drivers/staging/lustre/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, 2015, 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/div64.h>
  38 #include <asm/page.h>
  39 #include "o2iblnd.h"
  40
  41 static lnd_t the_o2iblnd;
  42
  43 struct kib_data kiblnd_data;
  44
  45 static __u32 kiblnd_cksum(void *ptr, int nob)
  46 {
  47         char *c = ptr;
  48         __u32 sum = 0;
  49
  50         while (nob-- > 0)
  51                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  52
  53         /* ensure I don't return 0 (== no checksum) */
  54         return !sum ? 1 : sum;
  55 }
  56
  57 static char *kiblnd_msgtype2str(int type)
  58 {
  59         switch (type) {
  60         case IBLND_MSG_CONNREQ:
  61                 return "CONNREQ";
  62
  63         case IBLND_MSG_CONNACK:
  64                 return "CONNACK";
  65
  66         case IBLND_MSG_NOOP:
  67                 return "NOOP";
  68
  69         case IBLND_MSG_IMMEDIATE:
  70                 return "IMMEDIATE";
  71
  72         case IBLND_MSG_PUT_REQ:
  73                 return "PUT_REQ";
  74
  75         case IBLND_MSG_PUT_NAK:
  76                 return "PUT_NAK";
  77
  78         case IBLND_MSG_PUT_ACK:
  79                 return "PUT_ACK";
  80
  81         case IBLND_MSG_PUT_DONE:
  82                 return "PUT_DONE";
  83
  84         case IBLND_MSG_GET_REQ:
  85                 return "GET_REQ";
  86
  87         case IBLND_MSG_GET_DONE:
  88                 return "GET_DONE";
  89
  90         default:
  91                 return "???";
  92         }
  93 }
  94
  95 static int kiblnd_msgtype2size(int type)
  96 {
  97         const int hdr_size = offsetof(struct kib_msg, ibm_u);
  98
  99         switch (type) {
 100         case IBLND_MSG_CONNREQ:
 101         case IBLND_MSG_CONNACK:
 102                 return hdr_size + sizeof(struct kib_connparams);
 103
 104         case IBLND_MSG_NOOP:
 105                 return hdr_size;
 106
 107         case IBLND_MSG_IMMEDIATE:
 108                 return offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[0]);
 109
 110         case IBLND_MSG_PUT_REQ:
 111                 return hdr_size + sizeof(struct kib_putreq_msg);
 112
 113         case IBLND_MSG_PUT_ACK:
 114                 return hdr_size + sizeof(struct kib_putack_msg);
 115
 116         case IBLND_MSG_GET_REQ:
 117                 return hdr_size + sizeof(struct kib_get_msg);
 118
 119         case IBLND_MSG_PUT_NAK:
 120         case IBLND_MSG_PUT_DONE:
 121         case IBLND_MSG_GET_DONE:
 122                 return hdr_size + sizeof(struct kib_completion_msg);
 123         default:
 124                 return -1;
 125         }
 126 }
 127
 128 static int kiblnd_unpack_rd(struct kib_msg *msg, int flip)
 129 {
 130         struct kib_rdma_desc *rd;
 131         int nob;
 132         int n;
 133         int i;
 134
 135         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 136                 msg->ibm_type == IBLND_MSG_PUT_ACK);
 137
 138         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 139                               &msg->ibm_u.get.ibgm_rd :
 140                               &msg->ibm_u.putack.ibpam_rd;
 141
 142         if (flip) {
 143                 __swab32s(&rd->rd_key);
 144                 __swab32s(&rd->rd_nfrags);
 145         }
 146
 147         n = rd->rd_nfrags;
 148
 149         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 150                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 151                        n, IBLND_MAX_RDMA_FRAGS);
 152                 return 1;
 153         }
 154
 155         nob = offsetof(struct kib_msg, ibm_u) +
 156               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 157
 158         if (msg->ibm_nob < nob) {
 159                 CERROR("Short %s: %d(%d)\n",
 160                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 161                 return 1;
 162         }
 163
 164         if (!flip)
 165                 return 0;
 166
 167         for (i = 0; i < n; i++) {
 168                 __swab32s(&rd->rd_frags[i].rf_nob);
 169                 __swab64s(&rd->rd_frags[i].rf_addr);
 170         }
 171
 172         return 0;
 173 }
 174
 175 void kiblnd_pack_msg(lnet_ni_t *ni, struct kib_msg *msg, int version,
 176                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 177 {
 178         struct kib_net *net = ni->ni_data;
 179
 180         /*
 181          * CAVEAT EMPTOR! all message fields not set here should have been
 182          * initialised previously.
 183          */
 184         msg->ibm_magic    = IBLND_MSG_MAGIC;
 185         msg->ibm_version  = version;
 186         /*   ibm_type */
 187         msg->ibm_credits  = credits;
 188         /*   ibm_nob */
 189         msg->ibm_cksum    = 0;
 190         msg->ibm_srcnid   = ni->ni_nid;
 191         msg->ibm_srcstamp = net->ibn_incarnation;
 192         msg->ibm_dstnid   = dstnid;
 193         msg->ibm_dststamp = dststamp;
 194
 195         if (*kiblnd_tunables.kib_cksum) {
 196                 /* NB ibm_cksum zero while computing cksum */
 197                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 198         }
 199 }
 200
 201 int kiblnd_unpack_msg(struct kib_msg *msg, int nob)
 202 {
 203         const int hdr_size = offsetof(struct kib_msg, ibm_u);
 204         __u32 msg_cksum;
 205         __u16 version;
 206         int msg_nob;
 207         int flip;
 208
 209         /* 6 bytes are enough to have received magic + version */
 210         if (nob < 6) {
 211                 CERROR("Short message: %d\n", nob);
 212                 return -EPROTO;
 213         }
 214
 215         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 216                 flip = 0;
 217         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 218                 flip = 1;
 219         } else {
 220                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 221                 return -EPROTO;
 222         }
 223
 224         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 225         if (version != IBLND_MSG_VERSION &&
 226             version != IBLND_MSG_VERSION_1) {
 227                 CERROR("Bad version: %x\n", version);
 228                 return -EPROTO;
 229         }
 230
 231         if (nob < hdr_size) {
 232                 CERROR("Short message: %d\n", nob);
 233                 return -EPROTO;
 234         }
 235
 236         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 237         if (msg_nob > nob) {
 238                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 239                 return -EPROTO;
 240         }
 241
 242         /*
 243          * checksum must be computed with ibm_cksum zero and BEFORE anything
 244          * gets flipped
 245          */
 246         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 247         msg->ibm_cksum = 0;
 248         if (msg_cksum &&
 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 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 kiblnd_create_peer(lnet_ni_t *ni, struct kib_peer **peerp, lnet_nid_t nid)
 315 {
 316         struct kib_peer *peer;
 317         struct kib_net *net = ni->ni_data;
 318         int cpt = lnet_cpt_of_nid(nid);
 319         unsigned long flags;
 320
 321         LASSERT(net);
 322         LASSERT(nid != LNET_NID_ANY);
 323
 324         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 325         if (!peer) {
 326                 CERROR("Cannot allocate peer\n");
 327                 return -ENOMEM;
 328         }
 329
 330         peer->ibp_ni = ni;
 331         peer->ibp_nid = nid;
 332         peer->ibp_error = 0;
 333         peer->ibp_last_alive = 0;
 334         peer->ibp_max_frags = kiblnd_cfg_rdma_frags(peer->ibp_ni);
 335         peer->ibp_queue_depth = ni->ni_peertxcredits;
 336         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 337
 338         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 339         INIT_LIST_HEAD(&peer->ibp_conns);
 340         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 341
 342         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 343
 344         /* always called with a ref on ni, which prevents ni being shutdown */
 345         LASSERT(!net->ibn_shutdown);
 346
 347         /* npeers only grows with the global lock held */
 348         atomic_inc(&net->ibn_npeers);
 349
 350         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 351
 352         *peerp = peer;
 353         return 0;
 354 }
 355
 356 void kiblnd_destroy_peer(struct kib_peer *peer)
 357 {
 358         struct kib_net *net = peer->ibp_ni->ni_data;
 359
 360         LASSERT(net);
 361         LASSERT(!atomic_read(&peer->ibp_refcount));
 362         LASSERT(!kiblnd_peer_active(peer));
 363         LASSERT(kiblnd_peer_idle(peer));
 364         LASSERT(list_empty(&peer->ibp_tx_queue));
 365
 366         LIBCFS_FREE(peer, sizeof(*peer));
 367
 368         /*
 369          * NB a peer's connections keep a reference on their peer until
 370          * they are destroyed, so we can be assured that _all_ state to do
 371          * with this peer has been cleaned up when its refcount drops to
 372          * zero.
 373          */
 374         atomic_dec(&net->ibn_npeers);
 375 }
 376
 377 struct kib_peer *kiblnd_find_peer_locked(lnet_nid_t nid)
 378 {
 379         /*
 380          * the caller is responsible for accounting the additional reference
 381          * that this creates
 382          */
 383         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 384         struct list_head *tmp;
 385         struct kib_peer *peer;
 386
 387         list_for_each(tmp, peer_list) {
 388                 peer = list_entry(tmp, struct kib_peer, ibp_list);
 389                 LASSERT(!kiblnd_peer_idle(peer));
 390
 391                 if (peer->ibp_nid != nid)
 392                         continue;
 393
 394                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 395                        peer, libcfs_nid2str(nid),
 396                        atomic_read(&peer->ibp_refcount),
 397                        peer->ibp_version);
 398                 return peer;
 399         }
 400         return NULL;
 401 }
 402
 403 void kiblnd_unlink_peer_locked(struct kib_peer *peer)
 404 {
 405         LASSERT(list_empty(&peer->ibp_conns));
 406
 407         LASSERT(kiblnd_peer_active(peer));
 408         list_del_init(&peer->ibp_list);
 409         /* lose peerlist's ref */
 410         kiblnd_peer_decref(peer);
 411 }
 412
 413 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 414                                 lnet_nid_t *nidp, int *count)
 415 {
 416         struct kib_peer *peer;
 417         struct list_head *ptmp;
 418         int i;
 419         unsigned long flags;
 420
 421         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 422
 423         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 424                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 425                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 426                         LASSERT(!kiblnd_peer_idle(peer));
 427
 428                         if (peer->ibp_ni != ni)
 429                                 continue;
 430
 431                         if (index-- > 0)
 432                                 continue;
 433
 434                         *nidp = peer->ibp_nid;
 435                         *count = atomic_read(&peer->ibp_refcount);
 436
 437                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 438                                                flags);
 439                         return 0;
 440                 }
 441         }
 442
 443         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 444         return -ENOENT;
 445 }
 446
 447 static void kiblnd_del_peer_locked(struct kib_peer *peer)
 448 {
 449         struct list_head *ctmp;
 450         struct list_head *cnxt;
 451         struct kib_conn *conn;
 452
 453         if (list_empty(&peer->ibp_conns)) {
 454                 kiblnd_unlink_peer_locked(peer);
 455         } else {
 456                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 457                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
 458
 459                         kiblnd_close_conn_locked(conn, 0);
 460                 }
 461                 /* NB closing peer's last conn unlinked it. */
 462         }
 463         /*
 464          * NB peer now unlinked; might even be freed if the peer table had the
 465          * last ref on it.
 466          */
 467 }
 468
 469 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 470 {
 471         LIST_HEAD(zombies);
 472         struct list_head *ptmp;
 473         struct list_head *pnxt;
 474         struct kib_peer *peer;
 475         int lo;
 476         int hi;
 477         int i;
 478         unsigned long flags;
 479         int rc = -ENOENT;
 480
 481         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 482
 483         if (nid != LNET_NID_ANY) {
 484                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 485                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 486         } else {
 487                 lo = 0;
 488                 hi = kiblnd_data.kib_peer_hash_size - 1;
 489         }
 490
 491         for (i = lo; i <= hi; i++) {
 492                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 493                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 494                         LASSERT(!kiblnd_peer_idle(peer));
 495
 496                         if (peer->ibp_ni != ni)
 497                                 continue;
 498
 499                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 500                                 continue;
 501
 502                         if (!list_empty(&peer->ibp_tx_queue)) {
 503                                 LASSERT(list_empty(&peer->ibp_conns));
 504
 505                                 list_splice_init(&peer->ibp_tx_queue,
 506                                                  &zombies);
 507                         }
 508
 509                         kiblnd_del_peer_locked(peer);
 510                         rc = 0;  /* matched something */
 511                 }
 512         }
 513
 514         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 515
 516         kiblnd_txlist_done(ni, &zombies, -EIO);
 517
 518         return rc;
 519 }
 520
 521 static struct kib_conn *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 522 {
 523         struct kib_peer *peer;
 524         struct list_head *ptmp;
 525         struct kib_conn *conn;
 526         struct list_head *ctmp;
 527         int i;
 528         unsigned long flags;
 529
 530         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 531
 532         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 533                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 534                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 535                         LASSERT(!kiblnd_peer_idle(peer));
 536
 537                         if (peer->ibp_ni != ni)
 538                                 continue;
 539
 540                         list_for_each(ctmp, &peer->ibp_conns) {
 541                                 if (index-- > 0)
 542                                         continue;
 543
 544                                 conn = list_entry(ctmp, struct kib_conn,
 545                                                   ibc_list);
 546                                 kiblnd_conn_addref(conn);
 547                                 read_unlock_irqrestore(
 548                                         &kiblnd_data.kib_global_lock,
 549                                         flags);
 550                                 return conn;
 551                         }
 552                 }
 553         }
 554
 555         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 556         return NULL;
 557 }
 558
 559 int kiblnd_translate_mtu(int value)
 560 {
 561         switch (value) {
 562         default:
 563                 return -1;
 564         case 0:
 565                 return 0;
 566         case 256:
 567                 return IB_MTU_256;
 568         case 512:
 569                 return IB_MTU_512;
 570         case 1024:
 571                 return IB_MTU_1024;
 572         case 2048:
 573                 return IB_MTU_2048;
 574         case 4096:
 575                 return IB_MTU_4096;
 576         }
 577 }
 578
 579 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 580 {
 581         int mtu;
 582
 583         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 584         if (!cmid->route.path_rec)
 585                 return;
 586
 587         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 588         LASSERT(mtu >= 0);
 589         if (mtu)
 590                 cmid->route.path_rec->mtu = mtu;
 591 }
 592
 593 static int kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
 594 {
 595         cpumask_t *mask;
 596         int vectors;
 597         int off;
 598         int i;
 599         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 600
 601         vectors = conn->ibc_cmid->device->num_comp_vectors;
 602         if (vectors <= 1)
 603                 return 0;
 604
 605         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 606         if (!mask)
 607                 return 0;
 608
 609         /* hash NID to CPU id in this partition... */
 610         off = do_div(nid, cpumask_weight(mask));
 611         for_each_cpu(i, mask) {
 612                 if (!off--)
 613                         return i % vectors;
 614         }
 615
 616         LBUG();
 617         return 1;
 618 }
 619
 620 struct kib_conn *kiblnd_create_conn(struct kib_peer *peer, struct rdma_cm_id *cmid,
 621                                int state, int version)
 622 {
 623         /*
 624          * CAVEAT EMPTOR:
 625          * If the new conn is created successfully it takes over the caller's
 626          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 627          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 628          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 629          * to destroy 'cmid' here since I'm called from the CM which still has
 630          * its ref on 'cmid').
 631          */
 632         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 633         struct kib_net *net = peer->ibp_ni->ni_data;
 634         struct kib_dev *dev;
 635         struct ib_qp_init_attr *init_qp_attr;
 636         struct kib_sched_info *sched;
 637         struct ib_cq_init_attr cq_attr = {};
 638         struct kib_conn *conn;
 639         struct ib_cq *cq;
 640         unsigned long flags;
 641         int cpt;
 642         int rc;
 643         int i;
 644
 645         LASSERT(net);
 646         LASSERT(!in_interrupt());
 647
 648         dev = net->ibn_dev;
 649
 650         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 651         sched = kiblnd_data.kib_scheds[cpt];
 652
 653         LASSERT(sched->ibs_nthreads > 0);
 654
 655         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 656                          sizeof(*init_qp_attr));
 657         if (!init_qp_attr) {
 658                 CERROR("Can't allocate qp_attr for %s\n",
 659                        libcfs_nid2str(peer->ibp_nid));
 660                 goto failed_0;
 661         }
 662
 663         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 664         if (!conn) {
 665                 CERROR("Can't allocate connection for %s\n",
 666                        libcfs_nid2str(peer->ibp_nid));
 667                 goto failed_1;
 668         }
 669
 670         conn->ibc_state = IBLND_CONN_INIT;
 671         conn->ibc_version = version;
 672         conn->ibc_peer = peer;            /* I take the caller's ref */
 673         cmid->context = conn;              /* for future CM callbacks */
 674         conn->ibc_cmid = cmid;
 675         conn->ibc_max_frags = peer->ibp_max_frags;
 676         conn->ibc_queue_depth = peer->ibp_queue_depth;
 677
 678         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 679         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 680         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 681         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 682         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 683         INIT_LIST_HEAD(&conn->ibc_active_txs);
 684         spin_lock_init(&conn->ibc_lock);
 685
 686         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 687                          sizeof(*conn->ibc_connvars));
 688         if (!conn->ibc_connvars) {
 689                 CERROR("Can't allocate in-progress connection state\n");
 690                 goto failed_2;
 691         }
 692
 693         write_lock_irqsave(glock, flags);
 694         if (dev->ibd_failover) {
 695                 write_unlock_irqrestore(glock, flags);
 696                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 697                 goto failed_2;
 698         }
 699
 700         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 701                 /* wakeup failover thread and teardown connection */
 702                 if (kiblnd_dev_can_failover(dev)) {
 703                         list_add_tail(&dev->ibd_fail_list,
 704                                       &kiblnd_data.kib_failed_devs);
 705                         wake_up(&kiblnd_data.kib_failover_waitq);
 706                 }
 707
 708                 write_unlock_irqrestore(glock, flags);
 709                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 710                        cmid->device->name, dev->ibd_ifname);
 711                 goto failed_2;
 712         }
 713
 714         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 715         conn->ibc_hdev = dev->ibd_hdev;
 716
 717         kiblnd_setup_mtu_locked(cmid);
 718
 719         write_unlock_irqrestore(glock, flags);
 720
 721         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 722                          IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 723         if (!conn->ibc_rxs) {
 724                 CERROR("Cannot allocate RX buffers\n");
 725                 goto failed_2;
 726         }
 727
 728         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 729                                 IBLND_RX_MSG_PAGES(conn));
 730         if (rc)
 731                 goto failed_2;
 732
 733         kiblnd_map_rx_descs(conn);
 734
 735         cq_attr.cqe = IBLND_CQ_ENTRIES(conn);
 736         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 737         cq = ib_create_cq(cmid->device,
 738                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 739                           &cq_attr);
 740         if (IS_ERR(cq)) {
 741                 CERROR("Failed to create CQ with %d CQEs: %ld\n",
 742                        IBLND_CQ_ENTRIES(conn), PTR_ERR(cq));
 743                 goto failed_2;
 744         }
 745
 746         conn->ibc_cq = cq;
 747
 748         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 749         if (rc) {
 750                 CERROR("Can't request completion notification: %d\n", rc);
 751                 goto failed_2;
 752         }
 753
 754         init_qp_attr->event_handler = kiblnd_qp_event;
 755         init_qp_attr->qp_context = conn;
 756         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(conn);
 757         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(conn);
 758         init_qp_attr->cap.max_send_sge = 1;
 759         init_qp_attr->cap.max_recv_sge = 1;
 760         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 761         init_qp_attr->qp_type = IB_QPT_RC;
 762         init_qp_attr->send_cq = cq;
 763         init_qp_attr->recv_cq = cq;
 764
 765         conn->ibc_sched = sched;
 766
 767         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 768         if (rc) {
 769                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 770                        rc, init_qp_attr->cap.max_send_wr,
 771                        init_qp_attr->cap.max_recv_wr);
 772                 goto failed_2;
 773         }
 774
 775         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 776
 777         /* 1 ref for caller and each rxmsg */
 778         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(conn));
 779         conn->ibc_nrx = IBLND_RX_MSGS(conn);
 780
 781         /* post receives */
 782         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
 783                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 784                                     IBLND_POSTRX_NO_CREDIT);
 785                 if (rc) {
 786                         CERROR("Can't post rxmsg: %d\n", rc);
 787
 788                         /* Make posted receives complete */
 789                         kiblnd_abort_receives(conn);
 790
 791                         /*
 792                          * correct # of posted buffers
 793                          * NB locking needed now I'm racing with completion
 794                          */
 795                         spin_lock_irqsave(&sched->ibs_lock, flags);
 796                         conn->ibc_nrx -= IBLND_RX_MSGS(conn) - i;
 797                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 798
 799                         /*
 800                          * cmid will be destroyed by CM(ofed) after cm_callback
 801                          * returned, so we can't refer it anymore
 802                          * (by kiblnd_connd()->kiblnd_destroy_conn)
 803                          */
 804                         rdma_destroy_qp(conn->ibc_cmid);
 805                         conn->ibc_cmid = NULL;
 806
 807                         /* Drop my own and unused rxbuffer refcounts */
 808                         while (i++ <= IBLND_RX_MSGS(conn))
 809                                 kiblnd_conn_decref(conn);
 810
 811                         return NULL;
 812                 }
 813         }
 814
 815         /* Init successful! */
 816         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 817                 state == IBLND_CONN_PASSIVE_WAIT);
 818         conn->ibc_state = state;
 819
 820         /* 1 more conn */
 821         atomic_inc(&net->ibn_nconns);
 822         return conn;
 823
 824  failed_2:
 825         kiblnd_destroy_conn(conn, true);
 826  failed_1:
 827         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 828  failed_0:
 829         return NULL;
 830 }
 831
 832 void kiblnd_destroy_conn(struct kib_conn *conn, bool free_conn)
 833 {
 834         struct rdma_cm_id *cmid = conn->ibc_cmid;
 835         struct kib_peer *peer = conn->ibc_peer;
 836         int rc;
 837
 838         LASSERT(!in_interrupt());
 839         LASSERT(!atomic_read(&conn->ibc_refcount));
 840         LASSERT(list_empty(&conn->ibc_early_rxs));
 841         LASSERT(list_empty(&conn->ibc_tx_noops));
 842         LASSERT(list_empty(&conn->ibc_tx_queue));
 843         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 844         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 845         LASSERT(list_empty(&conn->ibc_active_txs));
 846         LASSERT(!conn->ibc_noops_posted);
 847         LASSERT(!conn->ibc_nsends_posted);
 848
 849         switch (conn->ibc_state) {
 850         default:
 851                 /* conn must be completely disengaged from the network */
 852                 LBUG();
 853
 854         case IBLND_CONN_DISCONNECTED:
 855                 /* connvars should have been freed already */
 856                 LASSERT(!conn->ibc_connvars);
 857                 break;
 858
 859         case IBLND_CONN_INIT:
 860                 break;
 861         }
 862
 863         /* conn->ibc_cmid might be destroyed by CM already */
 864         if (cmid && cmid->qp)
 865                 rdma_destroy_qp(cmid);
 866
 867         if (conn->ibc_cq) {
 868                 rc = ib_destroy_cq(conn->ibc_cq);
 869                 if (rc)
 870                         CWARN("Error destroying CQ: %d\n", rc);
 871         }
 872
 873         if (conn->ibc_rx_pages)
 874                 kiblnd_unmap_rx_descs(conn);
 875
 876         if (conn->ibc_rxs) {
 877                 LIBCFS_FREE(conn->ibc_rxs,
 878                             IBLND_RX_MSGS(conn) * sizeof(struct kib_rx));
 879         }
 880
 881         if (conn->ibc_connvars)
 882                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 883
 884         if (conn->ibc_hdev)
 885                 kiblnd_hdev_decref(conn->ibc_hdev);
 886
 887         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 888         if (conn->ibc_state != IBLND_CONN_INIT) {
 889                 struct kib_net *net = peer->ibp_ni->ni_data;
 890
 891                 kiblnd_peer_decref(peer);
 892                 rdma_destroy_id(cmid);
 893                 atomic_dec(&net->ibn_nconns);
 894         }
 895
 896         LIBCFS_FREE(conn, sizeof(*conn));
 897 }
 898
 899 int kiblnd_close_peer_conns_locked(struct kib_peer *peer, int why)
 900 {
 901         struct kib_conn *conn;
 902         struct list_head *ctmp;
 903         struct list_head *cnxt;
 904         int count = 0;
 905
 906         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 907                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 908
 909                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 910                        libcfs_nid2str(peer->ibp_nid),
 911                        conn->ibc_version, why);
 912
 913                 kiblnd_close_conn_locked(conn, why);
 914                 count++;
 915         }
 916
 917         return count;
 918 }
 919
 920 int kiblnd_close_stale_conns_locked(struct kib_peer *peer,
 921                                     int version, __u64 incarnation)
 922 {
 923         struct kib_conn *conn;
 924         struct list_head *ctmp;
 925         struct list_head *cnxt;
 926         int count = 0;
 927
 928         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 929                 conn = list_entry(ctmp, struct kib_conn, ibc_list);
 930
 931                 if (conn->ibc_version     == version &&
 932                     conn->ibc_incarnation == incarnation)
 933                         continue;
 934
 935                 CDEBUG(D_NET,
 936                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 937                        libcfs_nid2str(peer->ibp_nid),
 938                        conn->ibc_version, conn->ibc_incarnation,
 939                        version, incarnation);
 940
 941                 kiblnd_close_conn_locked(conn, -ESTALE);
 942                 count++;
 943         }
 944
 945         return count;
 946 }
 947
 948 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 949 {
 950         struct kib_peer *peer;
 951         struct list_head *ptmp;
 952         struct list_head *pnxt;
 953         int lo;
 954         int hi;
 955         int i;
 956         unsigned long flags;
 957         int count = 0;
 958
 959         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 960
 961         if (nid != LNET_NID_ANY) {
 962                 lo = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 963                 hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 964         } else {
 965                 lo = 0;
 966                 hi = kiblnd_data.kib_peer_hash_size - 1;
 967         }
 968
 969         for (i = lo; i <= hi; i++) {
 970                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 971                         peer = list_entry(ptmp, struct kib_peer, ibp_list);
 972                         LASSERT(!kiblnd_peer_idle(peer));
 973
 974                         if (peer->ibp_ni != ni)
 975                                 continue;
 976
 977                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 978                                 continue;
 979
 980                         count += kiblnd_close_peer_conns_locked(peer, 0);
 981                 }
 982         }
 983
 984         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 985
 986         /* wildcards always succeed */
 987         if (nid == LNET_NID_ANY)
 988                 return 0;
 989
 990         return !count ? -ENOENT : 0;
 991 }
 992
 993 static int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
 994 {
 995         struct libcfs_ioctl_data *data = arg;
 996         int rc = -EINVAL;
 997
 998         switch (cmd) {
 999         case IOC_LIBCFS_GET_PEER: {
1000                 lnet_nid_t nid = 0;
1001                 int count = 0;
1002
1003                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1004                                           &nid, &count);
1005                 data->ioc_nid   = nid;
1006                 data->ioc_count = count;
1007                 break;
1008         }
1009
1010         case IOC_LIBCFS_DEL_PEER: {
1011                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1012                 break;
1013         }
1014         case IOC_LIBCFS_GET_CONN: {
1015                 struct kib_conn *conn;
1016
1017                 rc = 0;
1018                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1019                 if (!conn) {
1020                         rc = -ENOENT;
1021                         break;
1022                 }
1023
1024                 LASSERT(conn->ibc_cmid);
1025                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1026                 if (!conn->ibc_cmid->route.path_rec)
1027                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1028                 else
1029                         data->ioc_u32[0] =
1030                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1031                 kiblnd_conn_decref(conn);
1032                 break;
1033         }
1034         case IOC_LIBCFS_CLOSE_CONNECTION: {
1035                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1036                 break;
1037         }
1038
1039         default:
1040                 break;
1041         }
1042
1043         return rc;
1044 }
1045
1046 static void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1047 {
1048         unsigned long last_alive = 0;
1049         unsigned long now = cfs_time_current();
1050         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1051         struct kib_peer *peer;
1052         unsigned long flags;
1053
1054         read_lock_irqsave(glock, flags);
1055
1056         peer = kiblnd_find_peer_locked(nid);
1057         if (peer)
1058                 last_alive = peer->ibp_last_alive;
1059
1060         read_unlock_irqrestore(glock, flags);
1061
1062         if (last_alive)
1063                 *when = last_alive;
1064
1065         /*
1066          * peer is not persistent in hash, trigger peer creation
1067          * and connection establishment with a NULL tx
1068          */
1069         if (!peer)
1070                 kiblnd_launch_tx(ni, NULL, nid);
1071
1072         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1073                libcfs_nid2str(nid), peer,
1074                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1075 }
1076
1077 static void kiblnd_free_pages(struct kib_pages *p)
1078 {
1079         int npages = p->ibp_npages;
1080         int i;
1081
1082         for (i = 0; i < npages; i++) {
1083                 if (p->ibp_pages[i])
1084                         __free_page(p->ibp_pages[i]);
1085         }
1086
1087         LIBCFS_FREE(p, offsetof(struct kib_pages, ibp_pages[npages]));
1088 }
1089
1090 int kiblnd_alloc_pages(struct kib_pages **pp, int cpt, int npages)
1091 {
1092         struct kib_pages *p;
1093         int i;
1094
1095         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1096                          offsetof(struct kib_pages, ibp_pages[npages]));
1097         if (!p) {
1098                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1099                 return -ENOMEM;
1100         }
1101
1102         memset(p, 0, offsetof(struct kib_pages, ibp_pages[npages]));
1103         p->ibp_npages = npages;
1104
1105         for (i = 0; i < npages; i++) {
1106                 p->ibp_pages[i] = alloc_pages_node(
1107                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1108                                     GFP_NOFS, 0);
1109                 if (!p->ibp_pages[i]) {
1110                         CERROR("Can't allocate page %d of %d\n", i, npages);
1111                         kiblnd_free_pages(p);
1112                         return -ENOMEM;
1113                 }
1114         }
1115
1116         *pp = p;
1117         return 0;
1118 }
1119
1120 void kiblnd_unmap_rx_descs(struct kib_conn *conn)
1121 {
1122         struct kib_rx *rx;
1123         int i;
1124
1125         LASSERT(conn->ibc_rxs);
1126         LASSERT(conn->ibc_hdev);
1127
1128         for (i = 0; i < IBLND_RX_MSGS(conn); i++) {
1129                 rx = &conn->ibc_rxs[i];
1130
1131                 LASSERT(rx->rx_nob >= 0); /* not posted */
1132
1133                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1134                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1135                                                           rx->rx_msgaddr),
1136                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1137         }
1138
1139         kiblnd_free_pages(conn->ibc_rx_pages);
1140
1141         conn->ibc_rx_pages = NULL;
1142 }
1143
1144 void kiblnd_map_rx_descs(struct kib_conn *conn)
1145 {
1146         struct kib_rx *rx;
1147         struct page *pg;
1148         int pg_off;
1149         int ipg;
1150         int i;
1151
1152         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn); i++) {
1153                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1154                 rx = &conn->ibc_rxs[i];
1155
1156                 rx->rx_conn = conn;
1157                 rx->rx_msg = (struct kib_msg *)(((char *)page_address(pg)) + pg_off);
1158
1159                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1160                                                        rx->rx_msg,
1161                                                        IBLND_MSG_SIZE,
1162                                                        DMA_FROM_DEVICE);
1163                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1164                                                   rx->rx_msgaddr));
1165                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1166
1167                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1168                        i, rx->rx_msg, rx->rx_msgaddr,
1169                        (__u64)(page_to_phys(pg) + pg_off));
1170
1171                 pg_off += IBLND_MSG_SIZE;
1172                 LASSERT(pg_off <= PAGE_SIZE);
1173
1174                 if (pg_off == PAGE_SIZE) {
1175                         pg_off = 0;
1176                         ipg++;
1177                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn));
1178                 }
1179         }
1180 }
1181
1182 static void kiblnd_unmap_tx_pool(struct kib_tx_pool *tpo)
1183 {
1184         struct kib_hca_dev *hdev = tpo->tpo_hdev;
1185         struct kib_tx *tx;
1186         int i;
1187
1188         LASSERT(!tpo->tpo_pool.po_allocated);
1189
1190         if (!hdev)
1191                 return;
1192
1193         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1194                 tx = &tpo->tpo_tx_descs[i];
1195                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1196                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1197                                                           tx->tx_msgaddr),
1198                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1199         }
1200
1201         kiblnd_hdev_decref(hdev);
1202         tpo->tpo_hdev = NULL;
1203 }
1204
1205 static struct kib_hca_dev *kiblnd_current_hdev(struct kib_dev *dev)
1206 {
1207         struct kib_hca_dev *hdev;
1208         unsigned long flags;
1209         int i = 0;
1210
1211         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1212         while (dev->ibd_failover) {
1213                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1214                 if (!(i++ % 50))
1215                         CDEBUG(D_NET, "%s: Wait for failover\n",
1216                                dev->ibd_ifname);
1217                 set_current_state(TASK_INTERRUPTIBLE);
1218                 schedule_timeout(cfs_time_seconds(1) / 100);
1219
1220                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1221         }
1222
1223         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1224         hdev = dev->ibd_hdev;
1225
1226         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1227
1228         return hdev;
1229 }
1230
1231 static void kiblnd_map_tx_pool(struct kib_tx_pool *tpo)
1232 {
1233         struct kib_pages *txpgs = tpo->tpo_tx_pages;
1234         struct kib_pool *pool = &tpo->tpo_pool;
1235         struct kib_net *net = pool->po_owner->ps_net;
1236         struct kib_dev *dev;
1237         struct page *page;
1238         struct kib_tx *tx;
1239         int page_offset;
1240         int ipage;
1241         int i;
1242
1243         LASSERT(net);
1244
1245         dev = net->ibn_dev;
1246
1247         /* pre-mapped messages are not bigger than 1 page */
1248         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1249
1250         /* No fancy arithmetic when we do the buffer calculations */
1251         CLASSERT(!(PAGE_SIZE % IBLND_MSG_SIZE));
1252
1253         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1254
1255         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1256                 page = txpgs->ibp_pages[ipage];
1257                 tx = &tpo->tpo_tx_descs[i];
1258
1259                 tx->tx_msg = (struct kib_msg *)(((char *)page_address(page)) +
1260                                            page_offset);
1261
1262                 tx->tx_msgaddr = kiblnd_dma_map_single(
1263                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1264                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1265                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1266                                                   tx->tx_msgaddr));
1267                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1268
1269                 list_add(&tx->tx_list, &pool->po_free_list);
1270
1271                 page_offset += IBLND_MSG_SIZE;
1272                 LASSERT(page_offset <= PAGE_SIZE);
1273
1274                 if (page_offset == PAGE_SIZE) {
1275                         page_offset = 0;
1276                         ipage++;
1277                         LASSERT(ipage <= txpgs->ibp_npages);
1278                 }
1279         }
1280 }
1281
1282 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd,
1283                                     int negotiated_nfrags)
1284 {
1285         struct kib_net *net = ni->ni_data;
1286         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
1287         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1288         __u16 nfrags;
1289         int mod;
1290
1291         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
1292         mod = tunables->lnd_map_on_demand;
1293         nfrags = (negotiated_nfrags != -1) ? negotiated_nfrags : mod;
1294
1295         LASSERT(hdev->ibh_mrs);
1296
1297         if (mod > 0 && nfrags <= rd->rd_nfrags)
1298                 return NULL;
1299
1300         return hdev->ibh_mrs;
1301 }
1302
1303 static void kiblnd_destroy_fmr_pool(struct kib_fmr_pool *fpo)
1304 {
1305         LASSERT(!fpo->fpo_map_count);
1306
1307         if (fpo->fpo_is_fmr) {
1308                 if (fpo->fmr.fpo_fmr_pool)
1309                         ib_destroy_fmr_pool(fpo->fmr.fpo_fmr_pool);
1310         } else {
1311                 struct kib_fast_reg_descriptor *frd, *tmp;
1312                 int i = 0;
1313
1314                 list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1315                                          frd_list) {
1316                         list_del(&frd->frd_list);
1317                         ib_dereg_mr(frd->frd_mr);
1318                         LIBCFS_FREE(frd, sizeof(*frd));
1319                         i++;
1320                 }
1321                 if (i < fpo->fast_reg.fpo_pool_size)
1322                         CERROR("FastReg pool still has %d regions registered\n",
1323                                fpo->fast_reg.fpo_pool_size - i);
1324         }
1325
1326         if (fpo->fpo_hdev)
1327                 kiblnd_hdev_decref(fpo->fpo_hdev);
1328
1329         LIBCFS_FREE(fpo, sizeof(*fpo));
1330 }
1331
1332 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1333 {
1334         struct kib_fmr_pool *fpo, *tmp;
1335
1336         list_for_each_entry_safe(fpo, tmp, head, fpo_list) {
1337                 list_del(&fpo->fpo_list);
1338                 kiblnd_destroy_fmr_pool(fpo);
1339         }
1340 }
1341
1342 static int
1343 kiblnd_fmr_pool_size(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1344                      int ncpts)
1345 {
1346         int size = tunables->lnd_fmr_pool_size / ncpts;
1347
1348         return max(IBLND_FMR_POOL, size);
1349 }
1350
1351 static int
1352 kiblnd_fmr_flush_trigger(struct lnet_ioctl_config_o2iblnd_tunables *tunables,
1353                          int ncpts)
1354 {
1355         int size = tunables->lnd_fmr_flush_trigger / ncpts;
1356
1357         return max(IBLND_FMR_POOL_FLUSH, size);
1358 }
1359
1360 static int kiblnd_alloc_fmr_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1361 {
1362         struct ib_fmr_pool_param param = {
1363                 .max_pages_per_fmr = LNET_MAX_PAYLOAD / PAGE_SIZE,
1364                 .page_shift        = PAGE_SHIFT,
1365                 .access            = (IB_ACCESS_LOCAL_WRITE |
1366                                       IB_ACCESS_REMOTE_WRITE),
1367                 .pool_size         = fps->fps_pool_size,
1368                 .dirty_watermark   = fps->fps_flush_trigger,
1369                 .flush_function    = NULL,
1370                 .flush_arg         = NULL,
1371                 .cache             = !!fps->fps_cache };
1372         int rc = 0;
1373
1374         fpo->fmr.fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd,
1375                                                    &param);
1376         if (IS_ERR(fpo->fmr.fpo_fmr_pool)) {
1377                 rc = PTR_ERR(fpo->fmr.fpo_fmr_pool);
1378                 if (rc != -ENOSYS)
1379                         CERROR("Failed to create FMR pool: %d\n", rc);
1380                 else
1381                         CERROR("FMRs are not supported\n");
1382         }
1383
1384         return rc;
1385 }
1386
1387 static int kiblnd_alloc_freg_pool(struct kib_fmr_poolset *fps, struct kib_fmr_pool *fpo)
1388 {
1389         struct kib_fast_reg_descriptor *frd, *tmp;
1390         int i, rc;
1391
1392         INIT_LIST_HEAD(&fpo->fast_reg.fpo_pool_list);
1393         fpo->fast_reg.fpo_pool_size = 0;
1394         for (i = 0; i < fps->fps_pool_size; i++) {
1395                 LIBCFS_CPT_ALLOC(frd, lnet_cpt_table(), fps->fps_cpt,
1396                                  sizeof(*frd));
1397                 if (!frd) {
1398                         CERROR("Failed to allocate a new fast_reg descriptor\n");
1399                         rc = -ENOMEM;
1400                         goto out;
1401                 }
1402
1403                 frd->frd_mr = ib_alloc_mr(fpo->fpo_hdev->ibh_pd,
1404                                           IB_MR_TYPE_MEM_REG,
1405                                           LNET_MAX_PAYLOAD / PAGE_SIZE);
1406                 if (IS_ERR(frd->frd_mr)) {
1407                         rc = PTR_ERR(frd->frd_mr);
1408                         CERROR("Failed to allocate ib_alloc_mr: %d\n", rc);
1409                         frd->frd_mr = NULL;
1410                         goto out_middle;
1411                 }
1412
1413                 frd->frd_valid = true;
1414
1415                 list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1416                 fpo->fast_reg.fpo_pool_size++;
1417         }
1418
1419         return 0;
1420
1421 out_middle:
1422         if (frd->frd_mr)
1423                 ib_dereg_mr(frd->frd_mr);
1424         LIBCFS_FREE(frd, sizeof(*frd));
1425
1426 out:
1427         list_for_each_entry_safe(frd, tmp, &fpo->fast_reg.fpo_pool_list,
1428                                  frd_list) {
1429                 list_del(&frd->frd_list);
1430                 ib_dereg_mr(frd->frd_mr);
1431                 LIBCFS_FREE(frd, sizeof(*frd));
1432         }
1433
1434         return rc;
1435 }
1436
1437 static int kiblnd_create_fmr_pool(struct kib_fmr_poolset *fps,
1438                                   struct kib_fmr_pool **pp_fpo)
1439 {
1440         struct kib_dev *dev = fps->fps_net->ibn_dev;
1441         struct ib_device_attr *dev_attr;
1442         struct kib_fmr_pool *fpo;
1443         int rc;
1444
1445         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1446         if (!fpo)
1447                 return -ENOMEM;
1448
1449         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1450         dev_attr = &fpo->fpo_hdev->ibh_ibdev->attrs;
1451
1452         /* Check for FMR or FastReg support */
1453         fpo->fpo_is_fmr = 0;
1454         if (fpo->fpo_hdev->ibh_ibdev->alloc_fmr &&
1455             fpo->fpo_hdev->ibh_ibdev->dealloc_fmr &&
1456             fpo->fpo_hdev->ibh_ibdev->map_phys_fmr &&
1457             fpo->fpo_hdev->ibh_ibdev->unmap_fmr) {
1458                 LCONSOLE_INFO("Using FMR for registration\n");
1459                 fpo->fpo_is_fmr = 1;
1460         } else if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
1461                 LCONSOLE_INFO("Using FastReg for registration\n");
1462         } else {
1463                 rc = -ENOSYS;
1464                 LCONSOLE_ERROR_MSG(rc, "IB device does not support FMRs nor FastRegs, can't register memory\n");
1465                 goto out_fpo;
1466         }
1467
1468         if (fpo->fpo_is_fmr)
1469                 rc = kiblnd_alloc_fmr_pool(fps, fpo);
1470         else
1471                 rc = kiblnd_alloc_freg_pool(fps, fpo);
1472         if (rc)
1473                 goto out_fpo;
1474
1475         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1476         fpo->fpo_owner = fps;
1477         *pp_fpo = fpo;
1478
1479         return 0;
1480
1481 out_fpo:
1482         kiblnd_hdev_decref(fpo->fpo_hdev);
1483         LIBCFS_FREE(fpo, sizeof(*fpo));
1484         return rc;
1485 }
1486
1487 static void kiblnd_fail_fmr_poolset(struct kib_fmr_poolset *fps,
1488                                     struct list_head *zombies)
1489 {
1490         if (!fps->fps_net) /* intialized? */
1491                 return;
1492
1493         spin_lock(&fps->fps_lock);
1494
1495         while (!list_empty(&fps->fps_pool_list)) {
1496                 struct kib_fmr_pool *fpo = list_entry(fps->fps_pool_list.next,
1497                                                  struct kib_fmr_pool, fpo_list);
1498                 fpo->fpo_failed = 1;
1499                 list_del(&fpo->fpo_list);
1500                 if (!fpo->fpo_map_count)
1501                         list_add(&fpo->fpo_list, zombies);
1502                 else
1503                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1504         }
1505
1506         spin_unlock(&fps->fps_lock);
1507 }
1508
1509 static void kiblnd_fini_fmr_poolset(struct kib_fmr_poolset *fps)
1510 {
1511         if (fps->fps_net) { /* initialized? */
1512                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1513                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1514         }
1515 }
1516
1517 static int
1518 kiblnd_init_fmr_poolset(struct kib_fmr_poolset *fps, int cpt, int ncpts,
1519                         struct kib_net *net,
1520                         struct lnet_ioctl_config_o2iblnd_tunables *tunables)
1521 {
1522         struct kib_fmr_pool *fpo;
1523         int rc;
1524
1525         memset(fps, 0, sizeof(*fps));
1526
1527         fps->fps_net = net;
1528         fps->fps_cpt = cpt;
1529
1530         fps->fps_pool_size = kiblnd_fmr_pool_size(tunables, ncpts);
1531         fps->fps_flush_trigger = kiblnd_fmr_flush_trigger(tunables, ncpts);
1532         fps->fps_cache = tunables->lnd_fmr_cache;
1533
1534         spin_lock_init(&fps->fps_lock);
1535         INIT_LIST_HEAD(&fps->fps_pool_list);
1536         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1537
1538         rc = kiblnd_create_fmr_pool(fps, &fpo);
1539         if (!rc)
1540                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1541
1542         return rc;
1543 }
1544
1545 static int kiblnd_fmr_pool_is_idle(struct kib_fmr_pool *fpo, unsigned long now)
1546 {
1547         if (fpo->fpo_map_count) /* still in use */
1548                 return 0;
1549         if (fpo->fpo_failed)
1550                 return 1;
1551         return cfs_time_aftereq(now, fpo->fpo_deadline);
1552 }
1553
1554 static int
1555 kiblnd_map_tx_pages(struct kib_tx *tx, struct kib_rdma_desc *rd)
1556 {
1557         __u64 *pages = tx->tx_pages;
1558         struct kib_hca_dev *hdev;
1559         int npages;
1560         int size;
1561         int i;
1562
1563         hdev = tx->tx_pool->tpo_hdev;
1564
1565         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
1566                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
1567                      size += hdev->ibh_page_size) {
1568                         pages[npages++] = (rd->rd_frags[i].rf_addr &
1569                                            hdev->ibh_page_mask) + size;
1570                 }
1571         }
1572
1573         return npages;
1574 }
1575
1576 void kiblnd_fmr_pool_unmap(struct kib_fmr *fmr, int status)
1577 {
1578         LIST_HEAD(zombies);
1579         struct kib_fmr_pool *fpo = fmr->fmr_pool;
1580         struct kib_fmr_poolset *fps;
1581         unsigned long now = cfs_time_current();
1582         struct kib_fmr_pool *tmp;
1583         int rc;
1584
1585         if (!fpo)
1586                 return;
1587
1588         fps = fpo->fpo_owner;
1589         if (fpo->fpo_is_fmr) {
1590                 if (fmr->fmr_pfmr) {
1591                         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1592                         LASSERT(!rc);
1593                         fmr->fmr_pfmr = NULL;
1594                 }
1595
1596                 if (status) {
1597                         rc = ib_flush_fmr_pool(fpo->fmr.fpo_fmr_pool);
1598                         LASSERT(!rc);
1599                 }
1600         } else {
1601                 struct kib_fast_reg_descriptor *frd = fmr->fmr_frd;
1602
1603                 if (frd) {
1604                         frd->frd_valid = false;
1605                         spin_lock(&fps->fps_lock);
1606                         list_add_tail(&frd->frd_list, &fpo->fast_reg.fpo_pool_list);
1607                         spin_unlock(&fps->fps_lock);
1608                         fmr->fmr_frd = NULL;
1609                 }
1610         }
1611         fmr->fmr_pool = NULL;
1612
1613         spin_lock(&fps->fps_lock);
1614         fpo->fpo_map_count--;  /* decref the pool */
1615
1616         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1617                 /* the first pool is persistent */
1618                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1619                         continue;
1620
1621                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1622                         list_move(&fpo->fpo_list, &zombies);
1623                         fps->fps_version++;
1624                 }
1625         }
1626         spin_unlock(&fps->fps_lock);
1627
1628         if (!list_empty(&zombies))
1629                 kiblnd_destroy_fmr_pool_list(&zombies);
1630 }
1631
1632 int kiblnd_fmr_pool_map(struct kib_fmr_poolset *fps, struct kib_tx *tx,
1633                         struct kib_rdma_desc *rd, __u32 nob, __u64 iov,
1634                         struct kib_fmr *fmr)
1635 {
1636         __u64 *pages = tx->tx_pages;
1637         bool is_rx = (rd != tx->tx_rd);
1638         bool tx_pages_mapped = 0;
1639         struct kib_fmr_pool *fpo;
1640         int npages = 0;
1641         __u64 version;
1642         int rc;
1643
1644  again:
1645         spin_lock(&fps->fps_lock);
1646         version = fps->fps_version;
1647         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1648                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1649                 fpo->fpo_map_count++;
1650
1651                 if (fpo->fpo_is_fmr) {
1652                         struct ib_pool_fmr *pfmr;
1653
1654                         spin_unlock(&fps->fps_lock);
1655
1656                         if (!tx_pages_mapped) {
1657                                 npages = kiblnd_map_tx_pages(tx, rd);
1658                                 tx_pages_mapped = 1;
1659                         }
1660
1661                         pfmr = ib_fmr_pool_map_phys(fpo->fmr.fpo_fmr_pool,
1662                                                     pages, npages, iov);
1663                         if (likely(!IS_ERR(pfmr))) {
1664                                 fmr->fmr_key = is_rx ? pfmr->fmr->rkey :
1665                                                        pfmr->fmr->lkey;
1666                                 fmr->fmr_frd = NULL;
1667                                 fmr->fmr_pfmr = pfmr;
1668                                 fmr->fmr_pool = fpo;
1669                                 return 0;
1670                         }
1671                         rc = PTR_ERR(pfmr);
1672                 } else {
1673                         if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
1674                                 struct kib_fast_reg_descriptor *frd;
1675                                 struct ib_reg_wr *wr;
1676                                 struct ib_mr *mr;
1677                                 int n;
1678
1679                                 frd = list_first_entry(&fpo->fast_reg.fpo_pool_list,
1680                                                        struct kib_fast_reg_descriptor,
1681                                                        frd_list);
1682                                 list_del(&frd->frd_list);
1683                                 spin_unlock(&fps->fps_lock);
1684
1685                                 mr = frd->frd_mr;
1686
1687                                 if (!frd->frd_valid) {
1688                                         __u32 key = is_rx ? mr->rkey : mr->lkey;
1689                                         struct ib_send_wr *inv_wr;
1690
1691                                         inv_wr = &frd->frd_inv_wr;
1692                                         memset(inv_wr, 0, sizeof(*inv_wr));
1693                                         inv_wr->opcode = IB_WR_LOCAL_INV;
1694                                         inv_wr->wr_id = IBLND_WID_MR;
1695                                         inv_wr->ex.invalidate_rkey = key;
1696
1697                                         /* Bump the key */
1698                                         key = ib_inc_rkey(key);
1699                                         ib_update_fast_reg_key(mr, key);
1700                                 }
1701
1702                                 n = ib_map_mr_sg(mr, tx->tx_frags,
1703                                                  tx->tx_nfrags, NULL, PAGE_SIZE);
1704                                 if (unlikely(n != tx->tx_nfrags)) {
1705                                         CERROR("Failed to map mr %d/%d elements\n",
1706                                                n, tx->tx_nfrags);
1707                                         return n < 0 ? n : -EINVAL;
1708                                 }
1709
1710                                 mr->iova = iov;
1711
1712                                 /* Prepare FastReg WR */
1713                                 wr = &frd->frd_fastreg_wr;
1714                                 memset(wr, 0, sizeof(*wr));
1715                                 wr->wr.opcode = IB_WR_REG_MR;
1716                                 wr->wr.wr_id = IBLND_WID_MR;
1717                                 wr->wr.num_sge = 0;
1718                                 wr->wr.send_flags = 0;
1719                                 wr->mr = mr;
1720                                 wr->key = is_rx ? mr->rkey : mr->lkey;
1721                                 wr->access = (IB_ACCESS_LOCAL_WRITE |
1722                                               IB_ACCESS_REMOTE_WRITE);
1723
1724                                 fmr->fmr_key = is_rx ? mr->rkey : mr->lkey;
1725                                 fmr->fmr_frd = frd;
1726                                 fmr->fmr_pfmr = NULL;
1727                                 fmr->fmr_pool = fpo;
1728                                 return 0;
1729                         }
1730                         spin_unlock(&fps->fps_lock);
1731                         rc = -EBUSY;
1732                 }
1733
1734                 spin_lock(&fps->fps_lock);
1735                 fpo->fpo_map_count--;
1736                 if (rc != -EAGAIN) {
1737                         spin_unlock(&fps->fps_lock);
1738                         return rc;
1739                 }
1740
1741                 /* EAGAIN and ... */
1742                 if (version != fps->fps_version) {
1743                         spin_unlock(&fps->fps_lock);
1744                         goto again;
1745                 }
1746         }
1747
1748         if (fps->fps_increasing) {
1749                 spin_unlock(&fps->fps_lock);
1750                 CDEBUG(D_NET, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1751                 schedule();
1752                 goto again;
1753         }
1754
1755         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1756                 /* someone failed recently */
1757                 spin_unlock(&fps->fps_lock);
1758                 return -EAGAIN;
1759         }
1760
1761         fps->fps_increasing = 1;
1762         spin_unlock(&fps->fps_lock);
1763
1764         CDEBUG(D_NET, "Allocate new FMR pool\n");
1765         rc = kiblnd_create_fmr_pool(fps, &fpo);
1766         spin_lock(&fps->fps_lock);
1767         fps->fps_increasing = 0;
1768         if (!rc) {
1769                 fps->fps_version++;
1770                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1771         } else {
1772                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1773         }
1774         spin_unlock(&fps->fps_lock);
1775
1776         goto again;
1777 }
1778
1779 static void kiblnd_fini_pool(struct kib_pool *pool)
1780 {
1781         LASSERT(list_empty(&pool->po_free_list));
1782         LASSERT(!pool->po_allocated);
1783
1784         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1785 }
1786
1787 static void kiblnd_init_pool(struct kib_poolset *ps, struct kib_pool *pool, int size)
1788 {
1789         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1790
1791         memset(pool, 0, sizeof(*pool));
1792         INIT_LIST_HEAD(&pool->po_free_list);
1793         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1794         pool->po_owner    = ps;
1795         pool->po_size     = size;
1796 }
1797
1798 static void kiblnd_destroy_pool_list(struct list_head *head)
1799 {
1800         struct kib_pool *pool;
1801
1802         while (!list_empty(head)) {
1803                 pool = list_entry(head->next, struct kib_pool, po_list);
1804                 list_del(&pool->po_list);
1805
1806                 LASSERT(pool->po_owner);
1807                 pool->po_owner->ps_pool_destroy(pool);
1808         }
1809 }
1810
1811 static void kiblnd_fail_poolset(struct kib_poolset *ps, struct list_head *zombies)
1812 {
1813         if (!ps->ps_net) /* intialized? */
1814                 return;
1815
1816         spin_lock(&ps->ps_lock);
1817         while (!list_empty(&ps->ps_pool_list)) {
1818                 struct kib_pool *po = list_entry(ps->ps_pool_list.next,
1819                                             struct kib_pool, po_list);
1820                 po->po_failed = 1;
1821                 list_del(&po->po_list);
1822                 if (!po->po_allocated)
1823                         list_add(&po->po_list, zombies);
1824                 else
1825                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1826         }
1827         spin_unlock(&ps->ps_lock);
1828 }
1829
1830 static void kiblnd_fini_poolset(struct kib_poolset *ps)
1831 {
1832         if (ps->ps_net) { /* initialized? */
1833                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1834                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1835         }
1836 }
1837
1838 static int kiblnd_init_poolset(struct kib_poolset *ps, int cpt,
1839                                struct kib_net *net, char *name, int size,
1840                                kib_ps_pool_create_t po_create,
1841                                kib_ps_pool_destroy_t po_destroy,
1842                                kib_ps_node_init_t nd_init,
1843                                kib_ps_node_fini_t nd_fini)
1844 {
1845         struct kib_pool *pool;
1846         int rc;
1847
1848         memset(ps, 0, sizeof(*ps));
1849
1850         ps->ps_cpt          = cpt;
1851         ps->ps_net          = net;
1852         ps->ps_pool_create  = po_create;
1853         ps->ps_pool_destroy = po_destroy;
1854         ps->ps_node_init    = nd_init;
1855         ps->ps_node_fini    = nd_fini;
1856         ps->ps_pool_size    = size;
1857         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1858             >= sizeof(ps->ps_name))
1859                 return -E2BIG;
1860         spin_lock_init(&ps->ps_lock);
1861         INIT_LIST_HEAD(&ps->ps_pool_list);
1862         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1863
1864         rc = ps->ps_pool_create(ps, size, &pool);
1865         if (!rc)
1866                 list_add(&pool->po_list, &ps->ps_pool_list);
1867         else
1868                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1869
1870         return rc;
1871 }
1872
1873 static int kiblnd_pool_is_idle(struct kib_pool *pool, unsigned long now)
1874 {
1875         if (pool->po_allocated) /* still in use */
1876                 return 0;
1877         if (pool->po_failed)
1878                 return 1;
1879         return cfs_time_aftereq(now, pool->po_deadline);
1880 }
1881
1882 void kiblnd_pool_free_node(struct kib_pool *pool, struct list_head *node)
1883 {
1884         LIST_HEAD(zombies);
1885         struct kib_poolset *ps = pool->po_owner;
1886         struct kib_pool *tmp;
1887         unsigned long now = cfs_time_current();
1888
1889         spin_lock(&ps->ps_lock);
1890
1891         if (ps->ps_node_fini)
1892                 ps->ps_node_fini(pool, node);
1893
1894         LASSERT(pool->po_allocated > 0);
1895         list_add(node, &pool->po_free_list);
1896         pool->po_allocated--;
1897
1898         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1899                 /* the first pool is persistent */
1900                 if (ps->ps_pool_list.next == &pool->po_list)
1901                         continue;
1902
1903                 if (kiblnd_pool_is_idle(pool, now))
1904                         list_move(&pool->po_list, &zombies);
1905         }
1906         spin_unlock(&ps->ps_lock);
1907
1908         if (!list_empty(&zombies))
1909                 kiblnd_destroy_pool_list(&zombies);
1910 }
1911
1912 struct list_head *kiblnd_pool_alloc_node(struct kib_poolset *ps)
1913 {
1914         struct list_head *node;
1915         struct kib_pool *pool;
1916         unsigned int interval = 1;
1917         unsigned long time_before;
1918         unsigned int trips = 0;
1919         int rc;
1920
1921  again:
1922         spin_lock(&ps->ps_lock);
1923         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1924                 if (list_empty(&pool->po_free_list))
1925                         continue;
1926
1927                 pool->po_allocated++;
1928                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1929                 node = pool->po_free_list.next;
1930                 list_del(node);
1931
1932                 if (ps->ps_node_init) {
1933                         /* still hold the lock */
1934                         ps->ps_node_init(pool, node);
1935                 }
1936                 spin_unlock(&ps->ps_lock);
1937                 return node;
1938         }
1939
1940         /* no available tx pool and ... */
1941         if (ps->ps_increasing) {
1942                 /* another thread is allocating a new pool */
1943                 spin_unlock(&ps->ps_lock);
1944                 trips++;
1945                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting %d HZs for her to complete. trips = %d\n",
1946                        ps->ps_name, interval, trips);
1947
1948                 set_current_state(TASK_INTERRUPTIBLE);
1949                 schedule_timeout(interval);
1950                 if (interval < cfs_time_seconds(1))
1951                         interval *= 2;
1952
1953                 goto again;
1954         }
1955
1956         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1957                 /* someone failed recently */
1958                 spin_unlock(&ps->ps_lock);
1959                 return NULL;
1960         }
1961
1962         ps->ps_increasing = 1;
1963         spin_unlock(&ps->ps_lock);
1964
1965         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1966         time_before = cfs_time_current();
1967         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1968         CDEBUG(D_NET, "ps_pool_create took %lu HZ to complete",
1969                cfs_time_current() - time_before);
1970
1971         spin_lock(&ps->ps_lock);
1972         ps->ps_increasing = 0;
1973         if (!rc) {
1974                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1975         } else {
1976                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1977                 CERROR("Can't allocate new %s pool because out of memory\n",
1978                        ps->ps_name);
1979         }
1980         spin_unlock(&ps->ps_lock);
1981
1982         goto again;
1983 }
1984
1985 static void kiblnd_destroy_tx_pool(struct kib_pool *pool)
1986 {
1987         struct kib_tx_pool *tpo = container_of(pool, struct kib_tx_pool, tpo_pool);
1988         int i;
1989
1990         LASSERT(!pool->po_allocated);
1991
1992         if (tpo->tpo_tx_pages) {
1993                 kiblnd_unmap_tx_pool(tpo);
1994                 kiblnd_free_pages(tpo->tpo_tx_pages);
1995         }
1996
1997         if (!tpo->tpo_tx_descs)
1998                 goto out;
1999
2000         for (i = 0; i < pool->po_size; i++) {
2001                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2002
2003                 list_del(&tx->tx_list);
2004                 if (tx->tx_pages)
2005                         LIBCFS_FREE(tx->tx_pages,
2006                                     LNET_MAX_IOV *
2007                                     sizeof(*tx->tx_pages));
2008                 if (tx->tx_frags)
2009                         LIBCFS_FREE(tx->tx_frags,
2010                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2011                                      sizeof(*tx->tx_frags));
2012                 if (tx->tx_wrq)
2013                         LIBCFS_FREE(tx->tx_wrq,
2014                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2015                                     sizeof(*tx->tx_wrq));
2016                 if (tx->tx_sge)
2017                         LIBCFS_FREE(tx->tx_sge,
2018                                     (1 + IBLND_MAX_RDMA_FRAGS) *
2019                                     sizeof(*tx->tx_sge));
2020                 if (tx->tx_rd)
2021                         LIBCFS_FREE(tx->tx_rd,
2022                                     offsetof(struct kib_rdma_desc,
2023                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
2024         }
2025
2026         LIBCFS_FREE(tpo->tpo_tx_descs,
2027                     pool->po_size * sizeof(struct kib_tx));
2028 out:
2029         kiblnd_fini_pool(pool);
2030         LIBCFS_FREE(tpo, sizeof(*tpo));
2031 }
2032
2033 static int kiblnd_tx_pool_size(int ncpts)
2034 {
2035         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
2036
2037         return max(IBLND_TX_POOL, ntx);
2038 }
2039
2040 static int kiblnd_create_tx_pool(struct kib_poolset *ps, int size,
2041                                  struct kib_pool **pp_po)
2042 {
2043         int i;
2044         int npg;
2045         struct kib_pool *pool;
2046         struct kib_tx_pool *tpo;
2047
2048         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
2049         if (!tpo) {
2050                 CERROR("Failed to allocate TX pool\n");
2051                 return -ENOMEM;
2052         }
2053
2054         pool = &tpo->tpo_pool;
2055         kiblnd_init_pool(ps, pool, size);
2056         tpo->tpo_tx_descs = NULL;
2057         tpo->tpo_tx_pages = NULL;
2058
2059         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
2060         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg)) {
2061                 CERROR("Can't allocate tx pages: %d\n", npg);
2062                 LIBCFS_FREE(tpo, sizeof(*tpo));
2063                 return -ENOMEM;
2064         }
2065
2066         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
2067                          size * sizeof(struct kib_tx));
2068         if (!tpo->tpo_tx_descs) {
2069                 CERROR("Can't allocate %d tx descriptors\n", size);
2070                 ps->ps_pool_destroy(pool);
2071                 return -ENOMEM;
2072         }
2073
2074         memset(tpo->tpo_tx_descs, 0, size * sizeof(struct kib_tx));
2075
2076         for (i = 0; i < size; i++) {
2077                 struct kib_tx *tx = &tpo->tpo_tx_descs[i];
2078
2079                 tx->tx_pool = tpo;
2080                 if (ps->ps_net->ibn_fmr_ps) {
2081                         LIBCFS_CPT_ALLOC(tx->tx_pages,
2082                                          lnet_cpt_table(), ps->ps_cpt,
2083                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
2084                         if (!tx->tx_pages)
2085                                 break;
2086                 }
2087
2088                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
2089                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2090                                  sizeof(*tx->tx_frags));
2091                 if (!tx->tx_frags)
2092                         break;
2093
2094                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS + 1);
2095
2096                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
2097                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2098                                  sizeof(*tx->tx_wrq));
2099                 if (!tx->tx_wrq)
2100                         break;
2101
2102                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
2103                                  (1 + IBLND_MAX_RDMA_FRAGS) *
2104                                  sizeof(*tx->tx_sge));
2105                 if (!tx->tx_sge)
2106                         break;
2107
2108                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
2109                                  offsetof(struct kib_rdma_desc,
2110                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
2111                 if (!tx->tx_rd)
2112                         break;
2113         }
2114
2115         if (i == size) {
2116                 kiblnd_map_tx_pool(tpo);
2117                 *pp_po = pool;
2118                 return 0;
2119         }
2120
2121         ps->ps_pool_destroy(pool);
2122         return -ENOMEM;
2123 }
2124
2125 static void kiblnd_tx_init(struct kib_pool *pool, struct list_head *node)
2126 {
2127         struct kib_tx_poolset *tps = container_of(pool->po_owner,
2128                                                   struct kib_tx_poolset,
2129                                                   tps_poolset);
2130         struct kib_tx *tx = list_entry(node, struct kib_tx, tx_list);
2131
2132         tx->tx_cookie = tps->tps_next_tx_cookie++;
2133 }
2134
2135 static void kiblnd_net_fini_pools(struct kib_net *net)
2136 {
2137         int i;
2138
2139         cfs_cpt_for_each(i, lnet_cpt_table()) {
2140                 struct kib_tx_poolset *tps;
2141                 struct kib_fmr_poolset *fps;
2142
2143                 if (net->ibn_tx_ps) {
2144                         tps = net->ibn_tx_ps[i];
2145                         kiblnd_fini_poolset(&tps->tps_poolset);
2146                 }
2147
2148                 if (net->ibn_fmr_ps) {
2149                         fps = net->ibn_fmr_ps[i];
2150                         kiblnd_fini_fmr_poolset(fps);
2151                 }
2152         }
2153
2154         if (net->ibn_tx_ps) {
2155                 cfs_percpt_free(net->ibn_tx_ps);
2156                 net->ibn_tx_ps = NULL;
2157         }
2158
2159         if (net->ibn_fmr_ps) {
2160                 cfs_percpt_free(net->ibn_fmr_ps);
2161                 net->ibn_fmr_ps = NULL;
2162         }
2163 }
2164
2165 static int kiblnd_net_init_pools(struct kib_net *net, lnet_ni_t *ni, __u32 *cpts,
2166                                  int ncpts)
2167 {
2168         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2169         unsigned long flags;
2170         int cpt;
2171         int rc;
2172         int i;
2173
2174         tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
2175
2176         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2177         if (!tunables->lnd_map_on_demand) {
2178                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2179                 goto create_tx_pool;
2180         }
2181
2182         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2183
2184         if (tunables->lnd_fmr_pool_size < *kiblnd_tunables.kib_ntx / 4) {
2185                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2186                        tunables->lnd_fmr_pool_size,
2187                        *kiblnd_tunables.kib_ntx / 4);
2188                 rc = -EINVAL;
2189                 goto failed;
2190         }
2191
2192         /*
2193          * TX pool must be created later than FMR, see LU-2268
2194          * for details
2195          */
2196         LASSERT(!net->ibn_tx_ps);
2197
2198         /*
2199          * premapping can fail if ibd_nmr > 1, so we always create
2200          * FMR pool and map-on-demand if premapping failed
2201          *
2202          * cfs_precpt_alloc is creating an array of struct kib_fmr_poolset
2203          * The number of struct kib_fmr_poolsets create is equal to the
2204          * number of CPTs that exist, i.e net->ibn_fmr_ps[cpt].
2205          */
2206         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2207                                            sizeof(struct kib_fmr_poolset));
2208         if (!net->ibn_fmr_ps) {
2209                 CERROR("Failed to allocate FMR pool array\n");
2210                 rc = -ENOMEM;
2211                 goto failed;
2212         }
2213
2214         for (i = 0; i < ncpts; i++) {
2215                 cpt = !cpts ? i : cpts[i];
2216                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, ncpts,
2217                                              net, tunables);
2218                 if (rc) {
2219                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2220                                cpt, rc);
2221                         goto failed;
2222                 }
2223         }
2224
2225         if (i > 0)
2226                 LASSERT(i == ncpts);
2227
2228  create_tx_pool:
2229         /*
2230          * cfs_precpt_alloc is creating an array of struct kib_tx_poolset
2231          * The number of struct kib_tx_poolsets create is equal to the
2232          * number of CPTs that exist, i.e net->ibn_tx_ps[cpt].
2233          */
2234         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2235                                           sizeof(struct kib_tx_poolset));
2236         if (!net->ibn_tx_ps) {
2237                 CERROR("Failed to allocate tx pool array\n");
2238                 rc = -ENOMEM;
2239                 goto failed;
2240         }
2241
2242         for (i = 0; i < ncpts; i++) {
2243                 cpt = !cpts ? i : cpts[i];
2244                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2245                                          cpt, net, "TX",
2246                                          kiblnd_tx_pool_size(ncpts),
2247                                          kiblnd_create_tx_pool,
2248                                          kiblnd_destroy_tx_pool,
2249                                          kiblnd_tx_init, NULL);
2250                 if (rc) {
2251                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2252                                cpt, rc);
2253                         goto failed;
2254                 }
2255         }
2256
2257         return 0;
2258  failed:
2259         kiblnd_net_fini_pools(net);
2260         LASSERT(rc);
2261         return rc;
2262 }
2263
2264 static int kiblnd_hdev_get_attr(struct kib_hca_dev *hdev)
2265 {
2266         /*
2267          * It's safe to assume a HCA can handle a page size
2268          * matching that of the native system
2269          */
2270         hdev->ibh_page_shift = PAGE_SHIFT;
2271         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2272         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2273
2274         hdev->ibh_mr_size = hdev->ibh_ibdev->attrs.max_mr_size;
2275         if (hdev->ibh_mr_size == ~0ULL) {
2276                 hdev->ibh_mr_shift = 64;
2277                 return 0;
2278         }
2279
2280         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2281         return -EINVAL;
2282 }
2283
2284 static void kiblnd_hdev_cleanup_mrs(struct kib_hca_dev *hdev)
2285 {
2286         if (!hdev->ibh_mrs)
2287                 return;
2288
2289         ib_dereg_mr(hdev->ibh_mrs);
2290
2291         hdev->ibh_mrs = NULL;
2292 }
2293
2294 void kiblnd_hdev_destroy(struct kib_hca_dev *hdev)
2295 {
2296         kiblnd_hdev_cleanup_mrs(hdev);
2297
2298         if (hdev->ibh_pd)
2299                 ib_dealloc_pd(hdev->ibh_pd);
2300
2301         if (hdev->ibh_cmid)
2302                 rdma_destroy_id(hdev->ibh_cmid);
2303
2304         LIBCFS_FREE(hdev, sizeof(*hdev));
2305 }
2306
2307 static int kiblnd_hdev_setup_mrs(struct kib_hca_dev *hdev)
2308 {
2309         struct ib_mr *mr;
2310         int rc;
2311         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2312
2313         rc = kiblnd_hdev_get_attr(hdev);
2314         if (rc)
2315                 return rc;
2316
2317         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2318         if (IS_ERR(mr)) {
2319                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2320                 kiblnd_hdev_cleanup_mrs(hdev);
2321                 return PTR_ERR(mr);
2322         }
2323
2324         hdev->ibh_mrs = mr;
2325
2326         return 0;
2327 }
2328
2329 /* DUMMY */
2330 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2331                                  struct rdma_cm_event *event)
2332 {
2333         return 0;
2334 }
2335
2336 static int kiblnd_dev_need_failover(struct kib_dev *dev)
2337 {
2338         struct rdma_cm_id *cmid;
2339         struct sockaddr_in srcaddr;
2340         struct sockaddr_in dstaddr;
2341         int rc;
2342
2343         if (!dev->ibd_hdev || /* initializing */
2344             !dev->ibd_hdev->ibh_cmid || /* listener is dead */
2345             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2346                 return 1;
2347
2348         /*
2349          * XXX: it's UGLY, but I don't have better way to find
2350          * ib-bonding HCA failover because:
2351          *
2352          * a. no reliable CM event for HCA failover...
2353          * b. no OFED API to get ib_device for current net_device...
2354          *
2355          * We have only two choices at this point:
2356          *
2357          * a. rdma_bind_addr(), it will conflict with listener cmid
2358          * b. rdma_resolve_addr() to zero addr
2359          */
2360         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2361                                      IB_QPT_RC);
2362         if (IS_ERR(cmid)) {
2363                 rc = PTR_ERR(cmid);
2364                 CERROR("Failed to create cmid for failover: %d\n", rc);
2365                 return rc;
2366         }
2367
2368         memset(&srcaddr, 0, sizeof(srcaddr));
2369         srcaddr.sin_family = AF_INET;
2370         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2371
2372         memset(&dstaddr, 0, sizeof(dstaddr));
2373         dstaddr.sin_family = AF_INET;
2374         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2375                                (struct sockaddr *)&dstaddr, 1);
2376         if (rc || !cmid->device) {
2377                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2378                        dev->ibd_ifname, &dev->ibd_ifip,
2379                        cmid->device, rc);
2380                 rdma_destroy_id(cmid);
2381                 return rc;
2382         }
2383
2384         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2385         rdma_destroy_id(cmid);
2386
2387         return rc;
2388 }
2389
2390 int kiblnd_dev_failover(struct kib_dev *dev)
2391 {
2392         LIST_HEAD(zombie_tpo);
2393         LIST_HEAD(zombie_ppo);
2394         LIST_HEAD(zombie_fpo);
2395         struct rdma_cm_id *cmid  = NULL;
2396         struct kib_hca_dev *hdev  = NULL;
2397         struct ib_pd *pd;
2398         struct kib_net *net;
2399         struct sockaddr_in addr;
2400         unsigned long flags;
2401         int rc = 0;
2402         int i;
2403
2404         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2405                 dev->ibd_can_failover || !dev->ibd_hdev);
2406
2407         rc = kiblnd_dev_need_failover(dev);
2408         if (rc <= 0)
2409                 goto out;
2410
2411         if (dev->ibd_hdev &&
2412             dev->ibd_hdev->ibh_cmid) {
2413                 /*
2414                  * XXX it's not good to close old listener at here,
2415                  * because we can fail to create new listener.
2416                  * But we have to close it now, otherwise rdma_bind_addr
2417                  * will return EADDRINUSE... How crap!
2418                  */
2419                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2420
2421                 cmid = dev->ibd_hdev->ibh_cmid;
2422                 /*
2423                  * make next schedule of kiblnd_dev_need_failover()
2424                  * return 1 for me
2425                  */
2426                 dev->ibd_hdev->ibh_cmid  = NULL;
2427                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2428
2429                 rdma_destroy_id(cmid);
2430         }
2431
2432         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2433                                      IB_QPT_RC);
2434         if (IS_ERR(cmid)) {
2435                 rc = PTR_ERR(cmid);
2436                 CERROR("Failed to create cmid for failover: %d\n", rc);
2437                 goto out;
2438         }
2439
2440         memset(&addr, 0, sizeof(addr));
2441         addr.sin_family      = AF_INET;
2442         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2443         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2444
2445         /* Bind to failover device or port */
2446         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2447         if (rc || !cmid->device) {
2448                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2449                        dev->ibd_ifname, &dev->ibd_ifip,
2450                        cmid->device, rc);
2451                 rdma_destroy_id(cmid);
2452                 goto out;
2453         }
2454
2455         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2456         if (!hdev) {
2457                 CERROR("Failed to allocate kib_hca_dev\n");
2458                 rdma_destroy_id(cmid);
2459                 rc = -ENOMEM;
2460                 goto out;
2461         }
2462
2463         atomic_set(&hdev->ibh_ref, 1);
2464         hdev->ibh_dev   = dev;
2465         hdev->ibh_cmid  = cmid;
2466         hdev->ibh_ibdev = cmid->device;
2467
2468         pd = ib_alloc_pd(cmid->device, 0);
2469         if (IS_ERR(pd)) {
2470                 rc = PTR_ERR(pd);
2471                 CERROR("Can't allocate PD: %d\n", rc);
2472                 goto out;
2473         }
2474
2475         hdev->ibh_pd = pd;
2476
2477         rc = rdma_listen(cmid, 0);
2478         if (rc) {
2479                 CERROR("Can't start new listener: %d\n", rc);
2480                 goto out;
2481         }
2482
2483         rc = kiblnd_hdev_setup_mrs(hdev);
2484         if (rc) {
2485                 CERROR("Can't setup device: %d\n", rc);
2486                 goto out;
2487         }
2488
2489         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2490
2491         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2492
2493         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2494                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2495                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2496                                             &zombie_tpo);
2497
2498                         if (net->ibn_fmr_ps)
2499                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2500                                                         &zombie_fpo);
2501                 }
2502         }
2503
2504         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2505  out:
2506         if (!list_empty(&zombie_tpo))
2507                 kiblnd_destroy_pool_list(&zombie_tpo);
2508         if (!list_empty(&zombie_ppo))
2509                 kiblnd_destroy_pool_list(&zombie_ppo);
2510         if (!list_empty(&zombie_fpo))
2511                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2512         if (hdev)
2513                 kiblnd_hdev_decref(hdev);
2514
2515         if (rc)
2516                 dev->ibd_failed_failover++;
2517         else
2518                 dev->ibd_failed_failover = 0;
2519
2520         return rc;
2521 }
2522
2523 void kiblnd_destroy_dev(struct kib_dev *dev)
2524 {
2525         LASSERT(!dev->ibd_nnets);
2526         LASSERT(list_empty(&dev->ibd_nets));
2527
2528         list_del(&dev->ibd_fail_list);
2529         list_del(&dev->ibd_list);
2530
2531         if (dev->ibd_hdev)
2532                 kiblnd_hdev_decref(dev->ibd_hdev);
2533
2534         LIBCFS_FREE(dev, sizeof(*dev));
2535 }
2536
2537 static struct kib_dev *kiblnd_create_dev(char *ifname)
2538 {
2539         struct net_device *netdev;
2540         struct kib_dev *dev;
2541         __u32 netmask;
2542         __u32 ip;
2543         int up;
2544         int rc;
2545
2546         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2547         if (rc) {
2548                 CERROR("Can't query IPoIB interface %s: %d\n",
2549                        ifname, rc);
2550                 return NULL;
2551         }
2552
2553         if (!up) {
2554                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2555                 return NULL;
2556         }
2557
2558         LIBCFS_ALLOC(dev, sizeof(*dev));
2559         if (!dev)
2560                 return NULL;
2561
2562         netdev = dev_get_by_name(&init_net, ifname);
2563         if (!netdev) {
2564                 dev->ibd_can_failover = 0;
2565         } else {
2566                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2567                 dev_put(netdev);
2568         }
2569
2570         INIT_LIST_HEAD(&dev->ibd_nets);
2571         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2572         INIT_LIST_HEAD(&dev->ibd_fail_list);
2573         dev->ibd_ifip = ip;
2574         strcpy(&dev->ibd_ifname[0], ifname);
2575
2576         /* initialize the device */
2577         rc = kiblnd_dev_failover(dev);
2578         if (rc) {
2579                 CERROR("Can't initialize device: %d\n", rc);
2580                 LIBCFS_FREE(dev, sizeof(*dev));
2581                 return NULL;
2582         }
2583
2584         list_add_tail(&dev->ibd_list, &kiblnd_data.kib_devs);
2585         return dev;
2586 }
2587
2588 static void kiblnd_base_shutdown(void)
2589 {
2590         struct kib_sched_info *sched;
2591         int i;
2592
2593         LASSERT(list_empty(&kiblnd_data.kib_devs));
2594
2595         switch (kiblnd_data.kib_init) {
2596         default:
2597                 LBUG();
2598
2599         case IBLND_INIT_ALL:
2600         case IBLND_INIT_DATA:
2601                 LASSERT(kiblnd_data.kib_peers);
2602                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2603                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2604                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2605                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2606                 LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
2607                 LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
2608
2609                 /* flag threads to terminate; wake and wait for them to die */
2610                 kiblnd_data.kib_shutdown = 1;
2611
2612                 /*
2613                  * NB: we really want to stop scheduler threads net by net
2614                  * instead of the whole module, this should be improved
2615                  * with dynamic configuration LNet
2616                  */
2617                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2618                         wake_up_all(&sched->ibs_waitq);
2619
2620                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2621                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2622
2623                 i = 2;
2624                 while (atomic_read(&kiblnd_data.kib_nthreads)) {
2625                         i++;
2626                         /* power of 2 ? */
2627                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2628                                "Waiting for %d threads to terminate\n",
2629                                atomic_read(&kiblnd_data.kib_nthreads));
2630                         set_current_state(TASK_UNINTERRUPTIBLE);
2631                         schedule_timeout(cfs_time_seconds(1));
2632                 }
2633
2634                 /* fall through */
2635
2636         case IBLND_INIT_NOTHING:
2637                 break;
2638         }
2639
2640         if (kiblnd_data.kib_peers) {
2641                 LIBCFS_FREE(kiblnd_data.kib_peers,
2642                             sizeof(struct list_head) *
2643                             kiblnd_data.kib_peer_hash_size);
2644         }
2645
2646         if (kiblnd_data.kib_scheds)
2647                 cfs_percpt_free(kiblnd_data.kib_scheds);
2648
2649         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2650         module_put(THIS_MODULE);
2651 }
2652
2653 static void kiblnd_shutdown(lnet_ni_t *ni)
2654 {
2655         struct kib_net *net = ni->ni_data;
2656         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2657         int i;
2658         unsigned long flags;
2659
2660         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2661
2662         if (!net)
2663                 goto out;
2664
2665         write_lock_irqsave(g_lock, flags);
2666         net->ibn_shutdown = 1;
2667         write_unlock_irqrestore(g_lock, flags);
2668
2669         switch (net->ibn_init) {
2670         default:
2671                 LBUG();
2672
2673         case IBLND_INIT_ALL:
2674                 /* nuke all existing peers within this net */
2675                 kiblnd_del_peer(ni, LNET_NID_ANY);
2676
2677                 /* Wait for all peer state to clean up */
2678                 i = 2;
2679                 while (atomic_read(&net->ibn_npeers)) {
2680                         i++;
2681                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2682                                "%s: waiting for %d peers to disconnect\n",
2683                                libcfs_nid2str(ni->ni_nid),
2684                                atomic_read(&net->ibn_npeers));
2685                         set_current_state(TASK_UNINTERRUPTIBLE);
2686                         schedule_timeout(cfs_time_seconds(1));
2687                 }
2688
2689                 kiblnd_net_fini_pools(net);
2690
2691                 write_lock_irqsave(g_lock, flags);
2692                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2693                 net->ibn_dev->ibd_nnets--;
2694                 list_del(&net->ibn_list);
2695                 write_unlock_irqrestore(g_lock, flags);
2696
2697                 /* fall through */
2698
2699         case IBLND_INIT_NOTHING:
2700                 LASSERT(!atomic_read(&net->ibn_nconns));
2701
2702                 if (net->ibn_dev && !net->ibn_dev->ibd_nnets)
2703                         kiblnd_destroy_dev(net->ibn_dev);
2704
2705                 break;
2706         }
2707
2708         net->ibn_init = IBLND_INIT_NOTHING;
2709         ni->ni_data = NULL;
2710
2711         LIBCFS_FREE(net, sizeof(*net));
2712
2713 out:
2714         if (list_empty(&kiblnd_data.kib_devs))
2715                 kiblnd_base_shutdown();
2716 }
2717
2718 static int kiblnd_base_startup(void)
2719 {
2720         struct kib_sched_info *sched;
2721         int rc;
2722         int i;
2723
2724         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2725
2726         try_module_get(THIS_MODULE);
2727         /* zero pointers, flags etc */
2728         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2729
2730         rwlock_init(&kiblnd_data.kib_global_lock);
2731
2732         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2733         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2734
2735         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2736         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2737                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2738         if (!kiblnd_data.kib_peers)
2739                 goto failed;
2740         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2741                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2742
2743         spin_lock_init(&kiblnd_data.kib_connd_lock);
2744         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2745         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2746         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
2747         INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
2748
2749         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2750         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2751
2752         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2753                                                   sizeof(*sched));
2754         if (!kiblnd_data.kib_scheds)
2755                 goto failed;
2756
2757         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2758                 int nthrs;
2759
2760                 spin_lock_init(&sched->ibs_lock);
2761                 INIT_LIST_HEAD(&sched->ibs_conns);
2762                 init_waitqueue_head(&sched->ibs_waitq);
2763
2764                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2765                 if (*kiblnd_tunables.kib_nscheds > 0) {
2766                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2767                 } else {
2768                         /*
2769                          * max to half of CPUs, another half is reserved for
2770                          * upper layer modules
2771                          */
2772                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2773                 }
2774
2775                 sched->ibs_nthreads_max = nthrs;
2776                 sched->ibs_cpt = i;
2777         }
2778
2779         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2780
2781         /* lists/ptrs/locks initialised */
2782         kiblnd_data.kib_init = IBLND_INIT_DATA;
2783         /*****************************************************/
2784
2785         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2786         if (rc) {
2787                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2788                 goto failed;
2789         }
2790
2791         if (*kiblnd_tunables.kib_dev_failover)
2792                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2793                                          "kiblnd_failover");
2794
2795         if (rc) {
2796                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2797                 goto failed;
2798         }
2799
2800         /* flag everything initialised */
2801         kiblnd_data.kib_init = IBLND_INIT_ALL;
2802         /*****************************************************/
2803
2804         return 0;
2805
2806  failed:
2807         kiblnd_base_shutdown();
2808         return -ENETDOWN;
2809 }
2810
2811 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2812 {
2813         int rc = 0;
2814         int nthrs;
2815         int i;
2816
2817         if (!sched->ibs_nthreads) {
2818                 if (*kiblnd_tunables.kib_nscheds > 0) {
2819                         nthrs = sched->ibs_nthreads_max;
2820                 } else {
2821                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2822                                                sched->ibs_cpt);
2823                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2824                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2825                 }
2826         } else {
2827                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2828                 /* increase one thread if there is new interface */
2829                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2830         }
2831
2832         for (i = 0; i < nthrs; i++) {
2833                 long id;
2834                 char name[20];
2835
2836                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2837                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2838                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2839                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2840                 if (!rc)
2841                         continue;
2842
2843                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2844                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2845                 break;
2846         }
2847
2848         sched->ibs_nthreads += i;
2849         return rc;
2850 }
2851
2852 static int kiblnd_dev_start_threads(struct kib_dev *dev, int newdev, __u32 *cpts,
2853                                     int ncpts)
2854 {
2855         int cpt;
2856         int rc;
2857         int i;
2858
2859         for (i = 0; i < ncpts; i++) {
2860                 struct kib_sched_info *sched;
2861
2862                 cpt = !cpts ? i : cpts[i];
2863                 sched = kiblnd_data.kib_scheds[cpt];
2864
2865                 if (!newdev && sched->ibs_nthreads > 0)
2866                         continue;
2867
2868                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2869                 if (rc) {
2870                         CERROR("Failed to start scheduler threads for %s\n",
2871                                dev->ibd_ifname);
2872                         return rc;
2873                 }
2874         }
2875         return 0;
2876 }
2877
2878 static struct kib_dev *kiblnd_dev_search(char *ifname)
2879 {
2880         struct kib_dev *alias = NULL;
2881         struct kib_dev *dev;
2882         char *colon;
2883         char *colon2;
2884
2885         colon = strchr(ifname, ':');
2886         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2887                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2888                         return dev;
2889
2890                 if (alias)
2891                         continue;
2892
2893                 colon2 = strchr(dev->ibd_ifname, ':');
2894                 if (colon)
2895                         *colon = 0;
2896                 if (colon2)
2897                         *colon2 = 0;
2898
2899                 if (!strcmp(&dev->ibd_ifname[0], ifname))
2900                         alias = dev;
2901
2902                 if (colon)
2903                         *colon = ':';
2904                 if (colon2)
2905                         *colon2 = ':';
2906         }
2907         return alias;
2908 }
2909
2910 static int kiblnd_startup(lnet_ni_t *ni)
2911 {
2912         char *ifname;
2913         struct kib_dev *ibdev = NULL;
2914         struct kib_net *net;
2915         struct timespec64 tv;
2916         unsigned long flags;
2917         int rc;
2918         int newdev;
2919
2920         LASSERT(ni->ni_lnd == &the_o2iblnd);
2921
2922         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2923                 rc = kiblnd_base_startup();
2924                 if (rc)
2925                         return rc;
2926         }
2927
2928         LIBCFS_ALLOC(net, sizeof(*net));
2929         ni->ni_data = net;
2930         if (!net)
2931                 goto net_failed;
2932
2933         ktime_get_real_ts64(&tv);
2934         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2935                                tv.tv_nsec / NSEC_PER_USEC;
2936
2937         rc = kiblnd_tunables_setup(ni);
2938         if (rc)
2939                 goto net_failed;
2940
2941         if (ni->ni_interfaces[0]) {
2942                 /* Use the IPoIB interface specified in 'networks=' */
2943
2944                 CLASSERT(LNET_MAX_INTERFACES > 1);
2945                 if (ni->ni_interfaces[1]) {
2946                         CERROR("Multiple interfaces not supported\n");
2947                         goto failed;
2948                 }
2949
2950                 ifname = ni->ni_interfaces[0];
2951         } else {
2952                 ifname = *kiblnd_tunables.kib_default_ipif;
2953         }
2954
2955         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2956                 CERROR("IPoIB interface name too long: %s\n", ifname);
2957                 goto failed;
2958         }
2959
2960         ibdev = kiblnd_dev_search(ifname);
2961
2962         newdev = !ibdev;
2963         /* hmm...create kib_dev even for alias */
2964         if (!ibdev || strcmp(&ibdev->ibd_ifname[0], ifname))
2965                 ibdev = kiblnd_create_dev(ifname);
2966
2967         if (!ibdev)
2968                 goto failed;
2969
2970         net->ibn_dev = ibdev;
2971         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2972
2973         rc = kiblnd_dev_start_threads(ibdev, newdev,
2974                                       ni->ni_cpts, ni->ni_ncpts);
2975         if (rc)
2976                 goto failed;
2977
2978         rc = kiblnd_net_init_pools(net, ni, ni->ni_cpts, ni->ni_ncpts);
2979         if (rc) {
2980                 CERROR("Failed to initialize NI pools: %d\n", rc);
2981                 goto failed;
2982         }
2983
2984         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2985         ibdev->ibd_nnets++;
2986         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2987         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2988
2989         net->ibn_init = IBLND_INIT_ALL;
2990
2991         return 0;
2992
2993 failed:
2994         if (!net->ibn_dev && ibdev)
2995                 kiblnd_destroy_dev(ibdev);
2996
2997 net_failed:
2998         kiblnd_shutdown(ni);
2999
3000         CDEBUG(D_NET, "kiblnd_startup failed\n");
3001         return -ENETDOWN;
3002 }
3003
3004 static lnd_t the_o2iblnd = {
3005         .lnd_type       = O2IBLND,
3006         .lnd_startup    = kiblnd_startup,
3007         .lnd_shutdown   = kiblnd_shutdown,
3008         .lnd_ctl        = kiblnd_ctl,
3009         .lnd_query      = kiblnd_query,
3010         .lnd_send       = kiblnd_send,
3011         .lnd_recv       = kiblnd_recv,
3012 };
3013
3014 static void __exit ko2iblnd_exit(void)
3015 {
3016         lnet_unregister_lnd(&the_o2iblnd);
3017 }
3018
3019 static int __init ko2iblnd_init(void)
3020 {
3021         CLASSERT(sizeof(struct kib_msg) <= IBLND_MSG_SIZE);
3022         CLASSERT(offsetof(struct kib_msg,
3023                           ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3024                           <= IBLND_MSG_SIZE);
3025         CLASSERT(offsetof(struct kib_msg,
3026                           ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
3027                           <= IBLND_MSG_SIZE);
3028
3029         kiblnd_tunables_init();
3030
3031         lnet_register_lnd(&the_o2iblnd);
3032
3033         return 0;
3034 }
3035
3036 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3037 MODULE_DESCRIPTION("OpenIB gen2 LNet Network Driver");
3038 MODULE_VERSION("2.7.0");
3039 MODULE_LICENSE("GPL");
3040
3041 module_init(ko2iblnd_init);
3042 module_exit(ko2iblnd_exit);