lnet/klnds/o2iblnd/o2iblnd.c

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