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 /*
 702  * Get the scheduler bound to this CPT. If the scheduler has no
 703  * threads, which means that the CPT has no CPUs, then grab the
 704  * next scheduler that we can use.
 705  *
 706  * This case would be triggered if a NUMA node is configured with
 707  * no associated CPUs.
 708  */
 709 static struct kib_sched_info *
 710 kiblnd_get_scheduler(int cpt)
 711 {
 712         struct kib_sched_info *sched;
 713         int i;
 714
 715         sched = kiblnd_data.kib_scheds[cpt];
 716
 717         if (sched->ibs_nthreads > 0)
 718                 return sched;
 719
 720         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 721                 if (sched->ibs_nthreads > 0) {
 722                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 723                                         cpt, sched->ibs_cpt);
 724                         return sched;
 725                 }
 726         }
 727
 728         return NULL;
 729 }
 730
 731 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 732 {
 733         /*
 734          * One WR for the LNet message
 735          * And ibc_max_frags for the transfer WRs
 736          */
 737         unsigned int ret = 1 + conn->ibc_max_frags;
 738
 739         /* account for a maximum of ibc_queue_depth in-flight transfers */
 740         ret *= conn->ibc_queue_depth;
 741         return ret;
 742 }
 743
 744 kib_conn_t *
 745 kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
 746                    int state, int version)
 747 {
 748         /* CAVEAT EMPTOR:
 749          * If the new conn is created successfully it takes over the caller's
 750          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 751          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 752          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 753          * to destroy 'cmid' here since I'm called from the CM which still has
 754          * its ref on 'cmid'). */
 755         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 756         kib_net_t              *net = peer_ni->ibp_ni->ni_data;
 757         kib_dev_t              *dev;
 758         struct ib_qp_init_attr *init_qp_attr;
 759         struct kib_sched_info   *sched;
 760 #ifdef HAVE_IB_CQ_INIT_ATTR
 761         struct ib_cq_init_attr  cq_attr = {};
 762 #endif
 763         kib_conn_t              *conn;
 764         struct ib_cq            *cq;
 765         unsigned long           flags;
 766         int                     cpt;
 767         int                     rc;
 768         int                     i;
 769
 770         LASSERT(net != NULL);
 771         LASSERT(!in_interrupt());
 772
 773         dev = net->ibn_dev;
 774
 775         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 776         sched = kiblnd_get_scheduler(cpt);
 777
 778         if (sched == NULL) {
 779                 CERROR("no schedulers available. node is unhealthy\n");
 780                 goto failed_0;
 781         }
 782
 783         /*
 784          * The cpt might have changed if we ended up selecting a non cpt
 785          * native scheduler. So use the scheduler's cpt instead.
 786          */
 787         cpt = sched->ibs_cpt;
 788
 789         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 790                          sizeof(*init_qp_attr));
 791         if (init_qp_attr == NULL) {
 792                 CERROR("Can't allocate qp_attr for %s\n",
 793                        libcfs_nid2str(peer_ni->ibp_nid));
 794                 goto failed_0;
 795         }
 796
 797         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 798         if (conn == NULL) {
 799                 CERROR("Can't allocate connection for %s\n",
 800                        libcfs_nid2str(peer_ni->ibp_nid));
 801                 goto failed_1;
 802         }
 803
 804         conn->ibc_state = IBLND_CONN_INIT;
 805         conn->ibc_version = version;
 806         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 807         cmid->context = conn;                   /* for future CM callbacks */
 808         conn->ibc_cmid = cmid;
 809         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 810         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 811
 812         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 813         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 814         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 815         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 816         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 817         INIT_LIST_HEAD(&conn->ibc_active_txs);
 818         spin_lock_init(&conn->ibc_lock);
 819
 820         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 821                          sizeof(*conn->ibc_connvars));
 822         if (conn->ibc_connvars == NULL) {
 823                 CERROR("Can't allocate in-progress connection state\n");
 824                 goto failed_2;
 825         }
 826
 827         write_lock_irqsave(glock, flags);
 828         if (dev->ibd_failover) {
 829                 write_unlock_irqrestore(glock, flags);
 830                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 831                 goto failed_2;
 832         }
 833
 834         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 835                 /* wakeup failover thread and teardown connection */
 836                 if (kiblnd_dev_can_failover(dev)) {
 837                         list_add_tail(&dev->ibd_fail_list,
 838                                       &kiblnd_data.kib_failed_devs);
 839                         wake_up(&kiblnd_data.kib_failover_waitq);
 840                 }
 841
 842                 write_unlock_irqrestore(glock, flags);
 843                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 844                        cmid->device->name, dev->ibd_ifname);
 845                 goto failed_2;
 846         }
 847
 848         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 849         conn->ibc_hdev = dev->ibd_hdev;
 850
 851         kiblnd_setup_mtu_locked(cmid);
 852
 853         write_unlock_irqrestore(glock, flags);
 854
 855         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 856                          IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
 857         if (conn->ibc_rxs == NULL) {
 858                 CERROR("Cannot allocate RX buffers\n");
 859                 goto failed_2;
 860         }
 861
 862         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 863                                 IBLND_RX_MSG_PAGES(conn));
 864         if (rc != 0)
 865                 goto failed_2;
 866
 867         kiblnd_map_rx_descs(conn);
 868
 869 #ifdef HAVE_IB_CQ_INIT_ATTR
 870         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 871         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 872         cq = ib_create_cq(cmid->device,
 873                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 874                           &cq_attr);
 875 #else
 876         cq = ib_create_cq(cmid->device,
 877                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 878                           IBLND_CQ_ENTRIES(conn),
 879                           kiblnd_get_completion_vector(conn, cpt));
 880 #endif
 881         if (IS_ERR(cq)) {
 882                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 883                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 884                 goto failed_2;
 885         }
 886
 887         conn->ibc_cq = cq;
 888
 889         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 890         if (rc != 0) {
 891                 CERROR("Can't request completion notification: %d\n", rc);
 892                 goto failed_2;
 893         }
 894
 895         init_qp_attr->event_handler = kiblnd_qp_event;
 896         init_qp_attr->qp_context = conn;
 897         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 898         init_qp_attr->cap.max_recv_sge = 1;
 899         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 900         init_qp_attr->qp_type = IB_QPT_RC;
 901         init_qp_attr->send_cq = cq;
 902         init_qp_attr->recv_cq = cq;
 903
 904         conn->ibc_sched = sched;
 905
 906         do {
 907                 init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 908                 init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 909
 910                 rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 911                 if (!rc || conn->ibc_queue_depth < 2)
 912                         break;
 913
 914                 conn->ibc_queue_depth--;
 915         } while (rc);
 916
 917         if (rc) {
 918                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 919                        "send_sge: %d, recv_sge: %d\n",
 920                        rc, init_qp_attr->cap.max_send_wr,
 921                        init_qp_attr->cap.max_recv_wr,
 922                        init_qp_attr->cap.max_send_sge,
 923                        init_qp_attr->cap.max_recv_sge);
 924                 goto failed_2;
 925         }
 926
 927         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 928                 CWARN("peer %s - queue depth reduced from %u to %u"
 929                       "  to allow for qp creation\n",
 930                       libcfs_nid2str(peer_ni->ibp_nid),
 931                       peer_ni->ibp_queue_depth,
 932                       conn->ibc_queue_depth);
 933
 934         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 935
 936         /* 1 ref for caller and each rxmsg */
 937         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 938         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 939
 940         /* post receives */
 941         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 942                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 943                 if (rc != 0) {
 944                         CERROR("Can't post rxmsg: %d\n", rc);
 945
 946                         /* Make posted receives complete */
 947                         kiblnd_abort_receives(conn);
 948
 949                         /* correct # of posted buffers
 950                          * NB locking needed now I'm racing with completion */
 951                         spin_lock_irqsave(&sched->ibs_lock, flags);
 952                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 953                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 954
 955                         /* cmid will be destroyed by CM(ofed) after cm_callback
 956                          * returned, so we can't refer it anymore
 957                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 958                         rdma_destroy_qp(conn->ibc_cmid);
 959                         conn->ibc_cmid = NULL;
 960
 961                         /* Drop my own and unused rxbuffer refcounts */
 962                         while (i++ <= IBLND_RX_MSGS(conn))
 963                                 kiblnd_conn_decref(conn);
 964
 965                         return NULL;
 966                 }
 967         }
 968
 969         /* Init successful! */
 970         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 971                  state == IBLND_CONN_PASSIVE_WAIT);
 972         conn->ibc_state = state;
 973
 974         /* 1 more conn */
 975         atomic_inc(&net->ibn_nconns);
 976         return conn;
 977
 978  failed_2:
 979         kiblnd_destroy_conn(conn, true);
 980  failed_1:
 981         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 982  failed_0:
 983         return NULL;
 984 }
 985
 986 void
 987 kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 988 {
 989         struct rdma_cm_id *cmid = conn->ibc_cmid;
 990         kib_peer_ni_t        *peer_ni = conn->ibc_peer;
 991         int                rc;
 992
 993         LASSERT (!in_interrupt());
 994         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 995         LASSERT(list_empty(&conn->ibc_early_rxs));
 996         LASSERT(list_empty(&conn->ibc_tx_noops));
 997         LASSERT(list_empty(&conn->ibc_tx_queue));
 998         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 999         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1000         LASSERT(list_empty(&conn->ibc_active_txs));
1001         LASSERT (conn->ibc_noops_posted == 0);
1002         LASSERT (conn->ibc_nsends_posted == 0);
1003
1004         switch (conn->ibc_state) {
1005         default:
1006                 /* conn must be completely disengaged from the network */
1007                 LBUG();
1008
1009         case IBLND_CONN_DISCONNECTED:
1010                 /* connvars should have been freed already */
1011                 LASSERT (conn->ibc_connvars == NULL);
1012                 break;
1013
1014         case IBLND_CONN_INIT:
1015                 break;
1016         }
1017
1018         /* conn->ibc_cmid might be destroyed by CM already */
1019         if (cmid != NULL && cmid->qp != NULL)
1020                 rdma_destroy_qp(cmid);
1021
1022         if (conn->ibc_cq != NULL) {
1023                 rc = ib_destroy_cq(conn->ibc_cq);
1024                 if (rc != 0)
1025                         CWARN("Error destroying CQ: %d\n", rc);
1026         }
1027
1028         if (conn->ibc_rx_pages != NULL)
1029                 kiblnd_unmap_rx_descs(conn);
1030
1031         if (conn->ibc_rxs != NULL) {
1032                 LIBCFS_FREE(conn->ibc_rxs,
1033                             IBLND_RX_MSGS(conn) * sizeof(kib_rx_t));
1034         }
1035
1036         if (conn->ibc_connvars != NULL)
1037                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1038
1039         if (conn->ibc_hdev != NULL)
1040                 kiblnd_hdev_decref(conn->ibc_hdev);
1041
1042         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1043         if (conn->ibc_state != IBLND_CONN_INIT) {
1044                 kib_net_t *net = peer_ni->ibp_ni->ni_data;
1045
1046                 kiblnd_peer_decref(peer_ni);
1047                 rdma_destroy_id(cmid);
1048                 atomic_dec(&net->ibn_nconns);
1049         }
1050
1051         if (free_conn)
1052                 LIBCFS_FREE(conn, sizeof(*conn));
1053 }
1054
1055 int
1056 kiblnd_close_peer_conns_locked(kib_peer_ni_t *peer_ni, int why)
1057 {
1058         kib_conn_t              *conn;
1059         struct list_head        *ctmp;
1060         struct list_head        *cnxt;
1061         int                     count = 0;
1062
1063         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1064                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1065
1066                 CDEBUG(D_NET, "Closing conn -> %s, "
1067                               "version: %x, reason: %d\n",
1068                        libcfs_nid2str(peer_ni->ibp_nid),
1069                        conn->ibc_version, why);
1070
1071                 kiblnd_close_conn_locked(conn, why);
1072                 count++;
1073         }
1074
1075         return count;
1076 }
1077
1078 int
1079 kiblnd_close_stale_conns_locked(kib_peer_ni_t *peer_ni,
1080                                 int version, __u64 incarnation)
1081 {
1082         kib_conn_t              *conn;
1083         struct list_head        *ctmp;
1084         struct list_head        *cnxt;
1085         int                     count = 0;
1086
1087         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1088                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
1089
1090                 if (conn->ibc_version     == version &&
1091                     conn->ibc_incarnation == incarnation)
1092                         continue;
1093
1094                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1095                               "incarnation:%#llx(%x, %#llx)\n",
1096                        libcfs_nid2str(peer_ni->ibp_nid),
1097                        conn->ibc_version, conn->ibc_incarnation,
1098                        version, incarnation);
1099
1100                 kiblnd_close_conn_locked(conn, -ESTALE);
1101                 count++;
1102         }
1103
1104         return count;
1105 }
1106
1107 static int
1108 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1109 {
1110         kib_peer_ni_t           *peer_ni;
1111         struct list_head        *ptmp;
1112         struct list_head        *pnxt;
1113         int                     lo;
1114         int                     hi;
1115         int                     i;
1116         unsigned long           flags;
1117         int                     count = 0;
1118
1119         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1120
1121         if (nid != LNET_NID_ANY)
1122                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1123         else {
1124                 lo = 0;
1125                 hi = kiblnd_data.kib_peer_hash_size - 1;
1126         }
1127
1128         for (i = lo; i <= hi; i++) {
1129                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1130
1131                         peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
1132                         LASSERT(!kiblnd_peer_idle(peer_ni));
1133
1134                         if (peer_ni->ibp_ni != ni)
1135                                 continue;
1136
1137                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1138                                 continue;
1139
1140                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1141                 }
1142         }
1143
1144         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1145
1146         /* wildcards always succeed */
1147         if (nid == LNET_NID_ANY)
1148                 return 0;
1149
1150         return (count == 0) ? -ENOENT : 0;
1151 }
1152
1153 static int
1154 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1155 {
1156         struct libcfs_ioctl_data *data = arg;
1157         int                       rc = -EINVAL;
1158
1159         switch(cmd) {
1160         case IOC_LIBCFS_GET_PEER: {
1161                 lnet_nid_t   nid = 0;
1162                 int          count = 0;
1163
1164                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1165                                           &nid, &count);
1166                 data->ioc_nid    = nid;
1167                 data->ioc_count  = count;
1168                 break;
1169         }
1170
1171         case IOC_LIBCFS_DEL_PEER: {
1172                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1173                 break;
1174         }
1175         case IOC_LIBCFS_GET_CONN: {
1176                 kib_conn_t *conn;
1177
1178                 rc = 0;
1179                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1180                 if (conn == NULL) {
1181                         rc = -ENOENT;
1182                         break;
1183                 }
1184
1185                 LASSERT(conn->ibc_cmid != NULL);
1186                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1187                 if (conn->ibc_cmid->route.path_rec == NULL)
1188                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1189                 else
1190                         data->ioc_u32[0] =
1191                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1192                 kiblnd_conn_decref(conn);
1193                 break;
1194         }
1195         case IOC_LIBCFS_CLOSE_CONNECTION: {
1196                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1197                 break;
1198         }
1199
1200         default:
1201                 break;
1202         }
1203
1204         return rc;
1205 }
1206
1207 static void
1208 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1209 {
1210         cfs_time_t      last_alive = 0;
1211         cfs_time_t      now = cfs_time_current();
1212         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
1213         kib_peer_ni_t   *peer_ni;
1214         unsigned long   flags;
1215
1216         read_lock_irqsave(glock, flags);
1217
1218         peer_ni = kiblnd_find_peer_locked(ni, nid);
1219         if (peer_ni != NULL)
1220                 last_alive = peer_ni->ibp_last_alive;
1221
1222         read_unlock_irqrestore(glock, flags);
1223
1224         if (last_alive != 0)
1225                 *when = last_alive;
1226
1227         /* peer_ni is not persistent in hash, trigger peer_ni creation
1228          * and connection establishment with a NULL tx */
1229         if (peer_ni == NULL)
1230                 kiblnd_launch_tx(ni, NULL, nid);
1231
1232         CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago\n",
1233                libcfs_nid2str(nid), peer_ni,
1234                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1235         return;
1236 }
1237
1238 static void
1239 kiblnd_free_pages(kib_pages_t *p)
1240 {
1241         int     npages = p->ibp_npages;
1242         int     i;
1243
1244         for (i = 0; i < npages; i++) {
1245                 if (p->ibp_pages[i] != NULL)
1246                         __free_page(p->ibp_pages[i]);
1247         }
1248
1249         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1250 }
1251
1252 int
1253 kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1254 {
1255         kib_pages_t     *p;
1256         int             i;
1257
1258         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1259                          offsetof(kib_pages_t, ibp_pages[npages]));
1260         if (p == NULL) {
1261                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1262                 return -ENOMEM;
1263         }
1264
1265         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1266         p->ibp_npages = npages;
1267
1268         for (i = 0; i < npages; i++) {
1269                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1270                                                      GFP_NOFS);
1271                 if (p->ibp_pages[i] == NULL) {
1272                         CERROR("Can't allocate page %d of %d\n", i, npages);
1273                         kiblnd_free_pages(p);
1274                         return -ENOMEM;
1275                 }
1276         }
1277
1278         *pp = p;
1279         return 0;
1280 }
1281
1282 void
1283 kiblnd_unmap_rx_descs(kib_conn_t *conn)
1284 {
1285         kib_rx_t *rx;
1286         int       i;
1287
1288         LASSERT (conn->ibc_rxs != NULL);
1289         LASSERT (conn->ibc_hdev != NULL);
1290
1291         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1292                 rx = &conn->ibc_rxs[i];
1293
1294                 LASSERT(rx->rx_nob >= 0); /* not posted */
1295
1296                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1297                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1298                                                           rx->rx_msgaddr),
1299                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1300         }
1301
1302         kiblnd_free_pages(conn->ibc_rx_pages);
1303
1304         conn->ibc_rx_pages = NULL;
1305 }
1306
1307 void
1308 kiblnd_map_rx_descs(kib_conn_t *conn)
1309 {
1310         kib_rx_t       *rx;
1311         struct page    *pg;
1312         int             pg_off;
1313         int             ipg;
1314         int             i;
1315
1316         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1317                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1318                 rx = &conn->ibc_rxs[i];
1319
1320                 rx->rx_conn = conn;
1321                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1322
1323                 rx->rx_msgaddr =
1324                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1325                                               rx->rx_msg, IBLND_MSG_SIZE,
1326                                               DMA_FROM_DEVICE);
1327                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1328                                                   rx->rx_msgaddr));
1329                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1330
1331                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1332                        i, rx->rx_msg, rx->rx_msgaddr,
1333                        (__u64)(page_to_phys(pg) + pg_off));
1334
1335                 pg_off += IBLND_MSG_SIZE;
1336                 LASSERT(pg_off <= PAGE_SIZE);
1337
1338                 if (pg_off == PAGE_SIZE) {
1339                         pg_off = 0;
1340                         ipg++;
1341                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1342                 }
1343         }
1344 }
1345
1346 static void
1347 kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1348 {
1349         kib_hca_dev_t  *hdev = tpo->tpo_hdev;
1350         kib_tx_t       *tx;
1351         int             i;
1352
1353         LASSERT (tpo->tpo_pool.po_allocated == 0);
1354
1355         if (hdev == NULL)
1356                 return;
1357
1358         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1359                 tx = &tpo->tpo_tx_descs[i];
1360                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1361                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1362                                                           tx->tx_msgaddr),
1363                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1364         }
1365
1366         kiblnd_hdev_decref(hdev);
1367         tpo->tpo_hdev = NULL;
1368 }
1369
1370 static kib_hca_dev_t *
1371 kiblnd_current_hdev(kib_dev_t *dev)
1372 {
1373         kib_hca_dev_t *hdev;
1374         unsigned long  flags;
1375         int            i = 0;
1376
1377         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1378         while (dev->ibd_failover) {
1379                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1380                 if (i++ % 50 == 0)
1381                         CDEBUG(D_NET, "%s: Wait for failover\n",
1382                                dev->ibd_ifname);
1383                 set_current_state(TASK_INTERRUPTIBLE);
1384                 schedule_timeout(cfs_time_seconds(1) / 100);
1385
1386                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1387         }
1388
1389         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1390         hdev = dev->ibd_hdev;
1391
1392         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1393
1394         return hdev;
1395 }
1396
1397 static void
1398 kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1399 {
1400         kib_pages_t    *txpgs = tpo->tpo_tx_pages;
1401         kib_pool_t     *pool  = &tpo->tpo_pool;
1402         kib_net_t      *net   = pool->po_owner->ps_net;
1403         kib_dev_t      *dev;
1404         struct page    *page;
1405         kib_tx_t       *tx;
1406         int             page_offset;
1407         int             ipage;
1408         int             i;
1409
1410         LASSERT (net != NULL);
1411
1412         dev = net->ibn_dev;
1413
1414         /* pre-mapped messages are not bigger than 1 page */
1415         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1416
1417         /* No fancy arithmetic when we do the buffer calculations */
1418         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1419
1420         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1421
1422         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1423                 page = txpgs->ibp_pages[ipage];
1424                 tx = &tpo->tpo_tx_descs[i];
1425
1426                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1427                                            page_offset);
1428
1429                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1430                                                        tx->tx_msg,
1431                                                        IBLND_MSG_SIZE,
1432                                                        DMA_TO_DEVICE);
1433                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1434                                                   tx->tx_msgaddr));
1435                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1436
1437                 list_add(&tx->tx_list, &pool->po_free_list);
1438
1439                 page_offset += IBLND_MSG_SIZE;
1440                 LASSERT(page_offset <= PAGE_SIZE);
1441
1442                 if (page_offset == PAGE_SIZE) {
1443                         page_offset = 0;
1444                         ipage++;
1445                         LASSERT(ipage <= txpgs->ibp_npages);
1446                 }
1447         }
1448 }
1449
1450 #ifdef HAVE_IB_GET_DMA_MR
1451 struct ib_mr *
1452 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1453                       int negotiated_nfrags)
1454 {
1455         kib_net_t     *net   = ni->ni_data;
1456         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
1457         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1458         int     mod;
1459         __u16   nfrags;
1460
1461         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1462         mod = tunables->lnd_map_on_demand;
1463         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1464
1465         LASSERT(hdev->ibh_mrs != NULL);
1466
1467         if (mod > 0 && nfrags <= rd->rd_nfrags)
1468                 return NULL;
1469
1470         return hdev->ibh_mrs;
1471 }
1472 #endif
1473
1474 static void
1475 kiblnd_destroy_fmr_pool(kib_fmr_pool_t *fpo)
1476 {
1477         LASSERT(fpo->fpo_map_count == 0);
1478
1479         if (fpo->fpo_is_fmr) {
1480                 if (fpo->fmr.fpo_fmr_pool)
1481                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1482         } else {
1483                 struct kib_fast_reg_descriptor *frd, *tmp;
1484                 int i = 0;
1485
1486                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1487                                          frd_list) {
1488                         list_del(&frd->frd_list);
1489 #ifndef HAVE_IB_MAP_MR_SG
1490                         ib_free_fast_reg_page_list(frd->frd_frpl);
1491 #endif
1492                         ib_dereg_mr(frd->frd_mr);
1493                         LIBCFS_FREE(frd, sizeof(*frd));
1494                         i++;
1495                 }
1496                 if (i < fpo->fast_reg.fpo_pool_size)
1497                         CERROR("FastReg pool still has %d regions registered\n",
1498                                 fpo->fast_reg.fpo_pool_size - i);
1499         }
1500
1501         if (fpo->fpo_hdev)
1502                 kiblnd_hdev_decref(fpo->fpo_hdev);
1503
1504         LIBCFS_FREE(fpo, sizeof(*fpo));
1505 }
1506
1507 static void
1508 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1509 {
1510         kib_fmr_pool_t *fpo, *tmp;
1511
1512         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1513                 list_del(&fpo->fpo_list);
1514                 kiblnd_destroy_fmr_pool(fpo);
1515         }
1516 }
1517
1518 static int
1519 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1520                      int ncpts)
1521 {
1522         int size = tunables->lnd_fmr_pool_size / ncpts;
1523
1524         return max(IBLND_FMR_POOL, size);
1525 }
1526
1527 static int
1528 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1529                          int ncpts)
1530 {
1531         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1532
1533         return max(IBLND_FMR_POOL_FLUSH, size);
1534 }
1535
1536 static int kiblnd_alloc_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1537 {
1538         struct ib_fmr_pool_param param = {
1539                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1540                 .page_shift        = PAGE_SHIFT,
1541                 .access            = (IB_ACCESS_LOCAL_WRITE |
1542                                       IB_ACCESS_REMOTE_WRITE),
1543                 .pool_size         = fps->fps_pool_size,
1544                 .dirty_watermark   = fps->fps_flush_trigger,
1545                 .flush_function    = NULL,
1546                 .flush_arg         = NULL,
1547                 .cache             = !!fps->fps_cache };
1548         int rc = 0;
1549
1550         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1551                                                    &param);
1552         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1553                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1554                 if (rc != -ENOSYS)
1555                         CERROR("Failed to create FMR pool: %d\n", rc);
1556                 else
1557                         CERROR("FMRs are not supported\n");
1558         }
1559
1560         return rc;
1561 }
1562
1563 static int kiblnd_alloc_freg_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t *fpo)
1564 {
1565         struct kib_fast_reg_descriptor *frd, *tmp;
1566         int i, rc;
1567
1568         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1569         fpo->fast_reg.fpo_pool_size = 0;
1570         for (i = 0; i < fps->fps_pool_size; i++) {
1571                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1572                                  sizeof(*frd));
1573                 if (!frd) {
1574                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1575                         rc = -ENOMEM;
1576                         goto out;
1577                 }
1578                 frd->frd_mr = NULL;
1579
1580 #ifndef HAVE_IB_MAP_MR_SG
1581                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1582                                                             LNET_MAX_PAYLOAD/PAGE_SIZE);
1583                 if (IS_ERR(frd->frd_frpl)) {
1584                         rc = PTR_ERR(frd->frd_frpl);
1585                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1586                                 rc);
1587                         frd->frd_frpl = NULL;
1588                         goto out_middle;
1589                 }
1590 #endif
1591
1592 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1593                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1594                                                    LNET_MAX_PAYLOAD/PAGE_SIZE);
1595 #else
1596                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1597                                           IB_MR_TYPE_MEM_REG,
1598                                           LNET_MAX_PAYLOAD/PAGE_SIZE);
1599 #endif
1600                 if (IS_ERR(frd->frd_mr)) {
1601                         rc = PTR_ERR(frd->frd_mr);
1602                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1603                         frd->frd_mr = NULL;
1604                         goto out_middle;
1605                 }
1606
1607                 /* There appears to be a bug in MLX5 code where you must
1608                  * invalidate the rkey of a new FastReg pool before first
1609                  * using it. Thus, I am marking the FRD invalid here. */
1610                 frd->frd_valid = false;
1611
1612                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1613                 fpo->fast_reg.fpo_pool_size++;
1614         }
1615
1616         return 0;
1617
1618 out_middle:
1619         if (frd->frd_mr)
1620                 ib_dereg_mr(frd->frd_mr);
1621 #ifndef HAVE_IB_MAP_MR_SG
1622         if (frd->frd_frpl)
1623                 ib_free_fast_reg_page_list(frd->frd_frpl);
1624 #endif
1625         LIBCFS_FREE(frd, sizeof(*frd));
1626
1627 out:
1628         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1629                                  frd_list) {
1630                 list_del(&frd->frd_list);
1631 #ifndef HAVE_IB_MAP_MR_SG
1632                 ib_free_fast_reg_page_list(frd->frd_frpl);
1633 #endif
1634                 ib_dereg_mr(frd->frd_mr);
1635                 LIBCFS_FREE(frd, sizeof(*frd));
1636         }
1637
1638         return rc;
1639 }
1640
1641 static int
1642 kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps, kib_fmr_pool_t **pp_fpo)
1643 {
1644         struct ib_device_attr *dev_attr;
1645         kib_dev_t *dev = fps->fps_net->ibn_dev;
1646         kib_fmr_pool_t *fpo;
1647         int rc;
1648
1649 #ifndef HAVE_IB_DEVICE_ATTRS
1650         dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
1651         if (!dev_attr)
1652                 return -ENOMEM;
1653 #endif
1654
1655         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1656         if (!fpo) {
1657                 rc = -ENOMEM;
1658                 goto out_dev_attr;
1659         }
1660
1661         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1662
1663 #ifdef HAVE_IB_DEVICE_ATTRS
1664         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1665 #else
1666         rc = ib_query_device(fpo->fpo_hdev->ibh_ibdev, dev_attr);
1667         if (rc) {
1668                 CERROR("Query device failed for %s: %d\n",
1669                         fpo->fpo_hdev->ibh_ibdev->name, rc);
1670                 goto out_dev_attr;
1671         }
1672 #endif
1673
1674         /* Check for FMR or FastReg support */
1675         fpo->fpo_is_fmr = 0;
1676         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1677             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1678             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1679             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1680                 LCONSOLE_INFO("Using FMR for registration\n");
1681                 fpo->fpo_is_fmr = 1;
1682         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1683                 LCONSOLE_INFO("Using FastReg for registration\n");
1684         } else {
1685                 rc = -ENOSYS;
1686                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1687                 goto out_dev_attr;
1688         }
1689
1690         if (fpo->fpo_is_fmr)
1691                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1692         else
1693                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1694         if (rc)
1695                 goto out_fpo;
1696
1697 #ifndef HAVE_IB_DEVICE_ATTRS
1698         kfree(dev_attr);
1699 #endif
1700         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1701         fpo->fpo_owner    = fps;
1702         *pp_fpo = fpo;
1703
1704         return 0;
1705
1706 out_fpo:
1707         kiblnd_hdev_decref(fpo->fpo_hdev);
1708         LIBCFS_FREE(fpo, sizeof(*fpo));
1709
1710 out_dev_attr:
1711 #ifndef HAVE_IB_DEVICE_ATTRS
1712         kfree(dev_attr);
1713 #endif
1714
1715         return rc;
1716 }
1717
1718 static void
1719 kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps, struct list_head *zombies)
1720 {
1721         if (fps->fps_net == NULL) /* intialized? */
1722                 return;
1723
1724         spin_lock(&fps->fps_lock);
1725
1726         while (!list_empty(&fps->fps_pool_list)) {
1727                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1728                                                  kib_fmr_pool_t, fpo_list);
1729                 fpo->fpo_failed = 1;
1730                 list_del(&fpo->fpo_list);
1731                 if (fpo->fpo_map_count == 0)
1732                         list_add(&fpo->fpo_list, zombies);
1733                 else
1734                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1735         }
1736
1737         spin_unlock(&fps->fps_lock);
1738 }
1739
1740 static void
1741 kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1742 {
1743         if (fps->fps_net != NULL) { /* initialized? */
1744                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1745                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1746         }
1747 }
1748
1749 static int
1750 kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt, int ncpts,
1751                         kib_net_t *net,
1752                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1753 {
1754         kib_fmr_pool_t *fpo;
1755         int             rc;
1756
1757         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1758
1759         fps->fps_net = net;
1760         fps->fps_cpt = cpt;
1761
1762         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1763         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1764         fps->fps_cache = tunables->lnd_fmr_cache;
1765
1766         spin_lock_init(&fps->fps_lock);
1767         INIT_LIST_HEAD(&fps->fps_pool_list);
1768         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1769
1770         rc = kiblnd_create_fmr_pool(fps, &fpo);
1771         if (rc == 0)
1772                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1773
1774         return rc;
1775 }
1776
1777 static int
1778 kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, cfs_time_t now)
1779 {
1780         if (fpo->fpo_map_count != 0) /* still in use */
1781                 return 0;
1782         if (fpo->fpo_failed)
1783                 return 1;
1784         return cfs_time_aftereq(now, fpo->fpo_deadline);
1785 }
1786
1787 static int
1788 kiblnd_map_tx_pages(kib_tx_t *tx, kib_rdma_desc_t *rd)
1789 {
1790         kib_hca_dev_t   *hdev;
1791         __u64           *pages = tx->tx_pages;
1792         int             npages;
1793         int             size;
1794         int             i;
1795
1796         hdev = tx->tx_pool->tpo_hdev;
1797
1798         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1799                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1800                         size += hdev->ibh_page_size) {
1801                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1802                                            hdev->ibh_page_mask) + size;
1803                 }
1804         }
1805
1806         return npages;
1807 }
1808
1809 void
1810 kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1811 {
1812         struct list_head   zombies = LIST_HEAD_INIT(zombies);
1813         kib_fmr_pool_t    *fpo = fmr->fmr_pool;
1814         kib_fmr_poolset_t *fps;
1815         cfs_time_t         now = cfs_time_current();
1816         kib_fmr_pool_t    *tmp;
1817         int                rc;
1818
1819         if (!fpo)
1820                 return;
1821
1822         fps = fpo->fpo_owner;
1823         if (fpo->fpo_is_fmr) {
1824                 if (fmr->fmr_pfmr) {
1825                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1826                         LASSERT(!rc);
1827                         fmr->fmr_pfmr = NULL;
1828                 }
1829
1830                 if (status) {
1831                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1832                         LASSERT(!rc);
1833                 }
1834         } else {
1835                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1836
1837                 if (frd) {
1838                         frd->frd_valid = false;
1839                         spin_lock(&fps->fps_lock);
1840                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1841                         spin_unlock(&fps->fps_lock);
1842                         fmr->fmr_frd = NULL;
1843                 }
1844         }
1845         fmr->fmr_pool = NULL;
1846
1847         spin_lock(&fps->fps_lock);
1848         fpo->fpo_map_count--;   /* decref the pool */
1849
1850         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1851                 /* the first pool is persistent */
1852                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1853                         continue;
1854
1855                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1856                         list_move(&fpo->fpo_list, &zombies);
1857                         fps->fps_version++;
1858                 }
1859         }
1860         spin_unlock(&fps->fps_lock);
1861
1862         if (!list_empty(&zombies))
1863                 kiblnd_destroy_fmr_pool_list(&zombies);
1864 }
1865
1866 int
1867 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
1868                     __u32 nob, __u64 iov, kib_fmr_t *fmr, bool *is_fastreg)
1869 {
1870         kib_fmr_pool_t *fpo;
1871         __u64 *pages = tx->tx_pages;
1872         __u64 version;
1873         bool is_rx = (rd != tx->tx_rd);
1874         bool tx_pages_mapped = 0;
1875         int npages = 0;
1876         int rc;
1877
1878 again:
1879         spin_lock(&fps->fps_lock);
1880         version = fps->fps_version;
1881         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1882                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1883                 fpo->fpo_map_count++;
1884
1885                 if (fpo->fpo_is_fmr) {
1886                         struct ib_pool_fmr *pfmr;
1887
1888                         *is_fastreg = 0;
1889                         spin_unlock(&fps->fps_lock);
1890
1891                         if (!tx_pages_mapped) {
1892                                 npages = kiblnd_map_tx_pages(tx, rd);
1893                                 tx_pages_mapped = 1;
1894                         }
1895
1896                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1897                                                     pages, npages, iov);
1898                         if (likely(!IS_ERR(pfmr))) {
1899                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1900                                                       : pfmr->fmr->lkey;
1901                                 fmr->fmr_frd  = NULL;
1902                                 fmr->fmr_pfmr = pfmr;
1903                                 fmr->fmr_pool = fpo;
1904                                 return 0;
1905                         }
1906                         rc = PTR_ERR(pfmr);
1907                 } else {
1908                         *is_fastreg = 1;
1909                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1910                                 struct kib_fast_reg_descriptor *frd;
1911 #ifdef HAVE_IB_MAP_MR_SG
1912                                 struct ib_reg_wr *wr;
1913                                 int n;
1914 #else
1915                                 struct ib_rdma_wr *wr;
1916                                 struct ib_fast_reg_page_list *frpl;
1917 #endif
1918                                 struct ib_mr *mr;
1919
1920                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1921                                                         struct kib_fast_reg_descriptor,
1922                                                         frd_list);
1923                                 list_del(&frd->frd_list);
1924                                 spin_unlock(&fps->fps_lock);
1925
1926 #ifndef HAVE_IB_MAP_MR_SG
1927                                 frpl = frd->frd_frpl;
1928 #endif
1929                                 mr   = frd->frd_mr;
1930
1931                                 if (!frd->frd_valid) {
1932                                         struct ib_rdma_wr *inv_wr;
1933                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1934
1935                                         inv_wr = &frd->frd_inv_wr;
1936                                         memset(inv_wr, 0, sizeof(*inv_wr));
1937
1938                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1939                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1940                                         inv_wr->wr.ex.invalidate_rkey = key;
1941
1942                                         /* Bump the key */
1943                                         key = ib_inc_rkey(key);
1944                                         ib_update_fast_reg_key(mr, key);
1945                                 }
1946
1947 #ifdef HAVE_IB_MAP_MR_SG
1948 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1949                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1950                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1951 #else
1952                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1953                                                  tx->tx_nfrags, PAGE_SIZE);
1954 #endif
1955                                 if (unlikely(n != tx->tx_nfrags)) {
1956                                         CERROR("Failed to map mr %d/%d "
1957                                                "elements\n", n, tx->tx_nfrags);
1958                                         return n < 0 ? n : -EINVAL;
1959                                 }
1960
1961                                 wr = &frd->frd_fastreg_wr;
1962                                 memset(wr, 0, sizeof(*wr));
1963
1964                                 wr->wr.opcode = IB_WR_REG_MR;
1965                                 wr->wr.wr_id  = IBLND_WID_MR;
1966                                 wr->wr.num_sge = 0;
1967                                 wr->wr.send_flags = 0;
1968                                 wr->mr = mr;
1969                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1970                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1971                                               IB_ACCESS_REMOTE_WRITE);
1972 #else
1973                                 if (!tx_pages_mapped) {
1974                                         npages = kiblnd_map_tx_pages(tx, rd);
1975                                         tx_pages_mapped = 1;
1976                                 }
1977
1978                                 LASSERT(npages <= frpl->max_page_list_len);
1979                                 memcpy(frpl->page_list, pages,
1980                                         sizeof(*pages) * npages);
1981
1982                                 /* Prepare FastReg WR */
1983                                 wr = &frd->frd_fastreg_wr;
1984                                 memset(wr, 0, sizeof(*wr));
1985
1986                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1987                                 wr->wr.wr_id  = IBLND_WID_MR;
1988
1989                                 wr->wr.wr.fast_reg.iova_start = iov;
1990                                 wr->wr.wr.fast_reg.page_list  = frpl;
1991                                 wr->wr.wr.fast_reg.page_list_len = npages;
1992                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1993                                 wr->wr.wr.fast_reg.length = nob;
1994                                 wr->wr.wr.fast_reg.rkey =
1995                                                 is_rx ? mr->rkey : mr->lkey;
1996                                 wr->wr.wr.fast_reg.access_flags =
1997                                                 (IB_ACCESS_LOCAL_WRITE |
1998                                                  IB_ACCESS_REMOTE_WRITE);
1999 #endif
2000
2001                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
2002                                 fmr->fmr_frd  = frd;
2003                                 fmr->fmr_pfmr = NULL;
2004                                 fmr->fmr_pool = fpo;
2005                                 return 0;
2006                         }
2007                         spin_unlock(&fps->fps_lock);
2008                         rc = -EAGAIN;
2009                 }
2010
2011                 spin_lock(&fps->fps_lock);
2012                 fpo->fpo_map_count--;
2013                 if (rc != -EAGAIN) {
2014                         spin_unlock(&fps->fps_lock);
2015                         return rc;
2016                 }
2017
2018                 /* EAGAIN and ... */
2019                 if (version != fps->fps_version) {
2020                         spin_unlock(&fps->fps_lock);
2021                         goto again;
2022                 }
2023         }
2024
2025         if (fps->fps_increasing) {
2026                 spin_unlock(&fps->fps_lock);
2027                 CDEBUG(D_NET, "Another thread is allocating new "
2028                        "FMR pool, waiting for her to complete\n");
2029                 schedule();
2030                 goto again;
2031
2032         }
2033
2034         if (cfs_time_before(cfs_time_current(), fps->fps_next_retry)) {
2035                 /* someone failed recently */
2036                 spin_unlock(&fps->fps_lock);
2037                 return -EAGAIN;
2038         }
2039
2040         fps->fps_increasing = 1;
2041         spin_unlock(&fps->fps_lock);
2042
2043         CDEBUG(D_NET, "Allocate new FMR pool\n");
2044         rc = kiblnd_create_fmr_pool(fps, &fpo);
2045         spin_lock(&fps->fps_lock);
2046         fps->fps_increasing = 0;
2047         if (rc == 0) {
2048                 fps->fps_version++;
2049                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2050         } else {
2051                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2052         }
2053         spin_unlock(&fps->fps_lock);
2054
2055         goto again;
2056 }
2057
2058 static void
2059 kiblnd_fini_pool(kib_pool_t *pool)
2060 {
2061         LASSERT(list_empty(&pool->po_free_list));
2062         LASSERT(pool->po_allocated == 0);
2063
2064         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2065 }
2066
2067 static void
2068 kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
2069 {
2070         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2071
2072         memset(pool, 0, sizeof(kib_pool_t));
2073         INIT_LIST_HEAD(&pool->po_free_list);
2074         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2075         pool->po_owner    = ps;
2076         pool->po_size     = size;
2077 }
2078
2079 static void
2080 kiblnd_destroy_pool_list(struct list_head *head)
2081 {
2082         kib_pool_t *pool;
2083
2084         while (!list_empty(head)) {
2085                 pool = list_entry(head->next, kib_pool_t, po_list);
2086                 list_del(&pool->po_list);
2087
2088                 LASSERT(pool->po_owner != NULL);
2089                 pool->po_owner->ps_pool_destroy(pool);
2090         }
2091 }
2092
2093 static void
2094 kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
2095 {
2096         if (ps->ps_net == NULL) /* intialized? */
2097                 return;
2098
2099         spin_lock(&ps->ps_lock);
2100         while (!list_empty(&ps->ps_pool_list)) {
2101                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
2102                                             kib_pool_t, po_list);
2103                 po->po_failed = 1;
2104                 list_del(&po->po_list);
2105                 if (po->po_allocated == 0)
2106                         list_add(&po->po_list, zombies);
2107                 else
2108                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2109         }
2110         spin_unlock(&ps->ps_lock);
2111 }
2112
2113 static void
2114 kiblnd_fini_poolset(kib_poolset_t *ps)
2115 {
2116         if (ps->ps_net != NULL) { /* initialized? */
2117                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2118                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2119         }
2120 }
2121
2122 static int
2123 kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
2124                     kib_net_t *net, char *name, int size,
2125                     kib_ps_pool_create_t po_create,
2126                     kib_ps_pool_destroy_t po_destroy,
2127                     kib_ps_node_init_t nd_init,
2128                     kib_ps_node_fini_t nd_fini)
2129 {
2130         kib_pool_t      *pool;
2131         int             rc;
2132
2133         memset(ps, 0, sizeof(kib_poolset_t));
2134
2135         ps->ps_cpt          = cpt;
2136         ps->ps_net          = net;
2137         ps->ps_pool_create  = po_create;
2138         ps->ps_pool_destroy = po_destroy;
2139         ps->ps_node_init    = nd_init;
2140         ps->ps_node_fini    = nd_fini;
2141         ps->ps_pool_size    = size;
2142         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2143             >= sizeof(ps->ps_name))
2144                 return -E2BIG;
2145         spin_lock_init(&ps->ps_lock);
2146         INIT_LIST_HEAD(&ps->ps_pool_list);
2147         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2148
2149         rc = ps->ps_pool_create(ps, size, &pool);
2150         if (rc == 0)
2151                 list_add(&pool->po_list, &ps->ps_pool_list);
2152         else
2153                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2154
2155         return rc;
2156 }
2157
2158 static int
2159 kiblnd_pool_is_idle(kib_pool_t *pool, cfs_time_t now)
2160 {
2161         if (pool->po_allocated != 0) /* still in use */
2162                 return 0;
2163         if (pool->po_failed)
2164                 return 1;
2165         return cfs_time_aftereq(now, pool->po_deadline);
2166 }
2167
2168 void
2169 kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
2170 {
2171         struct list_head zombies = LIST_HEAD_INIT(zombies);
2172         kib_poolset_t   *ps = pool->po_owner;
2173         kib_pool_t      *tmp;
2174         cfs_time_t       now = cfs_time_current();
2175
2176         spin_lock(&ps->ps_lock);
2177
2178         if (ps->ps_node_fini != NULL)
2179                 ps->ps_node_fini(pool, node);
2180
2181         LASSERT(pool->po_allocated > 0);
2182         list_add(node, &pool->po_free_list);
2183         pool->po_allocated--;
2184
2185         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2186                 /* the first pool is persistent */
2187                 if (ps->ps_pool_list.next == &pool->po_list)
2188                         continue;
2189
2190                 if (kiblnd_pool_is_idle(pool, now))
2191                         list_move(&pool->po_list, &zombies);
2192         }
2193         spin_unlock(&ps->ps_lock);
2194
2195         if (!list_empty(&zombies))
2196                 kiblnd_destroy_pool_list(&zombies);
2197 }
2198
2199 struct list_head *
2200 kiblnd_pool_alloc_node(kib_poolset_t *ps)
2201 {
2202         struct list_head        *node;
2203         kib_pool_t              *pool;
2204         int                     rc;
2205         unsigned int            interval = 1;
2206         cfs_time_t              time_before;
2207         unsigned int            trips = 0;
2208
2209 again:
2210         spin_lock(&ps->ps_lock);
2211         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2212                 if (list_empty(&pool->po_free_list))
2213                         continue;
2214
2215                 pool->po_allocated++;
2216                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
2217                 node = pool->po_free_list.next;
2218                 list_del(node);
2219
2220                 if (ps->ps_node_init != NULL) {
2221                         /* still hold the lock */
2222                         ps->ps_node_init(pool, node);
2223                 }
2224                 spin_unlock(&ps->ps_lock);
2225                 return node;
2226         }
2227
2228         /* no available tx pool and ... */
2229         if (ps->ps_increasing) {
2230                 /* another thread is allocating a new pool */
2231                 spin_unlock(&ps->ps_lock);
2232                 trips++;
2233                 CDEBUG(D_NET, "Another thread is allocating new "
2234                        "%s pool, waiting %d HZs for her to complete."
2235                        "trips = %d\n",
2236                        ps->ps_name, interval, trips);
2237
2238                 set_current_state(TASK_INTERRUPTIBLE);
2239                 schedule_timeout(interval);
2240                 if (interval < cfs_time_seconds(1))
2241                         interval *= 2;
2242
2243                 goto again;
2244         }
2245
2246         if (cfs_time_before(cfs_time_current(), ps->ps_next_retry)) {
2247                 /* someone failed recently */
2248                 spin_unlock(&ps->ps_lock);
2249                 return NULL;
2250         }
2251
2252         ps->ps_increasing = 1;
2253         spin_unlock(&ps->ps_lock);
2254
2255         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2256         time_before = cfs_time_current();
2257         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2258         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
2259                cfs_time_current() - time_before);
2260
2261         spin_lock(&ps->ps_lock);
2262         ps->ps_increasing = 0;
2263         if (rc == 0) {
2264                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2265         } else {
2266                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
2267                 CERROR("Can't allocate new %s pool because out of memory\n",
2268                        ps->ps_name);
2269         }
2270         spin_unlock(&ps->ps_lock);
2271
2272         goto again;
2273 }
2274
2275 static void
2276 kiblnd_destroy_tx_pool(kib_pool_t *pool)
2277 {
2278         kib_tx_pool_t  *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
2279         int             i;
2280
2281         LASSERT (pool->po_allocated == 0);
2282
2283         if (tpo->tpo_tx_pages != NULL) {
2284                 kiblnd_unmap_tx_pool(tpo);
2285                 kiblnd_free_pages(tpo->tpo_tx_pages);
2286         }
2287
2288         if (tpo->tpo_tx_descs == NULL)
2289                 goto out;
2290
2291         for (i = 0; i < pool->po_size; i++) {
2292                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2293                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2294
2295                 list_del(&tx->tx_list);
2296                 if (tx->tx_pages != NULL)
2297                         LIBCFS_FREE(tx->tx_pages,
2298                                     LNET_MAX_IOV *
2299                                     sizeof(*tx->tx_pages));
2300                 if (tx->tx_frags != NULL)
2301                         LIBCFS_FREE(tx->tx_frags,
2302                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2303                                     sizeof(*tx->tx_frags));
2304                 if (tx->tx_wrq != NULL)
2305                         LIBCFS_FREE(tx->tx_wrq,
2306                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2307                                     sizeof(*tx->tx_wrq));
2308                 if (tx->tx_sge != NULL)
2309                         LIBCFS_FREE(tx->tx_sge,
2310                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2311                                     sizeof(*tx->tx_sge));
2312                 if (tx->tx_rd != NULL)
2313                         LIBCFS_FREE(tx->tx_rd,
2314                                     offsetof(kib_rdma_desc_t,
2315                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2316         }
2317
2318         LIBCFS_FREE(tpo->tpo_tx_descs,
2319                     pool->po_size * sizeof(kib_tx_t));
2320 out:
2321         kiblnd_fini_pool(pool);
2322         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2323 }
2324
2325 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2326 {
2327         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2328         int ntx;
2329
2330         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2331         ntx = tunables->lnd_ntx / ncpts;
2332
2333         return max(IBLND_TX_POOL, ntx);
2334 }
2335
2336 static int
2337 kiblnd_create_tx_pool(kib_poolset_t *ps, int size, kib_pool_t **pp_po)
2338 {
2339         int            i;
2340         int            npg;
2341         kib_pool_t    *pool;
2342         kib_tx_pool_t *tpo;
2343
2344         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2345         if (tpo == NULL) {
2346                 CERROR("Failed to allocate TX pool\n");
2347                 return -ENOMEM;
2348         }
2349
2350         pool = &tpo->tpo_pool;
2351         kiblnd_init_pool(ps, pool, size);
2352         tpo->tpo_tx_descs = NULL;
2353         tpo->tpo_tx_pages = NULL;
2354
2355         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2356         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2357                 CERROR("Can't allocate tx pages: %d\n", npg);
2358                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
2359                 return -ENOMEM;
2360         }
2361
2362         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2363                          size * sizeof(kib_tx_t));
2364         if (tpo->tpo_tx_descs == NULL) {
2365                 CERROR("Can't allocate %d tx descriptors\n", size);
2366                 ps->ps_pool_destroy(pool);
2367                 return -ENOMEM;
2368         }
2369
2370         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
2371
2372         for (i = 0; i < size; i++) {
2373                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
2374                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2375
2376                 tx->tx_pool = tpo;
2377                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2378                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2379                                          lnet_cpt_table(), ps->ps_cpt,
2380                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2381                         if (tx->tx_pages == NULL)
2382                                 break;
2383                 }
2384
2385                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2386                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2387                                  sizeof(*tx->tx_frags));
2388                 if (tx->tx_frags == NULL)
2389                         break;
2390
2391                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2392
2393                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2394                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2395                                  sizeof(*tx->tx_wrq));
2396                 if (tx->tx_wrq == NULL)
2397                         break;
2398
2399                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2400                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2401                                  sizeof(*tx->tx_sge));
2402                 if (tx->tx_sge == NULL)
2403                         break;
2404
2405                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2406                                  offsetof(kib_rdma_desc_t,
2407                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2408                 if (tx->tx_rd == NULL)
2409                         break;
2410         }
2411
2412         if (i == size) {
2413                 kiblnd_map_tx_pool(tpo);
2414                 *pp_po = pool;
2415                 return 0;
2416         }
2417
2418         ps->ps_pool_destroy(pool);
2419         return -ENOMEM;
2420 }
2421
2422 static void
2423 kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
2424 {
2425         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
2426                                              tps_poolset);
2427         kib_tx_t         *tx  = list_entry(node, kib_tx_t, tx_list);
2428
2429         tx->tx_cookie = tps->tps_next_tx_cookie++;
2430 }
2431
2432 static void
2433 kiblnd_net_fini_pools(kib_net_t *net)
2434 {
2435         int     i;
2436
2437         cfs_cpt_for_each(i, lnet_cpt_table()) {
2438                 kib_tx_poolset_t        *tps;
2439                 kib_fmr_poolset_t       *fps;
2440
2441                 if (net->ibn_tx_ps != NULL) {
2442                         tps = net->ibn_tx_ps[i];
2443                         kiblnd_fini_poolset(&tps->tps_poolset);
2444                 }
2445
2446                 if (net->ibn_fmr_ps != NULL) {
2447                         fps = net->ibn_fmr_ps[i];
2448                         kiblnd_fini_fmr_poolset(fps);
2449                 }
2450         }
2451
2452         if (net->ibn_tx_ps != NULL) {
2453                 cfs_percpt_free(net->ibn_tx_ps);
2454                 net->ibn_tx_ps = NULL;
2455         }
2456
2457         if (net->ibn_fmr_ps != NULL) {
2458                 cfs_percpt_free(net->ibn_fmr_ps);
2459                 net->ibn_fmr_ps = NULL;
2460         }
2461 }
2462
2463 static int
2464 kiblnd_net_init_pools(kib_net_t *net, struct lnet_ni *ni, __u32 *cpts,
2465                       int ncpts)
2466 {
2467         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2468 #ifdef HAVE_IB_GET_DMA_MR
2469         unsigned long   flags;
2470 #endif
2471         int             cpt;
2472         int             rc;
2473         int             i;
2474
2475         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2476
2477 #ifdef HAVE_IB_GET_DMA_MR
2478         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2479         if (tunables->lnd_map_on_demand == 0) {
2480                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2481                                            flags);
2482                 goto create_tx_pool;
2483         }
2484
2485         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2486 #endif
2487
2488         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2489                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2490                        tunables->lnd_fmr_pool_size,
2491                        tunables->lnd_ntx / 4);
2492                 rc = -EINVAL;
2493                 goto failed;
2494         }
2495
2496         /* TX pool must be created later than FMR, see LU-2268
2497          * for details */
2498         LASSERT(net->ibn_tx_ps == NULL);
2499
2500         /* premapping can fail if ibd_nmr > 1, so we always create
2501          * FMR pool and map-on-demand if premapping failed */
2502
2503         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2504                                            sizeof(kib_fmr_poolset_t));
2505         if (net->ibn_fmr_ps == NULL) {
2506                 CERROR("Failed to allocate FMR pool array\n");
2507                 rc = -ENOMEM;
2508                 goto failed;
2509         }
2510
2511         for (i = 0; i < ncpts; i++) {
2512                 cpt = (cpts == NULL) ? i : cpts[i];
2513                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2514                                              net, tunables);
2515                 if (rc != 0) {
2516                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2517                                cpt, rc);
2518                         goto failed;
2519                 }
2520         }
2521
2522         if (i > 0)
2523                 LASSERT(i == ncpts);
2524
2525 #ifdef HAVE_IB_GET_DMA_MR
2526  create_tx_pool:
2527 #endif
2528         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2529                                           sizeof(kib_tx_poolset_t));
2530         if (net->ibn_tx_ps == NULL) {
2531                 CERROR("Failed to allocate tx pool array\n");
2532                 rc = -ENOMEM;
2533                 goto failed;
2534         }
2535
2536         for (i = 0; i < ncpts; i++) {
2537                 cpt = (cpts == NULL) ? i : cpts[i];
2538                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2539                                          cpt, net, "TX",
2540                                          kiblnd_tx_pool_size(ni, ncpts),
2541                                          kiblnd_create_tx_pool,
2542                                          kiblnd_destroy_tx_pool,
2543                                          kiblnd_tx_init, NULL);
2544                 if (rc != 0) {
2545                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2546                                cpt, rc);
2547                         goto failed;
2548                 }
2549         }
2550
2551         return 0;
2552  failed:
2553         kiblnd_net_fini_pools(net);
2554         LASSERT(rc != 0);
2555         return rc;
2556 }
2557
2558 static int
2559 kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2560 {
2561 #ifndef HAVE_IB_DEVICE_ATTRS
2562         struct ib_device_attr *attr;
2563         int                    rc;
2564 #endif
2565
2566         /* It's safe to assume a HCA can handle a page size
2567          * matching that of the native system */
2568         hdev->ibh_page_shift = PAGE_SHIFT;
2569         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2570         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2571
2572 #ifdef HAVE_IB_DEVICE_ATTRS
2573         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2574 #else
2575         LIBCFS_ALLOC(attr, sizeof(*attr));
2576         if (attr == NULL) {
2577                 CERROR("Out of memory\n");
2578                 return -ENOMEM;
2579         }
2580
2581         rc = ib_query_device(hdev->ibh_ibdev, attr);
2582         if (rc == 0)
2583                 hdev->ibh_mr_size = attr->max_mr_size;
2584
2585         LIBCFS_FREE(attr, sizeof(*attr));
2586
2587         if (rc != 0) {
2588                 CERROR("Failed to query IB device: %d\n", rc);
2589                 return rc;
2590         }
2591 #endif
2592
2593         if (hdev->ibh_mr_size == ~0ULL) {
2594                 hdev->ibh_mr_shift = 64;
2595                 return 0;
2596         }
2597
2598         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2599         return -EINVAL;
2600 }
2601
2602 #ifdef HAVE_IB_GET_DMA_MR
2603 static void
2604 kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2605 {
2606         if (hdev->ibh_mrs == NULL)
2607                 return;
2608
2609         ib_dereg_mr(hdev->ibh_mrs);
2610
2611         hdev->ibh_mrs = NULL;
2612 }
2613 #endif
2614
2615 void
2616 kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2617 {
2618 #ifdef HAVE_IB_GET_DMA_MR
2619         kiblnd_hdev_cleanup_mrs(hdev);
2620 #endif
2621
2622         if (hdev->ibh_pd != NULL)
2623                 ib_dealloc_pd(hdev->ibh_pd);
2624
2625         if (hdev->ibh_cmid != NULL)
2626                 rdma_destroy_id(hdev->ibh_cmid);
2627
2628         LIBCFS_FREE(hdev, sizeof(*hdev));
2629 }
2630
2631 #ifdef HAVE_IB_GET_DMA_MR
2632 static int
2633 kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2634 {
2635         struct ib_mr *mr;
2636         int           rc;
2637         int           acflags = IB_ACCESS_LOCAL_WRITE |
2638                                 IB_ACCESS_REMOTE_WRITE;
2639
2640         rc = kiblnd_hdev_get_attr(hdev);
2641         if (rc != 0)
2642                 return rc;
2643
2644         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2645         if (IS_ERR(mr)) {
2646                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2647                 kiblnd_hdev_cleanup_mrs(hdev);
2648                 return PTR_ERR(mr);
2649         }
2650
2651         hdev->ibh_mrs = mr;
2652
2653         return 0;
2654 }
2655 #endif
2656
2657 static int
2658 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2659 {       /* DUMMY */
2660         return 0;
2661 }
2662
2663 static int
2664 kiblnd_dev_need_failover(kib_dev_t *dev)
2665 {
2666         struct rdma_cm_id  *cmid;
2667         struct sockaddr_in  srcaddr;
2668         struct sockaddr_in  dstaddr;
2669         int                 rc;
2670
2671         if (dev->ibd_hdev == NULL || /* initializing */
2672             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2673             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2674                 return 1;
2675
2676         /* XXX: it's UGLY, but I don't have better way to find
2677          * ib-bonding HCA failover because:
2678          *
2679          * a. no reliable CM event for HCA failover...
2680          * b. no OFED API to get ib_device for current net_device...
2681          *
2682          * We have only two choices at this point:
2683          *
2684          * a. rdma_bind_addr(), it will conflict with listener cmid
2685          * b. rdma_resolve_addr() to zero addr */
2686         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2687                                      IB_QPT_RC);
2688         if (IS_ERR(cmid)) {
2689                 rc = PTR_ERR(cmid);
2690                 CERROR("Failed to create cmid for failover: %d\n", rc);
2691                 return rc;
2692         }
2693
2694         memset(&srcaddr, 0, sizeof(srcaddr));
2695         srcaddr.sin_family      = AF_INET;
2696         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2697
2698         memset(&dstaddr, 0, sizeof(dstaddr));
2699         dstaddr.sin_family = AF_INET;
2700         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2701                                (struct sockaddr *)&dstaddr, 1);
2702         if (rc != 0 || cmid->device == NULL) {
2703                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2704                        dev->ibd_ifname, &dev->ibd_ifip,
2705                        cmid->device, rc);
2706                 rdma_destroy_id(cmid);
2707                 return rc;
2708         }
2709
2710         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2711         rdma_destroy_id(cmid);
2712         return rc;
2713 }
2714
2715 int
2716 kiblnd_dev_failover(kib_dev_t *dev)
2717 {
2718         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2719         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2720         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2721         struct rdma_cm_id  *cmid  = NULL;
2722         kib_hca_dev_t      *hdev  = NULL;
2723         kib_hca_dev_t      *old;
2724         struct ib_pd       *pd;
2725         kib_net_t          *net;
2726         struct sockaddr_in  addr;
2727         unsigned long       flags;
2728         int                 rc = 0;
2729         int                 i;
2730
2731         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2732                  dev->ibd_can_failover ||
2733                  dev->ibd_hdev == NULL);
2734
2735         rc = kiblnd_dev_need_failover(dev);
2736         if (rc <= 0)
2737                 goto out;
2738
2739         if (dev->ibd_hdev != NULL &&
2740             dev->ibd_hdev->ibh_cmid != NULL) {
2741                 /* XXX it's not good to close old listener at here,
2742                  * because we can fail to create new listener.
2743                  * But we have to close it now, otherwise rdma_bind_addr
2744                  * will return EADDRINUSE... How crap! */
2745                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2746
2747                 cmid = dev->ibd_hdev->ibh_cmid;
2748                 /* make next schedule of kiblnd_dev_need_failover()
2749                  * return 1 for me */
2750                 dev->ibd_hdev->ibh_cmid  = NULL;
2751                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2752
2753                 rdma_destroy_id(cmid);
2754         }
2755
2756         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2757                                      IB_QPT_RC);
2758         if (IS_ERR(cmid)) {
2759                 rc = PTR_ERR(cmid);
2760                 CERROR("Failed to create cmid for failover: %d\n", rc);
2761                 goto out;
2762         }
2763
2764         memset(&addr, 0, sizeof(addr));
2765         addr.sin_family      = AF_INET;
2766         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2767         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2768
2769         /* Bind to failover device or port */
2770         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2771         if (rc != 0 || cmid->device == NULL) {
2772                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2773                        dev->ibd_ifname, &dev->ibd_ifip,
2774                        cmid->device, rc);
2775                 rdma_destroy_id(cmid);
2776                 goto out;
2777         }
2778
2779         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2780         if (hdev == NULL) {
2781                 CERROR("Failed to allocate kib_hca_dev\n");
2782                 rdma_destroy_id(cmid);
2783                 rc = -ENOMEM;
2784                 goto out;
2785         }
2786
2787         atomic_set(&hdev->ibh_ref, 1);
2788         hdev->ibh_dev   = dev;
2789         hdev->ibh_cmid  = cmid;
2790         hdev->ibh_ibdev = cmid->device;
2791
2792 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2793         pd = ib_alloc_pd(cmid->device, 0);
2794 #else
2795         pd = ib_alloc_pd(cmid->device);
2796 #endif
2797         if (IS_ERR(pd)) {
2798                 rc = PTR_ERR(pd);
2799                 CERROR("Can't allocate PD: %d\n", rc);
2800                 goto out;
2801         }
2802
2803         hdev->ibh_pd = pd;
2804
2805         rc = rdma_listen(cmid, 0);
2806         if (rc != 0) {
2807                 CERROR("Can't start new listener: %d\n", rc);
2808                 goto out;
2809         }
2810
2811 #ifdef HAVE_IB_GET_DMA_MR
2812         rc = kiblnd_hdev_setup_mrs(hdev);
2813         if (rc != 0) {
2814                 CERROR("Can't setup device: %d\n", rc);
2815                 goto out;
2816         }
2817 #else
2818         rc = kiblnd_hdev_get_attr(hdev);
2819         if (rc != 0) {
2820                 CERROR("Can't get device attributes: %d\n", rc);
2821                 goto out;
2822         }
2823 #endif
2824
2825         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2826
2827         old = dev->ibd_hdev;
2828         dev->ibd_hdev = hdev;   /* take over the refcount */
2829         hdev = old;
2830
2831         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2832                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2833                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2834                                             &zombie_tpo);
2835
2836                         if (net->ibn_fmr_ps != NULL)
2837                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2838                                                         &zombie_fpo);
2839                 }
2840         }
2841
2842         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2843  out:
2844         if (!list_empty(&zombie_tpo))
2845                 kiblnd_destroy_pool_list(&zombie_tpo);
2846         if (!list_empty(&zombie_ppo))
2847                 kiblnd_destroy_pool_list(&zombie_ppo);
2848         if (!list_empty(&zombie_fpo))
2849                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2850         if (hdev != NULL)
2851                 kiblnd_hdev_decref(hdev);
2852
2853         if (rc != 0)
2854                 dev->ibd_failed_failover++;
2855         else
2856                 dev->ibd_failed_failover = 0;
2857
2858         return rc;
2859 }
2860
2861 void
2862 kiblnd_destroy_dev (kib_dev_t *dev)
2863 {
2864         LASSERT (dev->ibd_nnets == 0);
2865         LASSERT(list_empty(&dev->ibd_nets));
2866
2867         list_del(&dev->ibd_fail_list);
2868         list_del(&dev->ibd_list);
2869
2870         if (dev->ibd_hdev != NULL)
2871                 kiblnd_hdev_decref(dev->ibd_hdev);
2872
2873         LIBCFS_FREE(dev, sizeof(*dev));
2874 }
2875
2876 static kib_dev_t *
2877 kiblnd_create_dev(char *ifname)
2878 {
2879         struct net_device *netdev;
2880         kib_dev_t         *dev;
2881         __u32              netmask;
2882         __u32              ip;
2883         int                up;
2884         int                rc;
2885
2886         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2887         if (rc != 0) {
2888                 CERROR("Can't query IPoIB interface %s: %d\n",
2889                        ifname, rc);
2890                 return NULL;
2891         }
2892
2893         if (!up) {
2894                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2895                 return NULL;
2896         }
2897
2898         LIBCFS_ALLOC(dev, sizeof(*dev));
2899         if (dev == NULL)
2900                 return NULL;
2901
2902         netdev = dev_get_by_name(&init_net, ifname);
2903         if (netdev == NULL) {
2904                 dev->ibd_can_failover = 0;
2905         } else {
2906                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2907                 dev_put(netdev);
2908         }
2909
2910         INIT_LIST_HEAD(&dev->ibd_nets);
2911         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2912         INIT_LIST_HEAD(&dev->ibd_fail_list);
2913         dev->ibd_ifip = ip;
2914         strcpy(&dev->ibd_ifname[0], ifname);
2915
2916         /* initialize the device */
2917         rc = kiblnd_dev_failover(dev);
2918         if (rc != 0) {
2919                 CERROR("Can't initialize device: %d\n", rc);
2920                 LIBCFS_FREE(dev, sizeof(*dev));
2921                 return NULL;
2922         }
2923
2924         list_add_tail(&dev->ibd_list,
2925                           &kiblnd_data.kib_devs);
2926         return dev;
2927 }
2928
2929 static void
2930 kiblnd_base_shutdown(void)
2931 {
2932         struct kib_sched_info   *sched;
2933         int                     i;
2934
2935         LASSERT(list_empty(&kiblnd_data.kib_devs));
2936
2937         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2938                atomic_read(&libcfs_kmemory));
2939
2940         switch (kiblnd_data.kib_init) {
2941         default:
2942                 LBUG();
2943
2944         case IBLND_INIT_ALL:
2945         case IBLND_INIT_DATA:
2946                 LASSERT (kiblnd_data.kib_peers != NULL);
2947                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2948                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2949                 }
2950                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2951                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2952                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2953                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2954
2955                 /* flag threads to terminate; wake and wait for them to die */
2956                 kiblnd_data.kib_shutdown = 1;
2957
2958                 /* NB: we really want to stop scheduler threads net by net
2959                  * instead of the whole module, this should be improved
2960                  * with dynamic configuration LNet */
2961                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2962                         wake_up_all(&sched->ibs_waitq);
2963
2964                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2965                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2966
2967                 i = 2;
2968                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2969                         i++;
2970                         /* power of 2? */
2971                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2972                                "Waiting for %d threads to terminate\n",
2973                                atomic_read(&kiblnd_data.kib_nthreads));
2974                         set_current_state(TASK_UNINTERRUPTIBLE);
2975                         schedule_timeout(cfs_time_seconds(1));
2976                 }
2977
2978                 /* fall through */
2979
2980         case IBLND_INIT_NOTHING:
2981                 break;
2982         }
2983
2984         if (kiblnd_data.kib_peers != NULL) {
2985                 LIBCFS_FREE(kiblnd_data.kib_peers,
2986                             sizeof(struct list_head) *
2987                             kiblnd_data.kib_peer_hash_size);
2988         }
2989
2990         if (kiblnd_data.kib_scheds != NULL)
2991                 cfs_percpt_free(kiblnd_data.kib_scheds);
2992
2993         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2994                atomic_read(&libcfs_kmemory));
2995
2996         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2997         module_put(THIS_MODULE);
2998 }
2999
3000 static void
3001 kiblnd_shutdown(struct lnet_ni *ni)
3002 {
3003         kib_net_t        *net = ni->ni_data;
3004         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
3005         int               i;
3006         unsigned long     flags;
3007
3008         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
3009
3010         if (net == NULL)
3011                 goto out;
3012
3013         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
3014                atomic_read(&libcfs_kmemory));
3015
3016         write_lock_irqsave(g_lock, flags);
3017         net->ibn_shutdown = 1;
3018         write_unlock_irqrestore(g_lock, flags);
3019
3020         switch (net->ibn_init) {
3021         default:
3022                 LBUG();
3023
3024         case IBLND_INIT_ALL:
3025                 /* nuke all existing peers within this net */
3026                 kiblnd_del_peer(ni, LNET_NID_ANY);
3027
3028                 /* Wait for all peer_ni state to clean up */
3029                 i = 2;
3030                 while (atomic_read(&net->ibn_npeers) != 0) {
3031                         i++;
3032                         /* power of 2? */
3033                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
3034                                "%s: waiting for %d peers to disconnect\n",
3035                                libcfs_nid2str(ni->ni_nid),
3036                                atomic_read(&net->ibn_npeers));
3037                         set_current_state(TASK_UNINTERRUPTIBLE);
3038                         schedule_timeout(cfs_time_seconds(1));
3039                 }
3040
3041                 kiblnd_net_fini_pools(net);
3042
3043                 write_lock_irqsave(g_lock, flags);
3044                 LASSERT(net->ibn_dev->ibd_nnets > 0);
3045                 net->ibn_dev->ibd_nnets--;
3046                 list_del(&net->ibn_list);
3047                 write_unlock_irqrestore(g_lock, flags);
3048
3049                 /* fall through */
3050
3051         case IBLND_INIT_NOTHING:
3052                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3053
3054                 if (net->ibn_dev != NULL &&
3055                     net->ibn_dev->ibd_nnets == 0)
3056                         kiblnd_destroy_dev(net->ibn_dev);
3057
3058                 break;
3059         }
3060
3061         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3062                atomic_read(&libcfs_kmemory));
3063
3064         net->ibn_init = IBLND_INIT_NOTHING;
3065         ni->ni_data = NULL;
3066
3067         LIBCFS_FREE(net, sizeof(*net));
3068
3069 out:
3070         if (list_empty(&kiblnd_data.kib_devs))
3071                 kiblnd_base_shutdown();
3072         return;
3073 }
3074
3075 static int
3076 kiblnd_base_startup(void)
3077 {
3078         struct kib_sched_info   *sched;
3079         int                     rc;
3080         int                     i;
3081
3082         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3083
3084         try_module_get(THIS_MODULE);
3085         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3086
3087         rwlock_init(&kiblnd_data.kib_global_lock);
3088
3089         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3090         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3091
3092         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3093         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3094                      sizeof(struct list_head) *
3095                      kiblnd_data.kib_peer_hash_size);
3096         if (kiblnd_data.kib_peers == NULL)
3097                 goto failed;
3098
3099         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3100                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3101
3102         spin_lock_init(&kiblnd_data.kib_connd_lock);
3103         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3104         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3105         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3106         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3107
3108         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3109         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3110
3111         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3112                                                   sizeof(*sched));
3113         if (kiblnd_data.kib_scheds == NULL)
3114                 goto failed;
3115
3116         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3117                 int     nthrs;
3118
3119                 spin_lock_init(&sched->ibs_lock);
3120                 INIT_LIST_HEAD(&sched->ibs_conns);
3121                 init_waitqueue_head(&sched->ibs_waitq);
3122
3123                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3124                 if (*kiblnd_tunables.kib_nscheds > 0) {
3125                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3126                 } else {
3127                         /* max to half of CPUs, another half is reserved for
3128                          * upper layer modules */
3129                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3130                 }
3131
3132                 sched->ibs_nthreads_max = nthrs;
3133                 sched->ibs_cpt = i;
3134         }
3135
3136         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3137
3138         /* lists/ptrs/locks initialised */
3139         kiblnd_data.kib_init = IBLND_INIT_DATA;
3140         /*****************************************************/
3141
3142         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3143         if (rc != 0) {
3144                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3145                 goto failed;
3146         }
3147
3148         if (*kiblnd_tunables.kib_dev_failover != 0)
3149                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
3150                                          "kiblnd_failover");
3151
3152         if (rc != 0) {
3153                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3154                 goto failed;
3155         }
3156
3157         /* flag everything initialised */
3158         kiblnd_data.kib_init = IBLND_INIT_ALL;
3159         /*****************************************************/
3160
3161         return 0;
3162
3163  failed:
3164         kiblnd_base_shutdown();
3165         return -ENETDOWN;
3166 }
3167
3168 static int
3169 kiblnd_start_schedulers(struct kib_sched_info *sched)
3170 {
3171         int     rc = 0;
3172         int     nthrs;
3173         int     i;
3174
3175         if (sched->ibs_nthreads == 0) {
3176                 if (*kiblnd_tunables.kib_nscheds > 0) {
3177                         nthrs = sched->ibs_nthreads_max;
3178                 } else {
3179                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3180                                                sched->ibs_cpt);
3181                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3182                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3183                 }
3184         } else {
3185                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3186                 /* increase one thread if there is new interface */
3187                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3188         }
3189
3190         for (i = 0; i < nthrs; i++) {
3191                 long    id;
3192                 char    name[20];
3193                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3194                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3195                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3196                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3197                 if (rc == 0)
3198                         continue;
3199
3200                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3201                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3202                 break;
3203         }
3204
3205         sched->ibs_nthreads += i;
3206         return rc;
3207 }
3208
3209 static int
3210 kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts, int ncpts)
3211 {
3212         int     cpt;
3213         int     rc;
3214         int     i;
3215
3216         for (i = 0; i < ncpts; i++) {
3217                 struct kib_sched_info *sched;
3218
3219                 cpt = (cpts == NULL) ? i : cpts[i];
3220                 sched = kiblnd_data.kib_scheds[cpt];
3221
3222                 if (!newdev && sched->ibs_nthreads > 0)
3223                         continue;
3224
3225                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3226                 if (rc != 0) {
3227                         CERROR("Failed to start scheduler threads for %s\n",
3228                                dev->ibd_ifname);
3229                         return rc;
3230                 }
3231         }
3232         return 0;
3233 }
3234
3235 static kib_dev_t *
3236 kiblnd_dev_search(char *ifname)
3237 {
3238         kib_dev_t       *alias = NULL;
3239         kib_dev_t       *dev;
3240         char            *colon;
3241         char            *colon2;
3242
3243         colon = strchr(ifname, ':');
3244         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3245                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3246                         return dev;
3247
3248                 if (alias != NULL)
3249                         continue;
3250
3251                 colon2 = strchr(dev->ibd_ifname, ':');
3252                 if (colon != NULL)
3253                         *colon = 0;
3254                 if (colon2 != NULL)
3255                         *colon2 = 0;
3256
3257                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3258                         alias = dev;
3259
3260                 if (colon != NULL)
3261                         *colon = ':';
3262                 if (colon2 != NULL)
3263                         *colon2 = ':';
3264         }
3265         return alias;
3266 }
3267
3268 static int
3269 kiblnd_startup(struct lnet_ni *ni)
3270 {
3271         char                     *ifname;
3272         kib_dev_t                *ibdev = NULL;
3273         kib_net_t                *net;
3274         unsigned long             flags;
3275         int                       rc;
3276         int                       newdev;
3277         int                       node_id;
3278
3279         LASSERT (ni->ni_net->net_lnd == &the_o2iblnd);
3280
3281         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3282                 rc = kiblnd_base_startup();
3283                 if (rc != 0)
3284                         return rc;
3285         }
3286
3287         LIBCFS_ALLOC(net, sizeof(*net));
3288         ni->ni_data = net;
3289         if (net == NULL)
3290                 goto failed;
3291
3292         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3293
3294         kiblnd_tunables_setup(ni);
3295
3296         if (ni->ni_interfaces[0] != NULL) {
3297                 /* Use the IPoIB interface specified in 'networks=' */
3298
3299                 CLASSERT(LNET_NUM_INTERFACES > 1);
3300                 if (ni->ni_interfaces[1] != NULL) {
3301                         CERROR("Multiple interfaces not supported\n");
3302                         goto failed;
3303                 }
3304
3305                 ifname = ni->ni_interfaces[0];
3306         } else {
3307                 ifname = *kiblnd_tunables.kib_default_ipif;
3308         }
3309
3310         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3311                 CERROR("IPoIB interface name too long: %s\n", ifname);
3312                 goto failed;
3313         }
3314
3315         ibdev = kiblnd_dev_search(ifname);
3316
3317         newdev = ibdev == NULL;
3318         /* hmm...create kib_dev even for alias */
3319         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
3320                 ibdev = kiblnd_create_dev(ifname);
3321
3322         if (ibdev == NULL)
3323                 goto failed;
3324
3325         node_id = dev_to_node(ibdev->ibd_hdev->ibh_ibdev->dma_device);
3326         ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
3327
3328         net->ibn_dev = ibdev;
3329         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3330
3331         rc = kiblnd_dev_start_threads(ibdev, newdev,
3332                                       ni->ni_cpts, ni->ni_ncpts);
3333         if (rc != 0)
3334                 goto failed;
3335
3336         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3337         if (rc != 0) {
3338                 CERROR("Failed to initialize NI pools: %d\n", rc);
3339                 goto failed;
3340         }
3341
3342         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3343         ibdev->ibd_nnets++;
3344         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3345         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3346
3347         net->ibn_init = IBLND_INIT_ALL;
3348
3349         return 0;
3350
3351 failed:
3352         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3353                 kiblnd_destroy_dev(ibdev);
3354
3355         kiblnd_shutdown(ni);
3356
3357         CDEBUG(D_NET, "kiblnd_startup failed\n");
3358         return -ENETDOWN;
3359 }
3360
3361 static struct lnet_lnd the_o2iblnd = {
3362         .lnd_type       = O2IBLND,
3363         .lnd_startup    = kiblnd_startup,
3364         .lnd_shutdown   = kiblnd_shutdown,
3365         .lnd_ctl        = kiblnd_ctl,
3366         .lnd_query      = kiblnd_query,
3367         .lnd_send       = kiblnd_send,
3368         .lnd_recv       = kiblnd_recv,
3369 };
3370
3371 static void __exit ko2iblnd_exit(void)
3372 {
3373         lnet_unregister_lnd(&the_o2iblnd);
3374 }
3375
3376 static int __init ko2iblnd_init(void)
3377 {
3378         int rc;
3379
3380         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
3381         CLASSERT(offsetof(kib_msg_t,
3382                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3383                  IBLND_MSG_SIZE);
3384         CLASSERT(offsetof(kib_msg_t,
3385                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3386                  <= IBLND_MSG_SIZE);
3387
3388         rc = kiblnd_tunables_init();
3389         if (rc != 0)
3390                 return rc;
3391
3392         lnet_register_lnd(&the_o2iblnd);
3393
3394         return 0;
3395 }
3396
3397 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3398 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3399 MODULE_VERSION("2.8.0");
3400 MODULE_LICENSE("GPL");
3401
3402 module_init(ko2iblnd_init);
3403 module_exit(ko2iblnd_exit);