_CONNREQ:
  64                 return "CONNREQ";
  65
  66         case IBLND_MSG_CONNACK:
  67                 return "CONNACK";
  68
  69         case IBLND_MSG_NOOP:
  70                 return "NOOP";
  71
  72         case IBLND_MSG_IMMEDIATE:
  73                 return "IMMEDIATE";
  74
  75         case IBLND_MSG_PUT_REQ:
  76                 return "PUT_REQ";
  77
  78         case IBLND_MSG_PUT_NAK:
  79                 return "PUT_NAK";
  80
  81         case IBLND_MSG_PUT_ACK:
  82                 return "PUT_ACK";
  83
  84         case IBLND_MSG_PUT_DONE:
  85                 return "PUT_DONE";
  86
  87         case IBLND_MSG_GET_REQ:
  88                 return "GET_REQ";
  89
  90         case IBLND_MSG_GET_DONE:
  91                 return "GET_DONE";
  92
  93         default:
  94                 return "???";
  95         }
  96 }
  97
  98 static int
  99 kiblnd_msgtype2size(int type)
 100 {
 101         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 102
 103         switch (type) {
 104         case IBLND_MSG_CONNREQ:
 105         case IBLND_MSG_CONNACK:
 106                 return hdr_size + sizeof(struct kib_connparams);
 107
 108         case IBLND_MSG_NOOP:
 109                 return hdr_size;
 110
 111         case IBLND_MSG_IMMEDIATE:
 112                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 113
 114         case IBLND_MSG_PUT_REQ:
 115                 return hdr_size + sizeof(struct kib_putreq_msg);
 116
 117         case IBLND_MSG_PUT_ACK:
 118                 return hdr_size + sizeof(struct kib_putack_msg);
 119
 120         case IBLND_MSG_GET_REQ:
 121                 return hdr_size + sizeof(struct kib_get_msg);
 122
 123         case IBLND_MSG_PUT_NAK:
 124         case IBLND_MSG_PUT_DONE:
 125         case IBLND_MSG_GET_DONE:
 126                 return hdr_size + sizeof(struct kib_completion_msg);
 127         default:
 128                 return -1;
 129         }
 130 }
 131
 132 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
 133 {
 134         struct kib_rdma_desc *rd;
 135         int                nob;
 136         int                n;
 137         int                i;
 138
 139         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 140                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 141
 142         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 143                               &msg->ibm_u.get.ibgm_rd :
 144                               &msg->ibm_u.putack.ibpam_rd;
 145
 146         if (flip) {
 147                 __swab32s(&rd->rd_key);
 148                 __swab32s(&rd->rd_nfrags);
 149         }
 150
 151         n = rd->rd_nfrags;
 152
 153         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 154                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 155                        n, IBLND_MAX_RDMA_FRAGS);
 156                 return 1;
 157         }
 158
 159         nob = offsetof(struct kib_msg, ibm_u) +
 160               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 161
 162         if (msg->ibm_nob < nob) {
 163                 CERROR("Short %s: %d(%d)\n",
 164                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 165                 return 1;
 166         }
 167
 168         if (!flip)
 169                 return 0;
 170
 171         for (i = 0; i < n; i++) {
 172                 __swab32s(&rd->rd_frags[i].rf_nob);
 173                 __swab64s(&rd->rd_frags[i].rf_addr);
 174         }
 175
 176         return 0;
 177 }
 178
 179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
 180                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         struct kib_net *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 kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 204 {
 205         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 206         __u32     msg_cksum;
 207         __u16     version;
 208         int       msg_nob;
 209         int       flip;
 210
 211         /* 6 bytes are enough to have received magic + version */
 212         if (nob < 6) {
 213                 CERROR("Short message: %d\n", nob);
 214                 return -EPROTO;
 215         }
 216
 217         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 218                 flip = 0;
 219         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 220                 flip = 1;
 221         } else {
 222                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 223                 return -EPROTO;
 224         }
 225
 226         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 227         if (version != IBLND_MSG_VERSION &&
 228             version != IBLND_MSG_VERSION_1) {
 229                 CERROR("Bad version: %x\n", version);
 230                 return -EPROTO;
 231         }
 232
 233         if (nob < hdr_size) {
 234                 CERROR("Short message: %d\n", nob);
 235                 return -EPROTO;
 236         }
 237
 238         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 239         if (msg_nob > nob) {
 240                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 241                 return -EPROTO;
 242         }
 243
 244         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 245          * gets flipped */
 246         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 247         msg->ibm_cksum = 0;
 248         if (msg_cksum != 0 &&
 249             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 250                 CERROR("Bad checksum\n");
 251                 return -EPROTO;
 252         }
 253
 254         msg->ibm_cksum = msg_cksum;
 255
 256         if (flip) {
 257                 /* leave magic unflipped as a clue to peer_ni endianness */
 258                 msg->ibm_version = version;
 259                 BUILD_BUG_ON(sizeof(msg->ibm_type) != 1);
 260                 BUILD_BUG_ON(sizeof(msg->ibm_credits) != 1);
 261                 msg->ibm_nob     = msg_nob;
 262                 __swab64s(&msg->ibm_srcnid);
 263                 __swab64s(&msg->ibm_srcstamp);
 264                 __swab64s(&msg->ibm_dstnid);
 265                 __swab64s(&msg->ibm_dststamp);
 266         }
 267
 268         if (msg->ibm_srcnid == LNET_NID_ANY) {
 269                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 270                 return -EPROTO;
 271         }
 272
 273         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 274                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 275                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 276                 return -EPROTO;
 277         }
 278
 279         switch (msg->ibm_type) {
 280         default:
 281                 CERROR("Unknown message type %x\n", msg->ibm_type);
 282                 return -EPROTO;
 283
 284         case IBLND_MSG_NOOP:
 285         case IBLND_MSG_IMMEDIATE:
 286         case IBLND_MSG_PUT_REQ:
 287                 break;
 288
 289         case IBLND_MSG_PUT_ACK:
 290         case IBLND_MSG_GET_REQ:
 291                 if (kiblnd_unpack_rd(msg, flip))
 292                         return -EPROTO;
 293                 break;
 294
 295         case IBLND_MSG_PUT_NAK:
 296         case IBLND_MSG_PUT_DONE:
 297         case IBLND_MSG_GET_DONE:
 298                 if (flip)
 299                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 300                 break;
 301
 302         case IBLND_MSG_CONNREQ:
 303         case IBLND_MSG_CONNACK:
 304                 if (flip) {
 305                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 306                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 307                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 308                 }
 309                 break;
 310         }
 311         return 0;
 312 }
 313
 314 int
 315 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
 316                    lnet_nid_t nid)
 317 {
 318         struct kib_peer_ni *peer_ni;
 319         struct kib_net *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(struct kib_peer_ni *peer_ni)
 360 {
 361         struct kib_net *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         if (atomic_dec_and_test(&net->ibn_npeers))
 376                 wake_up_var(&net->ibn_npeers);
 377 }
 378
 379 struct kib_peer_ni *
 380 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 381 {
 382         /* the caller is responsible for accounting the additional reference
 383          * that this creates */
 384         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 385         struct list_head        *tmp;
 386         struct kib_peer_ni              *peer_ni;
 387
 388         list_for_each(tmp, peer_list) {
 389
 390                 peer_ni = list_entry(tmp, struct kib_peer_ni, ibp_list);
 391                 LASSERT(!kiblnd_peer_idle(peer_ni));
 392
 393                 /*
 394                  * Match a peer if its NID and the NID of the local NI it
 395                  * communicates over are the same. Otherwise don't match
 396                  * the peer, which will result in a new lnd peer being
 397                  * created.
 398                  */
 399                 if (peer_ni->ibp_nid != nid ||
 400                     peer_ni->ibp_ni->ni_nid != ni->ni_nid)
 401                         continue;
 402
 403                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 404                        peer_ni, libcfs_nid2str(nid),
 405                        atomic_read(&peer_ni->ibp_refcount),
 406                        peer_ni->ibp_version);
 407                 return peer_ni;
 408         }
 409         return NULL;
 410 }
 411
 412 void
 413 kiblnd_unlink_peer_locked(struct kib_peer_ni *peer_ni)
 414 {
 415         LASSERT(list_empty(&peer_ni->ibp_conns));
 416
 417         LASSERT (kiblnd_peer_active(peer_ni));
 418         list_del_init(&peer_ni->ibp_list);
 419         /* lose peerlist's ref */
 420         kiblnd_peer_decref(peer_ni);
 421 }
 422
 423 static int
 424 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
 425                      lnet_nid_t *nidp, int *count)
 426 {
 427         struct kib_peer_ni              *peer_ni;
 428         struct list_head        *ptmp;
 429         int                      i;
 430         unsigned long            flags;
 431
 432         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 433
 434         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 435
 436                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 437
 438                         peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
 439                         LASSERT(!kiblnd_peer_idle(peer_ni));
 440
 441                         if (peer_ni->ibp_ni != ni)
 442                                 continue;
 443
 444                         if (index-- > 0)
 445                                 continue;
 446
 447                         *nidp = peer_ni->ibp_nid;
 448                         *count = atomic_read(&peer_ni->ibp_refcount);
 449
 450                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 451                                                flags);
 452                         return 0;
 453                 }
 454         }
 455
 456         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 457         return -ENOENT;
 458 }
 459
 460 static void
 461 kiblnd_del_peer_locked(struct kib_peer_ni *peer_ni)
 462 {
 463         struct list_head *ctmp;
 464         struct list_head *cnxt;
 465         struct kib_conn *conn;
 466
 467         if (list_empty(&peer_ni->ibp_conns)) {
 468                 kiblnd_unlink_peer_locked(peer_ni);
 469         } else {
 470                 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
 471                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
 472
 473                         kiblnd_close_conn_locked(conn, 0);
 474                 }
 475                 /* NB closing peer_ni's last conn unlinked it. */
 476         }
 477         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 478          * last ref on it. */
 479 }
 480
 481 static int
 482 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 483 {
 484         LIST_HEAD(zombies);
 485         struct list_head        *ptmp;
 486         struct list_head        *pnxt;
 487         struct kib_peer_ni              *peer_ni;
 488         int                     lo;
 489         int                     hi;
 490         int                     i;
 491         unsigned long           flags;
 492         int                     rc = -ENOENT;
 493
 494         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 495
 496         if (nid != LNET_NID_ANY) {
 497                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 498         } else {
 499                 lo = 0;
 500                 hi = kiblnd_data.kib_peer_hash_size - 1;
 501         }
 502
 503         for (i = lo; i <= hi; i++) {
 504                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 505                         peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
 506                         LASSERT(!kiblnd_peer_idle(peer_ni));
 507
 508                         if (peer_ni->ibp_ni != ni)
 509                                 continue;
 510
 511                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 512                                 continue;
 513
 514                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 515                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 516
 517                                 list_splice_init(&peer_ni->ibp_tx_queue,
 518                                                  &zombies);
 519                         }
 520
 521                         kiblnd_del_peer_locked(peer_ni);
 522                         rc = 0;         /* matched something */
 523                 }
 524         }
 525
 526         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 527
 528         kiblnd_txlist_done(&zombies, -EIO, LNET_MSG_STATUS_LOCAL_ERROR);
 529
 530         return rc;
 531 }
 532
 533 static struct kib_conn *
 534 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 535 {
 536         struct kib_peer_ni              *peer_ni;
 537         struct list_head        *ptmp;
 538         struct kib_conn *conn;
 539         struct list_head        *ctmp;
 540         int                     i;
 541         unsigned long           flags;
 542
 543         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 544
 545         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 546                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 547
 548                         peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
 549                         LASSERT(!kiblnd_peer_idle(peer_ni));
 550
 551                         if (peer_ni->ibp_ni != ni)
 552                                 continue;
 553
 554                         list_for_each(ctmp, &peer_ni->ibp_conns) {
 555                                 if (index-- > 0)
 556                                         continue;
 557
 558                                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 559                                 kiblnd_conn_addref(conn);
 560                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 561                                                        flags);
 562                                 return conn;
 563                         }
 564                 }
 565         }
 566
 567         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 568         return NULL;
 569 }
 570
 571 static void
 572 kiblnd_debug_rx(struct kib_rx *rx)
 573 {
 574         CDEBUG(D_CONSOLE, "      %p msg_type %x cred %d\n",
 575                rx, rx->rx_msg->ibm_type,
 576                rx->rx_msg->ibm_credits);
 577 }
 578
 579 static void
 580 kiblnd_debug_tx(struct kib_tx *tx)
 581 {
 582         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 583                "cookie %#llx msg %s%s type %x cred %d\n",
 584                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 585                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 586                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 587                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 588                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 589 }
 590
 591 void
 592 kiblnd_debug_conn(struct kib_conn *conn)
 593 {
 594         struct list_head        *tmp;
 595         int                     i;
 596
 597         spin_lock(&conn->ibc_lock);
 598
 599         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 600                atomic_read(&conn->ibc_refcount), conn,
 601                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 602         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 603                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 604                conn->ibc_nsends_posted, conn->ibc_credits,
 605                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 606         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 607
 608         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 609         list_for_each(tmp, &conn->ibc_early_rxs)
 610                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 611
 612         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 613         list_for_each(tmp, &conn->ibc_tx_noops)
 614                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 615
 616         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 617         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 618                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 619
 620         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 621         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 622                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 623
 624         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 625         list_for_each(tmp, &conn->ibc_tx_queue)
 626                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 627
 628         CDEBUG(D_CONSOLE, "   active_txs:\n");
 629         list_for_each(tmp, &conn->ibc_active_txs)
 630                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 631
 632         CDEBUG(D_CONSOLE, "   rxs:\n");
 633         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 634                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 635
 636         spin_unlock(&conn->ibc_lock);
 637 }
 638
 639 int
 640 kiblnd_translate_mtu(int value)
 641 {
 642         switch (value) {
 643         default:
 644                 return -1;
 645         case 0:
 646                 return 0;
 647         case 256:
 648                 return IB_MTU_256;
 649         case 512:
 650                 return IB_MTU_512;
 651         case 1024:
 652                 return IB_MTU_1024;
 653         case 2048:
 654                 return IB_MTU_2048;
 655         case 4096:
 656                 return IB_MTU_4096;
 657         }
 658 }
 659
 660 static void
 661 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 662 {
 663         int           mtu;
 664
 665         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 666         if (cmid->route.path_rec == NULL)
 667                 return;
 668
 669         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 670         LASSERT (mtu >= 0);
 671         if (mtu != 0)
 672                 cmid->route.path_rec->mtu = mtu;
 673 }
 674
 675 static int
 676 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 677 {
 678         cpumask_var_t   *mask;
 679         int             vectors;
 680         int             off;
 681         int             i;
 682         lnet_nid_t      ibp_nid;
 683
 684         vectors = conn->ibc_cmid->device->num_comp_vectors;
 685         if (vectors <= 1)
 686                 return 0;
 687
 688         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 689
 690         /* hash NID to CPU id in this partition... */
 691         ibp_nid = conn->ibc_peer->ibp_nid;
 692         off = do_div(ibp_nid, cpumask_weight(*mask));
 693         for_each_cpu(i, *mask) {
 694                 if (off-- == 0)
 695                         return i % vectors;
 696         }
 697
 698         LBUG();
 699         return 1;
 700 }
 701
 702 /*
 703  * Get the scheduler bound to this CPT. If the scheduler has no
 704  * threads, which means that the CPT has no CPUs, then grab the
 705  * next scheduler that we can use.
 706  *
 707  * This case would be triggered if a NUMA node is configured with
 708  * no associated CPUs.
 709  */
 710 static struct kib_sched_info *
 711 kiblnd_get_scheduler(int cpt)
 712 {
 713         struct kib_sched_info *sched;
 714         int i;
 715
 716         sched = kiblnd_data.kib_scheds[cpt];
 717
 718         if (sched->ibs_nthreads > 0)
 719                 return sched;
 720
 721         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 722                 if (sched->ibs_nthreads > 0) {
 723                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 724                                         cpt, sched->ibs_cpt);
 725                         return sched;
 726                 }
 727         }
 728
 729         return NULL;
 730 }
 731
 732 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 733 {
 734         /*
 735          * One WR for the LNet message
 736          * And ibc_max_frags for the transfer WRs
 737          */
 738         int ret;
 739         int multiplier = 1 + conn->ibc_max_frags;
 740         enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
 741
 742         /* FastReg needs two extra WRs for map and invalidate */
 743         if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
 744                 multiplier += 2;
 745
 746         /* account for a maximum of ibc_queue_depth in-flight transfers */
 747         ret = multiplier * conn->ibc_queue_depth;
 748
 749         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 750                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 751                        "request size %d larger than maximum %d device "
 752                        "can handle\n", conn->ibc_queue_depth, ret,
 753                        conn->ibc_hdev->ibh_max_qp_wr);
 754                 conn->ibc_queue_depth =
 755                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 756         }
 757
 758         /* don't go beyond the maximum the device can handle */
 759         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 760 }
 761
 762 struct kib_conn *
 763 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 764                    int state, int version)
 765 {
 766         /* CAVEAT EMPTOR:
 767          * If the new conn is created successfully it takes over the caller's
 768          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 769          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 770          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 771          * to destroy 'cmid' here since I'm called from the CM which still has
 772          * its ref on 'cmid'). */
 773         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 774         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 775         struct kib_dev *dev;
 776         struct ib_qp_init_attr init_qp_attr = {};
 777         struct kib_sched_info   *sched;
 778 #ifdef HAVE_IB_CQ_INIT_ATTR
 779         struct ib_cq_init_attr  cq_attr = {};
 780 #endif
 781         struct kib_conn *conn;
 782         struct ib_cq            *cq;
 783         unsigned long           flags;
 784         int                     cpt;
 785         int                     rc;
 786         int                     i;
 787
 788         LASSERT(net != NULL);
 789         LASSERT(!in_interrupt());
 790
 791         dev = net->ibn_dev;
 792
 793         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 794         sched = kiblnd_get_scheduler(cpt);
 795
 796         if (sched == NULL) {
 797                 CERROR("no schedulers available. node is unhealthy\n");
 798                 goto failed_0;
 799         }
 800
 801         /*
 802          * The cpt might have changed if we ended up selecting a non cpt
 803          * native scheduler. So use the scheduler's cpt instead.
 804          */
 805         cpt = sched->ibs_cpt;
 806
 807         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 808         if (conn == NULL) {
 809                 CERROR("Can't allocate connection for %s\n",
 810                        libcfs_nid2str(peer_ni->ibp_nid));
 811                 goto failed_0;
 812         }
 813
 814         conn->ibc_state = IBLND_CONN_INIT;
 815         conn->ibc_version = version;
 816         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 817         cmid->context = conn;                   /* for future CM callbacks */
 818         conn->ibc_cmid = cmid;
 819         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 820         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 821         conn->ibc_rxs = NULL;
 822         conn->ibc_rx_pages = NULL;
 823
 824         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 825         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 826         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 827         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 828         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 829         INIT_LIST_HEAD(&conn->ibc_active_txs);
 830         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 831         spin_lock_init(&conn->ibc_lock);
 832
 833         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 834                          sizeof(*conn->ibc_connvars));
 835         if (conn->ibc_connvars == NULL) {
 836                 CERROR("Can't allocate in-progress connection state\n");
 837                 goto failed_2;
 838         }
 839
 840         write_lock_irqsave(glock, flags);
 841         if (dev->ibd_failover) {
 842                 write_unlock_irqrestore(glock, flags);
 843                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 844                 goto failed_2;
 845         }
 846
 847         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 848                 /* wakeup failover thread and teardown connection */
 849                 if (kiblnd_dev_can_failover(dev)) {
 850                         list_add_tail(&dev->ibd_fail_list,
 851                                       &kiblnd_data.kib_failed_devs);
 852                         wake_up(&kiblnd_data.kib_failover_waitq);
 853                 }
 854
 855                 write_unlock_irqrestore(glock, flags);
 856                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 857                        cmid->device->name, dev->ibd_ifname);
 858                 goto failed_2;
 859         }
 860
 861         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 862         conn->ibc_hdev = dev->ibd_hdev;
 863
 864         kiblnd_setup_mtu_locked(cmid);
 865
 866         write_unlock_irqrestore(glock, flags);
 867
 868 #ifdef HAVE_IB_CQ_INIT_ATTR
 869         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 870         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 871         cq = ib_create_cq(cmid->device,
 872                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 873                           &cq_attr);
 874 #else
 875         cq = ib_create_cq(cmid->device,
 876                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 877                           IBLND_CQ_ENTRIES(conn),
 878                           kiblnd_get_completion_vector(conn, cpt));
 879 #endif
 880         if (IS_ERR(cq)) {
 881                 /*
 882                  * on MLX-5 (possibly MLX-4 as well) this error could be
 883                  * hit if the concurrent_sends and/or peer_tx_credits is set
 884                  * too high. Or due to an MLX-5 bug which tries to
 885                  * allocate 256kb via kmalloc for WR cookie array
 886                  */
 887                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 888                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 889                 goto failed_2;
 890         }
 891
 892         conn->ibc_cq = cq;
 893
 894         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 895         if (rc != 0) {
 896                 CERROR("Can't request completion notification: %d\n", rc);
 897                 goto failed_2;
 898         }
 899
 900         init_qp_attr.event_handler = kiblnd_qp_event;
 901         init_qp_attr.qp_context = conn;
 902         init_qp_attr.cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 903         init_qp_attr.cap.max_recv_sge = 1;
 904         init_qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 905         init_qp_attr.qp_type = IB_QPT_RC;
 906         init_qp_attr.send_cq = cq;
 907         init_qp_attr.recv_cq = cq;
 908         /*
 909          * kiblnd_send_wrs() can change the connection's queue depth if
 910          * the maximum work requests for the device is maxed out
 911          */
 912         init_qp_attr.cap.max_send_wr = kiblnd_send_wrs(conn);
 913         init_qp_attr.cap.max_recv_wr = IBLND_RECV_WRS(conn);
 914
 915         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, &init_qp_attr);
 916         if (rc) {
 917                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 918                        "send_sge: %d, recv_sge: %d\n",
 919                        rc, init_qp_attr.cap.max_send_wr,
 920                        init_qp_attr.cap.max_recv_wr,
 921                        init_qp_attr.cap.max_send_sge,
 922                        init_qp_attr.cap.max_recv_sge);
 923                 goto failed_2;
 924         }
 925
 926         conn->ibc_sched = sched;
 927
 928         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 929                 CWARN("peer %s - queue depth reduced from %u to %u"
 930                       "  to allow for qp creation\n",
 931                       libcfs_nid2str(peer_ni->ibp_nid),
 932                       peer_ni->ibp_queue_depth,
 933                       conn->ibc_queue_depth);
 934
 935         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 936                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 937         if (conn->ibc_rxs == NULL) {
 938                 CERROR("Cannot allocate RX buffers\n");
 939                 goto failed_2;
 940         }
 941
 942         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 943                                 IBLND_RX_MSG_PAGES(conn));
 944         if (rc != 0)
 945                 goto failed_2;
 946
 947         kiblnd_map_rx_descs(conn);
 948
 949         /* 1 ref for caller and each rxmsg */
 950         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 951         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 952
 953         /* post receives */
 954         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 955                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 956                 if (rc != 0) {
 957                         CERROR("Can't post rxmsg: %d\n", rc);
 958
 959                         /* Make posted receives complete */
 960                         kiblnd_abort_receives(conn);
 961
 962                         /* correct # of posted buffers
 963                          * NB locking needed now I'm racing with completion */
 964                         spin_lock_irqsave(&sched->ibs_lock, flags);
 965                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 966                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 967
 968                         /* cmid will be destroyed by CM(ofed) after cm_callback
 969                          * returned, so we can't refer it anymore
 970                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 971                         rdma_destroy_qp(conn->ibc_cmid);
 972                         conn->ibc_cmid = NULL;
 973
 974                         /* Drop my own and unused rxbuffer refcounts */
 975                         while (i++ <= IBLND_RX_MSGS(conn))
 976                                 kiblnd_conn_decref(conn);
 977
 978                         return NULL;
 979                 }
 980         }
 981
 982         /* Init successful! */
 983         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 984                  state == IBLND_CONN_PASSIVE_WAIT);
 985         conn->ibc_state = state;
 986
 987         /* 1 more conn */
 988         atomic_inc(&net->ibn_nconns);
 989         return conn;
 990
 991  failed_2:
 992         kiblnd_destroy_conn(conn);
 993         LIBCFS_FREE(conn, sizeof(*conn));
 994  failed_0:
 995         return NULL;
 996 }
 997
 998 void
 999 kiblnd_destroy_conn(struct kib_conn *conn)
1000 {
1001         struct rdma_cm_id *cmid = conn->ibc_cmid;
1002         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1003
1004         LASSERT (!in_interrupt());
1005         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1006         LASSERT(list_empty(&conn->ibc_early_rxs));
1007         LASSERT(list_empty(&conn->ibc_tx_noops));
1008         LASSERT(list_empty(&conn->ibc_tx_queue));
1009         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1010         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1011         LASSERT(list_empty(&conn->ibc_active_txs));
1012         LASSERT (conn->ibc_noops_posted == 0);
1013         LASSERT (conn->ibc_nsends_posted == 0);
1014
1015         switch (conn->ibc_state) {
1016         default:
1017                 /* conn must be completely disengaged from the network */
1018                 LBUG();
1019
1020         case IBLND_CONN_DISCONNECTED:
1021                 /* connvars should have been freed already */
1022                 LASSERT (conn->ibc_connvars == NULL);
1023                 break;
1024
1025         case IBLND_CONN_INIT:
1026                 break;
1027         }
1028
1029         /* conn->ibc_cmid might be destroyed by CM already */
1030         if (cmid != NULL && cmid->qp != NULL)
1031                 rdma_destroy_qp(cmid);
1032
1033         if (conn->ibc_cq)
1034                 ib_destroy_cq(conn->ibc_cq);
1035
1036         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1037                            LNET_MSG_STATUS_OK);
1038
1039         if (conn->ibc_rx_pages != NULL)
1040                 kiblnd_unmap_rx_descs(conn);
1041
1042         if (conn->ibc_rxs != NULL)
1043                 CFS_FREE_PTR_ARRAY(conn->ibc_rxs, IBLND_RX_MSGS(conn));
1044
1045         if (conn->ibc_connvars != NULL)
1046                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1047
1048         if (conn->ibc_hdev != NULL)
1049                 kiblnd_hdev_decref(conn->ibc_hdev);
1050
1051         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1052         if (conn->ibc_state != IBLND_CONN_INIT) {
1053                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1054
1055                 kiblnd_peer_decref(peer_ni);
1056                 rdma_destroy_id(cmid);
1057                 atomic_dec(&net->ibn_nconns);
1058         }
1059 }
1060
1061 int
1062 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1063 {
1064         struct kib_conn *conn;
1065         struct list_head        *ctmp;
1066         struct list_head        *cnxt;
1067         int                     count = 0;
1068
1069         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1070                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1071
1072                 CDEBUG(D_NET, "Closing conn -> %s, "
1073                               "version: %x, reason: %d\n",
1074                        libcfs_nid2str(peer_ni->ibp_nid),
1075                        conn->ibc_version, why);
1076
1077                 kiblnd_close_conn_locked(conn, why);
1078                 count++;
1079         }
1080
1081         return count;
1082 }
1083
1084 int
1085 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1086                                 int version, __u64 incarnation)
1087 {
1088         struct kib_conn *conn;
1089         struct list_head        *ctmp;
1090         struct list_head        *cnxt;
1091         int                     count = 0;
1092
1093         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1094                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1095
1096                 if (conn->ibc_version     == version &&
1097                     conn->ibc_incarnation == incarnation)
1098                         continue;
1099
1100                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1101                               "incarnation:%#llx(%x, %#llx)\n",
1102                        libcfs_nid2str(peer_ni->ibp_nid),
1103                        conn->ibc_version, conn->ibc_incarnation,
1104                        version, incarnation);
1105
1106                 kiblnd_close_conn_locked(conn, -ESTALE);
1107                 count++;
1108         }
1109
1110         return count;
1111 }
1112
1113 static int
1114 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1115 {
1116         struct kib_peer_ni              *peer_ni;
1117         struct list_head        *ptmp;
1118         struct list_head        *pnxt;
1119         int                     lo;
1120         int                     hi;
1121         int                     i;
1122         unsigned long           flags;
1123         int                     count = 0;
1124
1125         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1126
1127         if (nid != LNET_NID_ANY)
1128                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1129         else {
1130                 lo = 0;
1131                 hi = kiblnd_data.kib_peer_hash_size - 1;
1132         }
1133
1134         for (i = lo; i <= hi; i++) {
1135                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1136
1137                         peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1138                         LASSERT(!kiblnd_peer_idle(peer_ni));
1139
1140                         if (peer_ni->ibp_ni != ni)
1141                                 continue;
1142
1143                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1144                                 continue;
1145
1146                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1147                 }
1148         }
1149
1150         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1151
1152         /* wildcards always succeed */
1153         if (nid == LNET_NID_ANY)
1154                 return 0;
1155
1156         return (count == 0) ? -ENOENT : 0;
1157 }
1158
1159 static int
1160 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1161 {
1162         struct libcfs_ioctl_data *data = arg;
1163         int                       rc = -EINVAL;
1164
1165         switch(cmd) {
1166         case IOC_LIBCFS_GET_PEER: {
1167                 lnet_nid_t   nid = 0;
1168                 int          count = 0;
1169
1170                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1171                                           &nid, &count);
1172                 data->ioc_nid    = nid;
1173                 data->ioc_count  = count;
1174                 break;
1175         }
1176
1177         case IOC_LIBCFS_DEL_PEER: {
1178                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1179                 break;
1180         }
1181         case IOC_LIBCFS_GET_CONN: {
1182                 struct kib_conn *conn;
1183
1184                 rc = 0;
1185                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1186                 if (conn == NULL) {
1187                         rc = -ENOENT;
1188                         break;
1189                 }
1190
1191                 LASSERT(conn->ibc_cmid != NULL);
1192                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1193                 if (conn->ibc_cmid->route.path_rec == NULL)
1194                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1195                 else
1196                         data->ioc_u32[0] =
1197                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1198                 kiblnd_conn_decref(conn);
1199                 break;
1200         }
1201         case IOC_LIBCFS_CLOSE_CONNECTION: {
1202                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1203                 break;
1204         }
1205
1206         default:
1207                 break;
1208         }
1209
1210         return rc;
1211 }
1212
1213 static void
1214 kiblnd_free_pages(struct kib_pages *p)
1215 {
1216         int     npages = p->ibp_npages;
1217         int     i;
1218
1219         for (i = 0; i < npages; i++) {
1220                 if (p->ibp_pages[i] != NULL)
1221                         __free_page(p->ibp_pages[i]);
1222         }
1223
1224         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1225 }
1226
1227 int
1228 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1229 {
1230         struct kib_pages *p;
1231         int i;
1232
1233         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1234                          offsetof(struct kib_pages, ibp_pages[npages]));
1235         if (p == NULL) {
1236                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1237                 return -ENOMEM;
1238         }
1239
1240         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1241         p->ibp_npages = npages;
1242
1243         for (i = 0; i < npages; i++) {
1244                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1245                                                      GFP_NOFS);
1246                 if (p->ibp_pages[i] == NULL) {
1247                         CERROR("Can't allocate page %d of %d\n", i, npages);
1248                         kiblnd_free_pages(p);
1249                         return -ENOMEM;
1250                 }
1251         }
1252
1253         *pp = p;
1254         return 0;
1255 }
1256
1257 void
1258 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1259 {
1260         struct kib_rx *rx;
1261         int       i;
1262
1263         LASSERT (conn->ibc_rxs != NULL);
1264         LASSERT (conn->ibc_hdev != NULL);
1265
1266         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1267                 rx = &conn->ibc_rxs[i];
1268
1269                 LASSERT(rx->rx_nob >= 0); /* not posted */
1270
1271                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1272                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1273                                                           rx->rx_msgaddr),
1274                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1275         }
1276
1277         kiblnd_free_pages(conn->ibc_rx_pages);
1278
1279         conn->ibc_rx_pages = NULL;
1280 }
1281
1282 void
1283 kiblnd_map_rx_descs(struct kib_conn *conn)
1284 {
1285         struct kib_rx *rx;
1286         struct page    *pg;
1287         int             pg_off;
1288         int             ipg;
1289         int             i;
1290
1291         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1292                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1293                 rx = &conn->ibc_rxs[i];
1294
1295                 rx->rx_conn = conn;
1296                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1297
1298                 rx->rx_msgaddr =
1299                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1300                                               rx->rx_msg, IBLND_MSG_SIZE,
1301                                               DMA_FROM_DEVICE);
1302                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1303                                                   rx->rx_msgaddr));
1304                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1305
1306                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1307                        i, rx->rx_msg, rx->rx_msgaddr,
1308                        (__u64)(page_to_phys(pg) + pg_off));
1309
1310                 pg_off += IBLND_MSG_SIZE;
1311                 LASSERT(pg_off <= PAGE_SIZE);
1312
1313                 if (pg_off == PAGE_SIZE) {
1314                         pg_off = 0;
1315                         ipg++;
1316                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1317                 }
1318         }
1319 }
1320
1321 static void
1322 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1323 {
1324         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1325         struct kib_tx *tx;
1326         int i;
1327
1328         LASSERT (tpo->tpo_pool.po_allocated == 0);
1329
1330         if (hdev == NULL)
1331                 return;
1332
1333         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1334                 tx = &tpo->tpo_tx_descs[i];
1335                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1336                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1337                                                           tx->tx_msgaddr),
1338                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1339         }
1340
1341         kiblnd_hdev_decref(hdev);
1342         tpo->tpo_hdev = NULL;
1343 }
1344
1345 static struct kib_hca_dev *
1346 kiblnd_current_hdev(struct kib_dev *dev)
1347 {
1348         struct kib_hca_dev *hdev;
1349         unsigned long  flags;
1350         int            i = 0;
1351
1352         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1353         while (dev->ibd_failover) {
1354                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1355                 if (i++ % 50 == 0)
1356                         CDEBUG(D_NET, "%s: Wait for failover\n",
1357                                dev->ibd_ifname);
1358                 schedule_timeout_interruptible(cfs_time_seconds(1) / 100);
1359
1360                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1361         }
1362
1363         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1364         hdev = dev->ibd_hdev;
1365
1366         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1367
1368         return hdev;
1369 }
1370
1371 static void
1372 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1373 {
1374         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1375         struct kib_pool *pool = &tpo->tpo_pool;
1376         struct kib_net      *net   = pool->po_owner->ps_net;
1377         struct kib_dev *dev;
1378         struct page *page;
1379         struct kib_tx *tx;
1380         int             page_offset;
1381         int             ipage;
1382         int             i;
1383
1384         LASSERT (net != NULL);
1385
1386         dev = net->ibn_dev;
1387
1388         /* pre-mapped messages are not bigger than 1 page */
1389         BUILD_BUG_ON(IBLND_MSG_SIZE > PAGE_SIZE);
1390
1391         /* No fancy arithmetic when we do the buffer calculations */
1392         BUILD_BUG_ON(PAGE_SIZE % IBLND_MSG_SIZE != 0);
1393
1394         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1395
1396         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1397                 page = txpgs->ibp_pages[ipage];
1398                 tx = &tpo->tpo_tx_descs[i];
1399
1400                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1401                                                 page_offset);
1402
1403                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1404                                                        tx->tx_msg,
1405                                                        IBLND_MSG_SIZE,
1406                                                        DMA_TO_DEVICE);
1407                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1408                                                   tx->tx_msgaddr));
1409                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1410
1411                 list_add(&tx->tx_list, &pool->po_free_list);
1412
1413                 page_offset += IBLND_MSG_SIZE;
1414                 LASSERT(page_offset <= PAGE_SIZE);
1415
1416                 if (page_offset == PAGE_SIZE) {
1417                         page_offset = 0;
1418                         ipage++;
1419                         LASSERT(ipage <= txpgs->ibp_npages);
1420                 }
1421         }
1422 }
1423
1424 static void
1425 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1426 {
1427         LASSERT(fpo->fpo_map_count == 0);
1428
1429         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1430                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1431         } else {
1432                 struct kib_fast_reg_descriptor *frd, *tmp;
1433                 int i = 0;
1434
1435                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1436                                          frd_list) {
1437                         list_del(&frd->frd_list);
1438 #ifndef HAVE_IB_MAP_MR_SG
1439                         ib_free_fast_reg_page_list(frd->frd_frpl);
1440 #endif
1441                         ib_dereg_mr(frd->frd_mr);
1442                         LIBCFS_FREE(frd, sizeof(*frd));
1443                         i++;
1444                 }
1445                 if (i < fpo->fast_reg.fpo_pool_size)
1446                         CERROR("FastReg pool still has %d regions registered\n",
1447                                 fpo->fast_reg.fpo_pool_size - i);
1448         }
1449
1450         if (fpo->fpo_hdev)
1451                 kiblnd_hdev_decref(fpo->fpo_hdev);
1452
1453         LIBCFS_FREE(fpo, sizeof(*fpo));
1454 }
1455
1456 static void
1457 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1458 {
1459         struct kib_fmr_pool *fpo, *tmp;
1460
1461         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1462                 list_del(&fpo->fpo_list);
1463                 kiblnd_destroy_fmr_pool(fpo);
1464         }
1465 }
1466
1467 static int
1468 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1469                      int ncpts)
1470 {
1471         int size = tunables->lnd_fmr_pool_size / ncpts;
1472
1473         return max(IBLND_FMR_POOL, size);
1474 }
1475
1476 static int
1477 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1478                          int ncpts)
1479 {
1480         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1481
1482         return max(IBLND_FMR_POOL_FLUSH, size);
1483 }
1484
1485 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1486                                  struct kib_fmr_pool *fpo)
1487 {
1488         struct ib_fmr_pool_param param = {
1489                 .max_pages_per_fmr = LNET_MAX_IOV,
1490                 .page_shift        = PAGE_SHIFT,
1491                 .access            = (IB_ACCESS_LOCAL_WRITE |
1492                                       IB_ACCESS_REMOTE_WRITE),
1493                 .pool_size         = fps->fps_pool_size,
1494                 .dirty_watermark   = fps->fps_flush_trigger,
1495                 .flush_function    = NULL,
1496                 .flush_arg         = NULL,
1497                 .cache             = !!fps->fps_cache };
1498         int rc = 0;
1499
1500         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1501                                                    &param);
1502         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1503                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1504                 if (rc != -ENOSYS)
1505                         CERROR("Failed to create FMR pool: %d\n", rc);
1506                 else
1507                         CERROR("FMRs are not supported\n");
1508         }
1509         fpo->fpo_is_fmr = true;
1510
1511         return rc;
1512 }
1513
1514 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1515                                   struct kib_fmr_pool *fpo,
1516                                   enum kib_dev_caps dev_caps)
1517 {
1518         struct kib_fast_reg_descriptor *frd, *tmp;
1519         int i, rc;
1520
1521         fpo->fpo_is_fmr = false;
1522
1523         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1524         fpo->fast_reg.fpo_pool_size = 0;
1525         for (i = 0; i < fps->fps_pool_size; i++) {
1526                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1527                                  sizeof(*frd));
1528                 if (!frd) {
1529                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1530                         rc = -ENOMEM;
1531                         goto out;
1532                 }
1533                 frd->frd_mr = NULL;
1534
1535 #ifndef HAVE_IB_MAP_MR_SG
1536                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1537                                                             LNET_MAX_IOV);
1538                 if (IS_ERR(frd->frd_frpl)) {
1539                         rc = PTR_ERR(frd->frd_frpl);
1540                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1541                                 rc);
1542                         frd->frd_frpl = NULL;
1543                         goto out_middle;
1544                 }
1545 #endif
1546
1547 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1548                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1549                                                    LNET_MAX_IOV);
1550 #else
1551                 /*
1552                  * it is expected to get here if this is an MLX-5 card.
1553                  * MLX-4 cards will always use FMR and MLX-5 cards will
1554                  * always use fast_reg. It turns out that some MLX-5 cards
1555                  * (possibly due to older FW versions) do not natively support
1556                  * gaps. So we will need to track them here.
1557                  */
1558                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1559 #ifdef IB_MR_TYPE_SG_GAPS
1560                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1561                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1562                                                 IB_MR_TYPE_SG_GAPS :
1563                                                 IB_MR_TYPE_MEM_REG,
1564 #else
1565                                                 IB_MR_TYPE_MEM_REG,
1566 #endif
1567                                           LNET_MAX_IOV);
1568                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1569                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1570                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1571 #endif
1572                 if (IS_ERR(frd->frd_mr)) {
1573                         rc = PTR_ERR(frd->frd_mr);
1574                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1575                         frd->frd_mr = NULL;
1576                         goto out_middle;
1577                 }
1578
1579                 /* There appears to be a bug in MLX5 code where you must
1580                  * invalidate the rkey of a new FastReg pool before first
1581                  * using it. Thus, I am marking the FRD invalid here. */
1582                 frd->frd_valid = false;
1583
1584                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1585                 fpo->fast_reg.fpo_pool_size++;
1586         }
1587
1588         return 0;
1589
1590 out_middle:
1591         if (frd->frd_mr)
1592                 ib_dereg_mr(frd->frd_mr);
1593 #ifndef HAVE_IB_MAP_MR_SG
1594         if (frd->frd_frpl)
1595                 ib_free_fast_reg_page_list(frd->frd_frpl);
1596 #endif
1597         LIBCFS_FREE(frd, sizeof(*frd));
1598
1599 out:
1600         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1601                                  frd_list) {
1602                 list_del(&frd->frd_list);
1603 #ifndef HAVE_IB_MAP_MR_SG
1604                 ib_free_fast_reg_page_list(frd->frd_frpl);
1605 #endif
1606                 ib_dereg_mr(frd->frd_mr);
1607                 LIBCFS_FREE(frd, sizeof(*frd));
1608         }
1609
1610         return rc;
1611 }
1612
1613 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1614                                   struct kib_fmr_pool **pp_fpo)
1615 {
1616         struct kib_dev *dev = fps->fps_net->ibn_dev;
1617         struct kib_fmr_pool *fpo;
1618         int rc;
1619
1620         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1621         if (!fpo) {
1622                 return -ENOMEM;
1623         }
1624         memset(fpo, 0, sizeof(*fpo));
1625
1626         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1627
1628         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1629                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1630         else
1631                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1632         if (rc)
1633                 goto out_fpo;
1634
1635         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1636         fpo->fpo_owner = fps;
1637         *pp_fpo = fpo;
1638
1639         return 0;
1640
1641 out_fpo:
1642         kiblnd_hdev_decref(fpo->fpo_hdev);
1643         LIBCFS_FREE(fpo, sizeof(*fpo));
1644         return rc;
1645 }
1646
1647 static void
1648 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1649 {
1650         if (fps->fps_net == NULL) /* intialized? */
1651                 return;
1652
1653         spin_lock(&fps->fps_lock);
1654
1655         while (!list_empty(&fps->fps_pool_list)) {
1656                 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1657                                                       struct kib_fmr_pool,
1658                                                       fpo_list);
1659
1660                 fpo->fpo_failed = 1;
1661                 if (fpo->fpo_map_count == 0)
1662                         list_move(&fpo->fpo_list, zombies);
1663                 else
1664                         list_move(&fpo->fpo_list, &fps->fps_failed_pool_list);
1665         }
1666
1667         spin_unlock(&fps->fps_lock);
1668 }
1669
1670 static void
1671 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1672 {
1673         if (fps->fps_net != NULL) { /* initialized? */
1674                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1675                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1676         }
1677 }
1678
1679 static int
1680 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1681                         struct kib_net *net,
1682                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1683 {
1684         struct kib_fmr_pool *fpo;
1685         int rc;
1686
1687         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1688
1689         fps->fps_net = net;
1690         fps->fps_cpt = cpt;
1691
1692         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1693         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1694         fps->fps_cache = tunables->lnd_fmr_cache;
1695
1696         spin_lock_init(&fps->fps_lock);
1697         INIT_LIST_HEAD(&fps->fps_pool_list);
1698         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1699
1700         rc = kiblnd_create_fmr_pool(fps, &fpo);
1701         if (rc == 0)
1702                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1703
1704         return rc;
1705 }
1706
1707 static int
1708 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1709 {
1710         if (fpo->fpo_map_count != 0) /* still in use */
1711                 return 0;
1712         if (fpo->fpo_failed)
1713                 return 1;
1714         return now >= fpo->fpo_deadline;
1715 }
1716
1717 static int
1718 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1719 {
1720         struct kib_hca_dev *hdev;
1721         __u64           *pages = tx->tx_pages;
1722         int             npages;
1723         int             size;
1724         int             i;
1725
1726         hdev = tx->tx_pool->tpo_hdev;
1727
1728         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1729                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1730                         size += hdev->ibh_page_size) {
1731                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1732                                            hdev->ibh_page_mask) + size;
1733                 }
1734         }
1735
1736         return npages;
1737 }
1738
1739 void
1740 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1741 {
1742         LIST_HEAD(zombies);
1743         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1744         struct kib_fmr_poolset *fps;
1745         time64_t now = ktime_get_seconds();
1746         struct kib_fmr_pool *tmp;
1747         int rc;
1748
1749         if (!fpo)
1750                 return;
1751
1752         fps = fpo->fpo_owner;
1753         if (fpo->fpo_is_fmr) {
1754                 if (fmr->fmr_pfmr) {
1755                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1756                         fmr->fmr_pfmr = NULL;
1757                 }
1758
1759                 if (status) {
1760                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1761                         LASSERT(!rc);
1762                 }
1763         } else {
1764                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1765
1766                 if (frd) {
1767                         frd->frd_valid = false;
1768                         spin_lock(&fps->fps_lock);
1769                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1770                         spin_unlock(&fps->fps_lock);
1771                         fmr->fmr_frd = NULL;
1772                 }
1773         }
1774         fmr->fmr_pool = NULL;
1775
1776         spin_lock(&fps->fps_lock);
1777         fpo->fpo_map_count--;   /* decref the pool */
1778
1779         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1780                 /* the first pool is persistent */
1781                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1782                         continue;
1783
1784                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1785                         list_move(&fpo->fpo_list, &zombies);
1786                         fps->fps_version++;
1787                 }
1788         }
1789         spin_unlock(&fps->fps_lock);
1790
1791         if (!list_empty(&zombies))
1792                 kiblnd_destroy_fmr_pool_list(&zombies);
1793 }
1794
1795 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1796                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1797                         struct kib_fmr *fmr)
1798 {
1799         struct kib_fmr_pool *fpo;
1800         __u64 *pages = tx->tx_pages;
1801         __u64 version;
1802         bool is_rx = (rd != tx->tx_rd);
1803         bool tx_pages_mapped = 0;
1804         int npages = 0;
1805         int rc;
1806
1807 again:
1808         spin_lock(&fps->fps_lock);
1809         version = fps->fps_version;
1810         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1811                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1812                 fpo->fpo_map_count++;
1813
1814                 if (fpo->fpo_is_fmr) {
1815                         struct ib_pool_fmr *pfmr;
1816
1817                         spin_unlock(&fps->fps_lock);
1818
1819                         if (!tx_pages_mapped) {
1820                                 npages = kiblnd_map_tx_pages(tx, rd);
1821                                 tx_pages_mapped = 1;
1822                         }
1823
1824                         pfmr = kib_fmr_pool_map(fpo->fmr.fpo_fmr_pool,
1825                                                 pages, npages, iov);
1826                         if (likely(!IS_ERR(pfmr))) {
1827                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1828                                                       : pfmr->fmr->lkey;
1829                                 fmr->fmr_frd  = NULL;
1830                                 fmr->fmr_pfmr = pfmr;
1831                                 fmr->fmr_pool = fpo;
1832                                 return 0;
1833                         }
1834                         rc = PTR_ERR(pfmr);
1835                 } else {
1836                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1837                                 struct kib_fast_reg_descriptor *frd;
1838 #ifdef HAVE_IB_MAP_MR_SG
1839                                 struct ib_reg_wr *wr;
1840                                 int n;
1841 #else
1842                                 struct ib_rdma_wr *wr;
1843                                 struct ib_fast_reg_page_list *frpl;
1844 #endif
1845                                 struct ib_mr *mr;
1846
1847                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1848                                                         struct kib_fast_reg_descriptor,
1849                                                         frd_list);
1850                                 list_del(&frd->frd_list);
1851                                 spin_unlock(&fps->fps_lock);
1852
1853 #ifndef HAVE_IB_MAP_MR_SG
1854                                 frpl = frd->frd_frpl;
1855 #endif
1856                                 mr   = frd->frd_mr;
1857
1858                                 if (!frd->frd_valid) {
1859                                         struct ib_rdma_wr *inv_wr;
1860                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1861
1862                                         inv_wr = &frd->frd_inv_wr;
1863                                         memset(inv_wr, 0, sizeof(*inv_wr));
1864
1865                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1866                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1867                                         inv_wr->wr.ex.invalidate_rkey = key;
1868
1869                                         /* Bump the key */
1870                                         key = ib_inc_rkey(key);
1871                                         ib_update_fast_reg_key(mr, key);
1872                                 }
1873
1874 #ifdef HAVE_IB_MAP_MR_SG
1875 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1876                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1877                                                  rd->rd_nfrags, NULL, PAGE_SIZE);
1878 #else
1879                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1880                                                  rd->rd_nfrags, PAGE_SIZE);
1881 #endif
1882                                 if (unlikely(n != rd->rd_nfrags)) {
1883                                         CERROR("Failed to map mr %d/%d "
1884                                                "elements\n", n, rd->rd_nfrags);
1885                                         return n < 0 ? n : -EINVAL;
1886                                 }
1887
1888                                 wr = &frd->frd_fastreg_wr;
1889                                 memset(wr, 0, sizeof(*wr));
1890
1891                                 wr->wr.opcode = IB_WR_REG_MR;
1892                                 wr->wr.wr_id  = IBLND_WID_MR;
1893                                 wr->wr.num_sge = 0;
1894                                 wr->wr.send_flags = 0;
1895                                 wr->mr = mr;
1896                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1897                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1898                                               IB_ACCESS_REMOTE_WRITE);
1899 #else
1900                                 if (!tx_pages_mapped) {
1901                                         npages = kiblnd_map_tx_pages(tx, rd);
1902                                         tx_pages_mapped = 1;
1903                                 }
1904
1905                                 LASSERT(npages <= frpl->max_page_list_len);
1906                                 memcpy(frpl->page_list, pages,
1907                                         sizeof(*pages) * npages);
1908
1909                                 /* Prepare FastReg WR */
1910                                 wr = &frd->frd_fastreg_wr;
1911                                 memset(wr, 0, sizeof(*wr));
1912
1913                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1914                                 wr->wr.wr_id  = IBLND_WID_MR;
1915
1916                                 wr->wr.wr.fast_reg.iova_start = iov;
1917                                 wr->wr.wr.fast_reg.page_list  = frpl;
1918                                 wr->wr.wr.fast_reg.page_list_len = npages;
1919                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1920                                 wr->wr.wr.fast_reg.length = nob;
1921                                 wr->wr.wr.fast_reg.rkey =
1922                                                 is_rx ? mr->rkey : mr->lkey;
1923                                 wr->wr.wr.fast_reg.access_flags =
1924                                                 (IB_ACCESS_LOCAL_WRITE |
1925                                                  IB_ACCESS_REMOTE_WRITE);
1926 #endif
1927
1928                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1929                                 fmr->fmr_frd  = frd;
1930                                 fmr->fmr_pfmr = NULL;
1931                                 fmr->fmr_pool = fpo;
1932                                 return 0;
1933                         }
1934                         spin_unlock(&fps->fps_lock);
1935                         rc = -EAGAIN;
1936                 }
1937
1938                 spin_lock(&fps->fps_lock);
1939                 fpo->fpo_map_count--;
1940                 if (rc != -EAGAIN) {
1941                         spin_unlock(&fps->fps_lock);
1942                         return rc;
1943                 }
1944
1945                 /* EAGAIN and ... */
1946                 if (version != fps->fps_version) {
1947                         spin_unlock(&fps->fps_lock);
1948                         goto again;
1949                 }
1950         }
1951
1952         if (fps->fps_increasing) {
1953                 spin_unlock(&fps->fps_lock);
1954                 CDEBUG(D_NET, "Another thread is allocating new "
1955                        "FMR pool, waiting for her to complete\n");
1956                 schedule();
1957                 goto again;
1958
1959         }
1960
1961         if (ktime_get_seconds() < fps->fps_next_retry) {
1962                 /* someone failed recently */
1963                 spin_unlock(&fps->fps_lock);
1964                 return -EAGAIN;
1965         }
1966
1967         fps->fps_increasing = 1;
1968         spin_unlock(&fps->fps_lock);
1969
1970         CDEBUG(D_NET, "Allocate new FMR pool\n");
1971         rc = kiblnd_create_fmr_pool(fps, &fpo);
1972         spin_lock(&fps->fps_lock);
1973         fps->fps_increasing = 0;
1974         if (rc == 0) {
1975                 fps->fps_version++;
1976                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1977         } else {
1978                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
1979         }
1980         spin_unlock(&fps->fps_lock);
1981
1982         goto again;
1983 }
1984
1985 static void
1986 kiblnd_fini_pool(struct kib_pool *pool)
1987 {
1988         LASSERT(list_empty(&pool->po_free_list));
1989         LASSERT(pool->po_allocated == 0);
1990
1991         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1992 }
1993
1994 static void
1995 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1996 {
1997         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1998
1999         memset(pool, 0, sizeof(struct kib_pool));
2000         INIT_LIST_HEAD(&pool->po_free_list);
2001         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2002         pool->po_owner = ps;
2003         pool->po_size = size;
2004 }
2005
2006 static void
2007 kiblnd_destroy_pool_list(struct list_head *head)
2008 {
2009         struct kib_pool *pool;
2010
2011         while (!list_empty(head)) {
2012                 pool = list_entry(head->next, struct kib_pool, po_list);
2013                 list_del(&pool->po_list);
2014
2015                 LASSERT(pool->po_owner != NULL);
2016                 pool->po_owner->ps_pool_destroy(pool);
2017         }
2018 }
2019
2020 static void
2021 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2022 {
2023         if (ps->ps_net == NULL) /* intialized? */
2024                 return;
2025
2026         spin_lock(&ps->ps_lock);
2027         while (!list_empty(&ps->ps_pool_list)) {
2028                 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2029                                                  struct kib_pool, po_list);
2030
2031                 po->po_failed = 1;
2032                 if (po->po_allocated == 0)
2033                         list_move(&po->po_list, zombies);
2034                 else
2035                         list_move(&po->po_list, &ps->ps_failed_pool_list);
2036         }
2037         spin_unlock(&ps->ps_lock);
2038 }
2039
2040 static void
2041 kiblnd_fini_poolset(struct kib_poolset *ps)
2042 {
2043         if (ps->ps_net != NULL) { /* initialized? */
2044                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2045                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2046         }
2047 }
2048
2049 static int
2050 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2051                     struct kib_net *net, char *name, int size,
2052                     kib_ps_pool_create_t po_create,
2053                     kib_ps_pool_destroy_t po_destroy,
2054                     kib_ps_node_init_t nd_init,
2055                     kib_ps_node_fini_t nd_fini)
2056 {
2057         struct kib_pool *pool;
2058         int rc;
2059
2060         memset(ps, 0, sizeof(struct kib_poolset));
2061
2062         ps->ps_cpt          = cpt;
2063         ps->ps_net          = net;
2064         ps->ps_pool_create  = po_create;
2065         ps->ps_pool_destroy = po_destroy;
2066         ps->ps_node_init    = nd_init;
2067         ps->ps_node_fini    = nd_fini;
2068         ps->ps_pool_size    = size;
2069         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2070             >= sizeof(ps->ps_name))
2071                 return -E2BIG;
2072         spin_lock_init(&ps->ps_lock);
2073         INIT_LIST_HEAD(&ps->ps_pool_list);
2074         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2075
2076         rc = ps->ps_pool_create(ps, size, &pool);
2077         if (rc == 0)
2078                 list_add(&pool->po_list, &ps->ps_pool_list);
2079         else
2080                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2081
2082         return rc;
2083 }
2084
2085 static int
2086 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2087 {
2088         if (pool->po_allocated != 0) /* still in use */
2089                 return 0;
2090         if (pool->po_failed)
2091                 return 1;
2092         return now >= pool->po_deadline;
2093 }
2094
2095 void
2096 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2097 {
2098         LIST_HEAD(zombies);
2099         struct kib_poolset *ps = pool->po_owner;
2100         struct kib_pool *tmp;
2101         time64_t now = ktime_get_seconds();
2102
2103         spin_lock(&ps->ps_lock);
2104
2105         if (ps->ps_node_fini != NULL)
2106                 ps->ps_node_fini(pool, node);
2107
2108         LASSERT(pool->po_allocated > 0);
2109         list_add(node, &pool->po_free_list);
2110         pool->po_allocated--;
2111
2112         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2113                 /* the first pool is persistent */
2114                 if (ps->ps_pool_list.next == &pool->po_list)
2115                         continue;
2116
2117                 if (kiblnd_pool_is_idle(pool, now))
2118                         list_move(&pool->po_list, &zombies);
2119         }
2120         spin_unlock(&ps->ps_lock);
2121
2122         if (!list_empty(&zombies))
2123                 kiblnd_destroy_pool_list(&zombies);
2124 }
2125
2126 struct list_head *
2127 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2128 {
2129         struct list_head        *node;
2130         struct kib_pool *pool;
2131         int                     rc;
2132         unsigned int            interval = 1;
2133         ktime_t time_before;
2134         unsigned int trips = 0;
2135
2136 again:
2137         spin_lock(&ps->ps_lock);
2138         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2139                 if (list_empty(&pool->po_free_list))
2140                         continue;
2141
2142                 pool->po_allocated++;
2143                 pool->po_deadline = ktime_get_seconds() +
2144                                     IBLND_POOL_DEADLINE;
2145                 node = pool->po_free_list.next;
2146                 list_del(node);
2147
2148                 if (ps->ps_node_init != NULL) {
2149                         /* still hold the lock */
2150                         ps->ps_node_init(pool, node);
2151                 }
2152                 spin_unlock(&ps->ps_lock);
2153                 return node;
2154         }
2155
2156         /* no available tx pool and ... */
2157         if (ps->ps_increasing) {
2158                 /* another thread is allocating a new pool */
2159                 spin_unlock(&ps->ps_lock);
2160                 trips++;
2161                 CDEBUG(D_NET,
2162                        "Another thread is allocating new %s pool, waiting %d jiffies for her to complete. trips = %d\n",
2163                        ps->ps_name, interval, trips);
2164
2165                 schedule_timeout_interruptible(interval);
2166                 if (interval < cfs_time_seconds(1))
2167                         interval *= 2;
2168
2169                 goto again;
2170         }
2171
2172         if (ktime_get_seconds() < ps->ps_next_retry) {
2173                 /* someone failed recently */
2174                 spin_unlock(&ps->ps_lock);
2175                 return NULL;
2176         }
2177
2178         ps->ps_increasing = 1;
2179         spin_unlock(&ps->ps_lock);
2180
2181         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2182         time_before = ktime_get();
2183         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2184         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2185                ktime_ms_delta(ktime_get(), time_before));
2186
2187         spin_lock(&ps->ps_lock);
2188         ps->ps_increasing = 0;
2189         if (rc == 0) {
2190                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2191         } else {
2192                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2193                 CERROR("Can't allocate new %s pool because out of memory\n",
2194                        ps->ps_name);
2195         }
2196         spin_unlock(&ps->ps_lock);
2197
2198         goto again;
2199 }
2200
2201 static void
2202 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2203 {
2204         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2205                                                tpo_pool);
2206         int i;
2207
2208         LASSERT (pool->po_allocated == 0);
2209
2210         if (tpo->tpo_tx_pages != NULL) {
2211                 kiblnd_unmap_tx_pool(tpo);
2212                 kiblnd_free_pages(tpo->tpo_tx_pages);
2213         }
2214
2215         if (tpo->tpo_tx_descs == NULL)
2216                 goto out;
2217
2218         for (i = 0; i < pool->po_size; i++) {
2219                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2220                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2221
2222                 list_del(&tx->tx_list);
2223                 if (tx->tx_pages != NULL)
2224                         CFS_FREE_PTR_ARRAY(tx->tx_pages, LNET_MAX_IOV);
2225                 if (tx->tx_frags != NULL)
2226                         CFS_FREE_PTR_ARRAY(tx->tx_frags,
2227                                            (1 + IBLND_MAX_RDMA_FRAGS));
2228                 if (tx->tx_wrq != NULL)
2229                         CFS_FREE_PTR_ARRAY(tx->tx_wrq,
2230                                            (1 + IBLND_MAX_RDMA_FRAGS));
2231                 if (tx->tx_sge != NULL)
2232                         CFS_FREE_PTR_ARRAY(tx->tx_sge,
2233                                            (1 + IBLND_MAX_RDMA_FRAGS) *
2234                                            wrq_sge);
2235                 if (tx->tx_rd != NULL)
2236                         LIBCFS_FREE(tx->tx_rd,
2237                                     offsetof(struct kib_rdma_desc,
2238                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2239         }
2240
2241         CFS_FREE_PTR_ARRAY(tpo->tpo_tx_descs, pool->po_size);
2242 out:
2243         kiblnd_fini_pool(pool);
2244         CFS_FREE_PTR(tpo);
2245 }
2246
2247 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2248 {
2249         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2250         int ntx;
2251
2252         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2253         ntx = tunables->lnd_ntx / ncpts;
2254
2255         return max(IBLND_TX_POOL, ntx);
2256 }
2257
2258 static int
2259 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2260 {
2261         int            i;
2262         int            npg;
2263         struct kib_pool *pool;
2264         struct kib_tx_pool *tpo;
2265
2266         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2267         if (tpo == NULL) {
2268                 CERROR("Failed to allocate TX pool\n");
2269                 return -ENOMEM;
2270         }
2271
2272         pool = &tpo->tpo_pool;
2273         kiblnd_init_pool(ps, pool, size);
2274         tpo->tpo_tx_descs = NULL;
2275         tpo->tpo_tx_pages = NULL;
2276
2277         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2278         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2279                 CERROR("Can't allocate tx pages: %d\n", npg);
2280                 CFS_FREE_PTR(tpo);
2281                 return -ENOMEM;
2282         }
2283
2284         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2285                          size * sizeof(struct kib_tx));
2286         if (tpo->tpo_tx_descs == NULL) {
2287                 CERROR("Can't allocate %d tx descriptors\n", size);
2288                 ps->ps_pool_destroy(pool);
2289                 return -ENOMEM;
2290         }
2291
2292         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2293
2294         for (i = 0; i < size; i++) {
2295                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2296                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2297
2298                 tx->tx_pool = tpo;
2299                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2300                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2301                                          lnet_cpt_table(), ps->ps_cpt,
2302                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2303                         if (tx->tx_pages == NULL)
2304                                 break;
2305                 }
2306
2307                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2308                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2309                                  sizeof(*tx->tx_frags));
2310                 if (tx->tx_frags == NULL)
2311                         break;
2312
2313                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2314
2315                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2316                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2317                                  sizeof(*tx->tx_wrq));
2318                 if (tx->tx_wrq == NULL)
2319                         break;
2320
2321                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2322                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2323                                  sizeof(*tx->tx_sge));
2324                 if (tx->tx_sge == NULL)
2325                         break;
2326
2327                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2328                                  offsetof(struct kib_rdma_desc,
2329                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2330                 if (tx->tx_rd == NULL)
2331                         break;
2332         }
2333
2334         if (i == size) {
2335                 kiblnd_map_tx_pool(tpo);
2336                 *pp_po = pool;
2337                 return 0;
2338         }
2339
2340         ps->ps_pool_destroy(pool);
2341         return -ENOMEM;
2342 }
2343
2344 static void
2345 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2346 {
2347         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2348                                                   struct kib_tx_poolset,
2349                                                   tps_poolset);
2350         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2351
2352         tx->tx_cookie = tps->tps_next_tx_cookie++;
2353 }
2354
2355 static void
2356 kiblnd_net_fini_pools(struct kib_net *net)
2357 {
2358         int     i;
2359
2360         cfs_cpt_for_each(i, lnet_cpt_table()) {
2361                 struct kib_tx_poolset *tps;
2362                 struct kib_fmr_poolset *fps;
2363
2364                 if (net->ibn_tx_ps != NULL) {
2365                         tps = net->ibn_tx_ps[i];
2366                         kiblnd_fini_poolset(&tps->tps_poolset);
2367                 }
2368
2369                 if (net->ibn_fmr_ps != NULL) {
2370                         fps = net->ibn_fmr_ps[i];
2371                         kiblnd_fini_fmr_poolset(fps);
2372                 }
2373         }
2374
2375         if (net->ibn_tx_ps != NULL) {
2376                 cfs_percpt_free(net->ibn_tx_ps);
2377                 net->ibn_tx_ps = NULL;
2378         }
2379
2380         if (net->ibn_fmr_ps != NULL) {
2381                 cfs_percpt_free(net->ibn_fmr_ps);
2382                 net->ibn_fmr_ps = NULL;
2383         }
2384 }
2385
2386 static int
2387 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2388                       int ncpts)
2389 {
2390         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2391 #ifdef HAVE_IB_GET_DMA_MR
2392         unsigned long   flags;
2393 #endif
2394         int             cpt;
2395         int             rc;
2396         int             i;
2397
2398         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2399
2400 #ifdef HAVE_IB_GET_DMA_MR
2401         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2402         /*
2403          * if lnd_map_on_demand is zero then we have effectively disabled
2404          * FMR or FastReg and we're using global memory regions
2405          * exclusively.
2406          */
2407         if (!tunables->lnd_map_on_demand) {
2408                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2409                                            flags);
2410                 goto create_tx_pool;
2411         }
2412
2413         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2414 #endif
2415
2416         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2417                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2418                        tunables->lnd_fmr_pool_size,
2419                        tunables->lnd_ntx / 4);
2420                 rc = -EINVAL;
2421                 goto failed;
2422         }
2423
2424         /* TX pool must be created later than FMR, see LU-2268
2425          * for details */
2426         LASSERT(net->ibn_tx_ps == NULL);
2427
2428         /* premapping can fail if ibd_nmr > 1, so we always create
2429          * FMR pool and map-on-demand if premapping failed */
2430
2431         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2432                                            sizeof(struct kib_fmr_poolset));
2433         if (net->ibn_fmr_ps == NULL) {
2434                 CERROR("Failed to allocate FMR pool array\n");
2435                 rc = -ENOMEM;
2436                 goto failed;
2437         }
2438
2439         for (i = 0; i < ncpts; i++) {
2440                 cpt = (cpts == NULL) ? i : cpts[i];
2441                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2442                                              net, tunables);
2443                 if (rc != 0) {
2444                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2445                                cpt, rc);
2446                         goto failed;
2447                 }
2448         }
2449
2450         if (i > 0)
2451                 LASSERT(i == ncpts);
2452
2453 #ifdef HAVE_IB_GET_DMA_MR
2454  create_tx_pool:
2455 #endif
2456         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2457                                           sizeof(struct kib_tx_poolset));
2458         if (net->ibn_tx_ps == NULL) {
2459                 CERROR("Failed to allocate tx pool array\n");
2460                 rc = -ENOMEM;
2461                 goto failed;
2462         }
2463
2464         for (i = 0; i < ncpts; i++) {
2465                 cpt = (cpts == NULL) ? i : cpts[i];
2466                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2467                                          cpt, net, "TX",
2468                                          kiblnd_tx_pool_size(ni, ncpts),
2469                                          kiblnd_create_tx_pool,
2470                                          kiblnd_destroy_tx_pool,
2471                                          kiblnd_tx_init, NULL);
2472                 if (rc != 0) {
2473                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2474                                cpt, rc);
2475                         goto failed;
2476                 }
2477         }
2478
2479         return 0;
2480  failed:
2481         kiblnd_net_fini_pools(net);
2482         LASSERT(rc != 0);
2483         return rc;
2484 }
2485
2486 static int
2487 kiblnd_port_get_attr(struct kib_hca_dev *hdev)
2488 {
2489         struct ib_port_attr *port_attr;
2490         int rc;
2491         unsigned long flags;
2492         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2493
2494         LIBCFS_ALLOC(port_attr, sizeof(*port_attr));
2495         if (port_attr == NULL) {
2496                 CDEBUG(D_NETERROR, "Out of memory\n");
2497                 return -ENOMEM;
2498         }
2499
2500         rc = ib_query_port(hdev->ibh_ibdev, hdev->ibh_port, port_attr);
2501
2502         write_lock_irqsave(g_lock, flags);
2503
2504         if (rc == 0)
2505                 hdev->ibh_state = port_attr->state == IB_PORT_ACTIVE
2506                                  ? IBLND_DEV_PORT_ACTIVE
2507                                  : IBLND_DEV_PORT_DOWN;
2508
2509         write_unlock_irqrestore(g_lock, flags);
2510         LIBCFS_FREE(port_attr, sizeof(*port_attr));
2511
2512         if (rc != 0) {
2513                 CDEBUG(D_NETERROR, "Failed to query IB port: %d\n", rc);
2514                 return rc;
2515         }
2516         return 0;
2517 }
2518
2519 static inline void
2520 kiblnd_set_ni_fatal_on(struct kib_hca_dev *hdev, int val)
2521 {
2522         struct kib_net  *net;
2523
2524         /* for health check */
2525         list_for_each_entry(net, &hdev->ibh_dev->ibd_nets, ibn_list) {
2526                 if (val)
2527                         CDEBUG(D_NETERROR, "Fatal device error for NI %s\n",
2528                                         libcfs_nid2str(net->ibn_ni->ni_nid));
2529                 atomic_set(&net->ibn_ni->ni_fatal_error_on, val);
2530         }
2531 }
2532
2533 void
2534 kiblnd_event_handler(struct ib_event_handler *handler, struct ib_event *event)
2535 {
2536         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2537         struct kib_hca_dev  *hdev;
2538         unsigned long flags;
2539
2540         hdev = container_of(handler, struct kib_hca_dev, ibh_event_handler);
2541
2542         write_lock_irqsave(g_lock, flags);
2543
2544         switch (event->event) {
2545         case IB_EVENT_DEVICE_FATAL:
2546                 CDEBUG(D_NET, "IB device fatal\n");
2547                 hdev->ibh_state = IBLND_DEV_FATAL;
2548                 kiblnd_set_ni_fatal_on(hdev, 1);
2549                 break;
2550         case IB_EVENT_PORT_ACTIVE:
2551                 CDEBUG(D_NET, "IB port active\n");
2552                 if (event->element.port_num == hdev->ibh_port) {
2553                         hdev->ibh_state = IBLND_DEV_PORT_ACTIVE;
2554                         kiblnd_set_ni_fatal_on(hdev, 0);
2555                 }
2556                 break;
2557         case IB_EVENT_PORT_ERR:
2558                 CDEBUG(D_NET, "IB port err\n");
2559                 if (event->element.port_num == hdev->ibh_port) {
2560                         hdev->ibh_state = IBLND_DEV_PORT_DOWN;
2561                         kiblnd_set_ni_fatal_on(hdev, 1);
2562                 }
2563                 break;
2564         default:
2565                 break;
2566         }
2567         write_unlock_irqrestore(g_lock, flags);
2568 }
2569
2570 static int
2571 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2572 {
2573         struct ib_device_attr *dev_attr;
2574         int rc = 0;
2575         int rc2 = 0;
2576
2577         /* It's safe to assume a HCA can handle a page size
2578          * matching that of the native system */
2579         hdev->ibh_page_shift = PAGE_SHIFT;
2580         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2581         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2582
2583 #ifndef HAVE_IB_DEVICE_ATTRS
2584         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2585         if (dev_attr == NULL) {
2586                 CERROR("Out of memory\n");
2587                 return -ENOMEM;
2588         }
2589
2590         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2591         if (rc != 0) {
2592                 CERROR("Failed to query IB device: %d\n", rc);
2593                 goto out_clean_attr;
2594         }
2595 #else
2596         dev_attr = &hdev->ibh_ibdev->attrs;
2597 #endif
2598
2599         hdev->ibh_mr_size = dev_attr->max_mr_size;
2600         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2601
2602         /* Setup device Memory Registration capabilities */
2603 #ifdef HAVE_IB_DEVICE_OPS
2604         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2605             hdev->ibh_ibdev->ops.dealloc_fmr &&
2606             hdev->ibh_ibdev->ops.map_phys_fmr &&
2607             hdev->ibh_ibdev->ops.unmap_fmr) {
2608 #else
2609         if (hdev->ibh_ibdev->alloc_fmr &&
2610             hdev->ibh_ibdev->dealloc_fmr &&
2611             hdev->ibh_ibdev->map_phys_fmr &&
2612             hdev->ibh_ibdev->unmap_fmr) {
2613 #endif
2614                 LCONSOLE_INFO("Using FMR for registration\n");
2615                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2616         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2617                 LCONSOLE_INFO("Using FastReg for registration\n");
2618                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2619 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2620 #ifdef IB_DEVICE_SG_GAPS_REG
2621                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2622                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2623 #endif
2624 #endif
2625         } else {
2626                 rc = -ENOSYS;
2627         }
2628
2629         rc2 = kiblnd_port_get_attr(hdev);
2630         if (rc2 != 0)
2631                 return rc2;
2632
2633         if (rc != 0)
2634                 rc = -EINVAL;
2635
2636 #ifndef HAVE_IB_DEVICE_ATTRS
2637 out_clean_attr:
2638         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2639 #endif
2640
2641         if (rc == -ENOSYS)
2642                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2643                        "register memory: %d\n", rc);
2644         else if (rc == -EINVAL)
2645                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2646         return rc;
2647 }
2648
2649 #ifdef HAVE_IB_GET_DMA_MR
2650 static void
2651 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2652 {
2653         if (hdev->ibh_mrs == NULL)
2654                 return;
2655
2656         ib_dereg_mr(hdev->ibh_mrs);
2657
2658         hdev->ibh_mrs = NULL;
2659 }
2660 #endif
2661
2662 void
2663 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2664 {
2665         if (hdev->ibh_event_handler.device != NULL)
2666                 ib_unregister_event_handler(&hdev->ibh_event_handler);
2667
2668 #ifdef HAVE_IB_GET_DMA_MR
2669         kiblnd_hdev_cleanup_mrs(hdev);
2670 #endif
2671
2672         if (hdev->ibh_pd != NULL)
2673                 ib_dealloc_pd(hdev->ibh_pd);
2674
2675         if (hdev->ibh_cmid != NULL)
2676                 rdma_destroy_id(hdev->ibh_cmid);
2677
2678         LIBCFS_FREE(hdev, sizeof(*hdev));
2679 }
2680
2681 #ifdef HAVE_IB_GET_DMA_MR
2682 static int
2683 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2684 {
2685         struct ib_mr *mr;
2686         int           acflags = IB_ACCESS_LOCAL_WRITE |
2687                                 IB_ACCESS_REMOTE_WRITE;
2688
2689         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2690         if (IS_ERR(mr)) {
2691                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2692                 kiblnd_hdev_cleanup_mrs(hdev);
2693                 return PTR_ERR(mr);
2694         }
2695
2696         hdev->ibh_mrs = mr;
2697
2698         return 0;
2699 }
2700 #endif
2701
2702 static int
2703 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2704 {       /* DUMMY */
2705         return 0;
2706 }
2707
2708 static int
2709 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2710 {
2711         struct rdma_cm_id  *cmid;
2712         struct sockaddr_in  srcaddr;
2713         struct sockaddr_in  dstaddr;
2714         int                 rc;
2715
2716         if (dev->ibd_hdev == NULL || /* initializing */
2717             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2718             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2719                 return 1;
2720
2721         /* XXX: it's UGLY, but I don't have better way to find
2722          * ib-bonding HCA failover because:
2723          *
2724          * a. no reliable CM event for HCA failover...
2725          * b. no OFED API to get ib_device for current net_device...
2726          *
2727          * We have only two choices at this point:
2728          *
2729          * a. rdma_bind_addr(), it will conflict with listener cmid
2730          * b. rdma_resolve_addr() to zero addr */
2731         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2732                                      RDMA_PS_TCP, IB_QPT_RC);
2733         if (IS_ERR(cmid)) {
2734                 rc = PTR_ERR(cmid);
2735                 CERROR("Failed to create cmid for failover: %d\n", rc);
2736                 return rc;
2737         }
2738
2739         memset(&srcaddr, 0, sizeof(srcaddr));
2740         srcaddr.sin_family      = AF_INET;
2741         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2742
2743         memset(&dstaddr, 0, sizeof(dstaddr));
2744         dstaddr.sin_family = AF_INET;
2745         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2746                                (struct sockaddr *)&dstaddr, 1);
2747         if (rc != 0 || cmid->device == NULL) {
2748                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2749                        dev->ibd_ifname, &dev->ibd_ifip,
2750                        cmid->device, rc);
2751                 rdma_destroy_id(cmid);
2752                 return rc;
2753         }
2754
2755         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2756         rdma_destroy_id(cmid);
2757         return rc;
2758 }
2759
2760 int
2761 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2762 {
2763         LIST_HEAD(zombie_tpo);
2764         LIST_HEAD(zombie_ppo);
2765         LIST_HEAD(zombie_fpo);
2766         struct rdma_cm_id  *cmid  = NULL;
2767         struct kib_hca_dev *hdev  = NULL;
2768         struct kib_hca_dev *old;
2769         struct ib_pd       *pd;
2770         struct kib_net *net;
2771         struct sockaddr_in  addr;
2772         unsigned long       flags;
2773         int                 rc = 0;
2774         int                 i;
2775
2776         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2777                  dev->ibd_can_failover ||
2778                  dev->ibd_hdev == NULL);
2779
2780         rc = kiblnd_dev_need_failover(dev, ns);
2781         if (rc <= 0)
2782                 goto out;
2783
2784         if (dev->ibd_hdev != NULL &&
2785             dev->ibd_hdev->ibh_cmid != NULL) {
2786                 /* XXX it's not good to close old listener at here,
2787                  * because we can fail to create new listener.
2788                  * But we have to close it now, otherwise rdma_bind_addr
2789                  * will return EADDRINUSE... How crap! */
2790                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2791
2792                 cmid = dev->ibd_hdev->ibh_cmid;
2793                 /* make next schedule of kiblnd_dev_need_failover()
2794                  * return 1 for me */
2795                 dev->ibd_hdev->ibh_cmid  = NULL;
2796                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2797
2798                 rdma_destroy_id(cmid);
2799         }
2800
2801         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2802                                      IB_QPT_RC);
2803         if (IS_ERR(cmid)) {
2804                 rc = PTR_ERR(cmid);
2805                 CERROR("Failed to create cmid for failover: %d\n", rc);
2806                 goto out;
2807         }
2808
2809         memset(&addr, 0, sizeof(addr));
2810         addr.sin_family      = AF_INET;
2811         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2812         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2813
2814         /* Bind to failover device or port */
2815         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2816         if (rc != 0 || cmid->device == NULL) {
2817                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2818                        dev->ibd_ifname, &dev->ibd_ifip,
2819                        cmid->device, rc);
2820                 rdma_destroy_id(cmid);
2821                 goto out;
2822         }
2823
2824         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2825         if (hdev == NULL) {
2826                 CERROR("Failed to allocate kib_hca_dev\n");
2827                 rdma_destroy_id(cmid);
2828                 rc = -ENOMEM;
2829                 goto out;
2830         }
2831
2832         atomic_set(&hdev->ibh_ref, 1);
2833         hdev->ibh_dev   = dev;
2834         hdev->ibh_cmid  = cmid;
2835         hdev->ibh_ibdev = cmid->device;
2836         hdev->ibh_port  = cmid->port_num;
2837
2838 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2839         pd = ib_alloc_pd(cmid->device, 0);
2840 #else
2841         pd = ib_alloc_pd(cmid->device);
2842 #endif
2843         if (IS_ERR(pd)) {
2844                 rc = PTR_ERR(pd);
2845                 CERROR("Can't allocate PD: %d\n", rc);
2846                 goto out;
2847         }
2848
2849         hdev->ibh_pd = pd;
2850
2851         rc = rdma_listen(cmid, 0);
2852         if (rc != 0) {
2853                 CERROR("Can't start new listener: %d\n", rc);
2854                 goto out;
2855         }
2856
2857         rc = kiblnd_hdev_get_attr(hdev);
2858         if (rc != 0) {
2859                 CERROR("Can't get device attributes: %d\n", rc);
2860                 goto out;
2861         }
2862
2863 #ifdef HAVE_IB_GET_DMA_MR
2864         rc = kiblnd_hdev_setup_mrs(hdev);
2865         if (rc != 0) {
2866                 CERROR("Can't setup device: %d\n", rc);
2867                 goto out;
2868         }
2869 #endif
2870
2871         INIT_IB_EVENT_HANDLER(&hdev->ibh_event_handler,
2872                                 hdev->ibh_ibdev, kiblnd_event_handler);
2873         ib_register_event_handler(&hdev->ibh_event_handler);
2874
2875         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2876
2877         old = dev->ibd_hdev;
2878         dev->ibd_hdev = hdev;   /* take over the refcount */
2879         hdev = old;
2880
2881         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2882                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2883                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2884                                             &zombie_tpo);
2885
2886                         if (net->ibn_fmr_ps != NULL)
2887                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2888                                                         &zombie_fpo);
2889                 }
2890         }
2891
2892         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2893  out:
2894         if (!list_empty(&zombie_tpo))
2895                 kiblnd_destroy_pool_list(&zombie_tpo);
2896         if (!list_empty(&zombie_ppo))
2897                 kiblnd_destroy_pool_list(&zombie_ppo);
2898         if (!list_empty(&zombie_fpo))
2899                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2900         if (hdev != NULL)
2901                 kiblnd_hdev_decref(hdev);
2902
2903         if (rc != 0)
2904                 dev->ibd_failed_failover++;
2905         else
2906                 dev->ibd_failed_failover = 0;
2907
2908         return rc;
2909 }
2910
2911 void
2912 kiblnd_destroy_dev(struct kib_dev *dev)
2913 {
2914         LASSERT(dev->ibd_nnets == 0);
2915         LASSERT(list_empty(&dev->ibd_nets));
2916
2917         list_del(&dev->ibd_fail_list);
2918         list_del(&dev->ibd_list);
2919
2920         if (dev->ibd_hdev != NULL)
2921                 kiblnd_hdev_decref(dev->ibd_hdev);
2922
2923         LIBCFS_FREE(dev, sizeof(*dev));
2924 }
2925
2926 static void
2927 kiblnd_base_shutdown(void)
2928 {
2929         struct kib_sched_info   *sched;
2930         int                     i;
2931
2932         LASSERT(list_empty(&kiblnd_data.kib_devs));
2933
2934         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2935                atomic_read(&libcfs_kmemory));
2936
2937         switch (kiblnd_data.kib_init) {
2938         default:
2939                 LBUG();
2940
2941         case IBLND_INIT_ALL:
2942         case IBLND_INIT_DATA:
2943                 LASSERT (kiblnd_data.kib_peers != NULL);
2944                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2945                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2946                 }
2947                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2948                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2949                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2950                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2951
2952                 /* flag threads to terminate; wake and wait for them to die */
2953                 kiblnd_data.kib_shutdown = 1;
2954
2955                 /* NB: we really want to stop scheduler threads net by net
2956                  * instead of the whole module, this should be improved
2957                  * with dynamic configuration LNet */
2958                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2959                         wake_up_all(&sched->ibs_waitq);
2960
2961                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2962                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2963
2964                 wait_var_event_warning(&kiblnd_data.kib_nthreads,
2965                                        !atomic_read(&kiblnd_data.kib_nthreads),
2966                                        "Waiting for %d threads to terminate\n",
2967                                        atomic_read(&kiblnd_data.kib_nthreads));
2968                 /* fall through */
2969
2970         case IBLND_INIT_NOTHING:
2971                 break;
2972         }
2973
2974         if (kiblnd_data.kib_peers)
2975                 CFS_FREE_PTR_ARRAY(kiblnd_data.kib_peers,
2976                                    kiblnd_data.kib_peer_hash_size);
2977
2978         if (kiblnd_data.kib_scheds != NULL)
2979                 cfs_percpt_free(kiblnd_data.kib_scheds);
2980
2981         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2982                atomic_read(&libcfs_kmemory));
2983
2984         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2985         module_put(THIS_MODULE);
2986 }
2987
2988 static void
2989 kiblnd_shutdown(struct lnet_ni *ni)
2990 {
2991         struct kib_net *net = ni->ni_data;
2992         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2993         unsigned long     flags;
2994
2995         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2996
2997         if (net == NULL)
2998                 goto out;
2999
3000         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3001                atomic_read(&libcfs_kmemory));
3002
3003         write_lock_irqsave(g_lock, flags);
3004         net->ibn_shutdown = 1;
3005         write_unlock_irqrestore(g_lock, flags);
3006
3007         switch (net->ibn_init) {
3008         default:
3009                 LBUG();
3010
3011         case IBLND_INIT_ALL:
3012                 /* nuke all existing peers within this net */
3013                 kiblnd_del_peer(ni, LNET_NID_ANY);
3014
3015                 /* Wait for all peer_ni state to clean up */
3016                 wait_var_event_warning(&net->ibn_npeers,
3017                                        atomic_read(&net->ibn_npeers) == 0,
3018                                        "%s: waiting for %d peers to disconnect\n",
3019                                        libcfs_nid2str(ni->ni_nid),
3020                                        atomic_read(&net->ibn_npeers));
3021
3022                 kiblnd_net_fini_pools(net);
3023
3024                 write_lock_irqsave(g_lock, flags);
3025                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3026                 net->ibn_dev->ibd_nnets--;
3027                 list_del(&net->ibn_list);
3028                 write_unlock_irqrestore(g_lock, flags);
3029
3030                 /* fall through */
3031
3032         case IBLND_INIT_NOTHING:
3033                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3034
3035                 if (net->ibn_dev != NULL &&
3036                     net->ibn_dev->ibd_nnets == 0)
3037                         kiblnd_destroy_dev(net->ibn_dev);
3038
3039                 break;
3040         }
3041
3042         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3043                atomic_read(&libcfs_kmemory));
3044
3045         net->ibn_init = IBLND_INIT_NOTHING;
3046         ni->ni_data = NULL;
3047
3048         LIBCFS_FREE(net, sizeof(*net));
3049
3050 out:
3051         if (list_empty(&kiblnd_data.kib_devs))
3052                 kiblnd_base_shutdown();
3053 }
3054
3055 static int
3056 kiblnd_base_startup(struct net *ns)
3057 {
3058         struct kib_sched_info   *sched;
3059         int                     rc;
3060         int                     i;
3061
3062         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3063
3064         if (!try_module_get(THIS_MODULE))
3065                 goto failed;
3066
3067         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3068
3069         rwlock_init(&kiblnd_data.kib_global_lock);
3070
3071         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3072         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3073
3074         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3075         CFS_ALLOC_PTR_ARRAY(kiblnd_data.kib_peers,
3076                             kiblnd_data.kib_peer_hash_size);
3077         if (kiblnd_data.kib_peers == NULL)
3078                 goto failed;
3079
3080         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3081                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3082
3083         spin_lock_init(&kiblnd_data.kib_connd_lock);
3084         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3085         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3086         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3087         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3088
3089         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3090         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3091
3092         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3093                                                   sizeof(*sched));
3094         if (kiblnd_data.kib_scheds == NULL)
3095                 goto failed;
3096
3097         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3098                 int     nthrs;
3099
3100                 spin_lock_init(&sched->ibs_lock);
3101                 INIT_LIST_HEAD(&sched->ibs_conns);
3102                 init_waitqueue_head(&sched->ibs_waitq);
3103
3104                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3105                 if (*kiblnd_tunables.kib_nscheds > 0) {
3106                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3107                 } else {
3108                         /* max to half of CPUs, another half is reserved for
3109                          * upper layer modules */
3110                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3111                 }
3112
3113                 sched->ibs_nthreads_max = nthrs;
3114                 sched->ibs_cpt = i;
3115         }
3116
3117         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3118
3119         /* lists/ptrs/locks initialised */
3120         kiblnd_data.kib_init = IBLND_INIT_DATA;
3121         /*****************************************************/
3122
3123         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3124         if (rc != 0) {
3125                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3126                 goto failed;
3127         }
3128
3129         if (*kiblnd_tunables.kib_dev_failover != 0)
3130                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3131                                          "kiblnd_failover");
3132
3133         if (rc != 0) {
3134                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3135                 goto failed;
3136         }
3137
3138         /* flag everything initialised */
3139         kiblnd_data.kib_init = IBLND_INIT_ALL;
3140         /*****************************************************/
3141
3142         return 0;
3143
3144  failed:
3145         kiblnd_base_shutdown();
3146         return -ENETDOWN;
3147 }
3148
3149 static int
3150 kiblnd_start_schedulers(struct kib_sched_info *sched)
3151 {
3152         int     rc = 0;
3153         int     nthrs;
3154         int     i;
3155
3156         if (sched->ibs_nthreads == 0) {
3157                 if (*kiblnd_tunables.kib_nscheds > 0) {
3158                         nthrs = sched->ibs_nthreads_max;
3159                 } else {
3160                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3161                                                sched->ibs_cpt);
3162                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3163                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3164                 }
3165         } else {
3166                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3167                 /* increase one thread if there is new interface */
3168                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3169         }
3170
3171         for (i = 0; i < nthrs; i++) {
3172                 long    id;
3173                 char    name[20];
3174                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3175                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3176                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3177                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3178                 if (rc == 0)
3179                         continue;
3180
3181                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3182                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3183                 break;
3184         }
3185
3186         sched->ibs_nthreads += i;
3187         return rc;
3188 }
3189
3190 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3191                                     int ncpts)
3192 {
3193         int     cpt;
3194         int     rc;
3195         int     i;
3196
3197         for (i = 0; i < ncpts; i++) {
3198                 struct kib_sched_info *sched;
3199
3200                 cpt = (cpts == NULL) ? i : cpts[i];
3201                 sched = kiblnd_data.kib_scheds[cpt];
3202
3203                 if (!newdev && sched->ibs_nthreads > 0)
3204                         continue;
3205
3206                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3207                 if (rc != 0) {
3208                         CERROR("Failed to start scheduler threads for %s\n",
3209                                dev->ibd_ifname);
3210                         return rc;
3211                 }
3212         }
3213         return 0;
3214 }
3215
3216 static struct kib_dev *
3217 kiblnd_dev_search(char *ifname)
3218 {
3219         struct kib_dev *alias = NULL;
3220         struct kib_dev *dev;
3221         char            *colon;
3222         char            *colon2;
3223
3224         colon = strchr(ifname, ':');
3225         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3226                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3227                         return dev;
3228
3229                 if (alias != NULL)
3230                         continue;
3231
3232                 colon2 = strchr(dev->ibd_ifname, ':');
3233                 if (colon != NULL)
3234                         *colon = 0;
3235                 if (colon2 != NULL)
3236                         *colon2 = 0;
3237
3238                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3239                         alias = dev;
3240
3241                 if (colon != NULL)
3242                         *colon = ':';
3243                 if (colon2 != NULL)
3244                         *colon2 = ':';
3245         }
3246         return alias;
3247 }
3248
3249 static int
3250 kiblnd_startup(struct lnet_ni *ni)
3251 {
3252         char *ifname = NULL;
3253         struct lnet_inetdev *ifaces = NULL;
3254         struct kib_dev *ibdev = NULL;
3255         struct kib_net *net = NULL;
3256         unsigned long flags;
3257         int rc;
3258         int i;
3259         bool newdev;
3260
3261         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3262
3263         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3264                 rc = kiblnd_base_startup(ni->ni_net_ns);
3265                 if (rc != 0)
3266                         return rc;
3267         }
3268
3269         LIBCFS_ALLOC(net, sizeof(*net));
3270         ni->ni_data = net;
3271         if (net == NULL) {
3272                 rc = -ENOMEM;
3273                 goto failed;
3274         }
3275
3276         net->ibn_ni = ni;
3277         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3278
3279         kiblnd_tunables_setup(ni);
3280
3281         /*
3282          * ni_interfaces is only to support legacy pre Multi-Rail
3283          * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3284          * IP to be treated as an unique 'struct ni' interfaces instead.
3285          */
3286         if (ni->ni_interfaces[0] != NULL) {
3287                 /* Use the IPoIB interface specified in 'networks=' */
3288                 if (ni->ni_interfaces[1] != NULL) {
3289                         CERROR("ko2iblnd: Multiple interfaces not supported\n");
3290                         rc = -EINVAL;
3291                         goto failed;
3292                 }
3293
3294                 ifname = ni->ni_interfaces[0];
3295         } else {
3296                 ifname = *kiblnd_tunables.kib_default_ipif;
3297         }
3298
3299         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3300                 CERROR("IPoIB interface name too long: %s\n", ifname);
3301                 rc = -E2BIG;
3302                 goto failed;
3303         }
3304
3305         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3306         if (rc < 0)
3307                 goto failed;
3308
3309         for (i = 0; i < rc; i++) {
3310                 if (strcmp(ifname, ifaces[i].li_name) == 0)
3311                         break;
3312         }
3313
3314         if (i == rc) {
3315                 CERROR("ko2iblnd: No matching interfaces\n");
3316                 rc = -ENOENT;
3317                 goto failed;
3318         }
3319
3320         ibdev = kiblnd_dev_search(ifname);
3321         newdev = ibdev == NULL;
3322         /* hmm...create kib_dev even for alias */
3323         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3324                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3325                 if (!ibdev) {
3326                         rc = -ENOMEM;
3327                         goto failed;
3328                 }
3329
3330                 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3331                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3332                         sizeof(ibdev->ibd_ifname));
3333                 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3334
3335                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3336                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3337                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3338
3339                 /* initialize the device */
3340                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3341                 if (rc) {
3342                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n",
3343                                rc);
3344                         goto failed;
3345                 }
3346
3347                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3348         }
3349
3350         net->ibn_dev = ibdev;
3351         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3352
3353         ni->ni_dev_cpt = ifaces[i].li_cpt;
3354
3355         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3356         if (rc != 0)
3357                 goto failed;
3358
3359         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3360         if (rc != 0) {
3361                 CERROR("Failed to initialize NI pools: %d\n", rc);
3362                 goto failed;
3363         }
3364
3365         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3366         ibdev->ibd_nnets++;
3367         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3368         /* for health check */
3369         if (ibdev->ibd_hdev->ibh_state == IBLND_DEV_PORT_DOWN)
3370                 kiblnd_set_ni_fatal_on(ibdev->ibd_hdev, 1);
3371         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3372
3373         net->ibn_init = IBLND_INIT_ALL;
3374
3375         return 0;
3376
3377 failed:
3378         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3379                 kiblnd_destroy_dev(ibdev);
3380
3381         kfree(ifaces);
3382         kiblnd_shutdown(ni);
3383
3384         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3385                ifname ? ifname : "", rc);
3386
3387         return -ENETDOWN;
3388 }
3389
3390 static const struct lnet_lnd the_o2iblnd = {
3391         .lnd_type       = O2IBLND,
3392         .lnd_startup    = kiblnd_startup,
3393         .lnd_shutdown   = kiblnd_shutdown,
3394         .lnd_ctl        = kiblnd_ctl,
3395         .lnd_send       = kiblnd_send,
3396         .lnd_recv       = kiblnd_recv,
3397 };
3398
3399 static void __exit ko2iblnd_exit(void)
3400 {
3401         lnet_unregister_lnd(&the_o2iblnd);
3402 }
3403
3404 static int __init ko2iblnd_init(void)
3405 {
3406         int rc;
3407
3408         BUILD_BUG_ON(sizeof(struct kib_msg) > IBLND_MSG_SIZE);
3409         BUILD_BUG_ON(offsetof(struct kib_msg,
3410                      ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3411                      IBLND_MSG_SIZE);
3412         BUILD_BUG_ON(offsetof(struct kib_msg,
3413                      ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) >
3414                      IBLND_MSG_SIZE);
3415
3416         rc = kiblnd_tunables_init();
3417         if (rc != 0)
3418                 return rc;
3419
3420         lnet_register_lnd(&the_o2iblnd);
3421
3422         return 0;
3423 }
3424
3425 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3426 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3427 MODULE_VERSION("2.8.0");
3428 MODULE_LICENSE("GPL");
3429
3430 module_init(ko2iblnd_init);
3431 module_exit(ko2iblnd_exit);