lnet/klnds/o2iblnd/o2iblnd.c

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