lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include <asm/page.h>
  38 #include "o2iblnd.h"
  39
  40 static struct lnet_lnd the_o2iblnd;
  41
  42 kib_data_t              kiblnd_data;
  43
  44 static __u32
  45 kiblnd_cksum (void *ptr, int nob)
  46 {
  47         char  *c  = ptr;
  48         __u32  sum = 0;
  49
  50         while (nob-- > 0)
  51                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  52
  53         /* ensure I don't return 0 (== no checksum) */
  54         return (sum == 0) ? 1 : sum;
  55 }
  56
  57 static char *
  58 kiblnd_msgtype2str(int type)
  59 {
  60         switch (type) {
  61         case IBLND_MSG_CONNREQ:
  62                 return "CONNREQ";
  63
  64         case IBLND_MSG_CONNACK:
  65                 return "CONNACK";
  66
  67         case IBLND_MSG_NOOP:
  68                 return "NOOP";
  69
  70         case IBLND_MSG_IMMEDIATE:
  71                 return "IMMEDIATE";
  72
  73         case IBLND_MSG_PUT_REQ:
  74                 return "PUT_REQ";
  75
  76         case IBLND_MSG_PUT_NAK:
  77                 return "PUT_NAK";
  78
  79         case IBLND_MSG_PUT_ACK:
  80                 return "PUT_ACK";
  81
  82         case IBLND_MSG_PUT_DONE:
  83                 return "PUT_DONE";
  84
  85         case IBLND_MSG_GET_REQ:
  86                 return "GET_REQ";
  87
  88         case IBLND_MSG_GET_DONE:
  89                 return "GET_DONE";
  90
  91         default:
  92                 return "???";
  93         }
  94 }
  95
  96 static int
  97 kiblnd_msgtype2size(int type)
  98 {
  99         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 100
 101         switch (type) {
 102         case IBLND_MSG_CONNREQ:
 103         case IBLND_MSG_CONNACK:
 104                 return hdr_size + sizeof(kib_connparams_t);
 105
 106         case IBLND_MSG_NOOP:
 107                 return hdr_size;
 108
 109         case IBLND_MSG_IMMEDIATE:
 110                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 111
 112         case IBLND_MSG_PUT_REQ:
 113                 return hdr_size + sizeof(kib_putreq_msg_t);
 114
 115         case IBLND_MSG_PUT_ACK:
 116                 return hdr_size + sizeof(kib_putack_msg_t);
 117
 118         case IBLND_MSG_GET_REQ:
 119                 return hdr_size + sizeof(kib_get_msg_t);
 120
 121         case IBLND_MSG_PUT_NAK:
 122         case IBLND_MSG_PUT_DONE:
 123         case IBLND_MSG_GET_DONE:
 124                 return hdr_size + sizeof(kib_completion_msg_t);
 125         default:
 126                 return -1;
 127         }
 128 }
 129
 130 static int
 131 kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 132 {
 133         kib_rdma_desc_t   *rd;
 134         int                nob;
 135         int                n;
 136         int                i;
 137
 138         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 139                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 140
 141         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 142                               &msg->ibm_u.get.ibgm_rd :
 143                               &msg->ibm_u.putack.ibpam_rd;
 144
 145         if (flip) {
 146                 __swab32s(&rd->rd_key);
 147                 __swab32s(&rd->rd_nfrags);
 148         }
 149
 150         n = rd->rd_nfrags;
 151
 152         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 153                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 154                        n, IBLND_MAX_RDMA_FRAGS);
 155                 return 1;
 156         }
 157
 158         nob = offsetof (kib_msg_t, ibm_u) +
 159               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 160
 161         if (msg->ibm_nob < nob) {
 162                 CERROR("Short %s: %d(%d)\n",
 163                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 164                 return 1;
 165         }
 166
 167         if (!flip)
 168                 return 0;
 169
 170         for (i = 0; i < n; i++) {
 171                 __swab32s(&rd->rd_frags[i].rf_nob);
 172                 __swab64s(&rd->rd_frags[i].rf_addr);
 173         }
 174
 175         return 0;
 176 }
 177
 178 void
 179 kiblnd_pack_msg(struct lnet_ni *ni, kib_msg_t *msg, int version,
 180                 int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         kib_net_t *net = ni->ni_data;
 183
 184         /* CAVEAT EMPTOR! all message fields not set here should have been
 185          * initialised previously. */
 186         msg->ibm_magic    = IBLND_MSG_MAGIC;
 187         msg->ibm_version  = version;
 188         /*   ibm_type */
 189         msg->ibm_credits  = credits;
 190         /*   ibm_nob */
 191         msg->ibm_cksum    = 0;
 192         msg->ibm_srcnid   = ni->ni_nid;
 193         msg->ibm_srcstamp = net->ibn_incarnation;
 194         msg->ibm_dstnid   = dstnid;
 195         msg->ibm_dststamp = dststamp;
 196
 197         if (*kiblnd_tunables.kib_cksum) {
 198                 /* NB ibm_cksum zero while computing cksum */
 199                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 200         }
 201 }
 202
 203 int
 204 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 205 {
 206         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 207         __u32     msg_cksum;
 208         __u16     version;
 209         int       msg_nob;
 210         int       flip;
 211
 212         /* 6 bytes are enough to have received magic + version */
 213         if (nob < 6) {
 214                 CERROR("Short message: %d\n", nob);
 215                 return -EPROTO;
 216         }
 217
 218         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 219                 flip = 0;
 220         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 221                 flip = 1;
 222         } else {
 223                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 224                 return -EPROTO;
 225         }
 226
 227         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 228         if (version != IBLND_MSG_VERSION &&
 229             version != IBLND_MSG_VERSION_1) {
 230                 CERROR("Bad version: %x\n", version);
 231                 return -EPROTO;
 232         }
 233
 234         if (nob < hdr_size) {
 235                 CERROR("Short message: %d\n", nob);
 236                 return -EPROTO;
 237         }
 238
 239         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 240         if (msg_nob > nob) {
 241                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 242                 return -EPROTO;
 243         }
 244
 245         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 246          * gets flipped */
 247         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 248         msg->ibm_cksum = 0;
 249         if (msg_cksum != 0 &&
 250             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 251                 CERROR("Bad checksum\n");
 252                 return -EPROTO;
 253         }
 254
 255         msg->ibm_cksum = msg_cksum;
 256
 257         if (flip) {
 258                 /* leave magic unflipped as a clue to peer_ni endianness */
 259                 msg->ibm_version = version;
 260                 CLASSERT (sizeof(msg->ibm_type) == 1);
 261                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 262                 msg->ibm_nob     = msg_nob;
 263                 __swab64s(&msg->ibm_srcnid);
 264                 __swab64s(&msg->ibm_srcstamp);
 265                 __swab64s(&msg->ibm_dstnid);
 266                 __swab64s(&msg->ibm_dststamp);
 267         }
 268
 269         if (msg->ibm_srcnid == LNET_NID_ANY) {
 270                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 271                 return -EPROTO;
 272         }
 273
 274         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 275                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 276                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 277                 return -EPROTO;
 278         }
 279
 280         switch (msg->ibm_type) {
 281         default:
 282                 CERROR("Unknown message type %x\n", msg->ibm_type);
 283                 return -EPROTO;
 284
 285         case IBLND_MSG_NOOP:
 286         case IBLND_MSG_IMMEDIATE:
 287         case IBLND_MSG_PUT_REQ:
 288                 break;
 289
 290         case IBLND_MSG_PUT_ACK:
 291         case IBLND_MSG_GET_REQ:
 292                 if (kiblnd_unpack_rd(msg, flip))
 293                         return -EPROTO;
 294                 break;
 295
 296         case IBLND_MSG_PUT_NAK:
 297         case IBLND_MSG_PUT_DONE:
 298         case IBLND_MSG_GET_DONE:
 299                 if (flip)
 300                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 301                 break;
 302
 303         case IBLND_MSG_CONNREQ:
 304         case IBLND_MSG_CONNACK:
 305                 if (flip) {
 306                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 307                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 308                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 309                 }
 310                 break;
 311         }
 312         return 0;
 313 }
 314
 315 int
 316 kiblnd_create_peer(struct lnet_ni *ni, kib_peer_ni_t **peerp, lnet_nid_t nid)
 317 {
 318         kib_peer_ni_t   *peer_ni;
 319         kib_net_t       *net = ni->ni_data;
 320         int             cpt = lnet_cpt_of_nid(nid, ni);
 321         unsigned long   flags;
 322
 323         LASSERT(net != NULL);
 324         LASSERT(nid != LNET_NID_ANY);
 325
 326         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 327         if (peer_ni == NULL) {
 328                 CERROR("Cannot allocate peer_ni\n");
 329                 return -ENOMEM;
 330         }
 331
 332         peer_ni->ibp_ni = ni;
 333         peer_ni->ibp_nid = nid;
 334         peer_ni->ibp_error = 0;
 335         peer_ni->ibp_last_alive = 0;
 336         peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
 337         peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
 338         atomic_set(&peer_ni->ibp_refcount, 1);  /* 1 ref for caller */
 339
 340         INIT_LIST_HEAD(&peer_ni->ibp_list);     /* not in the peer_ni table yet */
 341         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 342         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 343
 344         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 345
 346         /* always called with a ref on ni, which prevents ni being shutdown */
 347         LASSERT(net->ibn_shutdown == 0);
 348
 349         /* npeers only grows with the global lock held */
 350         atomic_inc(&net->ibn_npeers);
 351
 352         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 353
 354         *peerp = peer_ni;
 355         return 0;
 356 }
 357
 358 void
 359 kiblnd_destroy_peer (kib_peer_ni_t *peer_ni)
 360 {
 361         kib_net_t *net = peer_ni->ibp_ni->ni_data;
 362
 363         LASSERT(net != NULL);
 364         LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
 365         LASSERT(!kiblnd_peer_active(peer_ni));
 366         LASSERT(kiblnd_peer_idle(peer_ni));
 367         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 368
 369         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 370
 371         /* NB a peer_ni's connections keep a reference on their peer_ni until
 372          * they are destroyed, so we can be assured that _all_ state to do
 373          * with this peer_ni has been cleaned up when its refcount drops to
 374          * zero. */
 375         atomic_dec(&net->ibn_npeers);
 376 }
 377
 378 kib_peer_ni_t *
 379 kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid)
 380 {
 381         /* the caller is responsible for accounting the additional reference
 382          * that this creates */
 383         struct list_head        *peer_list = kiblnd_nid2peerlist(nid);
 384         struct list_head        *tmp;
 385         kib_peer_ni_t           *peer_ni;
 386
 387         list_for_each(tmp, peer_list) {
 388
 389                 peer_ni = list_entry(tmp, kib_peer_ni_t, ibp_list);
 390                 LASSERT(!kiblnd_peer_idle(peer_ni));
 391
 392                 /*
 393                  * Match a peer if its NID and the NID of the local NI it
 394                  * communicates over are the same. Otherwise don't match
 395                  * the peer, which will result in a new lnd peer being
 396                  * created.
 397                  */
 398                 if (peer_ni->ibp_nid != nid ||
 399                     peer_ni->ibp_ni->ni_nid != ni->ni_nid)
 400                         continue;
 401
 402                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d) version: %x\n",
 403                        peer_ni, libcfs_nid2str(nid),
 404                        atomic_read(&peer_ni->ibp_refcount),
 405                        peer_ni->ibp_version);
 406                 return peer_ni;
 407         }
 408         return NULL;
 409 }
 410
 411 void
 412 kiblnd_unlink_peer_locked (kib_peer_ni_t *peer_ni)
 413 {
 414         LASSERT(list_empty(&peer_ni->ibp_conns));
 415
 416         LASSERT (kiblnd_peer_active(peer_ni));
 417         list_del_init(&peer_ni->ibp_list);
 418         /* lose peerlist's ref */
 419         kiblnd_peer_decref(peer_ni);
 420 }
 421
 422 static int
 423 kiblnd_get_peer_info(struct lnet_ni *ni, int index,
 424                      lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_ni_t           *peer_ni;
 427         struct list_head        *ptmp;
 428         int                      i;
 429         unsigned long            flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 438                         LASSERT(!kiblnd_peer_idle(peer_ni));
 439
 440                         if (peer_ni->ibp_ni != ni)
 441                                 continue;
 442
 443                         if (index-- > 0)
 444                                 continue;
 445
 446                         *nidp = peer_ni->ibp_nid;
 447                         *count = atomic_read(&peer_ni->ibp_refcount);
 448
 449                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 450                                                flags);
 451                         return 0;
 452                 }
 453         }
 454
 455         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 456         return -ENOENT;
 457 }
 458
 459 static void
 460 kiblnd_del_peer_locked (kib_peer_ni_t *peer_ni)
 461 {
 462         struct list_head        *ctmp;
 463         struct list_head        *cnxt;
 464         kib_conn_t              *conn;
 465
 466         if (list_empty(&peer_ni->ibp_conns)) {
 467                 kiblnd_unlink_peer_locked(peer_ni);
 468         } else {
 469                 list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
 470                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 471
 472                         kiblnd_close_conn_locked(conn, 0);
 473                 }
 474                 /* NB closing peer_ni's last conn unlinked it. */
 475         }
 476         /* NB peer_ni now unlinked; might even be freed if the peer_ni table had the
 477          * last ref on it. */
 478 }
 479
 480 static int
 481 kiblnd_del_peer(struct lnet_ni *ni, lnet_nid_t nid)
 482 {
 483         struct list_head        zombies = LIST_HEAD_INIT(zombies);
 484         struct list_head        *ptmp;
 485         struct list_head        *pnxt;
 486         kib_peer_ni_t           *peer_ni;
 487         int                     lo;
 488         int                     hi;
 489         int                     i;
 490         unsigned long           flags;
 491         int                     rc = -ENOENT;
 492
 493         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 494
 495         if (nid != LNET_NID_ANY) {
 496                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 497         } else {
 498                 lo = 0;
 499                 hi = kiblnd_data.kib_peer_hash_size - 1;
 500         }
 501
 502         for (i = lo; i <= hi; i++) {
 503                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 504                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 505                         LASSERT(!kiblnd_peer_idle(peer_ni));
 506
 507                         if (peer_ni->ibp_ni != ni)
 508                                 continue;
 509
 510                         if (!(nid == LNET_NID_ANY || peer_ni->ibp_nid == nid))
 511                                 continue;
 512
 513                         if (!list_empty(&peer_ni->ibp_tx_queue)) {
 514                                 LASSERT(list_empty(&peer_ni->ibp_conns));
 515
 516                                 list_splice_init(&peer_ni->ibp_tx_queue,
 517                                                  &zombies);
 518                         }
 519
 520                         kiblnd_del_peer_locked(peer_ni);
 521                         rc = 0;         /* matched something */
 522                 }
 523         }
 524
 525         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 526
 527         kiblnd_txlist_done(&zombies, -EIO);
 528
 529         return rc;
 530 }
 531
 532 static kib_conn_t *
 533 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 534 {
 535         kib_peer_ni_t           *peer_ni;
 536         struct list_head        *ptmp;
 537         kib_conn_t              *conn;
 538         struct list_head        *ctmp;
 539         int                     i;
 540         unsigned long           flags;
 541
 542         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 543
 544         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 545                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 546
 547                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
 548                         LASSERT(!kiblnd_peer_idle(peer_ni));
 549
 550                         if (peer_ni->ibp_ni != ni)
 551                                 continue;
 552
 553                         list_for_each(ctmp, &peer_ni->ibp_conns) {
 554                                 if (index-- > 0)
 555                                         continue;
 556
 557                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 558                                 kiblnd_conn_addref(conn);
 559                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 560                                                        flags);
 561                                 return conn;
 562                         }
 563                 }
 564         }
 565
 566         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 567         return NULL;
 568 }
 569
 570 static void
 571 kiblnd_debug_rx (kib_rx_t *rx)
 572 {
 573         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 574                rx, rx->rx_status, rx->rx_msg->ibm_type,
 575                rx->rx_msg->ibm_credits);
 576 }
 577
 578 static void
 579 kiblnd_debug_tx (kib_tx_t *tx)
 580 {
 581         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lld "
 582                "cookie %#llx msg %s%s type %x cred %d\n",
 583                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 584                tx->tx_status, ktime_to_ns(tx->tx_deadline), tx->tx_cookie,
 585                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 586                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 587                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 588 }
 589
 590 void
 591 kiblnd_debug_conn (kib_conn_t *conn)
 592 {
 593         struct list_head        *tmp;
 594         int                     i;
 595
 596         spin_lock(&conn->ibc_lock);
 597
 598         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 599                atomic_read(&conn->ibc_refcount), conn,
 600                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 601         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 602                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 603                conn->ibc_nsends_posted, conn->ibc_credits,
 604                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 605         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 606
 607         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 608         list_for_each(tmp, &conn->ibc_early_rxs)
 609                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 610
 611         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 612         list_for_each(tmp, &conn->ibc_tx_noops)
 613                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 614
 615         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 616         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 617                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 620         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   active_txs:\n");
 628         list_for_each(tmp, &conn->ibc_active_txs)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   rxs:\n");
 632         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 633                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 634
 635         spin_unlock(&conn->ibc_lock);
 636 }
 637
 638 int
 639 kiblnd_translate_mtu(int value)
 640 {
 641         switch (value) {
 642         default:
 643                 return -1;
 644         case 0:
 645                 return 0;
 646         case 256:
 647                 return IB_MTU_256;
 648         case 512:
 649                 return IB_MTU_512;
 650         case 1024:
 651                 return IB_MTU_1024;
 652         case 2048:
 653                 return IB_MTU_2048;
 654         case 4096:
 655                 return IB_MTU_4096;
 656         }
 657 }
 658
 659 static void
 660 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 661 {
 662         int           mtu;
 663
 664         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 665         if (cmid->route.path_rec == NULL)
 666                 return;
 667
 668         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 669         LASSERT (mtu >= 0);
 670         if (mtu != 0)
 671                 cmid->route.path_rec->mtu = mtu;
 672 }
 673
 674 static int
 675 kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 676 {
 677         cpumask_t       *mask;
 678         int             vectors;
 679         int             off;
 680         int             i;
 681         lnet_nid_t      ibp_nid;
 682
 683         vectors = conn->ibc_cmid->device->num_comp_vectors;
 684         if (vectors <= 1)
 685                 return 0;
 686
 687         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 688
 689         /* hash NID to CPU id in this partition... */
 690         ibp_nid = conn->ibc_peer->ibp_nid;
 691         off = do_div(ibp_nid, cpumask_weight(mask));
 692         for_each_cpu(i, mask) {
 693                 if (off-- == 0)
 694                         return i % vectors;
 695         }
 696
 697         LBUG();
 698         return 1;
 699 }
 700
 701 /*
 702  * Get the scheduler bound to this CPT. If the scheduler has no
 703  * threads, which means that the CPT has no CPUs, then grab the
 704  * next scheduler that we can use.
 705  *
 706  * This case would be triggered if a NUMA node is configured with
 707  * no associated CPUs.
 708  */
 709 static struct kib_sched_info *
 710 kiblnd_get_scheduler(int cpt)
 711 {
 712         struct kib_sched_info *sched;
 713         int i;
 714
 715         sched = kiblnd_data.kib_scheds[cpt];
 716
 717         if (sched->ibs_nthreads > 0)
 718                 return sched;
 719
 720         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 721                 if (sched->ibs_nthreads > 0) {
 722                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 723                                         cpt, sched->ibs_cpt);
 724                         return sched;
 725                 }
 726         }
 727
 728         return NULL;
 729 }
 730
 731 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 732 {
 733         /*
 734          * One WR for the LNet message
 735          * And ibc_max_frags for the transfer WRs
 736          */
 737         unsigned int ret = 1 + conn->ibc_max_frags;
 738         __u32 dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
 739
 740         /* FastReg needs two extra WRs for map and invalidate */
 741         if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
 742                 ret += 2;
 743
 744         /* account for a maximum of ibc_queue_depth in-flight transfers */
 745         ret *= conn->ibc_queue_depth;
 746         return ret;
 747 }
 748
 749 kib_conn_t *
 750 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 751                    int state, int version)
 752 {
 753         /* CAVEAT EMPTOR:
 754          * If the new conn is created successfully it takes over the caller's
 755          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 756          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 757          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 758          * to destroy 'cmid' here since I'm called from the CM which still has
 759          * its ref on 'cmid'). */
 760         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 761         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 762         kib_dev_t              *dev;
 763         struct ib_qp_init_attr *init_qp_attr;
 764         struct kib_sched_info   *sched;
 765 #ifdef HAVE_IB_CQ_INIT_ATTR
 766         struct ib_cq_init_attr  cq_attr = {};
 767 #endif
 768         kib_conn_t              *conn;
 769         struct ib_cq            *cq;
 770         unsigned long           flags;
 771         int                     cpt;
 772         int                     rc;
 773         int                     i;
 774
 775         LASSERT(net != NULL);
 776         LASSERT(!in_interrupt());
 777
 778         dev = net->ibn_dev;
 779
 780         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 781         sched = kiblnd_get_scheduler(cpt);
 782
 783         if (sched == NULL) {
 784                 CERROR("no schedulers available. node is unhealthy\n");
 785                 goto failed_0;
 786         }
 787
 788         /*
 789          * The cpt might have changed if we ended up selecting a non cpt
 790          * native scheduler. So use the scheduler's cpt instead.
 791          */
 792         cpt = sched->ibs_cpt;
 793
 794         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 795                          sizeof(*init_qp_attr));
 796         if (init_qp_attr == NULL) {
 797                 CERROR("Can't allocate qp_attr for %s\n",
 798                        libcfs_nid2str(peer_ni->ibp_nid));
 799                 goto failed_0;
 800         }
 801
 802         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 803         if (conn == NULL) {
 804                 CERROR("Can't allocate connection for %s\n",
 805                        libcfs_nid2str(peer_ni->ibp_nid));
 806                 goto failed_1;
 807         }
 808
 809         conn->ibc_state = IBLND_CONN_INIT;
 810         conn->ibc_version = version;
 811         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 812         cmid->context = conn;                   /* for future CM callbacks */
 813         conn->ibc_cmid = cmid;
 814         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 815         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 816
 817         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 818         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 819         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 820         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 821         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 822         INIT_LIST_HEAD(&conn->ibc_active_txs);
 823         spin_lock_init(&conn->ibc_lock);
 824
 825         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 826                          sizeof(*conn->ibc_connvars));
 827         if (conn->ibc_connvars == NULL) {
 828                 CERROR("Can't allocate in-progress connection state\n");
 829                 goto failed_2;
 830         }
 831
 832         write_lock_irqsave(glock, flags);
 833         if (dev->ibd_failover) {
 834                 write_unlock_irqrestore(glock, flags);
 835                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 836                 goto failed_2;
 837         }
 838
 839         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 840                 /* wakeup failover thread and teardown connection */
 841                 if (kiblnd_dev_can_failover(dev)) {
 842                         list_add_tail(&dev->ibd_fail_list,
 843                                       &kiblnd_data.kib_failed_devs);
 844                         wake_up(&kiblnd_data.kib_failover_waitq);
 845                 }
 846
 847                 write_unlock_irqrestore(glock, flags);
 848                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 849                        cmid->device->name, dev->ibd_ifname);
 850                 goto failed_2;
 851         }
 852
 853         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 854         conn->ibc_hdev = dev->ibd_hdev;
 855
 856         kiblnd_setup_mtu_locked(cmid);
 857
 858         write_unlock_irqrestore(glock, flags);
 859
 860         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 861                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 862         if (conn->ibc_rxs == NULL) {
 863                 CERROR("Cannot allocate RX buffers\n");
 864                 goto failed_2;
 865         }
 866
 867         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 868                                 IBLND_RX_MSG_PAGES(conn));
 869         if (rc != 0)
 870                 goto failed_2;
 871
 872         kiblnd_map_rx_descs(conn);
 873
 874 #ifdef HAVE_IB_CQ_INIT_ATTR
 875         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 876         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 877         cq = ib_create_cq(cmid->device,
 878                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 879                           &cq_attr);
 880 #else
 881         cq = ib_create_cq(cmid->device,
 882                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 883                           IBLND_CQ_ENTRIES(conn),
 884                           kiblnd_get_completion_vector(conn, cpt));
 885 #endif
 886         if (IS_ERR(cq)) {
 887                 /*
 888                  * on MLX-5 (possibly MLX-4 as well) this error could be
 889                  * hit if the concurrent_sends and/or peer_tx_credits is set
 890                  * too high. Or due to an MLX-5 bug which tries to
 891                  * allocate 256kb via kmalloc for WR cookie array
 892                  */
 893                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 894                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 895                 goto failed_2;
 896         }
 897
 898         conn->ibc_cq = cq;
 899
 900         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 901         if (rc != 0) {
 902                 CERROR("Can't request completion notification: %d\n", rc);
 903                 goto failed_2;
 904         }
 905
 906         init_qp_attr->event_handler = kiblnd_qp_event;
 907         init_qp_attr->qp_context = conn;
 908         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 909         init_qp_attr->cap.max_recv_sge = 1;
 910         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 911         init_qp_attr->qp_type = IB_QPT_RC;
 912         init_qp_attr->send_cq = cq;
 913         init_qp_attr->recv_cq = cq;
 914
 915         conn->ibc_sched = sched;
 916
 917         do {
 918                 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 919                 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 920
 921                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 922                 if (!rc || conn->ibc_queue_depth < 2)
 923                         break;
 924
 925                 conn->ibc_queue_depth--;
 926         } while (rc);
 927
 928         if (rc) {
 929                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 930                        "send_sge: %d, recv_sge: %d\n",
 931                        rc, init_qp_attr->cap.max_send_wr,
 932                        init_qp_attr->cap.max_recv_wr,
 933                        init_qp_attr->cap.max_send_sge,
 934                        init_qp_attr->cap.max_recv_sge);
 935                 goto failed_2;
 936         }
 937
 938         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 939                 CWARN("peer %s - queue depth reduced from %u to %u"
 940                       "  to allow for qp creation\n",
 941                       libcfs_nid2str(peer_ni->ibp_nid),
 942                       peer_ni->ibp_queue_depth,
 943                       conn->ibc_queue_depth);
 944
 945         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 946
 947         /* 1 ref for caller and each rxmsg */
 948         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 949         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 950
 951         /* post receives */
 952         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 953                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 954                 if (rc != 0) {
 955                         CERROR("Can't post rxmsg: %d\n", rc);
 956
 957                         /* Make posted receives complete */
 958                         kiblnd_abort_receives(conn);
 959
 960                         /* correct # of posted buffers
 961                          * NB locking needed now I'm racing with completion */
 962                         spin_lock_irqsave(&sched->ibs_lock, flags);
 963                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 964                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 965
 966                         /* cmid will be destroyed by CM(ofed) after cm_callback
 967                          * returned, so we can't refer it anymore
 968                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 969                         rdma_destroy_qp(conn->ibc_cmid);
 970                         conn->ibc_cmid = NULL;
 971
 972                         /* Drop my own and unused rxbuffer refcounts */
 973                         while (i++ <= IBLND_RX_MSGS(conn))
 974                                 kiblnd_conn_decref(conn);
 975
 976                         return NULL;
 977                 }
 978         }
 979
 980         /* Init successful! */
 981         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 982                  state == IBLND_CONN_PASSIVE_WAIT);
 983         conn->ibc_state = state;
 984
 985         /* 1 more conn */
 986         atomic_inc(&net->ibn_nconns);
 987         return conn;
 988
 989  failed_2:
 990         kiblnd_destroy_conn(conn);
 991         LIBCFS_FREE(conn, sizeof(*conn));
 992  failed_1:
 993         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 994  failed_0:
 995         return NULL;
 996 }
 997
 998 void
 999 kiblnd_destroy_conn(kib_conn_t *conn)
1000 {
1001         struct rdma_cm_id *cmid = conn->ibc_cmid;
1002         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
1003         int                rc;
1004
1005         LASSERT (!in_interrupt());
1006         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1007         LASSERT(list_empty(&conn->ibc_early_rxs));
1008         LASSERT(list_empty(&conn->ibc_tx_noops));
1009         LASSERT(list_empty(&conn->ibc_tx_queue));
1010         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1011         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1012         LASSERT(list_empty(&conn->ibc_active_txs));
1013         LASSERT (conn->ibc_noops_posted == 0);
1014         LASSERT (conn->ibc_nsends_posted == 0);
1015
1016         switch (conn->ibc_state) {
1017         default:
1018                 /* conn must be completely disengaged from the network */
1019                 LBUG();
1020
1021         case IBLND_CONN_DISCONNECTED:
1022                 /* connvars should have been freed already */
1023                 LASSERT (conn->ibc_connvars == NULL);
1024                 break;
1025
1026         case IBLND_CONN_INIT:
1027                 break;
1028         }
1029
1030         /* conn->ibc_cmid might be destroyed by CM already */
1031         if (cmid != NULL && cmid->qp != NULL)
1032                 rdma_destroy_qp(cmid);
1033
1034         if (conn->ibc_cq != NULL) {
1035                 rc = ib_destroy_cq(conn->ibc_cq);
1036                 if (rc != 0)
1037                         CWARN("Error destroying CQ: %d\n", rc);
1038         }
1039
1040         if (conn->ibc_rx_pages != NULL)
1041                 kiblnd_unmap_rx_descs(conn);
1042
1043         if (conn->ibc_rxs != NULL) {
1044                 LIBCFS_FREE(conn->ibc_rxs,
1045                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1046         }
1047
1048         if (conn->ibc_connvars != NULL)
1049                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1050
1051         if (conn->ibc_hdev != NULL)
1052                 kiblnd_hdev_decref(conn->ibc_hdev);
1053
1054         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1055         if (conn->ibc_state != IBLND_CONN_INIT) {
1056                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1057
1058                 kiblnd_peer_decref(peer_ni);
1059                 rdma_destroy_id(cmid);
1060                 atomic_dec(&net->ibn_nconns);
1061         }
1062 }
1063
1064 int
1065 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1066 {
1067         kib_conn_t              *conn;
1068         struct list_head        *ctmp;
1069         struct list_head        *cnxt;
1070         int                     count = 0;
1071
1072         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1073                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1074
1075                 CDEBUG(D_NET, "Closing conn -> %s, "
1076                               "version: %x, reason: %d\n",
1077                        libcfs_nid2str(peer_ni->ibp_nid),
1078                        conn->ibc_version, why);
1079
1080                 kiblnd_close_conn_locked(conn, why);
1081                 count++;
1082         }
1083
1084         return count;
1085 }
1086
1087 int
1088 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1089                                 int version, __u64 incarnation)
1090 {
1091         kib_conn_t              *conn;
1092         struct list_head        *ctmp;
1093         struct list_head        *cnxt;
1094         int                     count = 0;
1095
1096         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1097                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1098
1099                 if (conn->ibc_version     == version &&
1100                     conn->ibc_incarnation == incarnation)
1101                         continue;
1102
1103                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1104                               "incarnation:%#llx(%x, %#llx)\n",
1105                        libcfs_nid2str(peer_ni->ibp_nid),
1106                        conn->ibc_version, conn->ibc_incarnation,
1107                        version, incarnation);
1108
1109                 kiblnd_close_conn_locked(conn, -ESTALE);
1110                 count++;
1111         }
1112
1113         return count;
1114 }
1115
1116 static int
1117 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1118 {
1119         kib_peer_ni_t           *peer_ni;
1120         struct list_head        *ptmp;
1121         struct list_head        *pnxt;
1122         int                     lo;
1123         int                     hi;
1124         int                     i;
1125         unsigned long           flags;
1126         int                     count = 0;
1127
1128         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1129
1130         if (nid != LNET_NID_ANY)
1131                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1132         else {
1133                 lo = 0;
1134                 hi = kiblnd_data.kib_peer_hash_size - 1;
1135         }
1136
1137         for (i = lo; i <= hi; i++) {
1138                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1139
1140                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1141                         LASSERT(!kiblnd_peer_idle(peer_ni));
1142
1143                         if (peer_ni->ibp_ni != ni)
1144                                 continue;
1145
1146                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1147                                 continue;
1148
1149                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1150                 }
1151         }
1152
1153         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1154
1155         /* wildcards always succeed */
1156         if (nid == LNET_NID_ANY)
1157                 return 0;
1158
1159         return (count == 0) ? -ENOENT : 0;
1160 }
1161
1162 static int
1163 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1164 {
1165         struct libcfs_ioctl_data *data = arg;
1166         int                       rc = -EINVAL;
1167
1168         switch(cmd) {
1169         case IOC_LIBCFS_GET_PEER: {
1170                 lnet_nid_t   nid = 0;
1171                 int          count = 0;
1172
1173                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1174                                           &nid, &count);
1175                 data->ioc_nid    = nid;
1176                 data->ioc_count  = count;
1177                 break;
1178         }
1179
1180         case IOC_LIBCFS_DEL_PEER: {
1181                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1182                 break;
1183         }
1184         case IOC_LIBCFS_GET_CONN: {
1185                 kib_conn_t *conn;
1186
1187                 rc = 0;
1188                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1189                 if (conn == NULL) {
1190                         rc = -ENOENT;
1191                         break;
1192                 }
1193
1194                 LASSERT(conn->ibc_cmid != NULL);
1195                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1196                 if (conn->ibc_cmid->route.path_rec == NULL)
1197                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1198                 else
1199                         data->ioc_u32[0] =
1200                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1201                 kiblnd_conn_decref(conn);
1202                 break;
1203         }
1204         case IOC_LIBCFS_CLOSE_CONNECTION: {
1205                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1206                 break;
1207         }
1208
1209         default:
1210                 break;
1211         }
1212
1213         return rc;
1214 }
1215
1216 static void
1217 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1218 {
1219         time64_t last_alive = 0;
1220         time64_t now = ktime_get_seconds();
1221         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1222         kib_peer_ni_t *peer_ni;
1223         unsigned long flags;
1224
1225         read_lock_irqsave(glock, flags);
1226
1227         peer_ni = kiblnd_find_peer_locked(ni, nid);
1228         if (peer_ni != NULL)
1229                 last_alive = peer_ni->ibp_last_alive;
1230
1231         read_unlock_irqrestore(glock, flags);
1232
1233         if (last_alive != 0)
1234                 *when = last_alive;
1235
1236         /* peer_ni is not persistent in hash, trigger peer_ni creation
1237          * and connection establishment with a NULL tx */
1238         if (peer_ni == NULL)
1239                 kiblnd_launch_tx(ni, NULL, nid);
1240
1241         CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1242                libcfs_nid2str(nid), peer_ni,
1243                last_alive ? now - last_alive : -1);
1244         return;
1245 }
1246
1247 static void
1248 kiblnd_free_pages(kib_pages_t *p)
1249 {
1250         int     npages = p->ibp_npages;
1251         int     i;
1252
1253         for (i = 0; i < npages; i++) {
1254                 if (p->ibp_pages[i] != NULL)
1255                         __free_page(p->ibp_pages[i]);
1256         }
1257
1258         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1259 }
1260
1261 int
1262 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1263 {
1264         kib_pages_t     *p;
1265         int             i;
1266
1267         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1268                          offsetof(kib_pages_t, ibp_pages[npages]));
1269         if (p == NULL) {
1270                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1271                 return -ENOMEM;
1272         }
1273
1274         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1275         p->ibp_npages = npages;
1276
1277         for (i = 0; i < npages; i++) {
1278                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1279                                                      GFP_NOFS);
1280                 if (p->ibp_pages[i] == NULL) {
1281                         CERROR("Can't allocate page %d of %d\n", i, npages);
1282                         kiblnd_free_pages(p);
1283                         return -ENOMEM;
1284                 }
1285         }
1286
1287         *pp = p;
1288         return 0;
1289 }
1290
1291 void
1292 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1293 {
1294         kib_rx_t *rx;
1295         int       i;
1296
1297         LASSERT (conn->ibc_rxs != NULL);
1298         LASSERT (conn->ibc_hdev != NULL);
1299
1300         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1301                 rx = &conn->ibc_rxs[i];
1302
1303                 LASSERT(rx->rx_nob >= 0); /* not posted */
1304
1305                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1306                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1307                                                           rx->rx_msgaddr),
1308                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1309         }
1310
1311         kiblnd_free_pages(conn->ibc_rx_pages);
1312
1313         conn->ibc_rx_pages = NULL;
1314 }
1315
1316 void
1317 kiblnd_map_rx_descs(kib_conn_t *conn)
1318 {
1319         kib_rx_t       *rx;
1320         struct page    *pg;
1321         int             pg_off;
1322         int             ipg;
1323         int             i;
1324
1325         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1326                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1327                 rx = &conn->ibc_rxs[i];
1328
1329                 rx->rx_conn = conn;
1330                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1331
1332                 rx->rx_msgaddr =
1333                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1334                                               rx->rx_msg, IBLND_MSG_SIZE,
1335                                               DMA_FROM_DEVICE);
1336                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1337                                                   rx->rx_msgaddr));
1338                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1339
1340                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1341                        i, rx->rx_msg, rx->rx_msgaddr,
1342                        (__u64)(page_to_phys(pg) + pg_off));
1343
1344                 pg_off += IBLND_MSG_SIZE;
1345                 LASSERT(pg_off <= PAGE_SIZE);
1346
1347                 if (pg_off == PAGE_SIZE) {
1348                         pg_off = 0;
1349                         ipg++;
1350                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1351                 }
1352         }
1353 }
1354
1355 static void
1356 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1357 {
1358         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1359         kib_tx_t       *tx;
1360         int             i;
1361
1362         LASSERT (tpo->tpo_pool.po_allocated == 0);
1363
1364         if (hdev == NULL)
1365                 return;
1366
1367         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1368                 tx = &tpo->tpo_tx_descs[i];
1369                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1370                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1371                                                           tx->tx_msgaddr),
1372                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1373         }
1374
1375         kiblnd_hdev_decref(hdev);
1376         tpo->tpo_hdev = NULL;
1377 }
1378
1379 static kib_hca_dev_t *
1380 kiblnd_current_hdev(kib_dev_t *dev)
1381 {
1382         kib_hca_dev_t *hdev;
1383         unsigned long  flags;
1384         int            i = 0;
1385
1386         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1387         while (dev->ibd_failover) {
1388                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1389                 if (i++ % 50 == 0)
1390                         CDEBUG(D_NET, "%s: Wait for failover\n",
1391                                dev->ibd_ifname);
1392                 set_current_state(TASK_INTERRUPTIBLE);
1393                 schedule_timeout(cfs_time_seconds(1) / 100);
1394
1395                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1396         }
1397
1398         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1399         hdev = dev->ibd_hdev;
1400
1401         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1402
1403         return hdev;
1404 }
1405
1406 static void
1407 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1408 {
1409         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1410         kib_pool_t     *pool  = &tpo->tpo_pool;
1411         kib_net_t      *net   = pool->po_owner->ps_net;
1412         kib_dev_t      *dev;
1413         struct page    *page;
1414         kib_tx_t       *tx;
1415         int             page_offset;
1416         int             ipage;
1417         int             i;
1418
1419         LASSERT (net != NULL);
1420
1421         dev = net->ibn_dev;
1422
1423         /* pre-mapped messages are not bigger than 1 page */
1424         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1425
1426         /* No fancy arithmetic when we do the buffer calculations */
1427         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1428
1429         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1430
1431         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1432                 page = txpgs->ibp_pages[ipage];
1433                 tx = &tpo->tpo_tx_descs[i];
1434
1435                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1436                                            page_offset);
1437
1438                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1439                                                        tx->tx_msg,
1440                                                        IBLND_MSG_SIZE,
1441                                                        DMA_TO_DEVICE);
1442                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1443                                                   tx->tx_msgaddr));
1444                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1445
1446                 list_add(&tx->tx_list, &pool->po_free_list);
1447
1448                 page_offset += IBLND_MSG_SIZE;
1449                 LASSERT(page_offset <= PAGE_SIZE);
1450
1451                 if (page_offset == PAGE_SIZE) {
1452                         page_offset = 0;
1453                         ipage++;
1454                         LASSERT(ipage <= txpgs->ibp_npages);
1455                 }
1456         }
1457 }
1458
1459 static void
1460 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1461 {
1462         LASSERT(fpo->fpo_map_count == 0);
1463
1464         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1465                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1466         } else {
1467                 struct kib_fast_reg_descriptor *frd, *tmp;
1468                 int i = 0;
1469
1470                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1471                                          frd_list) {
1472                         list_del(&frd->frd_list);
1473 #ifndef HAVE_IB_MAP_MR_SG
1474                         ib_free_fast_reg_page_list(frd->frd_frpl);
1475 #endif
1476                         ib_dereg_mr(frd->frd_mr);
1477                         LIBCFS_FREE(frd, sizeof(*frd));
1478                         i++;
1479                 }
1480                 if (i < fpo->fast_reg.fpo_pool_size)
1481                         CERROR("FastReg pool still has %d regions registered\n",
1482                                 fpo->fast_reg.fpo_pool_size - i);
1483         }
1484
1485         if (fpo->fpo_hdev)
1486                 kiblnd_hdev_decref(fpo->fpo_hdev);
1487
1488         LIBCFS_FREE(fpo, sizeof(*fpo));
1489 }
1490
1491 static void
1492 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1493 {
1494         kib_fmr_pool_t *fpo, *tmp;
1495
1496         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1497                 list_del(&fpo->fpo_list);
1498                 kiblnd_destroy_fmr_pool(fpo);
1499         }
1500 }
1501
1502 static int
1503 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1504                      int ncpts)
1505 {
1506         int size = tunables->lnd_fmr_pool_size / ncpts;
1507
1508         return max(IBLND_FMR_POOL, size);
1509 }
1510
1511 static int
1512 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1513                          int ncpts)
1514 {
1515         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1516
1517         return max(IBLND_FMR_POOL_FLUSH, size);
1518 }
1519
1520 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1521 {
1522         struct ib_fmr_pool_param param = {
1523                 .max_pages_per_fmr = LNET_MAX_IOV,
1524                 .page_shift        = PAGE_SHIFT,
1525                 .access            = (IB_ACCESS_LOCAL_WRITE |
1526                                       IB_ACCESS_REMOTE_WRITE),
1527                 .pool_size         = fps->fps_pool_size,
1528                 .dirty_watermark   = fps->fps_flush_trigger,
1529                 .flush_function    = NULL,
1530                 .flush_arg         = NULL,
1531                 .cache             = !!fps->fps_cache };
1532         int rc = 0;
1533
1534         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1535                                                    &param);
1536         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1537                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1538                 if (rc != -ENOSYS)
1539                         CERROR("Failed to create FMR pool: %d\n", rc);
1540                 else
1541                         CERROR("FMRs are not supported\n");
1542         }
1543         fpo->fpo_is_fmr = true;
1544
1545         return rc;
1546 }
1547
1548 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo,
1549                                   __u32 dev_caps)
1550 {
1551         struct kib_fast_reg_descriptor *frd, *tmp;
1552         int i, rc;
1553
1554         fpo->fpo_is_fmr = false;
1555
1556         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1557         fpo->fast_reg.fpo_pool_size = 0;
1558         for (i = 0; i < fps->fps_pool_size; i++) {
1559                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1560                                  sizeof(*frd));
1561                 if (!frd) {
1562                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1563                         rc = -ENOMEM;
1564                         goto out;
1565                 }
1566                 frd->frd_mr = NULL;
1567
1568 #ifndef HAVE_IB_MAP_MR_SG
1569                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1570                                                             LNET_MAX_IOV);
1571                 if (IS_ERR(frd->frd_frpl)) {
1572                         rc = PTR_ERR(frd->frd_frpl);
1573                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1574                                 rc);
1575                         frd->frd_frpl = NULL;
1576                         goto out_middle;
1577                 }
1578 #endif
1579
1580 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1581                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1582                                                    LNET_MAX_IOV);
1583 #else
1584                 /*
1585                  * it is expected to get here if this is an MLX-5 card.
1586                  * MLX-4 cards will always use FMR and MLX-5 cards will
1587                  * always use fast_reg. It turns out that some MLX-5 cards
1588                  * (possibly due to older FW versions) do not natively support
1589                  * gaps. So we will need to track them here.
1590                  */
1591                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1592 #ifdef IB_MR_TYPE_SG_GAPS
1593                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1594                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1595                                                 IB_MR_TYPE_SG_GAPS :
1596                                                 IB_MR_TYPE_MEM_REG,
1597 #else
1598                                                 IB_MR_TYPE_MEM_REG,
1599 #endif
1600                                           LNET_MAX_IOV);
1601                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1602                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1603                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1604 #endif
1605                 if (IS_ERR(frd->frd_mr)) {
1606                         rc = PTR_ERR(frd->frd_mr);
1607                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1608                         frd->frd_mr = NULL;
1609                         goto out_middle;
1610                 }
1611
1612                 /* There appears to be a bug in MLX5 code where you must
1613                  * invalidate the rkey of a new FastReg pool before first
1614                  * using it. Thus, I am marking the FRD invalid here. */
1615                 frd->frd_valid = false;
1616
1617                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1618                 fpo->fast_reg.fpo_pool_size++;
1619         }
1620
1621         return 0;
1622
1623 out_middle:
1624         if (frd->frd_mr)
1625                 ib_dereg_mr(frd->frd_mr);
1626 #ifndef HAVE_IB_MAP_MR_SG
1627         if (frd->frd_frpl)
1628                 ib_free_fast_reg_page_list(frd->frd_frpl);
1629 #endif
1630         LIBCFS_FREE(frd, sizeof(*frd));
1631
1632 out:
1633         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1634                                  frd_list) {
1635                 list_del(&frd->frd_list);
1636 #ifndef HAVE_IB_MAP_MR_SG
1637                 ib_free_fast_reg_page_list(frd->frd_frpl);
1638 #endif
1639                 ib_dereg_mr(frd->frd_mr);
1640                 LIBCFS_FREE(frd, sizeof(*frd));
1641         }
1642
1643         return rc;
1644 }
1645
1646 static int
1647 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1648 {
1649         kib_dev_t *dev = fps->fps_net->ibn_dev;
1650         kib_fmr_pool_t *fpo;
1651         int rc;
1652
1653         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1654         if (!fpo) {
1655                 return -ENOMEM;
1656         }
1657         memset(fpo, 0, sizeof(*fpo));
1658
1659         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1660
1661         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1662                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1663         else
1664                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1665         if (rc)
1666                 goto out_fpo;
1667
1668         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1669         fpo->fpo_owner = fps;
1670         *pp_fpo = fpo;
1671
1672         return 0;
1673
1674 out_fpo:
1675         kiblnd_hdev_decref(fpo->fpo_hdev);
1676         LIBCFS_FREE(fpo, sizeof(*fpo));
1677         return rc;
1678 }
1679
1680 static void
1681 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1682 {
1683         if (fps->fps_net == NULL) /* intialized? */
1684                 return;
1685
1686         spin_lock(&fps->fps_lock);
1687
1688         while (!list_empty(&fps->fps_pool_list)) {
1689                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1690                                                  kib_fmr_pool_t, fpo_list);
1691                 fpo->fpo_failed = 1;
1692                 list_del(&fpo->fpo_list);
1693                 if (fpo->fpo_map_count == 0)
1694                         list_add(&fpo->fpo_list, zombies);
1695                 else
1696                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1697         }
1698
1699         spin_unlock(&fps->fps_lock);
1700 }
1701
1702 static void
1703 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1704 {
1705         if (fps->fps_net != NULL) { /* initialized? */
1706                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1707                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1708         }
1709 }
1710
1711 static int
1712 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1713                         kib_net_t *net,
1714                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1715 {
1716         kib_fmr_pool_t *fpo;
1717         int             rc;
1718
1719         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1720
1721         fps->fps_net = net;
1722         fps->fps_cpt = cpt;
1723
1724         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1725         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1726         fps->fps_cache = tunables->lnd_fmr_cache;
1727
1728         spin_lock_init(&fps->fps_lock);
1729         INIT_LIST_HEAD(&fps->fps_pool_list);
1730         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1731
1732         rc = kiblnd_create_fmr_pool(fps, &fpo);
1733         if (rc == 0)
1734                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1735
1736         return rc;
1737 }
1738
1739 static int
1740 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, time64_t now)
1741 {
1742         if (fpo->fpo_map_count != 0) /* still in use */
1743                 return 0;
1744         if (fpo->fpo_failed)
1745                 return 1;
1746         return now >= fpo->fpo_deadline;
1747 }
1748
1749 static int
1750 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1751 {
1752         kib_hca_dev_t   *hdev;
1753         __u64           *pages = tx->tx_pages;
1754         int             npages;
1755         int             size;
1756         int             i;
1757
1758         hdev = tx->tx_pool->tpo_hdev;
1759
1760         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1761                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1762                         size += hdev->ibh_page_size) {
1763                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1764                                            hdev->ibh_page_mask) + size;
1765                 }
1766         }
1767
1768         return npages;
1769 }
1770
1771 void
1772 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1773 {
1774         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1775         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1776         kib_fmr_poolset_t *fps;
1777         time64_t now = ktime_get_seconds();
1778         kib_fmr_pool_t *tmp;
1779         int rc;
1780
1781         if (!fpo)
1782                 return;
1783
1784         fps = fpo->fpo_owner;
1785         if (fpo->fpo_is_fmr) {
1786                 if (fmr->fmr_pfmr) {
1787                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1788                         LASSERT(!rc);
1789                         fmr->fmr_pfmr = NULL;
1790                 }
1791
1792                 if (status) {
1793                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1794                         LASSERT(!rc);
1795                 }
1796         } else {
1797                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1798
1799                 if (frd) {
1800                         frd->frd_valid = false;
1801                         spin_lock(&fps->fps_lock);
1802                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1803                         spin_unlock(&fps->fps_lock);
1804                         fmr->fmr_frd = NULL;
1805                 }
1806         }
1807         fmr->fmr_pool = NULL;
1808
1809         spin_lock(&fps->fps_lock);
1810         fpo->fpo_map_count--;   /* decref the pool */
1811
1812         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1813                 /* the first pool is persistent */
1814                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1815                         continue;
1816
1817                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1818                         list_move(&fpo->fpo_list, &zombies);
1819                         fps->fps_version++;
1820                 }
1821         }
1822         spin_unlock(&fps->fps_lock);
1823
1824         if (!list_empty(&zombies))
1825                 kiblnd_destroy_fmr_pool_list(&zombies);
1826 }
1827
1828 int
1829 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1830                     __u32 nob, __u64 iov, kib_fmr_t *fmr)
1831 {
1832         kib_fmr_pool_t *fpo;
1833         __u64 *pages = tx->tx_pages;
1834         __u64 version;
1835         bool is_rx = (rd != tx->tx_rd);
1836         bool tx_pages_mapped = 0;
1837         int npages = 0;
1838         int rc;
1839
1840 again:
1841         spin_lock(&fps->fps_lock);
1842         version = fps->fps_version;
1843         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1844                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1845                 fpo->fpo_map_count++;
1846
1847                 if (fpo->fpo_is_fmr) {
1848                         struct ib_pool_fmr *pfmr;
1849
1850                         spin_unlock(&fps->fps_lock);
1851
1852                         if (!tx_pages_mapped) {
1853                                 npages = kiblnd_map_tx_pages(tx, rd);
1854                                 tx_pages_mapped = 1;
1855                         }
1856
1857                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1858                                                     pages, npages, iov);
1859                         if (likely(!IS_ERR(pfmr))) {
1860                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1861                                                       : pfmr->fmr->lkey;
1862                                 fmr->fmr_frd  = NULL;
1863                                 fmr->fmr_pfmr = pfmr;
1864                                 fmr->fmr_pool = fpo;
1865                                 return 0;
1866                         }
1867                         rc = PTR_ERR(pfmr);
1868                 } else {
1869                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1870                                 struct kib_fast_reg_descriptor *frd;
1871 #ifdef HAVE_IB_MAP_MR_SG
1872                                 struct ib_reg_wr *wr;
1873                                 int n;
1874 #else
1875                                 struct ib_rdma_wr *wr;
1876                                 struct ib_fast_reg_page_list *frpl;
1877 #endif
1878                                 struct ib_mr *mr;
1879
1880                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1881                                                         struct kib_fast_reg_descriptor,
1882                                                         frd_list);
1883                                 list_del(&frd->frd_list);
1884                                 spin_unlock(&fps->fps_lock);
1885
1886 #ifndef HAVE_IB_MAP_MR_SG
1887                                 frpl = frd->frd_frpl;
1888 #endif
1889                                 mr   = frd->frd_mr;
1890
1891                                 if (!frd->frd_valid) {
1892                                         struct ib_rdma_wr *inv_wr;
1893                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1894
1895                                         inv_wr = &frd->frd_inv_wr;
1896                                         memset(inv_wr, 0, sizeof(*inv_wr));
1897
1898                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1899                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1900                                         inv_wr->wr.ex.invalidate_rkey = key;
1901
1902                                         /* Bump the key */
1903                                         key = ib_inc_rkey(key);
1904                                         ib_update_fast_reg_key(mr, key);
1905                                 }
1906
1907 #ifdef HAVE_IB_MAP_MR_SG
1908 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1909                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1910                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1911 #else
1912                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1913                                                  tx->tx_nfrags, PAGE_SIZE);
1914 #endif
1915                                 if (unlikely(n != tx->tx_nfrags)) {
1916                                         CERROR("Failed to map mr %d/%d "
1917                                                "elements\n", n, tx->tx_nfrags);
1918                                         return n < 0 ? n : -EINVAL;
1919                                 }
1920
1921                                 wr = &frd->frd_fastreg_wr;
1922                                 memset(wr, 0, sizeof(*wr));
1923
1924                                 wr->wr.opcode = IB_WR_REG_MR;
1925                                 wr->wr.wr_id  = IBLND_WID_MR;
1926                                 wr->wr.num_sge = 0;
1927                                 wr->wr.send_flags = 0;
1928                                 wr->mr = mr;
1929                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1930                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1931                                               IB_ACCESS_REMOTE_WRITE);
1932 #else
1933                                 if (!tx_pages_mapped) {
1934                                         npages = kiblnd_map_tx_pages(tx, rd);
1935                                         tx_pages_mapped = 1;
1936                                 }
1937
1938                                 LASSERT(npages <= frpl->max_page_list_len);
1939                                 memcpy(frpl->page_list, pages,
1940                                         sizeof(*pages) * npages);
1941
1942                                 /* Prepare FastReg WR */
1943                                 wr = &frd->frd_fastreg_wr;
1944                                 memset(wr, 0, sizeof(*wr));
1945
1946                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1947                                 wr->wr.wr_id  = IBLND_WID_MR;
1948
1949                                 wr->wr.wr.fast_reg.iova_start = iov;
1950                                 wr->wr.wr.fast_reg.page_list  = frpl;
1951                                 wr->wr.wr.fast_reg.page_list_len = npages;
1952                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1953                                 wr->wr.wr.fast_reg.length = nob;
1954                                 wr->wr.wr.fast_reg.rkey =
1955                                                 is_rx ? mr->rkey : mr->lkey;
1956                                 wr->wr.wr.fast_reg.access_flags =
1957                                                 (IB_ACCESS_LOCAL_WRITE |
1958                                                  IB_ACCESS_REMOTE_WRITE);
1959 #endif
1960
1961                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1962                                 fmr->fmr_frd  = frd;
1963                                 fmr->fmr_pfmr = NULL;
1964                                 fmr->fmr_pool = fpo;
1965                                 return 0;
1966                         }
1967                         spin_unlock(&fps->fps_lock);
1968                         rc = -EAGAIN;
1969                 }
1970
1971                 spin_lock(&fps->fps_lock);
1972                 fpo->fpo_map_count--;
1973                 if (rc != -EAGAIN) {
1974                         spin_unlock(&fps->fps_lock);
1975                         return rc;
1976                 }
1977
1978                 /* EAGAIN and ... */
1979                 if (version != fps->fps_version) {
1980                         spin_unlock(&fps->fps_lock);
1981                         goto again;
1982                 }
1983         }
1984
1985         if (fps->fps_increasing) {
1986                 spin_unlock(&fps->fps_lock);
1987                 CDEBUG(D_NET, "Another thread is allocating new "
1988                        "FMR pool, waiting for her to complete\n");
1989                 schedule();
1990                 goto again;
1991
1992         }
1993
1994         if (ktime_get_seconds() < fps->fps_next_retry) {
1995                 /* someone failed recently */
1996                 spin_unlock(&fps->fps_lock);
1997                 return -EAGAIN;
1998         }
1999
2000         fps->fps_increasing = 1;
2001         spin_unlock(&fps->fps_lock);
2002
2003         CDEBUG(D_NET, "Allocate new FMR pool\n");
2004         rc = kiblnd_create_fmr_pool(fps, &fpo);
2005         spin_lock(&fps->fps_lock);
2006         fps->fps_increasing = 0;
2007         if (rc == 0) {
2008                 fps->fps_version++;
2009                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2010         } else {
2011                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2012         }
2013         spin_unlock(&fps->fps_lock);
2014
2015         goto again;
2016 }
2017
2018 static void
2019 kiblnd_fini_pool(kib_pool_t *pool)
2020 {
2021         LASSERT(list_empty(&pool->po_free_list));
2022         LASSERT(pool->po_allocated == 0);
2023
2024         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2025 }
2026
2027 static void
2028 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2029 {
2030         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2031
2032         memset(pool, 0, sizeof(kib_pool_t));
2033         INIT_LIST_HEAD(&pool->po_free_list);
2034         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2035         pool->po_owner = ps;
2036         pool->po_size = size;
2037 }
2038
2039 static void
2040 kiblnd_destroy_pool_list(struct list_head *head)
2041 {
2042         kib_pool_t *pool;
2043
2044         while (!list_empty(head)) {
2045                 pool = list_entry(head->next, kib_pool_t, po_list);
2046                 list_del(&pool->po_list);
2047
2048                 LASSERT(pool->po_owner != NULL);
2049                 pool->po_owner->ps_pool_destroy(pool);
2050         }
2051 }
2052
2053 static void
2054 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2055 {
2056         if (ps->ps_net == NULL) /* intialized? */
2057                 return;
2058
2059         spin_lock(&ps->ps_lock);
2060         while (!list_empty(&ps->ps_pool_list)) {
2061                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2062                                             kib_pool_t, po_list);
2063                 po->po_failed = 1;
2064                 list_del(&po->po_list);
2065                 if (po->po_allocated == 0)
2066                         list_add(&po->po_list, zombies);
2067                 else
2068                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2069         }
2070         spin_unlock(&ps->ps_lock);
2071 }
2072
2073 static void
2074 kiblnd_fini_poolset(kib_poolset_t *ps)
2075 {
2076         if (ps->ps_net != NULL) { /* initialized? */
2077                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2078                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2079         }
2080 }
2081
2082 static int
2083 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2084                     kib_net_t *net, char *name, int size,
2085                     kib_ps_pool_create_t po_create,
2086                     kib_ps_pool_destroy_t po_destroy,
2087                     kib_ps_node_init_t nd_init,
2088                     kib_ps_node_fini_t nd_fini)
2089 {
2090         kib_pool_t      *pool;
2091         int             rc;
2092
2093         memset(ps, 0, sizeof(kib_poolset_t));
2094
2095         ps->ps_cpt          = cpt;
2096         ps->ps_net          = net;
2097         ps->ps_pool_create  = po_create;
2098         ps->ps_pool_destroy = po_destroy;
2099         ps->ps_node_init    = nd_init;
2100         ps->ps_node_fini    = nd_fini;
2101         ps->ps_pool_size    = size;
2102         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2103             >= sizeof(ps->ps_name))
2104                 return -E2BIG;
2105         spin_lock_init(&ps->ps_lock);
2106         INIT_LIST_HEAD(&ps->ps_pool_list);
2107         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2108
2109         rc = ps->ps_pool_create(ps, size, &pool);
2110         if (rc == 0)
2111                 list_add(&pool->po_list, &ps->ps_pool_list);
2112         else
2113                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2114
2115         return rc;
2116 }
2117
2118 static int
2119 kiblnd_pool_is_idle(kib_pool_t *pool, time64_t now)
2120 {
2121         if (pool->po_allocated != 0) /* still in use */
2122                 return 0;
2123         if (pool->po_failed)
2124                 return 1;
2125         return now >= pool->po_deadline;
2126 }
2127
2128 void
2129 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2130 {
2131         struct list_head zombies = LIST_HEAD_INIT(zombies);
2132         kib_poolset_t *ps = pool->po_owner;
2133         kib_pool_t *tmp;
2134         time64_t now = ktime_get_seconds();
2135
2136         spin_lock(&ps->ps_lock);
2137
2138         if (ps->ps_node_fini != NULL)
2139                 ps->ps_node_fini(pool, node);
2140
2141         LASSERT(pool->po_allocated > 0);
2142         list_add(node, &pool->po_free_list);
2143         pool->po_allocated--;
2144
2145         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2146                 /* the first pool is persistent */
2147                 if (ps->ps_pool_list.next == &pool->po_list)
2148                         continue;
2149
2150                 if (kiblnd_pool_is_idle(pool, now))
2151                         list_move(&pool->po_list, &zombies);
2152         }
2153         spin_unlock(&ps->ps_lock);
2154
2155         if (!list_empty(&zombies))
2156                 kiblnd_destroy_pool_list(&zombies);
2157 }
2158
2159 struct list_head *
2160 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2161 {
2162         struct list_head        *node;
2163         kib_pool_t              *pool;
2164         int                     rc;
2165         unsigned int            interval = 1;
2166         ktime_t time_before;
2167         unsigned int trips = 0;
2168
2169 again:
2170         spin_lock(&ps->ps_lock);
2171         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2172                 if (list_empty(&pool->po_free_list))
2173                         continue;
2174
2175                 pool->po_allocated++;
2176                 pool->po_deadline = ktime_get_seconds() +
2177                                     IBLND_POOL_DEADLINE;
2178                 node = pool->po_free_list.next;
2179                 list_del(node);
2180
2181                 if (ps->ps_node_init != NULL) {
2182                         /* still hold the lock */
2183                         ps->ps_node_init(pool, node);
2184                 }
2185                 spin_unlock(&ps->ps_lock);
2186                 return node;
2187         }
2188
2189         /* no available tx pool and ... */
2190         if (ps->ps_increasing) {
2191                 /* another thread is allocating a new pool */
2192                 spin_unlock(&ps->ps_lock);
2193                 trips++;
2194                 CDEBUG(D_NET, "Another thread is allocating new "
2195                        "%s pool, waiting %d HZs for her to complete."
2196                        "trips = %d\n",
2197                        ps->ps_name, interval, trips);
2198
2199                 set_current_state(TASK_INTERRUPTIBLE);
2200                 schedule_timeout(interval);
2201                 if (interval < cfs_time_seconds(1))
2202                         interval *= 2;
2203
2204                 goto again;
2205         }
2206
2207         if (ktime_get_seconds() < ps->ps_next_retry) {
2208                 /* someone failed recently */
2209                 spin_unlock(&ps->ps_lock);
2210                 return NULL;
2211         }
2212
2213         ps->ps_increasing = 1;
2214         spin_unlock(&ps->ps_lock);
2215
2216         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2217         time_before = ktime_get();
2218         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2219         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2220                ktime_ms_delta(ktime_get(), time_before));
2221
2222         spin_lock(&ps->ps_lock);
2223         ps->ps_increasing = 0;
2224         if (rc == 0) {
2225                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2226         } else {
2227                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2228                 CERROR("Can't allocate new %s pool because out of memory\n",
2229                        ps->ps_name);
2230         }
2231         spin_unlock(&ps->ps_lock);
2232
2233         goto again;
2234 }
2235
2236 static void
2237 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2238 {
2239         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2240         int             i;
2241
2242         LASSERT (pool->po_allocated == 0);
2243
2244         if (tpo->tpo_tx_pages != NULL) {
2245                 kiblnd_unmap_tx_pool(tpo);
2246                 kiblnd_free_pages(tpo->tpo_tx_pages);
2247         }
2248
2249         if (tpo->tpo_tx_descs == NULL)
2250                 goto out;
2251
2252         for (i = 0; i < pool->po_size; i++) {
2253                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2254                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2255
2256                 list_del(&tx->tx_list);
2257                 if (tx->tx_pages != NULL)
2258                         LIBCFS_FREE(tx->tx_pages,
2259                                     LNET_MAX_IOV *
2260                                     sizeof(*tx->tx_pages));
2261                 if (tx->tx_frags != NULL)
2262                         LIBCFS_FREE(tx->tx_frags,
2263                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2264                                     sizeof(*tx->tx_frags));
2265                 if (tx->tx_wrq != NULL)
2266                         LIBCFS_FREE(tx->tx_wrq,
2267                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2268                                     sizeof(*tx->tx_wrq));
2269                 if (tx->tx_sge != NULL)
2270                         LIBCFS_FREE(tx->tx_sge,
2271                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2272                                     sizeof(*tx->tx_sge));
2273                 if (tx->tx_rd != NULL)
2274                         LIBCFS_FREE(tx->tx_rd,
2275                                     offsetof(kib_rdma_desc_t,
2276                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2277         }
2278
2279         LIBCFS_FREE(tpo->tpo_tx_descs,
2280                     pool->po_size * sizeof(kib_tx_t));
2281 out:
2282         kiblnd_fini_pool(pool);
2283         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2284 }
2285
2286 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2287 {
2288         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2289         int ntx;
2290
2291         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2292         ntx = tunables->lnd_ntx / ncpts;
2293
2294         return max(IBLND_TX_POOL, ntx);
2295 }
2296
2297 static int
2298 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2299 {
2300         int            i;
2301         int            npg;
2302         kib_pool_t    *pool;
2303         kib_tx_pool_t *tpo;
2304
2305         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2306         if (tpo == NULL) {
2307                 CERROR("Failed to allocate TX pool\n");
2308                 return -ENOMEM;
2309         }
2310
2311         pool = &tpo->tpo_pool;
2312         kiblnd_init_pool(ps, pool, size);
2313         tpo->tpo_tx_descs = NULL;
2314         tpo->tpo_tx_pages = NULL;
2315
2316         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2317         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2318                 CERROR("Can't allocate tx pages: %d\n", npg);
2319                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2320                 return -ENOMEM;
2321         }
2322
2323         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2324                          size * sizeof(kib_tx_t));
2325         if (tpo->tpo_tx_descs == NULL) {
2326                 CERROR("Can't allocate %d tx descriptors\n", size);
2327                 ps->ps_pool_destroy(pool);
2328                 return -ENOMEM;
2329         }
2330
2331         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2332
2333         for (i = 0; i < size; i++) {
2334                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2335                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2336
2337                 tx->tx_pool = tpo;
2338                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2339                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2340                                          lnet_cpt_table(), ps->ps_cpt,
2341                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2342                         if (tx->tx_pages == NULL)
2343                                 break;
2344                 }
2345
2346                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2347                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2348                                  sizeof(*tx->tx_frags));
2349                 if (tx->tx_frags == NULL)
2350                         break;
2351
2352                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2353
2354                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2355                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2356                                  sizeof(*tx->tx_wrq));
2357                 if (tx->tx_wrq == NULL)
2358                         break;
2359
2360                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2361                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2362                                  sizeof(*tx->tx_sge));
2363                 if (tx->tx_sge == NULL)
2364                         break;
2365
2366                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2367                                  offsetof(kib_rdma_desc_t,
2368                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2369                 if (tx->tx_rd == NULL)
2370                         break;
2371         }
2372
2373         if (i == size) {
2374                 kiblnd_map_tx_pool(tpo);
2375                 *pp_po = pool;
2376                 return 0;
2377         }
2378
2379         ps->ps_pool_destroy(pool);
2380         return -ENOMEM;
2381 }
2382
2383 static void
2384 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2385 {
2386         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2387                                              tps_poolset);
2388         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2389
2390         tx->tx_cookie = tps->tps_next_tx_cookie++;
2391 }
2392
2393 static void
2394 kiblnd_net_fini_pools(kib_net_t *net)
2395 {
2396         int     i;
2397
2398         cfs_cpt_for_each(i, lnet_cpt_table()) {
2399                 kib_tx_poolset_t        *tps;
2400                 kib_fmr_poolset_t       *fps;
2401
2402                 if (net->ibn_tx_ps != NULL) {
2403                         tps = net->ibn_tx_ps[i];
2404                         kiblnd_fini_poolset(&tps->tps_poolset);
2405                 }
2406
2407                 if (net->ibn_fmr_ps != NULL) {
2408                         fps = net->ibn_fmr_ps[i];
2409                         kiblnd_fini_fmr_poolset(fps);
2410                 }
2411         }
2412
2413         if (net->ibn_tx_ps != NULL) {
2414                 cfs_percpt_free(net->ibn_tx_ps);
2415                 net->ibn_tx_ps = NULL;
2416         }
2417
2418         if (net->ibn_fmr_ps != NULL) {
2419                 cfs_percpt_free(net->ibn_fmr_ps);
2420                 net->ibn_fmr_ps = NULL;
2421         }
2422 }
2423
2424 static int
2425 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2426                       int ncpts)
2427 {
2428         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2429 #ifdef HAVE_IB_GET_DMA_MR
2430         unsigned long   flags;
2431 #endif
2432         int             cpt;
2433         int             rc;
2434         int             i;
2435
2436         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2437
2438 #ifdef HAVE_IB_GET_DMA_MR
2439         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2440         /*
2441          * if lnd_map_on_demand is zero then we have effectively disabled
2442          * FMR or FastReg and we're using global memory regions
2443          * exclusively.
2444          */
2445         if (!tunables->lnd_map_on_demand) {
2446                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2447                                            flags);
2448                 goto create_tx_pool;
2449         }
2450
2451         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2452 #endif
2453
2454         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2455                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2456                        tunables->lnd_fmr_pool_size,
2457                        tunables->lnd_ntx / 4);
2458                 rc = -EINVAL;
2459                 goto failed;
2460         }
2461
2462         /* TX pool must be created later than FMR, see LU-2268
2463          * for details */
2464         LASSERT(net->ibn_tx_ps == NULL);
2465
2466         /* premapping can fail if ibd_nmr > 1, so we always create
2467          * FMR pool and map-on-demand if premapping failed */
2468
2469         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2470                                            sizeof(kib_fmr_poolset_t));
2471         if (net->ibn_fmr_ps == NULL) {
2472                 CERROR("Failed to allocate FMR pool array\n");
2473                 rc = -ENOMEM;
2474                 goto failed;
2475         }
2476
2477         for (i = 0; i < ncpts; i++) {
2478                 cpt = (cpts == NULL) ? i : cpts[i];
2479                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2480                                              net, tunables);
2481                 if (rc != 0) {
2482                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2483                                cpt, rc);
2484                         goto failed;
2485                 }
2486         }
2487
2488         if (i > 0)
2489                 LASSERT(i == ncpts);
2490
2491 #ifdef HAVE_IB_GET_DMA_MR
2492  create_tx_pool:
2493 #endif
2494         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2495                                           sizeof(kib_tx_poolset_t));
2496         if (net->ibn_tx_ps == NULL) {
2497                 CERROR("Failed to allocate tx pool array\n");
2498                 rc = -ENOMEM;
2499                 goto failed;
2500         }
2501
2502         for (i = 0; i < ncpts; i++) {
2503                 cpt = (cpts == NULL) ? i : cpts[i];
2504                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2505                                          cpt, net, "TX",
2506                                          kiblnd_tx_pool_size(ni, ncpts),
2507                                          kiblnd_create_tx_pool,
2508                                          kiblnd_destroy_tx_pool,
2509                                          kiblnd_tx_init, NULL);
2510                 if (rc != 0) {
2511                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2512                                cpt, rc);
2513                         goto failed;
2514                 }
2515         }
2516
2517         return 0;
2518  failed:
2519         kiblnd_net_fini_pools(net);
2520         LASSERT(rc != 0);
2521         return rc;
2522 }
2523
2524 static int
2525 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2526 {
2527         struct ib_device_attr *dev_attr;
2528         int rc = 0;
2529
2530         /* It's safe to assume a HCA can handle a page size
2531          * matching that of the native system */
2532         hdev->ibh_page_shift = PAGE_SHIFT;
2533         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2534         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2535
2536 #ifndef HAVE_IB_DEVICE_ATTRS
2537         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2538         if (dev_attr == NULL) {
2539                 CERROR("Out of memory\n");
2540                 return -ENOMEM;
2541         }
2542
2543         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2544         if (rc != 0) {
2545                 CERROR("Failed to query IB device: %d\n", rc);
2546                 goto out_clean_attr;
2547         }
2548 #else
2549         dev_attr = &hdev->ibh_ibdev->attrs;
2550 #endif
2551
2552         hdev->ibh_mr_size = dev_attr->max_mr_size;
2553
2554         /* Setup device Memory Registration capabilities */
2555         if (hdev->ibh_ibdev->alloc_fmr &&
2556             hdev->ibh_ibdev->dealloc_fmr &&
2557             hdev->ibh_ibdev->map_phys_fmr &&
2558             hdev->ibh_ibdev->unmap_fmr) {
2559                 LCONSOLE_INFO("Using FMR for registration\n");
2560                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2561         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2562                 LCONSOLE_INFO("Using FastReg for registration\n");
2563                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2564 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2565 #ifdef IB_DEVICE_SG_GAPS_REG
2566                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2567                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2568 #endif
2569 #endif
2570         } else {
2571                 rc = -ENOSYS;
2572         }
2573
2574         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2575                 hdev->ibh_mr_shift = 64;
2576         else if (rc != 0)
2577                 rc = -EINVAL;
2578
2579 #ifndef HAVE_IB_DEVICE_ATTRS
2580 out_clean_attr:
2581         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2582 #endif
2583
2584         if (rc == -ENOSYS)
2585                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2586                        "register memory: %d\n", rc);
2587         else if (rc == -EINVAL)
2588                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2589         return rc;
2590 }
2591
2592 #ifdef HAVE_IB_GET_DMA_MR
2593 static void
2594 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2595 {
2596         if (hdev->ibh_mrs == NULL)
2597                 return;
2598
2599         ib_dereg_mr(hdev->ibh_mrs);
2600
2601         hdev->ibh_mrs = NULL;
2602 }
2603 #endif
2604
2605 void
2606 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2607 {
2608 #ifdef HAVE_IB_GET_DMA_MR
2609         kiblnd_hdev_cleanup_mrs(hdev);
2610 #endif
2611
2612         if (hdev->ibh_pd != NULL)
2613                 ib_dealloc_pd(hdev->ibh_pd);
2614
2615         if (hdev->ibh_cmid != NULL)
2616                 rdma_destroy_id(hdev->ibh_cmid);
2617
2618         LIBCFS_FREE(hdev, sizeof(*hdev));
2619 }
2620
2621 #ifdef HAVE_IB_GET_DMA_MR
2622 static int
2623 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2624 {
2625         struct ib_mr *mr;
2626         int           acflags = IB_ACCESS_LOCAL_WRITE |
2627                                 IB_ACCESS_REMOTE_WRITE;
2628
2629         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2630         if (IS_ERR(mr)) {
2631                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2632                 kiblnd_hdev_cleanup_mrs(hdev);
2633                 return PTR_ERR(mr);
2634         }
2635
2636         hdev->ibh_mrs = mr;
2637
2638         return 0;
2639 }
2640 #endif
2641
2642 static int
2643 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2644 {       /* DUMMY */
2645         return 0;
2646 }
2647
2648 static int
2649 kiblnd_dev_need_failover(kib_dev_t *dev)
2650 {
2651         struct rdma_cm_id  *cmid;
2652         struct sockaddr_in  srcaddr;
2653         struct sockaddr_in  dstaddr;
2654         int                 rc;
2655
2656         if (dev->ibd_hdev == NULL || /* initializing */
2657             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2658             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2659                 return 1;
2660
2661         /* XXX: it's UGLY, but I don't have better way to find
2662          * ib-bonding HCA failover because:
2663          *
2664          * a. no reliable CM event for HCA failover...
2665          * b. no OFED API to get ib_device for current net_device...
2666          *
2667          * We have only two choices at this point:
2668          *
2669          * a. rdma_bind_addr(), it will conflict with listener cmid
2670          * b. rdma_resolve_addr() to zero addr */
2671         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2672                                      IB_QPT_RC);
2673         if (IS_ERR(cmid)) {
2674                 rc = PTR_ERR(cmid);
2675                 CERROR("Failed to create cmid for failover: %d\n", rc);
2676                 return rc;
2677         }
2678
2679         memset(&srcaddr, 0, sizeof(srcaddr));
2680         srcaddr.sin_family      = AF_INET;
2681         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2682
2683         memset(&dstaddr, 0, sizeof(dstaddr));
2684         dstaddr.sin_family = AF_INET;
2685         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2686                                (struct sockaddr *)&dstaddr, 1);
2687         if (rc != 0 || cmid->device == NULL) {
2688                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2689                        dev->ibd_ifname, &dev->ibd_ifip,
2690                        cmid->device, rc);
2691                 rdma_destroy_id(cmid);
2692                 return rc;
2693         }
2694
2695         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2696         rdma_destroy_id(cmid);
2697         return rc;
2698 }
2699
2700 int
2701 kiblnd_dev_failover(kib_dev_t *dev)
2702 {
2703         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2704         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2705         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2706         struct rdma_cm_id  *cmid  = NULL;
2707         kib_hca_dev_t      *hdev  = NULL;
2708         kib_hca_dev_t      *old;
2709         struct ib_pd       *pd;
2710         kib_net_t          *net;
2711         struct sockaddr_in  addr;
2712         unsigned long       flags;
2713         int                 rc = 0;
2714         int                 i;
2715
2716         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2717                  dev->ibd_can_failover ||
2718                  dev->ibd_hdev == NULL);
2719
2720         rc = kiblnd_dev_need_failover(dev);
2721         if (rc <= 0)
2722                 goto out;
2723
2724         if (dev->ibd_hdev != NULL &&
2725             dev->ibd_hdev->ibh_cmid != NULL) {
2726                 /* XXX it's not good to close old listener at here,
2727                  * because we can fail to create new listener.
2728                  * But we have to close it now, otherwise rdma_bind_addr
2729                  * will return EADDRINUSE... How crap! */
2730                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2731
2732                 cmid = dev->ibd_hdev->ibh_cmid;
2733                 /* make next schedule of kiblnd_dev_need_failover()
2734                  * return 1 for me */
2735                 dev->ibd_hdev->ibh_cmid  = NULL;
2736                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2737
2738                 rdma_destroy_id(cmid);
2739         }
2740
2741         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2742                                      IB_QPT_RC);
2743         if (IS_ERR(cmid)) {
2744                 rc = PTR_ERR(cmid);
2745                 CERROR("Failed to create cmid for failover: %d\n", rc);
2746                 goto out;
2747         }
2748
2749         memset(&addr, 0, sizeof(addr));
2750         addr.sin_family      = AF_INET;
2751         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2752         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2753
2754         /* Bind to failover device or port */
2755         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2756         if (rc != 0 || cmid->device == NULL) {
2757                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2758                        dev->ibd_ifname, &dev->ibd_ifip,
2759                        cmid->device, rc);
2760                 rdma_destroy_id(cmid);
2761                 goto out;
2762         }
2763
2764         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2765         if (hdev == NULL) {
2766                 CERROR("Failed to allocate kib_hca_dev\n");
2767                 rdma_destroy_id(cmid);
2768                 rc = -ENOMEM;
2769                 goto out;
2770         }
2771
2772         atomic_set(&hdev->ibh_ref, 1);
2773         hdev->ibh_dev   = dev;
2774         hdev->ibh_cmid  = cmid;
2775         hdev->ibh_ibdev = cmid->device;
2776
2777 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2778         pd = ib_alloc_pd(cmid->device, 0);
2779 #else
2780         pd = ib_alloc_pd(cmid->device);
2781 #endif
2782         if (IS_ERR(pd)) {
2783                 rc = PTR_ERR(pd);
2784                 CERROR("Can't allocate PD: %d\n", rc);
2785                 goto out;
2786         }
2787
2788         hdev->ibh_pd = pd;
2789
2790         rc = rdma_listen(cmid, 0);
2791         if (rc != 0) {
2792                 CERROR("Can't start new listener: %d\n", rc);
2793                 goto out;
2794         }
2795
2796         rc = kiblnd_hdev_get_attr(hdev);
2797         if (rc != 0) {
2798                 CERROR("Can't get device attributes: %d\n", rc);
2799                 goto out;
2800         }
2801
2802 #ifdef HAVE_IB_GET_DMA_MR
2803         rc = kiblnd_hdev_setup_mrs(hdev);
2804         if (rc != 0) {
2805                 CERROR("Can't setup device: %d\n", rc);
2806                 goto out;
2807         }
2808 #endif
2809
2810         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2811
2812         old = dev->ibd_hdev;
2813         dev->ibd_hdev = hdev;   /* take over the refcount */
2814         hdev = old;
2815
2816         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2817                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2818                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2819                                             &zombie_tpo);
2820
2821                         if (net->ibn_fmr_ps != NULL)
2822                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2823                                                         &zombie_fpo);
2824                 }
2825         }
2826
2827         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2828  out:
2829         if (!list_empty(&zombie_tpo))
2830                 kiblnd_destroy_pool_list(&zombie_tpo);
2831         if (!list_empty(&zombie_ppo))
2832                 kiblnd_destroy_pool_list(&zombie_ppo);
2833         if (!list_empty(&zombie_fpo))
2834                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2835         if (hdev != NULL)
2836                 kiblnd_hdev_decref(hdev);
2837
2838         if (rc != 0)
2839                 dev->ibd_failed_failover++;
2840         else
2841                 dev->ibd_failed_failover = 0;
2842
2843         return rc;
2844 }
2845
2846 void
2847 kiblnd_destroy_dev (kib_dev_t *dev)
2848 {
2849         LASSERT (dev->ibd_nnets == 0);
2850         LASSERT(list_empty(&dev->ibd_nets));
2851
2852         list_del(&dev->ibd_fail_list);
2853         list_del(&dev->ibd_list);
2854
2855         if (dev->ibd_hdev != NULL)
2856                 kiblnd_hdev_decref(dev->ibd_hdev);
2857
2858         LIBCFS_FREE(dev, sizeof(*dev));
2859 }
2860
2861 static kib_dev_t *
2862 kiblnd_create_dev(char *ifname)
2863 {
2864         struct net_device *netdev;
2865         kib_dev_t         *dev;
2866         __u32              netmask;
2867         __u32              ip;
2868         int                up;
2869         int                rc;
2870
2871         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2872         if (rc != 0) {
2873                 CERROR("Can't query IPoIB interface %s: %d\n",
2874                        ifname, rc);
2875                 return NULL;
2876         }
2877
2878         if (!up) {
2879                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2880                 return NULL;
2881         }
2882
2883         LIBCFS_ALLOC(dev, sizeof(*dev));
2884         if (dev == NULL)
2885                 return NULL;
2886
2887         netdev = dev_get_by_name(&init_net, ifname);
2888         if (netdev == NULL) {
2889                 dev->ibd_can_failover = 0;
2890         } else {
2891                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2892                 dev_put(netdev);
2893         }
2894
2895         INIT_LIST_HEAD(&dev->ibd_nets);
2896         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2897         INIT_LIST_HEAD(&dev->ibd_fail_list);
2898         dev->ibd_ifip = ip;
2899         strcpy(&dev->ibd_ifname[0], ifname);
2900
2901         /* initialize the device */
2902         rc = kiblnd_dev_failover(dev);
2903         if (rc != 0) {
2904                 CERROR("Can't initialize device: %d\n", rc);
2905                 LIBCFS_FREE(dev, sizeof(*dev));
2906                 return NULL;
2907         }
2908
2909         list_add_tail(&dev->ibd_list,
2910                           &kiblnd_data.kib_devs);
2911         return dev;
2912 }
2913
2914 static void
2915 kiblnd_base_shutdown(void)
2916 {
2917         struct kib_sched_info   *sched;
2918         int                     i;
2919
2920         LASSERT(list_empty(&kiblnd_data.kib_devs));
2921
2922         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2923                atomic_read(&libcfs_kmemory));
2924
2925         switch (kiblnd_data.kib_init) {
2926         default:
2927                 LBUG();
2928
2929         case IBLND_INIT_ALL:
2930         case IBLND_INIT_DATA:
2931                 LASSERT (kiblnd_data.kib_peers != NULL);
2932                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2933                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2934                 }
2935                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2936                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2937                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2938                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2939
2940                 /* flag threads to terminate; wake and wait for them to die */
2941                 kiblnd_data.kib_shutdown = 1;
2942
2943                 /* NB: we really want to stop scheduler threads net by net
2944                  * instead of the whole module, this should be improved
2945                  * with dynamic configuration LNet */
2946                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2947                         wake_up_all(&sched->ibs_waitq);
2948
2949                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2950                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2951
2952                 i = 2;
2953                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2954                         i++;
2955                         /* power of 2? */
2956                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2957                                "Waiting for %d threads to terminate\n",
2958                                atomic_read(&kiblnd_data.kib_nthreads));
2959                         set_current_state(TASK_UNINTERRUPTIBLE);
2960                         schedule_timeout(cfs_time_seconds(1));
2961                 }
2962
2963                 /* fall through */
2964
2965         case IBLND_INIT_NOTHING:
2966                 break;
2967         }
2968
2969         if (kiblnd_data.kib_peers != NULL) {
2970                 LIBCFS_FREE(kiblnd_data.kib_peers,
2971                             sizeof(struct list_head) *
2972                             kiblnd_data.kib_peer_hash_size);
2973         }
2974
2975         if (kiblnd_data.kib_scheds != NULL)
2976                 cfs_percpt_free(kiblnd_data.kib_scheds);
2977
2978         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2979                atomic_read(&libcfs_kmemory));
2980
2981         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2982         module_put(THIS_MODULE);
2983 }
2984
2985 static void
2986 kiblnd_shutdown(struct lnet_ni *ni)
2987 {
2988         kib_net_t        *net = ni->ni_data;
2989         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2990         int               i;
2991         unsigned long     flags;
2992
2993         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2994
2995         if (net == NULL)
2996                 goto out;
2997
2998         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2999                atomic_read(&libcfs_kmemory));
3000
3001         write_lock_irqsave(g_lock, flags);
3002         net->ibn_shutdown = 1;
3003         write_unlock_irqrestore(g_lock, flags);
3004
3005         switch (net->ibn_init) {
3006         default:
3007                 LBUG();
3008
3009         case IBLND_INIT_ALL:
3010                 /* nuke all existing peers within this net */
3011                 kiblnd_del_peer(ni, LNET_NID_ANY);
3012
3013                 /* Wait for all peer_ni state to clean up */
3014                 i = 2;
3015                 while (atomic_read(&net->ibn_npeers) != 0) {
3016                         i++;
3017                         /* power of 2? */
3018                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3019                                "%s: waiting for %d peers to disconnect\n",
3020                                libcfs_nid2str(ni->ni_nid),
3021                                atomic_read(&net->ibn_npeers));
3022                         set_current_state(TASK_UNINTERRUPTIBLE);
3023                         schedule_timeout(cfs_time_seconds(1));
3024                 }
3025
3026                 kiblnd_net_fini_pools(net);
3027
3028                 write_lock_irqsave(g_lock, flags);
3029                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3030                 net->ibn_dev->ibd_nnets--;
3031                 list_del(&net->ibn_list);
3032                 write_unlock_irqrestore(g_lock, flags);
3033
3034                 /* fall through */
3035
3036         case IBLND_INIT_NOTHING:
3037                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3038
3039                 if (net->ibn_dev != NULL &&
3040                     net->ibn_dev->ibd_nnets == 0)
3041                         kiblnd_destroy_dev(net->ibn_dev);
3042
3043                 break;
3044         }
3045
3046         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3047                atomic_read(&libcfs_kmemory));
3048
3049         net->ibn_init = IBLND_INIT_NOTHING;
3050         ni->ni_data = NULL;
3051
3052         LIBCFS_FREE(net, sizeof(*net));
3053
3054 out:
3055         if (list_empty(&kiblnd_data.kib_devs))
3056                 kiblnd_base_shutdown();
3057         return;
3058 }
3059
3060 static int
3061 kiblnd_base_startup(void)
3062 {
3063         struct kib_sched_info   *sched;
3064         int                     rc;
3065         int                     i;
3066
3067         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3068
3069         try_module_get(THIS_MODULE);
3070         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3071
3072         rwlock_init(&kiblnd_data.kib_global_lock);
3073
3074         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3075         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3076
3077         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3078         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3079                      sizeof(struct list_head) *
3080                      kiblnd_data.kib_peer_hash_size);
3081         if (kiblnd_data.kib_peers == NULL)
3082                 goto failed;
3083
3084         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3085                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3086
3087         spin_lock_init(&kiblnd_data.kib_connd_lock);
3088         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3089         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3090         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3091         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3092
3093         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3094         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3095
3096         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3097                                                   sizeof(*sched));
3098         if (kiblnd_data.kib_scheds == NULL)
3099                 goto failed;
3100
3101         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3102                 int     nthrs;
3103
3104                 spin_lock_init(&sched->ibs_lock);
3105                 INIT_LIST_HEAD(&sched->ibs_conns);
3106                 init_waitqueue_head(&sched->ibs_waitq);
3107
3108                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3109                 if (*kiblnd_tunables.kib_nscheds > 0) {
3110                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3111                 } else {
3112                         /* max to half of CPUs, another half is reserved for
3113                          * upper layer modules */
3114                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3115                 }
3116
3117                 sched->ibs_nthreads_max = nthrs;
3118                 sched->ibs_cpt = i;
3119         }
3120
3121         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3122
3123         /* lists/ptrs/locks initialised */
3124         kiblnd_data.kib_init = IBLND_INIT_DATA;
3125         /*****************************************************/
3126
3127         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3128         if (rc != 0) {
3129                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3130                 goto failed;
3131         }
3132
3133         if (*kiblnd_tunables.kib_dev_failover != 0)
3134                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3135                                          "kiblnd_failover");
3136
3137         if (rc != 0) {
3138                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3139                 goto failed;
3140         }
3141
3142         /* flag everything initialised */
3143         kiblnd_data.kib_init = IBLND_INIT_ALL;
3144         /*****************************************************/
3145
3146         return 0;
3147
3148  failed:
3149         kiblnd_base_shutdown();
3150         return -ENETDOWN;
3151 }
3152
3153 static int
3154 kiblnd_start_schedulers(struct kib_sched_info *sched)
3155 {
3156         int     rc = 0;
3157         int     nthrs;
3158         int     i;
3159
3160         if (sched->ibs_nthreads == 0) {
3161                 if (*kiblnd_tunables.kib_nscheds > 0) {
3162                         nthrs = sched->ibs_nthreads_max;
3163                 } else {
3164                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3165                                                sched->ibs_cpt);
3166                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3167                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3168                 }
3169         } else {
3170                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3171                 /* increase one thread if there is new interface */
3172                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3173         }
3174
3175         for (i = 0; i < nthrs; i++) {
3176                 long    id;
3177                 char    name[20];
3178                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3179                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3180                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3181                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3182                 if (rc == 0)
3183                         continue;
3184
3185                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3186                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3187                 break;
3188         }
3189
3190         sched->ibs_nthreads += i;
3191         return rc;
3192 }
3193
3194 static int
3195 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3196 {
3197         int     cpt;
3198         int     rc;
3199         int     i;
3200
3201         for (i = 0; i < ncpts; i++) {
3202                 struct kib_sched_info *sched;
3203
3204                 cpt = (cpts == NULL) ? i : cpts[i];
3205                 sched = kiblnd_data.kib_scheds[cpt];
3206
3207                 if (!newdev && sched->ibs_nthreads > 0)
3208                         continue;
3209
3210                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3211                 if (rc != 0) {
3212                         CERROR("Failed to start scheduler threads for %s\n",
3213                                dev->ibd_ifname);
3214                         return rc;
3215                 }
3216         }
3217         return 0;
3218 }
3219
3220 static kib_dev_t *
3221 kiblnd_dev_search(char *ifname)
3222 {
3223         kib_dev_t       *alias = NULL;
3224         kib_dev_t       *dev;
3225         char            *colon;
3226         char            *colon2;
3227
3228         colon = strchr(ifname, ':');
3229         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3230                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3231                         return dev;
3232
3233                 if (alias != NULL)
3234                         continue;
3235
3236                 colon2 = strchr(dev->ibd_ifname, ':');
3237                 if (colon != NULL)
3238                         *colon = 0;
3239                 if (colon2 != NULL)
3240                         *colon2 = 0;
3241
3242                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3243                         alias = dev;
3244
3245                 if (colon != NULL)
3246                         *colon = ':';
3247                 if (colon2 != NULL)
3248                         *colon2 = ':';
3249         }
3250         return alias;
3251 }
3252
3253 static int
3254 kiblnd_startup(struct lnet_ni *ni)
3255 {
3256         char                     *ifname;
3257         kib_dev_t                *ibdev = NULL;
3258         kib_net_t                *net;
3259         unsigned long             flags;
3260         int                       rc;
3261         int                       newdev;
3262         int                       node_id;
3263
3264         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3265
3266         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3267                 rc = kiblnd_base_startup();
3268                 if (rc != 0)
3269                         return rc;
3270         }
3271
3272         LIBCFS_ALLOC(net, sizeof(*net));
3273         ni->ni_data = net;
3274         if (net == NULL)
3275                 goto failed;
3276
3277         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3278
3279         kiblnd_tunables_setup(ni);
3280
3281         if (ni->ni_interfaces[0] != NULL) {
3282                 /* Use the IPoIB interface specified in 'networks=' */
3283
3284                 CLASSERT(LNET_INTERFACES_NUM > 1);
3285                 if (ni->ni_interfaces[1] != NULL) {
3286                         CERROR("Multiple interfaces not supported\n");
3287                         goto failed;
3288                 }
3289
3290                 ifname = ni->ni_interfaces[0];
3291         } else {
3292                 ifname = *kiblnd_tunables.kib_default_ipif;
3293         }
3294
3295         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3296                 CERROR("IPoIB interface name too long: %s\n", ifname);
3297                 goto failed;
3298         }
3299
3300         ibdev = kiblnd_dev_search(ifname);
3301
3302         newdev = ibdev == NULL;
3303         /* hmm...create kib_dev even for alias */
3304         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3305                 ibdev = kiblnd_create_dev(ifname);
3306
3307         if (ibdev == NULL)
3308                 goto failed;
3309
3310         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3311         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3312
3313         net->ibn_dev = ibdev;
3314         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3315
3316         rc = kiblnd_dev_start_threads(ibdev, newdev,
3317                                       ni->ni_cpts, ni->ni_ncpts);
3318         if (rc != 0)
3319                 goto failed;
3320
3321         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3322         if (rc != 0) {
3323                 CERROR("Failed to initialize NI pools: %d\n", rc);
3324                 goto failed;
3325         }
3326
3327         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3328         ibdev->ibd_nnets++;
3329         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3330         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3331
3332         net->ibn_init = IBLND_INIT_ALL;
3333
3334         return 0;
3335
3336 failed:
3337         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3338                 kiblnd_destroy_dev(ibdev);
3339
3340         kiblnd_shutdown(ni);
3341
3342         CDEBUG(D_NET, "kiblnd_startup failed\n");
3343         return -ENETDOWN;
3344 }
3345
3346 static struct lnet_lnd the_o2iblnd = {
3347         .lnd_type       = O2IBLND,
3348         .lnd_startup    = kiblnd_startup,
3349         .lnd_shutdown   = kiblnd_shutdown,
3350         .lnd_ctl        = kiblnd_ctl,
3351         .lnd_query      = kiblnd_query,
3352         .lnd_send       = kiblnd_send,
3353         .lnd_recv       = kiblnd_recv,
3354 };
3355
3356 static void __exit ko2iblnd_exit(void)
3357 {
3358         lnet_unregister_lnd(&the_o2iblnd);
3359 }
3360
3361 static int __init ko2iblnd_init(void)
3362 {
3363         int rc;
3364
3365         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3366         CLASSERT(offsetof(kib_msg_t,
3367                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3368                  IBLND_MSG_SIZE);
3369         CLASSERT(offsetof(kib_msg_t,
3370                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3371                  <= IBLND_MSG_SIZE);
3372
3373         rc = kiblnd_tunables_init();
3374         if (rc != 0)
3375                 return rc;
3376
3377         lnet_register_lnd(&the_o2iblnd);
3378
3379         return 0;
3380 }
3381
3382 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3383 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3384 MODULE_VERSION("2.8.0");
3385 MODULE_LICENSE("GPL");
3386
3387 module_init(ko2iblnd_init);
3388 module_exit(ko2iblnd_exit);