lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include <asm/page.h>
  38 #include <linux/inetdevice.h>
  39
  40 #include "o2iblnd.h"
  41
  42 static struct lnet_lnd the_o2iblnd;
  43
  44 struct kib_data kiblnd_data;
  45
  46 static __u32
  47 kiblnd_cksum (void *ptr, int nob)
  48 {
  49         char  *c  = ptr;
  50         __u32  sum = 0;
  51
  52         while (nob-- > 0)
  53                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  54
  55         /* ensure I don't return 0 (== no checksum) */
  56         return (sum == 0) ? 1 : sum;
  57 }
  58
  59 static char *
  60 kiblnd_msgtype2str(int type)
  61 {
  62         switch (type) {
  63         case IBLND_MSG_CONNREQ:
  64                 return "CONNREQ";
  65
  66         case IBLND_MSG_CONNACK:
  67                 return "CONNACK";
  68
  69         case IBLND_MSG_NOOP:
  70                 return "NOOP";
  71
  72         case IBLND_MSG_IMMEDIATE:
  73                 return "IMMEDIATE";
  74
  75         case IBLND_MSG_PUT_REQ:
  76                 return "PUT_REQ";
  77
  78         case IBLND_MSG_PUT_NAK:
  79                 return "PUT_NAK";
  80
  81         case IBLND_MSG_PUT_ACK:
  82                 return "PUT_ACK";
  83
  84         case IBLND_MSG_PUT_DONE:
  85                 return "PUT_DONE";
  86
  87         case IBLND_MSG_GET_REQ:
  88                 return "GET_REQ";
  89
  90         case IBLND_MSG_GET_DONE:
  91                 return "GET_DONE";
  92
  93         default:
  94                 return "???";
  95         }
  96 }
  97
  98 static int
  99 kiblnd_msgtype2size(int type)
 100 {
 101         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 102
 103         switch (type) {
 104         case IBLND_MSG_CONNREQ:
 105         case IBLND_MSG_CONNACK:
 106                 return hdr_size + sizeof(struct kib_connparams);
 107
 108         case IBLND_MSG_NOOP:
 109                 return hdr_size;
 110
 111         case IBLND_MSG_IMMEDIATE:
 112                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 113
 114         case IBLND_MSG_PUT_REQ:
 115                 return hdr_size + sizeof(struct kib_putreq_msg);
 116
 117         case IBLND_MSG_PUT_ACK:
 118                 return hdr_size + sizeof(struct kib_putack_msg);
 119
 120         case IBLND_MSG_GET_REQ:
 121                 return hdr_size + sizeof(struct kib_get_msg);
 122
 123         case IBLND_MSG_PUT_NAK:
 124         case IBLND_MSG_PUT_DONE:
 125         case IBLND_MSG_GET_DONE:
 126                 return hdr_size + sizeof(struct kib_completion_msg);
 127         default:
 128                 return -1;
 129         }
 130 }
 131
 132 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
 133 {
 134         struct kib_rdma_desc *rd;
 135         int                nob;
 136         int                n;
 137         int                i;
 138
 139         LASSERT (msg->ibm_type == IBLND_MSG_GET_REQ ||
 140                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 141
 142         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 143                               &msg->ibm_u.get.ibgm_rd :
 144                               &msg->ibm_u.putack.ibpam_rd;
 145
 146         if (flip) {
 147                 __swab32s(&rd->rd_key);
 148                 __swab32s(&rd->rd_nfrags);
 149         }
 150
 151         n = rd->rd_nfrags;
 152
 153         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 154                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 155                        n, IBLND_MAX_RDMA_FRAGS);
 156                 return 1;
 157         }
 158
 159         nob = offsetof(struct kib_msg, ibm_u) +
 160               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 161
 162         if (msg->ibm_nob < nob) {
 163                 CERROR("Short %s: %d(%d)\n",
 164                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 165                 return 1;
 166         }
 167
 168         if (!flip)
 169                 return 0;
 170
 171         for (i = 0; i < n; i++) {
 172                 __swab32s(&rd->rd_frags[i].rf_nob);
 173                 __swab64s(&rd->rd_frags[i].rf_addr);
 174         }
 175
 176         return 0;
 177 }
 178
 179 void kiblnd_pack_msg(struct lnet_ni *ni, struct kib_msg *msg, int version,
 180                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 181 {
 182         struct kib_net *net = ni->ni_data;
 183
 184         /* CAVEAT EMPTOR! all message fields not set here should have been
 185          * initialised previously. */
 186         msg->ibm_magic    = IBLND_MSG_MAGIC;
 187         msg->ibm_version  = version;
 188         /*   ibm_type */
 189         msg->ibm_credits  = credits;
 190         /*   ibm_nob */
 191         msg->ibm_cksum    = 0;
 192         msg->ibm_srcnid   = ni->ni_nid;
 193         msg->ibm_srcstamp = net->ibn_incarnation;
 194         msg->ibm_dstnid   = dstnid;
 195         msg->ibm_dststamp = dststamp;
 196
 197         if (*kiblnd_tunables.kib_cksum) {
 198                 /* NB ibm_cksum zero while computing cksum */
 199                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 200         }
 201 }
 202
 203 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 204 {
 205         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 206         __u32     msg_cksum;
 207         __u16     version;
 208         int       msg_nob;
 209         int       flip;
 210
 211         /* 6 bytes are enough to have received magic + version */
 212         if (nob < 6) {
 213                 CERROR("Short message: %d\n", nob);
 214                 return -EPROTO;
 215         }
 216
 217         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 218                 flip = 0;
 219         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 220                 flip = 1;
 221         } else {
 222                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 223                 return -EPROTO;
 224         }
 225
 226         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 227         if (version != IBLND_MSG_VERSION &&
 228             version != IBLND_MSG_VERSION_1) {
 229                 CERROR("Bad version: %x\n", version);
 230                 return -EPROTO;
 231         }
 232
 233         if (nob < hdr_size) {
 234                 CERROR("Short message: %d\n", nob);
 235                 return -EPROTO;
 236         }
 237
 238         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 239         if (msg_nob > nob) {
 240                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 241                 return -EPROTO;
 242         }
 243
 244         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 245          * gets flipped */
 246         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 247         msg->ibm_cksum = 0;
 248         if (msg_cksum != 0 &&
 249             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 250                 CERROR("Bad checksum\n");
 251                 return -EPROTO;
 252         }
 253
 254         msg->ibm_cksum = msg_cksum;
 255
 256         if (flip) {
 257                 /* leave magic unflipped as a clue to peer_ni endianness */
 258                 msg->ibm_version = version;
 259                 CLASSERT (sizeof(msg->ibm_type) == 1);
 260                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 261                 msg->ibm_nob     = msg_nob;
 262                 __swab64s(&msg->ibm_srcnid);
 263                 __swab64s(&msg->ibm_srcstamp);
 264                 __swab64s(&msg->ibm_dstnid);
 265                 __swab64s(&msg->ibm_dststamp);
 266         }
 267
 268         if (msg->ibm_srcnid == LNET_NID_ANY) {
 269                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 270                 return -EPROTO;
 271         }
 272
 273         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 274                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 275                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 276                 return -EPROTO;
 277         }
 278
 279         switch (msg->ibm_type) {
 280         default:
 281                 CERROR("Unknown message type %x\n", msg->ibm_type);
 282                 return -EPROTO;
 283
 284         case IBLND_MSG_NOOP:
 285         case IBLND_MSG_IMMEDIATE:
 286         case IBLND_MSG_PUT_REQ:
 287                 break;
 288
 289         case IBLND_MSG_PUT_ACK:
 290         case IBLND_MSG_GET_REQ:
 291                 if (kiblnd_unpack_rd(msg, flip))
 292                         return -EPROTO;
 293                 break;
 294
 295         case IBLND_MSG_PUT_NAK:
 296         case IBLND_MSG_PUT_DONE:
 297         case IBLND_MSG_GET_DONE:
 298                 if (flip)
 299                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 300                 break;
 301
 302         case IBLND_MSG_CONNREQ:
 303         case IBLND_MSG_CONNACK:
 304                 if (flip) {
 305                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 306                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 307                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 308                 }
 309                 break;
 310         }
 311         return 0;
 312 }
 313
 314 int
 315 kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
 316                    lnet_nid_t nid)
 317 {
 318         struct kib_peer_ni *peer_ni;
 319         struct kib_net *net = ni->ni_data;
 320         int cpt = lnet_cpt_of_nid(nid, ni);
 321         unsigned long flags;
 322
 323         LASSERT(net != NULL);
 324         LASSERT(nid != LNET_NID_ANY);
 325
 326         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 327         if (peer_ni == NULL) {
 328                 CERROR("Cannot allocate peer_ni\n");
 329                 return -ENOMEM;
 330         }
 331
 332         peer_ni->ibp_ni = ni;
 333         peer_ni->ibp_nid = nid;
 334         peer_ni->ibp_error = 0;
 335         peer_ni->ibp_last_alive = 0;
 336         peer_ni->ibp_max_frags = IBLND_MAX_RDMA_FRAGS;
 337         peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
 338         atomic_set(&peer_ni->ibp_refcount, 1);  /* 1 ref for caller */
 339
 340         INIT_LIST_HEAD(&peer_ni->ibp_list);     /* not in the peer_ni table yet */
 341         INIT_LIST_HEAD(&peer_ni->ibp_conns);
 342         INIT_LIST_HEAD(&peer_ni->ibp_tx_queue);
 343
 344         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 345
 346         /* always called with a ref on ni, which prevents ni being shutdown */
 347         LASSERT(net->ibn_shutdown == 0);
 348
 349         /* npeers only grows with the global lock held */
 350         atomic_inc(&net->ibn_npeers);
 351
 352         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 353
 354         *peerp = peer_ni;
 355         return 0;
 356 }
 357
 358 void
 359 kiblnd_destroy_peer(struct kib_peer_ni *peer_ni)
 360 {
 361         struct kib_net *net = peer_ni->ibp_ni->ni_data;
 362
 363         LASSERT(net != NULL);
 364         LASSERT (atomic_read(&peer_ni->ibp_refcount) == 0);
 365         LASSERT(!kiblnd_peer_active(peer_ni));
 366         LASSERT(kiblnd_peer_idle(peer_ni));
 367         LASSERT(list_empty(&peer_ni->ibp_tx_queue));
 368
 369         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 370
 371         /* NB a peer_ni's connections keep a reference on their peer_ni until
 372          * they are destroyed, so we can be assured that _all_ state to do
 373          * with this peer_ni has been cleaned up when its refcount drops to
 374          * zero. */
 375         atomic_dec(&net->ibn_npeers);
 376 }
 377
 378 struct kib_peer_ni *
 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         struct kib_peer_ni              *peer_ni;
 386
 387         list_for_each(tmp, peer_list) {
 388
 389                 peer_ni = list_entry(tmp, struct kib_peer_ni, 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(struct kib_peer_ni *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         struct kib_peer_ni              *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, struct kib_peer_ni, 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(struct kib_peer_ni *peer_ni)
 461 {
 462         struct list_head *ctmp;
 463         struct list_head *cnxt;
 464         struct kib_conn *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, struct kib_conn, 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         struct kib_peer_ni              *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, struct kib_peer_ni, 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, LNET_MSG_STATUS_LOCAL_ERROR);
 528
 529         return rc;
 530 }
 531
 532 static struct kib_conn *
 533 kiblnd_get_conn_by_idx(struct lnet_ni *ni, int index)
 534 {
 535         struct kib_peer_ni              *peer_ni;
 536         struct list_head        *ptmp;
 537         struct kib_conn *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, struct kib_peer_ni, 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, struct kib_conn, 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(struct kib_rx *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(struct kib_tx *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(struct kib_conn *conn)
 592 {
 593         struct list_head        *tmp;
 594         int                     i;
 595
 596         spin_lock(&conn->ibc_lock);
 597
 598         CDEBUG(D_CONSOLE, "conn[%d] %p [version %x] -> %s:\n",
 599                atomic_read(&conn->ibc_refcount), conn,
 600                conn->ibc_version, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 601         CDEBUG(D_CONSOLE, "   state %d nposted %d/%d cred %d o_cred %d "
 602                " r_cred %d\n", conn->ibc_state, conn->ibc_noops_posted,
 603                conn->ibc_nsends_posted, conn->ibc_credits,
 604                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 605         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 606
 607         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 608         list_for_each(tmp, &conn->ibc_early_rxs)
 609                 kiblnd_debug_rx(list_entry(tmp, struct kib_rx, rx_list));
 610
 611         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 612         list_for_each(tmp, &conn->ibc_tx_noops)
 613                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 614
 615         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 616         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 617                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 620         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 621                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue)
 625                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   active_txs:\n");
 628         list_for_each(tmp, &conn->ibc_active_txs)
 629                 kiblnd_debug_tx(list_entry(tmp, struct kib_tx, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   rxs:\n");
 632         for (i = 0; i < IBLND_RX_MSGS(conn); i++)
 633                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 634
 635         spin_unlock(&conn->ibc_lock);
 636 }
 637
 638 int
 639 kiblnd_translate_mtu(int value)
 640 {
 641         switch (value) {
 642         default:
 643                 return -1;
 644         case 0:
 645                 return 0;
 646         case 256:
 647                 return IB_MTU_256;
 648         case 512:
 649                 return IB_MTU_512;
 650         case 1024:
 651                 return IB_MTU_1024;
 652         case 2048:
 653                 return IB_MTU_2048;
 654         case 4096:
 655                 return IB_MTU_4096;
 656         }
 657 }
 658
 659 static void
 660 kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 661 {
 662         int           mtu;
 663
 664         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 665         if (cmid->route.path_rec == NULL)
 666                 return;
 667
 668         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 669         LASSERT (mtu >= 0);
 670         if (mtu != 0)
 671                 cmid->route.path_rec->mtu = mtu;
 672 }
 673
 674 static int
 675 kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 676 {
 677         cpumask_t       *mask;
 678         int             vectors;
 679         int             off;
 680         int             i;
 681         lnet_nid_t      ibp_nid;
 682
 683         vectors = conn->ibc_cmid->device->num_comp_vectors;
 684         if (vectors <= 1)
 685                 return 0;
 686
 687         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 688
 689         /* hash NID to CPU id in this partition... */
 690         ibp_nid = conn->ibc_peer->ibp_nid;
 691         off = do_div(ibp_nid, cpumask_weight(mask));
 692         for_each_cpu(i, mask) {
 693                 if (off-- == 0)
 694                         return i % vectors;
 695         }
 696
 697         LBUG();
 698         return 1;
 699 }
 700
 701 /*
 702  * Get the scheduler bound to this CPT. If the scheduler has no
 703  * threads, which means that the CPT has no CPUs, then grab the
 704  * next scheduler that we can use.
 705  *
 706  * This case would be triggered if a NUMA node is configured with
 707  * no associated CPUs.
 708  */
 709 static struct kib_sched_info *
 710 kiblnd_get_scheduler(int cpt)
 711 {
 712         struct kib_sched_info *sched;
 713         int i;
 714
 715         sched = kiblnd_data.kib_scheds[cpt];
 716
 717         if (sched->ibs_nthreads > 0)
 718                 return sched;
 719
 720         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
 721                 if (sched->ibs_nthreads > 0) {
 722                         CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 723                                         cpt, sched->ibs_cpt);
 724                         return sched;
 725                 }
 726         }
 727
 728         return NULL;
 729 }
 730
 731 static unsigned int kiblnd_send_wrs(struct kib_conn *conn)
 732 {
 733         /*
 734          * One WR for the LNet message
 735          * And ibc_max_frags for the transfer WRs
 736          */
 737         int ret;
 738         int multiplier = 1 + conn->ibc_max_frags;
 739         enum kib_dev_caps dev_caps = conn->ibc_hdev->ibh_dev->ibd_dev_caps;
 740
 741         /* FastReg needs two extra WRs for map and invalidate */
 742         if (dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)
 743                 multiplier += 2;
 744
 745         /* account for a maximum of ibc_queue_depth in-flight transfers */
 746         ret = multiplier * conn->ibc_queue_depth;
 747
 748         if (ret > conn->ibc_hdev->ibh_max_qp_wr) {
 749                 CDEBUG(D_NET, "peer_credits %u will result in send work "
 750                        "request size %d larger than maximum %d device "
 751                        "can handle\n", conn->ibc_queue_depth, ret,
 752                        conn->ibc_hdev->ibh_max_qp_wr);
 753                 conn->ibc_queue_depth =
 754                         conn->ibc_hdev->ibh_max_qp_wr / multiplier;
 755         }
 756
 757         /* don't go beyond the maximum the device can handle */
 758         return min(ret, conn->ibc_hdev->ibh_max_qp_wr);
 759 }
 760
 761 struct kib_conn *
 762 kiblnd_create_conn(struct kib_peer_ni *peer_ni, struct rdma_cm_id *cmid,
 763                    int state, int version)
 764 {
 765         /* CAVEAT EMPTOR:
 766          * If the new conn is created successfully it takes over the caller's
 767          * ref on 'peer_ni'.  It also "owns" 'cmid' and destroys it when it itself
 768          * is destroyed.  On failure, the caller's ref on 'peer_ni' remains and
 769          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 770          * to destroy 'cmid' here since I'm called from the CM which still has
 771          * its ref on 'cmid'). */
 772         rwlock_t               *glock = &kiblnd_data.kib_global_lock;
 773         struct kib_net              *net = peer_ni->ibp_ni->ni_data;
 774         struct kib_dev *dev;
 775         struct ib_qp_init_attr *init_qp_attr;
 776         struct kib_sched_info   *sched;
 777 #ifdef HAVE_IB_CQ_INIT_ATTR
 778         struct ib_cq_init_attr  cq_attr = {};
 779 #endif
 780         struct kib_conn *conn;
 781         struct ib_cq            *cq;
 782         unsigned long           flags;
 783         int                     cpt;
 784         int                     rc;
 785         int                     i;
 786
 787         LASSERT(net != NULL);
 788         LASSERT(!in_interrupt());
 789
 790         dev = net->ibn_dev;
 791
 792         cpt = lnet_cpt_of_nid(peer_ni->ibp_nid, peer_ni->ibp_ni);
 793         sched = kiblnd_get_scheduler(cpt);
 794
 795         if (sched == NULL) {
 796                 CERROR("no schedulers available. node is unhealthy\n");
 797                 goto failed_0;
 798         }
 799
 800         /*
 801          * The cpt might have changed if we ended up selecting a non cpt
 802          * native scheduler. So use the scheduler's cpt instead.
 803          */
 804         cpt = sched->ibs_cpt;
 805
 806         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 807                          sizeof(*init_qp_attr));
 808         if (init_qp_attr == NULL) {
 809                 CERROR("Can't allocate qp_attr for %s\n",
 810                        libcfs_nid2str(peer_ni->ibp_nid));
 811                 goto failed_0;
 812         }
 813
 814         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 815         if (conn == NULL) {
 816                 CERROR("Can't allocate connection for %s\n",
 817                        libcfs_nid2str(peer_ni->ibp_nid));
 818                 goto failed_1;
 819         }
 820
 821         conn->ibc_state = IBLND_CONN_INIT;
 822         conn->ibc_version = version;
 823         conn->ibc_peer = peer_ni;                       /* I take the caller's ref */
 824         cmid->context = conn;                   /* for future CM callbacks */
 825         conn->ibc_cmid = cmid;
 826         conn->ibc_max_frags = peer_ni->ibp_max_frags;
 827         conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 828
 829         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 830         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 831         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 832         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 833         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 834         INIT_LIST_HEAD(&conn->ibc_active_txs);
 835         INIT_LIST_HEAD(&conn->ibc_zombie_txs);
 836         spin_lock_init(&conn->ibc_lock);
 837
 838         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 839                          sizeof(*conn->ibc_connvars));
 840         if (conn->ibc_connvars == NULL) {
 841                 CERROR("Can't allocate in-progress connection state\n");
 842                 goto failed_2;
 843         }
 844
 845         write_lock_irqsave(glock, flags);
 846         if (dev->ibd_failover) {
 847                 write_unlock_irqrestore(glock, flags);
 848                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 849                 goto failed_2;
 850         }
 851
 852         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 853                 /* wakeup failover thread and teardown connection */
 854                 if (kiblnd_dev_can_failover(dev)) {
 855                         list_add_tail(&dev->ibd_fail_list,
 856                                       &kiblnd_data.kib_failed_devs);
 857                         wake_up(&kiblnd_data.kib_failover_waitq);
 858                 }
 859
 860                 write_unlock_irqrestore(glock, flags);
 861                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 862                        cmid->device->name, dev->ibd_ifname);
 863                 goto failed_2;
 864         }
 865
 866         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 867         conn->ibc_hdev = dev->ibd_hdev;
 868
 869         kiblnd_setup_mtu_locked(cmid);
 870
 871         write_unlock_irqrestore(glock, flags);
 872
 873         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 874                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 875         if (conn->ibc_rxs == NULL) {
 876                 CERROR("Cannot allocate RX buffers\n");
 877                 goto failed_2;
 878         }
 879
 880         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 881                                 IBLND_RX_MSG_PAGES(conn));
 882         if (rc != 0)
 883                 goto failed_2;
 884
 885         kiblnd_map_rx_descs(conn);
 886
 887 #ifdef HAVE_IB_CQ_INIT_ATTR
 888         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 889         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 890         cq = ib_create_cq(cmid->device,
 891                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 892                           &cq_attr);
 893 #else
 894         cq = ib_create_cq(cmid->device,
 895                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 896                           IBLND_CQ_ENTRIES(conn),
 897                           kiblnd_get_completion_vector(conn, cpt));
 898 #endif
 899         if (IS_ERR(cq)) {
 900                 /*
 901                  * on MLX-5 (possibly MLX-4 as well) this error could be
 902                  * hit if the concurrent_sends and/or peer_tx_credits is set
 903                  * too high. Or due to an MLX-5 bug which tries to
 904                  * allocate 256kb via kmalloc for WR cookie array
 905                  */
 906                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 907                         IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 908                 goto failed_2;
 909         }
 910
 911         conn->ibc_cq = cq;
 912
 913         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 914         if (rc != 0) {
 915                 CERROR("Can't request completion notification: %d\n", rc);
 916                 goto failed_2;
 917         }
 918
 919         init_qp_attr->event_handler = kiblnd_qp_event;
 920         init_qp_attr->qp_context = conn;
 921         init_qp_attr->cap.max_send_sge = *kiblnd_tunables.kib_wrq_sge;
 922         init_qp_attr->cap.max_recv_sge = 1;
 923         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 924         init_qp_attr->qp_type = IB_QPT_RC;
 925         init_qp_attr->send_cq = cq;
 926         init_qp_attr->recv_cq = cq;
 927         /*
 928          * kiblnd_send_wrs() can change the connection's queue depth if
 929          * the maximum work requests for the device is maxed out
 930          */
 931         init_qp_attr->cap.max_send_wr = kiblnd_send_wrs(conn);
 932         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 933
 934         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 935         if (rc) {
 936                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d, "
 937                        "send_sge: %d, recv_sge: %d\n",
 938                        rc, init_qp_attr->cap.max_send_wr,
 939                        init_qp_attr->cap.max_recv_wr,
 940                        init_qp_attr->cap.max_send_sge,
 941                        init_qp_attr->cap.max_recv_sge);
 942                 goto failed_2;
 943         }
 944
 945         conn->ibc_sched = sched;
 946
 947         if (conn->ibc_queue_depth != peer_ni->ibp_queue_depth)
 948                 CWARN("peer %s - queue depth reduced from %u to %u"
 949                       "  to allow for qp creation\n",
 950                       libcfs_nid2str(peer_ni->ibp_nid),
 951                       peer_ni->ibp_queue_depth,
 952                       conn->ibc_queue_depth);
 953
 954         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 955
 956         /* 1 ref for caller and each rxmsg */
 957         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 958         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 959
 960         /* post receives */
 961         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 962                 rc = kiblnd_post_rx(&conn->ibc_rxs[i], IBLND_POSTRX_NO_CREDIT);
 963                 if (rc != 0) {
 964                         CERROR("Can't post rxmsg: %d\n", rc);
 965
 966                         /* Make posted receives complete */
 967                         kiblnd_abort_receives(conn);
 968
 969                         /* correct # of posted buffers
 970                          * NB locking needed now I'm racing with completion */
 971                         spin_lock_irqsave(&sched->ibs_lock, flags);
 972                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 973                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 974
 975                         /* cmid will be destroyed by CM(ofed) after cm_callback
 976                          * returned, so we can't refer it anymore
 977                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 978                         rdma_destroy_qp(conn->ibc_cmid);
 979                         conn->ibc_cmid = NULL;
 980
 981                         /* Drop my own and unused rxbuffer refcounts */
 982                         while (i++ <= IBLND_RX_MSGS(conn))
 983                                 kiblnd_conn_decref(conn);
 984
 985                         return NULL;
 986                 }
 987         }
 988
 989         /* Init successful! */
 990         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 991                  state == IBLND_CONN_PASSIVE_WAIT);
 992         conn->ibc_state = state;
 993
 994         /* 1 more conn */
 995         atomic_inc(&net->ibn_nconns);
 996         return conn;
 997
 998  failed_2:
 999         kiblnd_destroy_conn(conn);
1000         LIBCFS_FREE(conn, sizeof(*conn));
1001  failed_1:
1002         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
1003  failed_0:
1004         return NULL;
1005 }
1006
1007 void
1008 kiblnd_destroy_conn(struct kib_conn *conn)
1009 {
1010         struct rdma_cm_id *cmid = conn->ibc_cmid;
1011         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1012
1013         LASSERT (!in_interrupt());
1014         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
1015         LASSERT(list_empty(&conn->ibc_early_rxs));
1016         LASSERT(list_empty(&conn->ibc_tx_noops));
1017         LASSERT(list_empty(&conn->ibc_tx_queue));
1018         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
1019         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
1020         LASSERT(list_empty(&conn->ibc_active_txs));
1021         LASSERT (conn->ibc_noops_posted == 0);
1022         LASSERT (conn->ibc_nsends_posted == 0);
1023
1024         switch (conn->ibc_state) {
1025         default:
1026                 /* conn must be completely disengaged from the network */
1027                 LBUG();
1028
1029         case IBLND_CONN_DISCONNECTED:
1030                 /* connvars should have been freed already */
1031                 LASSERT (conn->ibc_connvars == NULL);
1032                 break;
1033
1034         case IBLND_CONN_INIT:
1035                 break;
1036         }
1037
1038         /* conn->ibc_cmid might be destroyed by CM already */
1039         if (cmid != NULL && cmid->qp != NULL)
1040                 rdma_destroy_qp(cmid);
1041
1042         if (conn->ibc_cq)
1043                 ib_destroy_cq(conn->ibc_cq);
1044
1045         kiblnd_txlist_done(&conn->ibc_zombie_txs, -ECONNABORTED,
1046                            LNET_MSG_STATUS_OK);
1047
1048         if (conn->ibc_rx_pages != NULL)
1049                 kiblnd_unmap_rx_descs(conn);
1050
1051         if (conn->ibc_rxs != NULL) {
1052                 LIBCFS_FREE(conn->ibc_rxs,
1053                             IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
1054         }
1055
1056         if (conn->ibc_connvars != NULL)
1057                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1058
1059         if (conn->ibc_hdev != NULL)
1060                 kiblnd_hdev_decref(conn->ibc_hdev);
1061
1062         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
1063         if (conn->ibc_state != IBLND_CONN_INIT) {
1064                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
1065
1066                 kiblnd_peer_decref(peer_ni);
1067                 rdma_destroy_id(cmid);
1068                 atomic_dec(&net->ibn_nconns);
1069         }
1070 }
1071
1072 int
1073 kiblnd_close_peer_conns_locked(struct kib_peer_ni *peer_ni, int why)
1074 {
1075         struct kib_conn *conn;
1076         struct list_head        *ctmp;
1077         struct list_head        *cnxt;
1078         int                     count = 0;
1079
1080         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1081                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1082
1083                 CDEBUG(D_NET, "Closing conn -> %s, "
1084                               "version: %x, reason: %d\n",
1085                        libcfs_nid2str(peer_ni->ibp_nid),
1086                        conn->ibc_version, why);
1087
1088                 kiblnd_close_conn_locked(conn, why);
1089                 count++;
1090         }
1091
1092         return count;
1093 }
1094
1095 int
1096 kiblnd_close_stale_conns_locked(struct kib_peer_ni *peer_ni,
1097                                 int version, __u64 incarnation)
1098 {
1099         struct kib_conn *conn;
1100         struct list_head        *ctmp;
1101         struct list_head        *cnxt;
1102         int                     count = 0;
1103
1104         list_for_each_safe(ctmp, cnxt, &peer_ni->ibp_conns) {
1105                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
1106
1107                 if (conn->ibc_version     == version &&
1108                     conn->ibc_incarnation == incarnation)
1109                         continue;
1110
1111                 CDEBUG(D_NET, "Closing stale conn -> %s version: %x, "
1112                               "incarnation:%#llx(%x, %#llx)\n",
1113                        libcfs_nid2str(peer_ni->ibp_nid),
1114                        conn->ibc_version, conn->ibc_incarnation,
1115                        version, incarnation);
1116
1117                 kiblnd_close_conn_locked(conn, -ESTALE);
1118                 count++;
1119         }
1120
1121         return count;
1122 }
1123
1124 static int
1125 kiblnd_close_matching_conns(struct lnet_ni *ni, lnet_nid_t nid)
1126 {
1127         struct kib_peer_ni              *peer_ni;
1128         struct list_head        *ptmp;
1129         struct list_head        *pnxt;
1130         int                     lo;
1131         int                     hi;
1132         int                     i;
1133         unsigned long           flags;
1134         int                     count = 0;
1135
1136         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1137
1138         if (nid != LNET_NID_ANY)
1139                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1140         else {
1141                 lo = 0;
1142                 hi = kiblnd_data.kib_peer_hash_size - 1;
1143         }
1144
1145         for (i = lo; i <= hi; i++) {
1146                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1147
1148                         peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
1149                         LASSERT(!kiblnd_peer_idle(peer_ni));
1150
1151                         if (peer_ni->ibp_ni != ni)
1152                                 continue;
1153
1154                         if (!(nid == LNET_NID_ANY || nid == peer_ni->ibp_nid))
1155                                 continue;
1156
1157                         count += kiblnd_close_peer_conns_locked(peer_ni, 0);
1158                 }
1159         }
1160
1161         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1162
1163         /* wildcards always succeed */
1164         if (nid == LNET_NID_ANY)
1165                 return 0;
1166
1167         return (count == 0) ? -ENOENT : 0;
1168 }
1169
1170 static int
1171 kiblnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1172 {
1173         struct libcfs_ioctl_data *data = arg;
1174         int                       rc = -EINVAL;
1175
1176         switch(cmd) {
1177         case IOC_LIBCFS_GET_PEER: {
1178                 lnet_nid_t   nid = 0;
1179                 int          count = 0;
1180
1181                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1182                                           &nid, &count);
1183                 data->ioc_nid    = nid;
1184                 data->ioc_count  = count;
1185                 break;
1186         }
1187
1188         case IOC_LIBCFS_DEL_PEER: {
1189                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1190                 break;
1191         }
1192         case IOC_LIBCFS_GET_CONN: {
1193                 struct kib_conn *conn;
1194
1195                 rc = 0;
1196                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1197                 if (conn == NULL) {
1198                         rc = -ENOENT;
1199                         break;
1200                 }
1201
1202                 LASSERT(conn->ibc_cmid != NULL);
1203                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1204                 if (conn->ibc_cmid->route.path_rec == NULL)
1205                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1206                 else
1207                         data->ioc_u32[0] =
1208                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1209                 kiblnd_conn_decref(conn);
1210                 break;
1211         }
1212         case IOC_LIBCFS_CLOSE_CONNECTION: {
1213                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1214                 break;
1215         }
1216
1217         default:
1218                 break;
1219         }
1220
1221         return rc;
1222 }
1223
1224 static void
1225 kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1226 {
1227         time64_t last_alive = 0;
1228         time64_t now = ktime_get_seconds();
1229         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1230         struct kib_peer_ni *peer_ni;
1231         unsigned long flags;
1232
1233         read_lock_irqsave(glock, flags);
1234
1235         peer_ni = kiblnd_find_peer_locked(ni, nid);
1236         if (peer_ni != NULL)
1237                 last_alive = peer_ni->ibp_last_alive;
1238
1239         read_unlock_irqrestore(glock, flags);
1240
1241         if (last_alive != 0)
1242                 *when = last_alive;
1243
1244         /* peer_ni is not persistent in hash, trigger peer_ni creation
1245          * and connection establishment with a NULL tx */
1246         if (peer_ni == NULL)
1247                 kiblnd_launch_tx(ni, NULL, nid);
1248
1249         CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago\n",
1250                libcfs_nid2str(nid), peer_ni,
1251                last_alive ? now - last_alive : -1);
1252         return;
1253 }
1254
1255 static void
1256 kiblnd_free_pages(struct kib_pages *p)
1257 {
1258         int     npages = p->ibp_npages;
1259         int     i;
1260
1261         for (i = 0; i < npages; i++) {
1262                 if (p->ibp_pages[i] != NULL)
1263                         __free_page(p->ibp_pages[i]);
1264         }
1265
1266         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1267 }
1268
1269 int
1270 kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1271 {
1272         struct kib_pages *p;
1273         int i;
1274
1275         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1276                          offsetof(struct kib_pages, ibp_pages[npages]));
1277         if (p == NULL) {
1278                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1279                 return -ENOMEM;
1280         }
1281
1282         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1283         p->ibp_npages = npages;
1284
1285         for (i = 0; i < npages; i++) {
1286                 p->ibp_pages[i] = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1287                                                      GFP_NOFS);
1288                 if (p->ibp_pages[i] == NULL) {
1289                         CERROR("Can't allocate page %d of %d\n", i, npages);
1290                         kiblnd_free_pages(p);
1291                         return -ENOMEM;
1292                 }
1293         }
1294
1295         *pp = p;
1296         return 0;
1297 }
1298
1299 void
1300 kiblnd_unmap_rx_descs(struct kib_conn *conn)
1301 {
1302         struct kib_rx *rx;
1303         int       i;
1304
1305         LASSERT (conn->ibc_rxs != NULL);
1306         LASSERT (conn->ibc_hdev != NULL);
1307
1308         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1309                 rx = &conn->ibc_rxs[i];
1310
1311                 LASSERT(rx->rx_nob >= 0); /* not posted */
1312
1313                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1314                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1315                                                           rx->rx_msgaddr),
1316                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1317         }
1318
1319         kiblnd_free_pages(conn->ibc_rx_pages);
1320
1321         conn->ibc_rx_pages = NULL;
1322 }
1323
1324 void
1325 kiblnd_map_rx_descs(struct kib_conn *conn)
1326 {
1327         struct kib_rx *rx;
1328         struct page    *pg;
1329         int             pg_off;
1330         int             ipg;
1331         int             i;
1332
1333         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1334                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1335                 rx = &conn->ibc_rxs[i];
1336
1337                 rx->rx_conn = conn;
1338                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1339
1340                 rx->rx_msgaddr =
1341                         kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1342                                               rx->rx_msg, IBLND_MSG_SIZE,
1343                                               DMA_FROM_DEVICE);
1344                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1345                                                   rx->rx_msgaddr));
1346                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1347
1348                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1349                        i, rx->rx_msg, rx->rx_msgaddr,
1350                        (__u64)(page_to_phys(pg) + pg_off));
1351
1352                 pg_off += IBLND_MSG_SIZE;
1353                 LASSERT(pg_off <= PAGE_SIZE);
1354
1355                 if (pg_off == PAGE_SIZE) {
1356                         pg_off = 0;
1357                         ipg++;
1358                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1359                 }
1360         }
1361 }
1362
1363 static void
1364 kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1365 {
1366         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1367         struct kib_tx *tx;
1368         int i;
1369
1370         LASSERT (tpo->tpo_pool.po_allocated == 0);
1371
1372         if (hdev == NULL)
1373                 return;
1374
1375         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1376                 tx = &tpo->tpo_tx_descs[i];
1377                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1378                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1379                                                           tx->tx_msgaddr),
1380                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1381         }
1382
1383         kiblnd_hdev_decref(hdev);
1384         tpo->tpo_hdev = NULL;
1385 }
1386
1387 static struct kib_hca_dev *
1388 kiblnd_current_hdev(struct kib_dev *dev)
1389 {
1390         struct kib_hca_dev *hdev;
1391         unsigned long  flags;
1392         int            i = 0;
1393
1394         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1395         while (dev->ibd_failover) {
1396                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1397                 if (i++ % 50 == 0)
1398                         CDEBUG(D_NET, "%s: Wait for failover\n",
1399                                dev->ibd_ifname);
1400                 set_current_state(TASK_INTERRUPTIBLE);
1401                 schedule_timeout(cfs_time_seconds(1) / 100);
1402
1403                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1404         }
1405
1406         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1407         hdev = dev->ibd_hdev;
1408
1409         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1410
1411         return hdev;
1412 }
1413
1414 static void
1415 kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1416 {
1417         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1418         struct kib_pool *pool = &tpo->tpo_pool;
1419         struct kib_net      *net   = pool->po_owner->ps_net;
1420         struct kib_dev *dev;
1421         struct page *page;
1422         struct kib_tx *tx;
1423         int             page_offset;
1424         int             ipage;
1425         int             i;
1426
1427         LASSERT (net != NULL);
1428
1429         dev = net->ibn_dev;
1430
1431         /* pre-mapped messages are not bigger than 1 page */
1432         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1433
1434         /* No fancy arithmetic when we do the buffer calculations */
1435         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1436
1437         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1438
1439         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1440                 page = txpgs->ibp_pages[ipage];
1441                 tx = &tpo->tpo_tx_descs[i];
1442
1443                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1444                                                 page_offset);
1445
1446                 tx->tx_msgaddr = kiblnd_dma_map_single(tpo->tpo_hdev->ibh_ibdev,
1447                                                        tx->tx_msg,
1448                                                        IBLND_MSG_SIZE,
1449                                                        DMA_TO_DEVICE);
1450                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1451                                                   tx->tx_msgaddr));
1452                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1453
1454                 list_add(&tx->tx_list, &pool->po_free_list);
1455
1456                 page_offset += IBLND_MSG_SIZE;
1457                 LASSERT(page_offset <= PAGE_SIZE);
1458
1459                 if (page_offset == PAGE_SIZE) {
1460                         page_offset = 0;
1461                         ipage++;
1462                         LASSERT(ipage <= txpgs->ibp_npages);
1463                 }
1464         }
1465 }
1466
1467 static void
1468 kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1469 {
1470         LASSERT(fpo->fpo_map_count == 0);
1471
1472         if (fpo->fpo_is_fmr && fpo->fmr.fpo_fmr_pool) {
1473                 ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1474         } else {
1475                 struct kib_fast_reg_descriptor *frd, *tmp;
1476                 int i = 0;
1477
1478                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1479                                          frd_list) {
1480                         list_del(&frd->frd_list);
1481 #ifndef HAVE_IB_MAP_MR_SG
1482                         ib_free_fast_reg_page_list(frd->frd_frpl);
1483 #endif
1484                         ib_dereg_mr(frd->frd_mr);
1485                         LIBCFS_FREE(frd, sizeof(*frd));
1486                         i++;
1487                 }
1488                 if (i < fpo->fast_reg.fpo_pool_size)
1489                         CERROR("FastReg pool still has %d regions registered\n",
1490                                 fpo->fast_reg.fpo_pool_size - i);
1491         }
1492
1493         if (fpo->fpo_hdev)
1494                 kiblnd_hdev_decref(fpo->fpo_hdev);
1495
1496         LIBCFS_FREE(fpo, sizeof(*fpo));
1497 }
1498
1499 static void
1500 kiblnd_destroy_fmr_pool_list(struct list_head *head)
1501 {
1502         struct kib_fmr_pool *fpo, *tmp;
1503
1504         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1505                 list_del(&fpo->fpo_list);
1506                 kiblnd_destroy_fmr_pool(fpo);
1507         }
1508 }
1509
1510 static int
1511 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1512                      int ncpts)
1513 {
1514         int size = tunables->lnd_fmr_pool_size / ncpts;
1515
1516         return max(IBLND_FMR_POOL, size);
1517 }
1518
1519 static int
1520 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1521                          int ncpts)
1522 {
1523         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1524
1525         return max(IBLND_FMR_POOL_FLUSH, size);
1526 }
1527
1528 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps,
1529                                  struct kib_fmr_pool *fpo)
1530 {
1531         struct ib_fmr_pool_param param = {
1532                 .max_pages_per_fmr = LNET_MAX_IOV,
1533                 .page_shift        = PAGE_SHIFT,
1534                 .access            = (IB_ACCESS_LOCAL_WRITE |
1535                                       IB_ACCESS_REMOTE_WRITE),
1536                 .pool_size         = fps->fps_pool_size,
1537                 .dirty_watermark   = fps->fps_flush_trigger,
1538                 .flush_function    = NULL,
1539                 .flush_arg         = NULL,
1540                 .cache             = !!fps->fps_cache };
1541         int rc = 0;
1542
1543         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1544                                                    &param);
1545         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1546                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1547                 if (rc != -ENOSYS)
1548                         CERROR("Failed to create FMR pool: %d\n", rc);
1549                 else
1550                         CERROR("FMRs are not supported\n");
1551         }
1552         fpo->fpo_is_fmr = true;
1553
1554         return rc;
1555 }
1556
1557 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps,
1558                                   struct kib_fmr_pool *fpo,
1559                                   enum kib_dev_caps dev_caps)
1560 {
1561         struct kib_fast_reg_descriptor *frd, *tmp;
1562         int i, rc;
1563
1564         fpo->fpo_is_fmr = false;
1565
1566         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1567         fpo->fast_reg.fpo_pool_size = 0;
1568         for (i = 0; i < fps->fps_pool_size; i++) {
1569                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1570                                  sizeof(*frd));
1571                 if (!frd) {
1572                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1573                         rc = -ENOMEM;
1574                         goto out;
1575                 }
1576                 frd->frd_mr = NULL;
1577
1578 #ifndef HAVE_IB_MAP_MR_SG
1579                 frd->frd_frpl = ib_alloc_fast_reg_page_list(fpo->fpo_hdev->ibh_ibdev,
1580                                                             LNET_MAX_IOV);
1581                 if (IS_ERR(frd->frd_frpl)) {
1582                         rc = PTR_ERR(frd->frd_frpl);
1583                         CERROR("Failed to allocate ib_fast_reg_page_list: %d\n",
1584                                 rc);
1585                         frd->frd_frpl = NULL;
1586                         goto out_middle;
1587                 }
1588 #endif
1589
1590 #ifdef HAVE_IB_ALLOC_FAST_REG_MR
1591                 frd->frd_mr = ib_alloc_fast_reg_mr(fpo->fpo_hdev->ibh_pd,
1592                                                    LNET_MAX_IOV);
1593 #else
1594                 /*
1595                  * it is expected to get here if this is an MLX-5 card.
1596                  * MLX-4 cards will always use FMR and MLX-5 cards will
1597                  * always use fast_reg. It turns out that some MLX-5 cards
1598                  * (possibly due to older FW versions) do not natively support
1599                  * gaps. So we will need to track them here.
1600                  */
1601                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1602 #ifdef IB_MR_TYPE_SG_GAPS
1603                                           ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1604                                            (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT)) ?
1605                                                 IB_MR_TYPE_SG_GAPS :
1606                                                 IB_MR_TYPE_MEM_REG,
1607 #else
1608                                                 IB_MR_TYPE_MEM_REG,
1609 #endif
1610                                           LNET_MAX_IOV);
1611                 if ((*kiblnd_tunables.kib_use_fastreg_gaps == 1) &&
1612                     (dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT))
1613                         CWARN("using IB_MR_TYPE_SG_GAPS, expect a performance drop\n");
1614 #endif
1615                 if (IS_ERR(frd->frd_mr)) {
1616                         rc = PTR_ERR(frd->frd_mr);
1617                         CERROR("Failed to allocate ib_fast_reg_mr: %d\n", rc);
1618                         frd->frd_mr = NULL;
1619                         goto out_middle;
1620                 }
1621
1622                 /* There appears to be a bug in MLX5 code where you must
1623                  * invalidate the rkey of a new FastReg pool before first
1624                  * using it. Thus, I am marking the FRD invalid here. */
1625                 frd->frd_valid = false;
1626
1627                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1628                 fpo->fast_reg.fpo_pool_size++;
1629         }
1630
1631         return 0;
1632
1633 out_middle:
1634         if (frd->frd_mr)
1635                 ib_dereg_mr(frd->frd_mr);
1636 #ifndef HAVE_IB_MAP_MR_SG
1637         if (frd->frd_frpl)
1638                 ib_free_fast_reg_page_list(frd->frd_frpl);
1639 #endif
1640         LIBCFS_FREE(frd, sizeof(*frd));
1641
1642 out:
1643         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1644                                  frd_list) {
1645                 list_del(&frd->frd_list);
1646 #ifndef HAVE_IB_MAP_MR_SG
1647                 ib_free_fast_reg_page_list(frd->frd_frpl);
1648 #endif
1649                 ib_dereg_mr(frd->frd_mr);
1650                 LIBCFS_FREE(frd, sizeof(*frd));
1651         }
1652
1653         return rc;
1654 }
1655
1656 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1657                                   struct kib_fmr_pool **pp_fpo)
1658 {
1659         struct kib_dev *dev = fps->fps_net->ibn_dev;
1660         struct kib_fmr_pool *fpo;
1661         int rc;
1662
1663         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1664         if (!fpo) {
1665                 return -ENOMEM;
1666         }
1667         memset(fpo, 0, sizeof(*fpo));
1668
1669         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1670
1671         if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
1672                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1673         else
1674                 rc = kiblnd_alloc_freg_pool(fps, fpo, dev->ibd_dev_caps);
1675         if (rc)
1676                 goto out_fpo;
1677
1678         fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1679         fpo->fpo_owner = fps;
1680         *pp_fpo = fpo;
1681
1682         return 0;
1683
1684 out_fpo:
1685         kiblnd_hdev_decref(fpo->fpo_hdev);
1686         LIBCFS_FREE(fpo, sizeof(*fpo));
1687         return rc;
1688 }
1689
1690 static void
1691 kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps, struct list_head *zombies)
1692 {
1693         if (fps->fps_net == NULL) /* intialized? */
1694                 return;
1695
1696         spin_lock(&fps->fps_lock);
1697
1698         while (!list_empty(&fps->fps_pool_list)) {
1699                 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1700                                                       struct kib_fmr_pool,
1701                                                       fpo_list);
1702
1703                 fpo->fpo_failed = 1;
1704                 list_del(&fpo->fpo_list);
1705                 if (fpo->fpo_map_count == 0)
1706                         list_add(&fpo->fpo_list, zombies);
1707                 else
1708                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1709         }
1710
1711         spin_unlock(&fps->fps_lock);
1712 }
1713
1714 static void
1715 kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1716 {
1717         if (fps->fps_net != NULL) { /* initialized? */
1718                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1719                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1720         }
1721 }
1722
1723 static int
1724 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1725                         struct kib_net *net,
1726                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1727 {
1728         struct kib_fmr_pool *fpo;
1729         int rc;
1730
1731         memset(fps, 0, sizeof(struct kib_fmr_poolset));
1732
1733         fps->fps_net = net;
1734         fps->fps_cpt = cpt;
1735
1736         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1737         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1738         fps->fps_cache = tunables->lnd_fmr_cache;
1739
1740         spin_lock_init(&fps->fps_lock);
1741         INIT_LIST_HEAD(&fps->fps_pool_list);
1742         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1743
1744         rc = kiblnd_create_fmr_pool(fps, &fpo);
1745         if (rc == 0)
1746                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1747
1748         return rc;
1749 }
1750
1751 static int
1752 kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, time64_t now)
1753 {
1754         if (fpo->fpo_map_count != 0) /* still in use */
1755                 return 0;
1756         if (fpo->fpo_failed)
1757                 return 1;
1758         return now >= fpo->fpo_deadline;
1759 }
1760
1761 static int
1762 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1763 {
1764         struct kib_hca_dev *hdev;
1765         __u64           *pages = tx->tx_pages;
1766         int             npages;
1767         int             size;
1768         int             i;
1769
1770         hdev = tx->tx_pool->tpo_hdev;
1771
1772         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1773                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1774                         size += hdev->ibh_page_size) {
1775                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1776                                            hdev->ibh_page_mask) + size;
1777                 }
1778         }
1779
1780         return npages;
1781 }
1782
1783 void
1784 kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1785 {
1786         struct list_head zombies = LIST_HEAD_INIT(zombies);
1787         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1788         struct kib_fmr_poolset *fps;
1789         time64_t now = ktime_get_seconds();
1790         struct kib_fmr_pool *tmp;
1791         int rc;
1792
1793         if (!fpo)
1794                 return;
1795
1796         fps = fpo->fpo_owner;
1797         if (fpo->fpo_is_fmr) {
1798                 if (fmr->fmr_pfmr) {
1799                         ib_fmr_pool_unmap(fmr->fmr_pfmr);
1800                         fmr->fmr_pfmr = NULL;
1801                 }
1802
1803                 if (status) {
1804                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1805                         LASSERT(!rc);
1806                 }
1807         } else {
1808                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1809
1810                 if (frd) {
1811                         frd->frd_valid = false;
1812                         spin_lock(&fps->fps_lock);
1813                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1814                         spin_unlock(&fps->fps_lock);
1815                         fmr->fmr_frd = NULL;
1816                 }
1817         }
1818         fmr->fmr_pool = NULL;
1819
1820         spin_lock(&fps->fps_lock);
1821         fpo->fpo_map_count--;   /* decref the pool */
1822
1823         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1824                 /* the first pool is persistent */
1825                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1826                         continue;
1827
1828                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1829                         list_move(&fpo->fpo_list, &zombies);
1830                         fps->fps_version++;
1831                 }
1832         }
1833         spin_unlock(&fps->fps_lock);
1834
1835         if (!list_empty(&zombies))
1836                 kiblnd_destroy_fmr_pool_list(&zombies);
1837 }
1838
1839 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1840                         struct kib_rdma_desc *rd, u32 nob, u64 iov,
1841                         struct kib_fmr *fmr)
1842 {
1843         struct kib_fmr_pool *fpo;
1844         __u64 *pages = tx->tx_pages;
1845         __u64 version;
1846         bool is_rx = (rd != tx->tx_rd);
1847         bool tx_pages_mapped = 0;
1848         int npages = 0;
1849         int rc;
1850
1851 again:
1852         spin_lock(&fps->fps_lock);
1853         version = fps->fps_version;
1854         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1855                 fpo->fpo_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
1856                 fpo->fpo_map_count++;
1857
1858                 if (fpo->fpo_is_fmr) {
1859                         struct ib_pool_fmr *pfmr;
1860
1861                         spin_unlock(&fps->fps_lock);
1862
1863                         if (!tx_pages_mapped) {
1864                                 npages = kiblnd_map_tx_pages(tx, rd);
1865                                 tx_pages_mapped = 1;
1866                         }
1867
1868                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1869                                                     pages, npages, iov);
1870                         if (likely(!IS_ERR(pfmr))) {
1871                                 fmr->fmr_key  = is_rx ? pfmr->fmr->rkey
1872                                                       : pfmr->fmr->lkey;
1873                                 fmr->fmr_frd  = NULL;
1874                                 fmr->fmr_pfmr = pfmr;
1875                                 fmr->fmr_pool = fpo;
1876                                 return 0;
1877                         }
1878                         rc = PTR_ERR(pfmr);
1879                 } else {
1880                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1881                                 struct kib_fast_reg_descriptor *frd;
1882 #ifdef HAVE_IB_MAP_MR_SG
1883                                 struct ib_reg_wr *wr;
1884                                 int n;
1885 #else
1886                                 struct ib_rdma_wr *wr;
1887                                 struct ib_fast_reg_page_list *frpl;
1888 #endif
1889                                 struct ib_mr *mr;
1890
1891                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1892                                                         struct kib_fast_reg_descriptor,
1893                                                         frd_list);
1894                                 list_del(&frd->frd_list);
1895                                 spin_unlock(&fps->fps_lock);
1896
1897 #ifndef HAVE_IB_MAP_MR_SG
1898                                 frpl = frd->frd_frpl;
1899 #endif
1900                                 mr   = frd->frd_mr;
1901
1902                                 if (!frd->frd_valid) {
1903                                         struct ib_rdma_wr *inv_wr;
1904                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1905
1906                                         inv_wr = &frd->frd_inv_wr;
1907                                         memset(inv_wr, 0, sizeof(*inv_wr));
1908
1909                                         inv_wr->wr.opcode = IB_WR_LOCAL_INV;
1910                                         inv_wr->wr.wr_id  = IBLND_WID_MR;
1911                                         inv_wr->wr.ex.invalidate_rkey = key;
1912
1913                                         /* Bump the key */
1914                                         key = ib_inc_rkey(key);
1915                                         ib_update_fast_reg_key(mr, key);
1916                                 }
1917
1918 #ifdef HAVE_IB_MAP_MR_SG
1919 #ifdef HAVE_IB_MAP_MR_SG_5ARGS
1920                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1921                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1922 #else
1923                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1924                                                  tx->tx_nfrags, PAGE_SIZE);
1925 #endif
1926                                 if (unlikely(n != tx->tx_nfrags)) {
1927                                         CERROR("Failed to map mr %d/%d "
1928                                                "elements\n", n, tx->tx_nfrags);
1929                                         return n < 0 ? n : -EINVAL;
1930                                 }
1931
1932                                 wr = &frd->frd_fastreg_wr;
1933                                 memset(wr, 0, sizeof(*wr));
1934
1935                                 wr->wr.opcode = IB_WR_REG_MR;
1936                                 wr->wr.wr_id  = IBLND_WID_MR;
1937                                 wr->wr.num_sge = 0;
1938                                 wr->wr.send_flags = 0;
1939                                 wr->mr = mr;
1940                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1941                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1942                                               IB_ACCESS_REMOTE_WRITE);
1943 #else
1944                                 if (!tx_pages_mapped) {
1945                                         npages = kiblnd_map_tx_pages(tx, rd);
1946                                         tx_pages_mapped = 1;
1947                                 }
1948
1949                                 LASSERT(npages <= frpl->max_page_list_len);
1950                                 memcpy(frpl->page_list, pages,
1951                                         sizeof(*pages) * npages);
1952
1953                                 /* Prepare FastReg WR */
1954                                 wr = &frd->frd_fastreg_wr;
1955                                 memset(wr, 0, sizeof(*wr));
1956
1957                                 wr->wr.opcode = IB_WR_FAST_REG_MR;
1958                                 wr->wr.wr_id  = IBLND_WID_MR;
1959
1960                                 wr->wr.wr.fast_reg.iova_start = iov;
1961                                 wr->wr.wr.fast_reg.page_list  = frpl;
1962                                 wr->wr.wr.fast_reg.page_list_len = npages;
1963                                 wr->wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1964                                 wr->wr.wr.fast_reg.length = nob;
1965                                 wr->wr.wr.fast_reg.rkey =
1966                                                 is_rx ? mr->rkey : mr->lkey;
1967                                 wr->wr.wr.fast_reg.access_flags =
1968                                                 (IB_ACCESS_LOCAL_WRITE |
1969                                                  IB_ACCESS_REMOTE_WRITE);
1970 #endif
1971
1972                                 fmr->fmr_key  = is_rx ? mr->rkey : mr->lkey;
1973                                 fmr->fmr_frd  = frd;
1974                                 fmr->fmr_pfmr = NULL;
1975                                 fmr->fmr_pool = fpo;
1976                                 return 0;
1977                         }
1978                         spin_unlock(&fps->fps_lock);
1979                         rc = -EAGAIN;
1980                 }
1981
1982                 spin_lock(&fps->fps_lock);
1983                 fpo->fpo_map_count--;
1984                 if (rc != -EAGAIN) {
1985                         spin_unlock(&fps->fps_lock);
1986                         return rc;
1987                 }
1988
1989                 /* EAGAIN and ... */
1990                 if (version != fps->fps_version) {
1991                         spin_unlock(&fps->fps_lock);
1992                         goto again;
1993                 }
1994         }
1995
1996         if (fps->fps_increasing) {
1997                 spin_unlock(&fps->fps_lock);
1998                 CDEBUG(D_NET, "Another thread is allocating new "
1999                        "FMR pool, waiting for her to complete\n");
2000                 schedule();
2001                 goto again;
2002
2003         }
2004
2005         if (ktime_get_seconds() < fps->fps_next_retry) {
2006                 /* someone failed recently */
2007                 spin_unlock(&fps->fps_lock);
2008                 return -EAGAIN;
2009         }
2010
2011         fps->fps_increasing = 1;
2012         spin_unlock(&fps->fps_lock);
2013
2014         CDEBUG(D_NET, "Allocate new FMR pool\n");
2015         rc = kiblnd_create_fmr_pool(fps, &fpo);
2016         spin_lock(&fps->fps_lock);
2017         fps->fps_increasing = 0;
2018         if (rc == 0) {
2019                 fps->fps_version++;
2020                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
2021         } else {
2022                 fps->fps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2023         }
2024         spin_unlock(&fps->fps_lock);
2025
2026         goto again;
2027 }
2028
2029 static void
2030 kiblnd_fini_pool(struct kib_pool *pool)
2031 {
2032         LASSERT(list_empty(&pool->po_free_list));
2033         LASSERT(pool->po_allocated == 0);
2034
2035         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
2036 }
2037
2038 static void
2039 kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
2040 {
2041         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
2042
2043         memset(pool, 0, sizeof(struct kib_pool));
2044         INIT_LIST_HEAD(&pool->po_free_list);
2045         pool->po_deadline = ktime_get_seconds() + IBLND_POOL_DEADLINE;
2046         pool->po_owner = ps;
2047         pool->po_size = size;
2048 }
2049
2050 static void
2051 kiblnd_destroy_pool_list(struct list_head *head)
2052 {
2053         struct kib_pool *pool;
2054
2055         while (!list_empty(head)) {
2056                 pool = list_entry(head->next, struct kib_pool, po_list);
2057                 list_del(&pool->po_list);
2058
2059                 LASSERT(pool->po_owner != NULL);
2060                 pool->po_owner->ps_pool_destroy(pool);
2061         }
2062 }
2063
2064 static void
2065 kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
2066 {
2067         if (ps->ps_net == NULL) /* intialized? */
2068                 return;
2069
2070         spin_lock(&ps->ps_lock);
2071         while (!list_empty(&ps->ps_pool_list)) {
2072                 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
2073                                                  struct kib_pool, po_list);
2074
2075                 po->po_failed = 1;
2076                 list_del(&po->po_list);
2077                 if (po->po_allocated == 0)
2078                         list_add(&po->po_list, zombies);
2079                 else
2080                         list_add(&po->po_list, &ps->ps_failed_pool_list);
2081         }
2082         spin_unlock(&ps->ps_lock);
2083 }
2084
2085 static void
2086 kiblnd_fini_poolset(struct kib_poolset *ps)
2087 {
2088         if (ps->ps_net != NULL) { /* initialized? */
2089                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
2090                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
2091         }
2092 }
2093
2094 static int
2095 kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
2096                     struct kib_net *net, char *name, int size,
2097                     kib_ps_pool_create_t po_create,
2098                     kib_ps_pool_destroy_t po_destroy,
2099                     kib_ps_node_init_t nd_init,
2100                     kib_ps_node_fini_t nd_fini)
2101 {
2102         struct kib_pool *pool;
2103         int rc;
2104
2105         memset(ps, 0, sizeof(struct kib_poolset));
2106
2107         ps->ps_cpt          = cpt;
2108         ps->ps_net          = net;
2109         ps->ps_pool_create  = po_create;
2110         ps->ps_pool_destroy = po_destroy;
2111         ps->ps_node_init    = nd_init;
2112         ps->ps_node_fini    = nd_fini;
2113         ps->ps_pool_size    = size;
2114         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
2115             >= sizeof(ps->ps_name))
2116                 return -E2BIG;
2117         spin_lock_init(&ps->ps_lock);
2118         INIT_LIST_HEAD(&ps->ps_pool_list);
2119         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
2120
2121         rc = ps->ps_pool_create(ps, size, &pool);
2122         if (rc == 0)
2123                 list_add(&pool->po_list, &ps->ps_pool_list);
2124         else
2125                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
2126
2127         return rc;
2128 }
2129
2130 static int
2131 kiblnd_pool_is_idle(struct kib_pool *pool, time64_t now)
2132 {
2133         if (pool->po_allocated != 0) /* still in use */
2134                 return 0;
2135         if (pool->po_failed)
2136                 return 1;
2137         return now >= pool->po_deadline;
2138 }
2139
2140 void
2141 kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
2142 {
2143         struct list_head zombies = LIST_HEAD_INIT(zombies);
2144         struct kib_poolset *ps = pool->po_owner;
2145         struct kib_pool *tmp;
2146         time64_t now = ktime_get_seconds();
2147
2148         spin_lock(&ps->ps_lock);
2149
2150         if (ps->ps_node_fini != NULL)
2151                 ps->ps_node_fini(pool, node);
2152
2153         LASSERT(pool->po_allocated > 0);
2154         list_add(node, &pool->po_free_list);
2155         pool->po_allocated--;
2156
2157         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
2158                 /* the first pool is persistent */
2159                 if (ps->ps_pool_list.next == &pool->po_list)
2160                         continue;
2161
2162                 if (kiblnd_pool_is_idle(pool, now))
2163                         list_move(&pool->po_list, &zombies);
2164         }
2165         spin_unlock(&ps->ps_lock);
2166
2167         if (!list_empty(&zombies))
2168                 kiblnd_destroy_pool_list(&zombies);
2169 }
2170
2171 struct list_head *
2172 kiblnd_pool_alloc_node(struct kib_poolset *ps)
2173 {
2174         struct list_head        *node;
2175         struct kib_pool *pool;
2176         int                     rc;
2177         unsigned int            interval = 1;
2178         ktime_t time_before;
2179         unsigned int trips = 0;
2180
2181 again:
2182         spin_lock(&ps->ps_lock);
2183         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
2184                 if (list_empty(&pool->po_free_list))
2185                         continue;
2186
2187                 pool->po_allocated++;
2188                 pool->po_deadline = ktime_get_seconds() +
2189                                     IBLND_POOL_DEADLINE;
2190                 node = pool->po_free_list.next;
2191                 list_del(node);
2192
2193                 if (ps->ps_node_init != NULL) {
2194                         /* still hold the lock */
2195                         ps->ps_node_init(pool, node);
2196                 }
2197                 spin_unlock(&ps->ps_lock);
2198                 return node;
2199         }
2200
2201         /* no available tx pool and ... */
2202         if (ps->ps_increasing) {
2203                 /* another thread is allocating a new pool */
2204                 spin_unlock(&ps->ps_lock);
2205                 trips++;
2206                 CDEBUG(D_NET, "Another thread is allocating new "
2207                        "%s pool, waiting %d HZs for her to complete."
2208                        "trips = %d\n",
2209                        ps->ps_name, interval, trips);
2210
2211                 set_current_state(TASK_INTERRUPTIBLE);
2212                 schedule_timeout(interval);
2213                 if (interval < cfs_time_seconds(1))
2214                         interval *= 2;
2215
2216                 goto again;
2217         }
2218
2219         if (ktime_get_seconds() < ps->ps_next_retry) {
2220                 /* someone failed recently */
2221                 spin_unlock(&ps->ps_lock);
2222                 return NULL;
2223         }
2224
2225         ps->ps_increasing = 1;
2226         spin_unlock(&ps->ps_lock);
2227
2228         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
2229         time_before = ktime_get();
2230         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
2231         CDEBUG(D_NET, "ps_pool_create took %lld ms to complete",
2232                ktime_ms_delta(ktime_get(), time_before));
2233
2234         spin_lock(&ps->ps_lock);
2235         ps->ps_increasing = 0;
2236         if (rc == 0) {
2237                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
2238         } else {
2239                 ps->ps_next_retry = ktime_get_seconds() + IBLND_POOL_RETRY;
2240                 CERROR("Can't allocate new %s pool because out of memory\n",
2241                        ps->ps_name);
2242         }
2243         spin_unlock(&ps->ps_lock);
2244
2245         goto again;
2246 }
2247
2248 static void
2249 kiblnd_destroy_tx_pool(struct kib_pool *pool)
2250 {
2251         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool,
2252                                                tpo_pool);
2253         int i;
2254
2255         LASSERT (pool->po_allocated == 0);
2256
2257         if (tpo->tpo_tx_pages != NULL) {
2258                 kiblnd_unmap_tx_pool(tpo);
2259                 kiblnd_free_pages(tpo->tpo_tx_pages);
2260         }
2261
2262         if (tpo->tpo_tx_descs == NULL)
2263                 goto out;
2264
2265         for (i = 0; i < pool->po_size; i++) {
2266                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2267                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2268
2269                 list_del(&tx->tx_list);
2270                 if (tx->tx_pages != NULL)
2271                         LIBCFS_FREE(tx->tx_pages,
2272                                     LNET_MAX_IOV *
2273                                     sizeof(*tx->tx_pages));
2274                 if (tx->tx_frags != NULL)
2275                         LIBCFS_FREE(tx->tx_frags,
2276                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2277                                     sizeof(*tx->tx_frags));
2278                 if (tx->tx_wrq != NULL)
2279                         LIBCFS_FREE(tx->tx_wrq,
2280                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2281                                     sizeof(*tx->tx_wrq));
2282                 if (tx->tx_sge != NULL)
2283                         LIBCFS_FREE(tx->tx_sge,
2284                                     (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2285                                     sizeof(*tx->tx_sge));
2286                 if (tx->tx_rd != NULL)
2287                         LIBCFS_FREE(tx->tx_rd,
2288                                     offsetof(struct kib_rdma_desc,
2289                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2290         }
2291
2292         LIBCFS_FREE(tpo->tpo_tx_descs,
2293                     pool->po_size * sizeof(struct kib_tx));
2294 out:
2295         kiblnd_fini_pool(pool);
2296         LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2297 }
2298
2299 static int kiblnd_tx_pool_size(struct lnet_ni *ni, int ncpts)
2300 {
2301         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2302         int ntx;
2303
2304         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2305         ntx = tunables->lnd_ntx / ncpts;
2306
2307         return max(IBLND_TX_POOL, ntx);
2308 }
2309
2310 static int
2311 kiblnd_create_tx_pool(struct kib_poolset *ps, int size, struct kib_pool **pp_po)
2312 {
2313         int            i;
2314         int            npg;
2315         struct kib_pool *pool;
2316         struct kib_tx_pool *tpo;
2317
2318         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2319         if (tpo == NULL) {
2320                 CERROR("Failed to allocate TX pool\n");
2321                 return -ENOMEM;
2322         }
2323
2324         pool = &tpo->tpo_pool;
2325         kiblnd_init_pool(ps, pool, size);
2326         tpo->tpo_tx_descs = NULL;
2327         tpo->tpo_tx_pages = NULL;
2328
2329         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2330         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
2331                 CERROR("Can't allocate tx pages: %d\n", npg);
2332                 LIBCFS_FREE(tpo, sizeof(struct kib_tx_pool));
2333                 return -ENOMEM;
2334         }
2335
2336         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2337                          size * sizeof(struct kib_tx));
2338         if (tpo->tpo_tx_descs == NULL) {
2339                 CERROR("Can't allocate %d tx descriptors\n", size);
2340                 ps->ps_pool_destroy(pool);
2341                 return -ENOMEM;
2342         }
2343
2344         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2345
2346         for (i = 0; i < size; i++) {
2347                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2348                 int       wrq_sge = *kiblnd_tunables.kib_wrq_sge;
2349
2350                 tx->tx_pool = tpo;
2351                 if (ps->ps_net->ibn_fmr_ps != NULL) {
2352                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2353                                          lnet_cpt_table(), ps->ps_cpt,
2354                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2355                         if (tx->tx_pages == NULL)
2356                                 break;
2357                 }
2358
2359                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2360                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2361                                  sizeof(*tx->tx_frags));
2362                 if (tx->tx_frags == NULL)
2363                         break;
2364
2365                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2366
2367                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2368                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2369                                  sizeof(*tx->tx_wrq));
2370                 if (tx->tx_wrq == NULL)
2371                         break;
2372
2373                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2374                                  (1 + IBLND_MAX_RDMA_FRAGS) * wrq_sge *
2375                                  sizeof(*tx->tx_sge));
2376                 if (tx->tx_sge == NULL)
2377                         break;
2378
2379                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2380                                  offsetof(struct kib_rdma_desc,
2381                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2382                 if (tx->tx_rd == NULL)
2383                         break;
2384         }
2385
2386         if (i == size) {
2387                 kiblnd_map_tx_pool(tpo);
2388                 *pp_po = pool;
2389                 return 0;
2390         }
2391
2392         ps->ps_pool_destroy(pool);
2393         return -ENOMEM;
2394 }
2395
2396 static void
2397 kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2398 {
2399         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2400                                                   struct kib_tx_poolset,
2401                                                   tps_poolset);
2402         struct kib_tx *tx  = list_entry(node, struct kib_tx, tx_list);
2403
2404         tx->tx_cookie = tps->tps_next_tx_cookie++;
2405 }
2406
2407 static void
2408 kiblnd_net_fini_pools(struct kib_net *net)
2409 {
2410         int     i;
2411
2412         cfs_cpt_for_each(i, lnet_cpt_table()) {
2413                 struct kib_tx_poolset *tps;
2414                 struct kib_fmr_poolset *fps;
2415
2416                 if (net->ibn_tx_ps != NULL) {
2417                         tps = net->ibn_tx_ps[i];
2418                         kiblnd_fini_poolset(&tps->tps_poolset);
2419                 }
2420
2421                 if (net->ibn_fmr_ps != NULL) {
2422                         fps = net->ibn_fmr_ps[i];
2423                         kiblnd_fini_fmr_poolset(fps);
2424                 }
2425         }
2426
2427         if (net->ibn_tx_ps != NULL) {
2428                 cfs_percpt_free(net->ibn_tx_ps);
2429                 net->ibn_tx_ps = NULL;
2430         }
2431
2432         if (net->ibn_fmr_ps != NULL) {
2433                 cfs_percpt_free(net->ibn_fmr_ps);
2434                 net->ibn_fmr_ps = NULL;
2435         }
2436 }
2437
2438 static int
2439 kiblnd_net_init_pools(struct kib_net *net, struct lnet_ni *ni, __u32 *cpts,
2440                       int ncpts)
2441 {
2442         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2443 #ifdef HAVE_IB_GET_DMA_MR
2444         unsigned long   flags;
2445 #endif
2446         int             cpt;
2447         int             rc;
2448         int             i;
2449
2450         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2451
2452 #ifdef HAVE_IB_GET_DMA_MR
2453         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2454         /*
2455          * if lnd_map_on_demand is zero then we have effectively disabled
2456          * FMR or FastReg and we're using global memory regions
2457          * exclusively.
2458          */
2459         if (!tunables->lnd_map_on_demand) {
2460                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2461                                            flags);
2462                 goto create_tx_pool;
2463         }
2464
2465         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2466 #endif
2467
2468         if (tunables->lnd_fmr_pool_size < tunables->lnd_ntx / 4) {
2469                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2470                        tunables->lnd_fmr_pool_size,
2471                        tunables->lnd_ntx / 4);
2472                 rc = -EINVAL;
2473                 goto failed;
2474         }
2475
2476         /* TX pool must be created later than FMR, see LU-2268
2477          * for details */
2478         LASSERT(net->ibn_tx_ps == NULL);
2479
2480         /* premapping can fail if ibd_nmr > 1, so we always create
2481          * FMR pool and map-on-demand if premapping failed */
2482
2483         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2484                                            sizeof(struct kib_fmr_poolset));
2485         if (net->ibn_fmr_ps == NULL) {
2486                 CERROR("Failed to allocate FMR pool array\n");
2487                 rc = -ENOMEM;
2488                 goto failed;
2489         }
2490
2491         for (i = 0; i < ncpts; i++) {
2492                 cpt = (cpts == NULL) ? i : cpts[i];
2493                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2494                                              net, tunables);
2495                 if (rc != 0) {
2496                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2497                                cpt, rc);
2498                         goto failed;
2499                 }
2500         }
2501
2502         if (i > 0)
2503                 LASSERT(i == ncpts);
2504
2505 #ifdef HAVE_IB_GET_DMA_MR
2506  create_tx_pool:
2507 #endif
2508         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2509                                           sizeof(struct kib_tx_poolset));
2510         if (net->ibn_tx_ps == NULL) {
2511                 CERROR("Failed to allocate tx pool array\n");
2512                 rc = -ENOMEM;
2513                 goto failed;
2514         }
2515
2516         for (i = 0; i < ncpts; i++) {
2517                 cpt = (cpts == NULL) ? i : cpts[i];
2518                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2519                                          cpt, net, "TX",
2520                                          kiblnd_tx_pool_size(ni, ncpts),
2521                                          kiblnd_create_tx_pool,
2522                                          kiblnd_destroy_tx_pool,
2523                                          kiblnd_tx_init, NULL);
2524                 if (rc != 0) {
2525                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2526                                cpt, rc);
2527                         goto failed;
2528                 }
2529         }
2530
2531         return 0;
2532  failed:
2533         kiblnd_net_fini_pools(net);
2534         LASSERT(rc != 0);
2535         return rc;
2536 }
2537
2538 static int
2539 kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2540 {
2541         struct ib_device_attr *dev_attr;
2542         int rc = 0;
2543
2544         /* It's safe to assume a HCA can handle a page size
2545          * matching that of the native system */
2546         hdev->ibh_page_shift = PAGE_SHIFT;
2547         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2548         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2549
2550 #ifndef HAVE_IB_DEVICE_ATTRS
2551         LIBCFS_ALLOC(dev_attr, sizeof(*dev_attr));
2552         if (dev_attr == NULL) {
2553                 CERROR("Out of memory\n");
2554                 return -ENOMEM;
2555         }
2556
2557         rc = ib_query_device(hdev->ibh_ibdev, dev_attr);
2558         if (rc != 0) {
2559                 CERROR("Failed to query IB device: %d\n", rc);
2560                 goto out_clean_attr;
2561         }
2562 #else
2563         dev_attr = &hdev->ibh_ibdev->attrs;
2564 #endif
2565
2566         hdev->ibh_mr_size = dev_attr->max_mr_size;
2567         hdev->ibh_max_qp_wr = dev_attr->max_qp_wr;
2568
2569         /* Setup device Memory Registration capabilities */
2570 #ifdef HAVE_IB_DEVICE_OPS
2571         if (hdev->ibh_ibdev->ops.alloc_fmr &&
2572             hdev->ibh_ibdev->ops.dealloc_fmr &&
2573             hdev->ibh_ibdev->ops.map_phys_fmr &&
2574             hdev->ibh_ibdev->ops.unmap_fmr) {
2575 #else
2576         if (hdev->ibh_ibdev->alloc_fmr &&
2577             hdev->ibh_ibdev->dealloc_fmr &&
2578             hdev->ibh_ibdev->map_phys_fmr &&
2579             hdev->ibh_ibdev->unmap_fmr) {
2580 #endif
2581                 LCONSOLE_INFO("Using FMR for registration\n");
2582                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FMR_ENABLED;
2583         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
2584                 LCONSOLE_INFO("Using FastReg for registration\n");
2585                 hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_ENABLED;
2586 #ifndef HAVE_IB_ALLOC_FAST_REG_MR
2587 #ifdef IB_DEVICE_SG_GAPS_REG
2588                 if (dev_attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
2589                         hdev->ibh_dev->ibd_dev_caps |= IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT;
2590 #endif
2591 #endif
2592         } else {
2593                 rc = -ENOSYS;
2594         }
2595
2596         if (rc == 0 && hdev->ibh_mr_size == ~0ULL)
2597                 hdev->ibh_mr_shift = 64;
2598         else if (rc != 0)
2599                 rc = -EINVAL;
2600
2601 #ifndef HAVE_IB_DEVICE_ATTRS
2602 out_clean_attr:
2603         LIBCFS_FREE(dev_attr, sizeof(*dev_attr));
2604 #endif
2605
2606         if (rc == -ENOSYS)
2607                 CERROR("IB device does not support FMRs nor FastRegs, can't "
2608                        "register memory: %d\n", rc);
2609         else if (rc == -EINVAL)
2610                 CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2611         return rc;
2612 }
2613
2614 #ifdef HAVE_IB_GET_DMA_MR
2615 static void
2616 kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2617 {
2618         if (hdev->ibh_mrs == NULL)
2619                 return;
2620
2621         ib_dereg_mr(hdev->ibh_mrs);
2622
2623         hdev->ibh_mrs = NULL;
2624 }
2625 #endif
2626
2627 void
2628 kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2629 {
2630 #ifdef HAVE_IB_GET_DMA_MR
2631         kiblnd_hdev_cleanup_mrs(hdev);
2632 #endif
2633
2634         if (hdev->ibh_pd != NULL)
2635                 ib_dealloc_pd(hdev->ibh_pd);
2636
2637         if (hdev->ibh_cmid != NULL)
2638                 rdma_destroy_id(hdev->ibh_cmid);
2639
2640         LIBCFS_FREE(hdev, sizeof(*hdev));
2641 }
2642
2643 #ifdef HAVE_IB_GET_DMA_MR
2644 static int
2645 kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2646 {
2647         struct ib_mr *mr;
2648         int           acflags = IB_ACCESS_LOCAL_WRITE |
2649                                 IB_ACCESS_REMOTE_WRITE;
2650
2651         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2652         if (IS_ERR(mr)) {
2653                 CERROR("Failed ib_get_dma_mr: %ld\n", PTR_ERR(mr));
2654                 kiblnd_hdev_cleanup_mrs(hdev);
2655                 return PTR_ERR(mr);
2656         }
2657
2658         hdev->ibh_mrs = mr;
2659
2660         return 0;
2661 }
2662 #endif
2663
2664 static int
2665 kiblnd_dummy_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2666 {       /* DUMMY */
2667         return 0;
2668 }
2669
2670 static int
2671 kiblnd_dev_need_failover(struct kib_dev *dev, struct net *ns)
2672 {
2673         struct rdma_cm_id  *cmid;
2674         struct sockaddr_in  srcaddr;
2675         struct sockaddr_in  dstaddr;
2676         int                 rc;
2677
2678         if (dev->ibd_hdev == NULL || /* initializing */
2679             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2680             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2681                 return 1;
2682
2683         /* XXX: it's UGLY, but I don't have better way to find
2684          * ib-bonding HCA failover because:
2685          *
2686          * a. no reliable CM event for HCA failover...
2687          * b. no OFED API to get ib_device for current net_device...
2688          *
2689          * We have only two choices at this point:
2690          *
2691          * a. rdma_bind_addr(), it will conflict with listener cmid
2692          * b. rdma_resolve_addr() to zero addr */
2693         cmid = kiblnd_rdma_create_id(ns, kiblnd_dummy_callback, dev,
2694                                      RDMA_PS_TCP, IB_QPT_RC);
2695         if (IS_ERR(cmid)) {
2696                 rc = PTR_ERR(cmid);
2697                 CERROR("Failed to create cmid for failover: %d\n", rc);
2698                 return rc;
2699         }
2700
2701         memset(&srcaddr, 0, sizeof(srcaddr));
2702         srcaddr.sin_family      = AF_INET;
2703         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2704
2705         memset(&dstaddr, 0, sizeof(dstaddr));
2706         dstaddr.sin_family = AF_INET;
2707         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2708                                (struct sockaddr *)&dstaddr, 1);
2709         if (rc != 0 || cmid->device == NULL) {
2710                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2711                        dev->ibd_ifname, &dev->ibd_ifip,
2712                        cmid->device, rc);
2713                 rdma_destroy_id(cmid);
2714                 return rc;
2715         }
2716
2717         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2718         rdma_destroy_id(cmid);
2719         return rc;
2720 }
2721
2722 int
2723 kiblnd_dev_failover(struct kib_dev *dev, struct net *ns)
2724 {
2725         struct list_head    zombie_tpo = LIST_HEAD_INIT(zombie_tpo);
2726         struct list_head    zombie_ppo = LIST_HEAD_INIT(zombie_ppo);
2727         struct list_head    zombie_fpo = LIST_HEAD_INIT(zombie_fpo);
2728         struct rdma_cm_id  *cmid  = NULL;
2729         struct kib_hca_dev *hdev  = NULL;
2730         struct kib_hca_dev *old;
2731         struct ib_pd       *pd;
2732         struct kib_net *net;
2733         struct sockaddr_in  addr;
2734         unsigned long       flags;
2735         int                 rc = 0;
2736         int                 i;
2737
2738         LASSERT (*kiblnd_tunables.kib_dev_failover > 1 ||
2739                  dev->ibd_can_failover ||
2740                  dev->ibd_hdev == NULL);
2741
2742         rc = kiblnd_dev_need_failover(dev, ns);
2743         if (rc <= 0)
2744                 goto out;
2745
2746         if (dev->ibd_hdev != NULL &&
2747             dev->ibd_hdev->ibh_cmid != NULL) {
2748                 /* XXX it's not good to close old listener at here,
2749                  * because we can fail to create new listener.
2750                  * But we have to close it now, otherwise rdma_bind_addr
2751                  * will return EADDRINUSE... How crap! */
2752                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2753
2754                 cmid = dev->ibd_hdev->ibh_cmid;
2755                 /* make next schedule of kiblnd_dev_need_failover()
2756                  * return 1 for me */
2757                 dev->ibd_hdev->ibh_cmid  = NULL;
2758                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2759
2760                 rdma_destroy_id(cmid);
2761         }
2762
2763         cmid = kiblnd_rdma_create_id(ns, kiblnd_cm_callback, dev, RDMA_PS_TCP,
2764                                      IB_QPT_RC);
2765         if (IS_ERR(cmid)) {
2766                 rc = PTR_ERR(cmid);
2767                 CERROR("Failed to create cmid for failover: %d\n", rc);
2768                 goto out;
2769         }
2770
2771         memset(&addr, 0, sizeof(addr));
2772         addr.sin_family      = AF_INET;
2773         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2774         addr.sin_port        = htons(*kiblnd_tunables.kib_service);
2775
2776         /* Bind to failover device or port */
2777         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2778         if (rc != 0 || cmid->device == NULL) {
2779                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2780                        dev->ibd_ifname, &dev->ibd_ifip,
2781                        cmid->device, rc);
2782                 rdma_destroy_id(cmid);
2783                 goto out;
2784         }
2785
2786         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2787         if (hdev == NULL) {
2788                 CERROR("Failed to allocate kib_hca_dev\n");
2789                 rdma_destroy_id(cmid);
2790                 rc = -ENOMEM;
2791                 goto out;
2792         }
2793
2794         atomic_set(&hdev->ibh_ref, 1);
2795         hdev->ibh_dev   = dev;
2796         hdev->ibh_cmid  = cmid;
2797         hdev->ibh_ibdev = cmid->device;
2798
2799 #ifdef HAVE_IB_ALLOC_PD_2ARGS
2800         pd = ib_alloc_pd(cmid->device, 0);
2801 #else
2802         pd = ib_alloc_pd(cmid->device);
2803 #endif
2804         if (IS_ERR(pd)) {
2805                 rc = PTR_ERR(pd);
2806                 CERROR("Can't allocate PD: %d\n", rc);
2807                 goto out;
2808         }
2809
2810         hdev->ibh_pd = pd;
2811
2812         rc = rdma_listen(cmid, 0);
2813         if (rc != 0) {
2814                 CERROR("Can't start new listener: %d\n", rc);
2815                 goto out;
2816         }
2817
2818         rc = kiblnd_hdev_get_attr(hdev);
2819         if (rc != 0) {
2820                 CERROR("Can't get device attributes: %d\n", rc);
2821                 goto out;
2822         }
2823
2824 #ifdef HAVE_IB_GET_DMA_MR
2825         rc = kiblnd_hdev_setup_mrs(hdev);
2826         if (rc != 0) {
2827                 CERROR("Can't setup device: %d\n", rc);
2828                 goto out;
2829         }
2830 #endif
2831
2832         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2833
2834         old = dev->ibd_hdev;
2835         dev->ibd_hdev = hdev;   /* take over the refcount */
2836         hdev = old;
2837
2838         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2839                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2840                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2841                                             &zombie_tpo);
2842
2843                         if (net->ibn_fmr_ps != NULL)
2844                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2845                                                         &zombie_fpo);
2846                 }
2847         }
2848
2849         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2850  out:
2851         if (!list_empty(&zombie_tpo))
2852                 kiblnd_destroy_pool_list(&zombie_tpo);
2853         if (!list_empty(&zombie_ppo))
2854                 kiblnd_destroy_pool_list(&zombie_ppo);
2855         if (!list_empty(&zombie_fpo))
2856                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2857         if (hdev != NULL)
2858                 kiblnd_hdev_decref(hdev);
2859
2860         if (rc != 0)
2861                 dev->ibd_failed_failover++;
2862         else
2863                 dev->ibd_failed_failover = 0;
2864
2865         return rc;
2866 }
2867
2868 void
2869 kiblnd_destroy_dev(struct kib_dev *dev)
2870 {
2871         LASSERT(dev->ibd_nnets == 0);
2872         LASSERT(list_empty(&dev->ibd_nets));
2873
2874         list_del(&dev->ibd_fail_list);
2875         list_del(&dev->ibd_list);
2876
2877         if (dev->ibd_hdev != NULL)
2878                 kiblnd_hdev_decref(dev->ibd_hdev);
2879
2880         LIBCFS_FREE(dev, sizeof(*dev));
2881 }
2882
2883 static void
2884 kiblnd_base_shutdown(void)
2885 {
2886         struct kib_sched_info   *sched;
2887         int                     i;
2888
2889         LASSERT(list_empty(&kiblnd_data.kib_devs));
2890
2891         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
2892                atomic_read(&libcfs_kmemory));
2893
2894         switch (kiblnd_data.kib_init) {
2895         default:
2896                 LBUG();
2897
2898         case IBLND_INIT_ALL:
2899         case IBLND_INIT_DATA:
2900                 LASSERT (kiblnd_data.kib_peers != NULL);
2901                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
2902                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2903                 }
2904                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2905                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2906                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2907                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2908
2909                 /* flag threads to terminate; wake and wait for them to die */
2910                 kiblnd_data.kib_shutdown = 1;
2911
2912                 /* NB: we really want to stop scheduler threads net by net
2913                  * instead of the whole module, this should be improved
2914                  * with dynamic configuration LNet */
2915                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2916                         wake_up_all(&sched->ibs_waitq);
2917
2918                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2919                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2920
2921                 i = 2;
2922                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2923                         i++;
2924                         /* power of 2? */
2925                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2926                                "Waiting for %d threads to terminate\n",
2927                                atomic_read(&kiblnd_data.kib_nthreads));
2928                         set_current_state(TASK_UNINTERRUPTIBLE);
2929                         schedule_timeout(cfs_time_seconds(1));
2930                 }
2931
2932                 /* fall through */
2933
2934         case IBLND_INIT_NOTHING:
2935                 break;
2936         }
2937
2938         if (kiblnd_data.kib_peers != NULL) {
2939                 LIBCFS_FREE(kiblnd_data.kib_peers,
2940                             sizeof(struct list_head) *
2941                             kiblnd_data.kib_peer_hash_size);
2942         }
2943
2944         if (kiblnd_data.kib_scheds != NULL)
2945                 cfs_percpt_free(kiblnd_data.kib_scheds);
2946
2947         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
2948                atomic_read(&libcfs_kmemory));
2949
2950         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2951         module_put(THIS_MODULE);
2952 }
2953
2954 static void
2955 kiblnd_shutdown(struct lnet_ni *ni)
2956 {
2957         struct kib_net *net = ni->ni_data;
2958         rwlock_t     *g_lock = &kiblnd_data.kib_global_lock;
2959         int               i;
2960         unsigned long     flags;
2961
2962         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2963
2964         if (net == NULL)
2965                 goto out;
2966
2967         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
2968                atomic_read(&libcfs_kmemory));
2969
2970         write_lock_irqsave(g_lock, flags);
2971         net->ibn_shutdown = 1;
2972         write_unlock_irqrestore(g_lock, flags);
2973
2974         switch (net->ibn_init) {
2975         default:
2976                 LBUG();
2977
2978         case IBLND_INIT_ALL:
2979                 /* nuke all existing peers within this net */
2980                 kiblnd_del_peer(ni, LNET_NID_ANY);
2981
2982                 /* Wait for all peer_ni state to clean up */
2983                 i = 2;
2984                 while (atomic_read(&net->ibn_npeers) != 0) {
2985                         i++;
2986                         /* power of 2? */
2987                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2988                                "%s: waiting for %d peers to disconnect\n",
2989                                libcfs_nid2str(ni->ni_nid),
2990                                atomic_read(&net->ibn_npeers));
2991                         set_current_state(TASK_UNINTERRUPTIBLE);
2992                         schedule_timeout(cfs_time_seconds(1));
2993                 }
2994
2995                 kiblnd_net_fini_pools(net);
2996
2997                 write_lock_irqsave(g_lock, flags);
2998                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2999                 net->ibn_dev->ibd_nnets--;
3000                 list_del(&net->ibn_list);
3001                 write_unlock_irqrestore(g_lock, flags);
3002
3003                 /* fall through */
3004
3005         case IBLND_INIT_NOTHING:
3006                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
3007
3008                 if (net->ibn_dev != NULL &&
3009                     net->ibn_dev->ibd_nnets == 0)
3010                         kiblnd_destroy_dev(net->ibn_dev);
3011
3012                 break;
3013         }
3014
3015         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
3016                atomic_read(&libcfs_kmemory));
3017
3018         net->ibn_init = IBLND_INIT_NOTHING;
3019         ni->ni_data = NULL;
3020
3021         LIBCFS_FREE(net, sizeof(*net));
3022
3023 out:
3024         if (list_empty(&kiblnd_data.kib_devs))
3025                 kiblnd_base_shutdown();
3026         return;
3027 }
3028
3029 static int
3030 kiblnd_base_startup(struct net *ns)
3031 {
3032         struct kib_sched_info   *sched;
3033         int                     rc;
3034         int                     i;
3035
3036         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
3037
3038         try_module_get(THIS_MODULE);
3039         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
3040
3041         rwlock_init(&kiblnd_data.kib_global_lock);
3042
3043         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
3044         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
3045
3046         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
3047         LIBCFS_ALLOC(kiblnd_data.kib_peers,
3048                      sizeof(struct list_head) *
3049                      kiblnd_data.kib_peer_hash_size);
3050         if (kiblnd_data.kib_peers == NULL)
3051                 goto failed;
3052
3053         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
3054                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
3055
3056         spin_lock_init(&kiblnd_data.kib_connd_lock);
3057         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
3058         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
3059         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
3060         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
3061
3062         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
3063         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
3064
3065         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
3066                                                   sizeof(*sched));
3067         if (kiblnd_data.kib_scheds == NULL)
3068                 goto failed;
3069
3070         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
3071                 int     nthrs;
3072
3073                 spin_lock_init(&sched->ibs_lock);
3074                 INIT_LIST_HEAD(&sched->ibs_conns);
3075                 init_waitqueue_head(&sched->ibs_waitq);
3076
3077                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
3078                 if (*kiblnd_tunables.kib_nscheds > 0) {
3079                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
3080                 } else {
3081                         /* max to half of CPUs, another half is reserved for
3082                          * upper layer modules */
3083                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3084                 }
3085
3086                 sched->ibs_nthreads_max = nthrs;
3087                 sched->ibs_cpt = i;
3088         }
3089
3090         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
3091
3092         /* lists/ptrs/locks initialised */
3093         kiblnd_data.kib_init = IBLND_INIT_DATA;
3094         /*****************************************************/
3095
3096         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
3097         if (rc != 0) {
3098                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
3099                 goto failed;
3100         }
3101
3102         if (*kiblnd_tunables.kib_dev_failover != 0)
3103                 rc = kiblnd_thread_start(kiblnd_failover_thread, ns,
3104                                          "kiblnd_failover");
3105
3106         if (rc != 0) {
3107                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
3108                 goto failed;
3109         }
3110
3111         /* flag everything initialised */
3112         kiblnd_data.kib_init = IBLND_INIT_ALL;
3113         /*****************************************************/
3114
3115         return 0;
3116
3117  failed:
3118         kiblnd_base_shutdown();
3119         return -ENETDOWN;
3120 }
3121
3122 static int
3123 kiblnd_start_schedulers(struct kib_sched_info *sched)
3124 {
3125         int     rc = 0;
3126         int     nthrs;
3127         int     i;
3128
3129         if (sched->ibs_nthreads == 0) {
3130                 if (*kiblnd_tunables.kib_nscheds > 0) {
3131                         nthrs = sched->ibs_nthreads_max;
3132                 } else {
3133                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
3134                                                sched->ibs_cpt);
3135                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
3136                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
3137                 }
3138         } else {
3139                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
3140                 /* increase one thread if there is new interface */
3141                 nthrs = (sched->ibs_nthreads < sched->ibs_nthreads_max);
3142         }
3143
3144         for (i = 0; i < nthrs; i++) {
3145                 long    id;
3146                 char    name[20];
3147                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
3148                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
3149                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
3150                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
3151                 if (rc == 0)
3152                         continue;
3153
3154                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
3155                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
3156                 break;
3157         }
3158
3159         sched->ibs_nthreads += i;
3160         return rc;
3161 }
3162
3163 static int kiblnd_dev_start_threads(struct kib_dev *dev, bool newdev, u32 *cpts,
3164                                     int ncpts)
3165 {
3166         int     cpt;
3167         int     rc;
3168         int     i;
3169
3170         for (i = 0; i < ncpts; i++) {
3171                 struct kib_sched_info *sched;
3172
3173                 cpt = (cpts == NULL) ? i : cpts[i];
3174                 sched = kiblnd_data.kib_scheds[cpt];
3175
3176                 if (!newdev && sched->ibs_nthreads > 0)
3177                         continue;
3178
3179                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
3180                 if (rc != 0) {
3181                         CERROR("Failed to start scheduler threads for %s\n",
3182                                dev->ibd_ifname);
3183                         return rc;
3184                 }
3185         }
3186         return 0;
3187 }
3188
3189 static struct kib_dev *
3190 kiblnd_dev_search(char *ifname)
3191 {
3192         struct kib_dev *alias = NULL;
3193         struct kib_dev *dev;
3194         char            *colon;
3195         char            *colon2;
3196
3197         colon = strchr(ifname, ':');
3198         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3199                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3200                         return dev;
3201
3202                 if (alias != NULL)
3203                         continue;
3204
3205                 colon2 = strchr(dev->ibd_ifname, ':');
3206                 if (colon != NULL)
3207                         *colon = 0;
3208                 if (colon2 != NULL)
3209                         *colon2 = 0;
3210
3211                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
3212                         alias = dev;
3213
3214                 if (colon != NULL)
3215                         *colon = ':';
3216                 if (colon2 != NULL)
3217                         *colon2 = ':';
3218         }
3219         return alias;
3220 }
3221
3222 static int
3223 kiblnd_startup(struct lnet_ni *ni)
3224 {
3225         char *ifname = NULL;
3226         struct lnet_inetdev *ifaces = NULL;
3227         struct kib_dev *ibdev = NULL;
3228         struct kib_net *net = NULL;
3229         unsigned long flags;
3230         int rc;
3231         int i;
3232         bool newdev;
3233
3234         LASSERT(ni->ni_net->net_lnd == &the_o2iblnd);
3235
3236         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
3237                 rc = kiblnd_base_startup(ni->ni_net_ns);
3238                 if (rc != 0)
3239                         return rc;
3240         }
3241
3242         LIBCFS_ALLOC(net, sizeof(*net));
3243         ni->ni_data = net;
3244         if (net == NULL) {
3245                 rc = -ENOMEM;
3246                 goto failed;
3247         }
3248
3249         net->ibn_incarnation = ktime_get_real_ns() / NSEC_PER_USEC;
3250
3251         kiblnd_tunables_setup(ni);
3252
3253         /*
3254          * ni_interfaces is only to support legacy pre Multi-Rail
3255          * tcp bonding for ksocklnd. Multi-Rail wants each secondary
3256          * IP to be treated as an unique 'struct ni' interfaces instead.
3257          */
3258         if (ni->ni_interfaces[0] != NULL) {
3259                 /* Use the IPoIB interface specified in 'networks=' */
3260                 if (ni->ni_interfaces[1] != NULL) {
3261                         CERROR("ko2iblnd: Multiple interfaces not supported\n");
3262                         rc = -EINVAL;
3263                         goto failed;
3264                 }
3265
3266                 ifname = ni->ni_interfaces[0];
3267         } else {
3268                 ifname = *kiblnd_tunables.kib_default_ipif;
3269         }
3270
3271         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
3272                 CERROR("IPoIB interface name too long: %s\n", ifname);
3273                 rc = -E2BIG;
3274                 goto failed;
3275         }
3276
3277         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
3278         if (rc < 0)
3279                 goto failed;
3280
3281         for (i = 0; i < rc; i++) {
3282                 if (strcmp(ifname, ifaces[i].li_name) == 0)
3283                         break;
3284         }
3285
3286         if (i == rc) {
3287                 CERROR("ko2iblnd: No matching interfaces\n");
3288                 rc = -ENOENT;
3289                 goto failed;
3290         }
3291
3292         ibdev = kiblnd_dev_search(ifname);
3293         newdev = ibdev == NULL;
3294         /* hmm...create kib_dev even for alias */
3295         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0) {
3296                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
3297                 if (!ibdev) {
3298                         rc = -ENOMEM;
3299                         goto failed;
3300                 }
3301
3302                 ibdev->ibd_ifip = ifaces[i].li_ipaddr;
3303                 strlcpy(ibdev->ibd_ifname, ifaces[i].li_name,
3304                         sizeof(ibdev->ibd_ifname));
3305                 ibdev->ibd_can_failover = !!(ifaces[i].li_flags & IFF_MASTER);
3306
3307                 INIT_LIST_HEAD(&ibdev->ibd_nets);
3308                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
3309                 INIT_LIST_HEAD(&ibdev->ibd_fail_list);
3310
3311                 /* initialize the device */
3312                 rc = kiblnd_dev_failover(ibdev, ni->ni_net_ns);
3313                 if (rc) {
3314                         CERROR("ko2iblnd: Can't initialize device: rc = %d\n", rc);
3315                         goto failed;
3316                 }
3317
3318                 list_add_tail(&ibdev->ibd_list, &kiblnd_data.kib_devs);
3319         }
3320
3321         net->ibn_dev = ibdev;
3322         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
3323
3324         ni->ni_dev_cpt = ifaces[i].li_cpt;
3325
3326         rc = kiblnd_dev_start_threads(ibdev, newdev, ni->ni_cpts, ni->ni_ncpts);
3327         if (rc != 0)
3328                 goto failed;
3329
3330         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
3331         if (rc != 0) {
3332                 CERROR("Failed to initialize NI pools: %d\n", rc);
3333                 goto failed;
3334         }
3335
3336         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3337         ibdev->ibd_nnets++;
3338         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
3339         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3340
3341         net->ibn_init = IBLND_INIT_ALL;
3342
3343         return 0;
3344
3345 failed:
3346         if (net != NULL && net->ibn_dev == NULL && ibdev != NULL)
3347                 kiblnd_destroy_dev(ibdev);
3348
3349         kfree(ifaces);
3350         kiblnd_shutdown(ni);
3351
3352         CDEBUG(D_NET, "Configuration of device %s failed: rc = %d\n",
3353                ifname ? ifname : "", rc);
3354
3355         return -ENETDOWN;
3356 }
3357
3358 static struct lnet_lnd the_o2iblnd = {
3359         .lnd_type       = O2IBLND,
3360         .lnd_startup    = kiblnd_startup,
3361         .lnd_shutdown   = kiblnd_shutdown,
3362         .lnd_ctl        = kiblnd_ctl,
3363         .lnd_query      = kiblnd_query,
3364         .lnd_send       = kiblnd_send,
3365         .lnd_recv       = kiblnd_recv,
3366 };
3367
3368 static void __exit ko2iblnd_exit(void)
3369 {
3370         lnet_unregister_lnd(&the_o2iblnd);
3371 }
3372
3373 static int __init ko2iblnd_init(void)
3374 {
3375         int rc;
3376
3377         CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3378         CLASSERT(offsetof(struct kib_msg,
3379                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS]) <=
3380                  IBLND_MSG_SIZE);
3381         CLASSERT(offsetof(struct kib_msg,
3382                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3383                  <= IBLND_MSG_SIZE);
3384
3385         rc = kiblnd_tunables_init();
3386         if (rc != 0)
3387                 return rc;
3388
3389         lnet_register_lnd(&the_o2iblnd);
3390
3391         return 0;
3392 }
3393
3394 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3395 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3396 MODULE_VERSION("2.8.0");
3397 MODULE_LICENSE("GPL");
3398
3399 module_init(ko2iblnd_init);
3400 module_exit(ko2iblnd_exit);