kernel/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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include <asm/div64.h>
  42 #include <asm/page.h>
  43 #include "o2iblnd.h"
  44
  45 static lnd_t the_o2iblnd = {
  46         .lnd_type     = O2IBLND,
  47         .lnd_startup  = kiblnd_startup,
  48         .lnd_shutdown = kiblnd_shutdown,
  49         .lnd_ctl      = kiblnd_ctl,
  50         .lnd_query    = kiblnd_query,
  51         .lnd_send     = kiblnd_send,
  52         .lnd_recv     = kiblnd_recv,
  53 };
  54
  55 kib_data_t kiblnd_data;
  56
  57 static __u32 kiblnd_cksum(void *ptr, int nob)
  58 {
  59         char *c = ptr;
  60         __u32 sum = 0;
  61
  62         while (nob-- > 0)
  63                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  64
  65         /* ensure I don't return 0 (== no checksum) */
  66         return (sum == 0) ? 1 : sum;
  67 }
  68
  69 static char *kiblnd_msgtype2str(int type)
  70 {
  71         switch (type) {
  72         case IBLND_MSG_CONNREQ:
  73                 return "CONNREQ";
  74
  75         case IBLND_MSG_CONNACK:
  76                 return "CONNACK";
  77
  78         case IBLND_MSG_NOOP:
  79                 return "NOOP";
  80
  81         case IBLND_MSG_IMMEDIATE:
  82                 return "IMMEDIATE";
  83
  84         case IBLND_MSG_PUT_REQ:
  85                 return "PUT_REQ";
  86
  87         case IBLND_MSG_PUT_NAK:
  88                 return "PUT_NAK";
  89
  90         case IBLND_MSG_PUT_ACK:
  91                 return "PUT_ACK";
  92
  93         case IBLND_MSG_PUT_DONE:
  94                 return "PUT_DONE";
  95
  96         case IBLND_MSG_GET_REQ:
  97                 return "GET_REQ";
  98
  99         case IBLND_MSG_GET_DONE:
 100                 return "GET_DONE";
 101
 102         default:
 103                 return "???";
 104         }
 105 }
 106
 107 static int kiblnd_msgtype2size(int type)
 108 {
 109         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 110
 111         switch (type) {
 112         case IBLND_MSG_CONNREQ:
 113         case IBLND_MSG_CONNACK:
 114                 return hdr_size + sizeof(kib_connparams_t);
 115
 116         case IBLND_MSG_NOOP:
 117                 return hdr_size;
 118
 119         case IBLND_MSG_IMMEDIATE:
 120                 return offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]);
 121
 122         case IBLND_MSG_PUT_REQ:
 123                 return hdr_size + sizeof(kib_putreq_msg_t);
 124
 125         case IBLND_MSG_PUT_ACK:
 126                 return hdr_size + sizeof(kib_putack_msg_t);
 127
 128         case IBLND_MSG_GET_REQ:
 129                 return hdr_size + sizeof(kib_get_msg_t);
 130
 131         case IBLND_MSG_PUT_NAK:
 132         case IBLND_MSG_PUT_DONE:
 133         case IBLND_MSG_GET_DONE:
 134                 return hdr_size + sizeof(kib_completion_msg_t);
 135         default:
 136                 return -1;
 137         }
 138 }
 139
 140 static int kiblnd_unpack_rd(kib_msg_t *msg, int flip)
 141 {
 142         kib_rdma_desc_t *rd;
 143         int nob;
 144         int n;
 145         int i;
 146
 147         LASSERT(msg->ibm_type == IBLND_MSG_GET_REQ ||
 148                  msg->ibm_type == IBLND_MSG_PUT_ACK);
 149
 150         rd = msg->ibm_type == IBLND_MSG_GET_REQ ?
 151                               &msg->ibm_u.get.ibgm_rd :
 152                               &msg->ibm_u.putack.ibpam_rd;
 153
 154         if (flip) {
 155                 __swab32s(&rd->rd_key);
 156                 __swab32s(&rd->rd_nfrags);
 157         }
 158
 159         n = rd->rd_nfrags;
 160
 161         if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 162                 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
 163                        n, IBLND_MAX_RDMA_FRAGS);
 164                 return 1;
 165         }
 166
 167         nob = offsetof(kib_msg_t, ibm_u) +
 168               kiblnd_rd_msg_size(rd, msg->ibm_type, n);
 169
 170         if (msg->ibm_nob < nob) {
 171                 CERROR("Short %s: %d(%d)\n",
 172                        kiblnd_msgtype2str(msg->ibm_type), msg->ibm_nob, nob);
 173                 return 1;
 174         }
 175
 176         if (!flip)
 177                 return 0;
 178
 179         for (i = 0; i < n; i++) {
 180                 __swab32s(&rd->rd_frags[i].rf_nob);
 181                 __swab64s(&rd->rd_frags[i].rf_addr);
 182         }
 183
 184         return 0;
 185 }
 186
 187 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
 188                      int credits, lnet_nid_t dstnid, __u64 dststamp)
 189 {
 190         kib_net_t *net = ni->ni_data;
 191
 192         /* CAVEAT EMPTOR! all message fields not set here should have been
 193          * initialised previously. */
 194         msg->ibm_magic    = IBLND_MSG_MAGIC;
 195         msg->ibm_version  = version;
 196         /*   ibm_type */
 197         msg->ibm_credits  = credits;
 198         /*   ibm_nob */
 199         msg->ibm_cksum    = 0;
 200         msg->ibm_srcnid   = ni->ni_nid;
 201         msg->ibm_srcstamp = net->ibn_incarnation;
 202         msg->ibm_dstnid   = dstnid;
 203         msg->ibm_dststamp = dststamp;
 204
 205         if (*kiblnd_tunables.kib_cksum) {
 206                 /* NB ibm_cksum zero while computing cksum */
 207                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
 208         }
 209 }
 210
 211 int kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 212 {
 213         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 214         __u32 msg_cksum;
 215         __u16 version;
 216         int msg_nob;
 217         int flip;
 218
 219         /* 6 bytes are enough to have received magic + version */
 220         if (nob < 6) {
 221                 CERROR("Short message: %d\n", nob);
 222                 return -EPROTO;
 223         }
 224
 225         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 226                 flip = 0;
 227         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 228                 flip = 1;
 229         } else {
 230                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 231                 return -EPROTO;
 232         }
 233
 234         version = flip ? __swab16(msg->ibm_version) : msg->ibm_version;
 235         if (version != IBLND_MSG_VERSION &&
 236             version != IBLND_MSG_VERSION_1) {
 237                 CERROR("Bad version: %x\n", version);
 238                 return -EPROTO;
 239         }
 240
 241         if (nob < hdr_size) {
 242                 CERROR("Short message: %d\n", nob);
 243                 return -EPROTO;
 244         }
 245
 246         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 247         if (msg_nob > nob) {
 248                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 249                 return -EPROTO;
 250         }
 251
 252         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 253          * gets flipped */
 254         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 255         msg->ibm_cksum = 0;
 256         if (msg_cksum != 0 &&
 257             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 258                 CERROR("Bad checksum\n");
 259                 return -EPROTO;
 260         }
 261
 262         msg->ibm_cksum = msg_cksum;
 263
 264         if (flip) {
 265                 /* leave magic unflipped as a clue to peer endianness */
 266                 msg->ibm_version = version;
 267                 CLASSERT(sizeof(msg->ibm_type) == 1);
 268                 CLASSERT(sizeof(msg->ibm_credits) == 1);
 269                 msg->ibm_nob     = msg_nob;
 270                 __swab64s(&msg->ibm_srcnid);
 271                 __swab64s(&msg->ibm_srcstamp);
 272                 __swab64s(&msg->ibm_dstnid);
 273                 __swab64s(&msg->ibm_dststamp);
 274         }
 275
 276         if (msg->ibm_srcnid == LNET_NID_ANY) {
 277                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 278                 return -EPROTO;
 279         }
 280
 281         if (msg_nob < kiblnd_msgtype2size(msg->ibm_type)) {
 282                 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg->ibm_type),
 283                        msg_nob, kiblnd_msgtype2size(msg->ibm_type));
 284                 return -EPROTO;
 285         }
 286
 287         switch (msg->ibm_type) {
 288         default:
 289                 CERROR("Unknown message type %x\n", msg->ibm_type);
 290                 return -EPROTO;
 291
 292         case IBLND_MSG_NOOP:
 293         case IBLND_MSG_IMMEDIATE:
 294         case IBLND_MSG_PUT_REQ:
 295                 break;
 296
 297         case IBLND_MSG_PUT_ACK:
 298         case IBLND_MSG_GET_REQ:
 299                 if (kiblnd_unpack_rd(msg, flip))
 300                         return -EPROTO;
 301                 break;
 302
 303         case IBLND_MSG_PUT_NAK:
 304         case IBLND_MSG_PUT_DONE:
 305         case IBLND_MSG_GET_DONE:
 306                 if (flip)
 307                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 308                 break;
 309
 310         case IBLND_MSG_CONNREQ:
 311         case IBLND_MSG_CONNACK:
 312                 if (flip) {
 313                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 314                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 315                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 316                 }
 317                 break;
 318         }
 319         return 0;
 320 }
 321
 322 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 323 {
 324         kib_peer_t *peer;
 325         kib_net_t *net = ni->ni_data;
 326         int cpt = lnet_cpt_of_nid(nid);
 327         unsigned long flags;
 328
 329         LASSERT(net != NULL);
 330         LASSERT(nid != LNET_NID_ANY);
 331
 332         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 333         if (peer == NULL) {
 334                 CERROR("Cannot allocate peer\n");
 335                 return -ENOMEM;
 336         }
 337
 338         memset(peer, 0, sizeof(*peer));  /* zero flags etc */
 339
 340         peer->ibp_ni = ni;
 341         peer->ibp_nid = nid;
 342         peer->ibp_error = 0;
 343         peer->ibp_last_alive = 0;
 344         atomic_set(&peer->ibp_refcount, 1);  /* 1 ref for caller */
 345
 346         INIT_LIST_HEAD(&peer->ibp_list);     /* not in the peer table yet */
 347         INIT_LIST_HEAD(&peer->ibp_conns);
 348         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 349
 350         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 351
 352         /* always called with a ref on ni, which prevents ni being shutdown */
 353         LASSERT(net->ibn_shutdown == 0);
 354
 355         /* npeers only grows with the global lock held */
 356         atomic_inc(&net->ibn_npeers);
 357
 358         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 359
 360         *peerp = peer;
 361         return 0;
 362 }
 363
 364 void kiblnd_destroy_peer(kib_peer_t *peer)
 365 {
 366         kib_net_t *net = peer->ibp_ni->ni_data;
 367
 368         LASSERT(net != NULL);
 369         LASSERT(atomic_read(&peer->ibp_refcount) == 0);
 370         LASSERT(!kiblnd_peer_active(peer));
 371         LASSERT(peer->ibp_connecting == 0);
 372         LASSERT(peer->ibp_accepting == 0);
 373         LASSERT(list_empty(&peer->ibp_conns));
 374         LASSERT(list_empty(&peer->ibp_tx_queue));
 375
 376         LIBCFS_FREE(peer, sizeof(*peer));
 377
 378         /* NB a peer's connections keep a reference on their peer until
 379          * they are destroyed, so we can be assured that _all_ state to do
 380          * with this peer has been cleaned up when its refcount drops to
 381          * zero. */
 382         atomic_dec(&net->ibn_npeers);
 383 }
 384
 385 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
 386 {
 387         /* the caller is responsible for accounting the additional reference
 388          * that this creates */
 389         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 390         struct list_head *tmp;
 391         kib_peer_t *peer;
 392
 393         list_for_each(tmp, peer_list) {
 394
 395                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 396
 397                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
 398                          peer->ibp_accepting > 0 ||
 399                          !list_empty(&peer->ibp_conns));  /* active conn */
 400
 401                 if (peer->ibp_nid != nid)
 402                         continue;
 403
 404                 CDEBUG(D_NET, "got peer [%p] -> %s (%d) version: %x\n",
 405                        peer, libcfs_nid2str(nid),
 406                        atomic_read(&peer->ibp_refcount),
 407                        peer->ibp_version);
 408                 return peer;
 409         }
 410         return NULL;
 411 }
 412
 413 void kiblnd_unlink_peer_locked(kib_peer_t *peer)
 414 {
 415         LASSERT(list_empty(&peer->ibp_conns));
 416
 417         LASSERT(kiblnd_peer_active(peer));
 418         list_del_init(&peer->ibp_list);
 419         /* lose peerlist's ref */
 420         kiblnd_peer_decref(peer);
 421 }
 422
 423 static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 424                                 lnet_nid_t *nidp, int *count)
 425 {
 426         kib_peer_t *peer;
 427         struct list_head *ptmp;
 428         int i;
 429         unsigned long flags;
 430
 431         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 432
 433         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 434
 435                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 436
 437                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 438                         LASSERT(peer->ibp_connecting > 0 ||
 439                                  peer->ibp_accepting > 0 ||
 440                                  !list_empty(&peer->ibp_conns));
 441
 442                         if (peer->ibp_ni != ni)
 443                                 continue;
 444
 445                         if (index-- > 0)
 446                                 continue;
 447
 448                         *nidp = peer->ibp_nid;
 449                         *count = atomic_read(&peer->ibp_refcount);
 450
 451                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 452                                                flags);
 453                         return 0;
 454                 }
 455         }
 456
 457         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 458         return -ENOENT;
 459 }
 460
 461 static void kiblnd_del_peer_locked(kib_peer_t *peer)
 462 {
 463         struct list_head *ctmp;
 464         struct list_head *cnxt;
 465         kib_conn_t *conn;
 466
 467         if (list_empty(&peer->ibp_conns)) {
 468                 kiblnd_unlink_peer_locked(peer);
 469         } else {
 470                 list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 471                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 472
 473                         kiblnd_close_conn_locked(conn, 0);
 474                 }
 475                 /* NB closing peer's last conn unlinked it. */
 476         }
 477         /* NB peer now unlinked; might even be freed if the peer table had the
 478          * last ref on it. */
 479 }
 480
 481 static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 482 {
 483         LIST_HEAD(zombies);
 484         struct list_head *ptmp;
 485         struct list_head *pnxt;
 486         kib_peer_t *peer;
 487         int lo;
 488         int hi;
 489         int i;
 490         unsigned long flags;
 491         int rc = -ENOENT;
 492
 493         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 494
 495         if (nid != LNET_NID_ANY) {
 496                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 497         } else {
 498                 lo = 0;
 499                 hi = kiblnd_data.kib_peer_hash_size - 1;
 500         }
 501
 502         for (i = lo; i <= hi; i++) {
 503                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 504                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 505                         LASSERT(peer->ibp_connecting > 0 ||
 506                                  peer->ibp_accepting > 0 ||
 507                                  !list_empty(&peer->ibp_conns));
 508
 509                         if (peer->ibp_ni != ni)
 510                                 continue;
 511
 512                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 513                                 continue;
 514
 515                         if (!list_empty(&peer->ibp_tx_queue)) {
 516                                 LASSERT(list_empty(&peer->ibp_conns));
 517
 518                                 list_splice_init(&peer->ibp_tx_queue,
 519                                                      &zombies);
 520                         }
 521
 522                         kiblnd_del_peer_locked(peer);
 523                         rc = 0;  /* matched something */
 524                 }
 525         }
 526
 527         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 528
 529         kiblnd_txlist_done(ni, &zombies, -EIO);
 530
 531         return rc;
 532 }
 533
 534 static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 535 {
 536         kib_peer_t *peer;
 537         struct list_head *ptmp;
 538         kib_conn_t *conn;
 539         struct list_head *ctmp;
 540         int i;
 541         unsigned long flags;
 542
 543         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 544
 545         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 546                 list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 547
 548                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 549                         LASSERT(peer->ibp_connecting > 0 ||
 550                                  peer->ibp_accepting > 0 ||
 551                                  !list_empty(&peer->ibp_conns));
 552
 553                         if (peer->ibp_ni != ni)
 554                                 continue;
 555
 556                         list_for_each(ctmp, &peer->ibp_conns) {
 557                                 if (index-- > 0)
 558                                         continue;
 559
 560                                 conn = list_entry(ctmp, kib_conn_t,
 561                                                       ibc_list);
 562                                 kiblnd_conn_addref(conn);
 563                                 read_unlock_irqrestore(
 564                                         &kiblnd_data.kib_global_lock,
 565                                         flags);
 566                                 return conn;
 567                         }
 568                 }
 569         }
 570
 571         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 572         return NULL;
 573 }
 574
 575 int kiblnd_translate_mtu(int value)
 576 {
 577         switch (value) {
 578         default:
 579                 return -1;
 580         case 0:
 581                 return 0;
 582         case 256:
 583                 return IB_MTU_256;
 584         case 512:
 585                 return IB_MTU_512;
 586         case 1024:
 587                 return IB_MTU_1024;
 588         case 2048:
 589                 return IB_MTU_2048;
 590         case 4096:
 591                 return IB_MTU_4096;
 592         }
 593 }
 594
 595 static void kiblnd_setup_mtu_locked(struct rdma_cm_id *cmid)
 596 {
 597         int mtu;
 598
 599         /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
 600         if (cmid->route.path_rec == NULL)
 601                 return;
 602
 603         mtu = kiblnd_translate_mtu(*kiblnd_tunables.kib_ib_mtu);
 604         LASSERT(mtu >= 0);
 605         if (mtu != 0)
 606                 cmid->route.path_rec->mtu = mtu;
 607 }
 608
 609 static int kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 610 {
 611         cpumask_t *mask;
 612         int vectors;
 613         int off;
 614         int i;
 615         lnet_nid_t nid = conn->ibc_peer->ibp_nid;
 616
 617         vectors = conn->ibc_cmid->device->num_comp_vectors;
 618         if (vectors <= 1)
 619                 return 0;
 620
 621         mask = cfs_cpt_cpumask(lnet_cpt_table(), cpt);
 622         if (mask == NULL)
 623                 return 0;
 624
 625         /* hash NID to CPU id in this partition... */
 626         off = do_div(nid, cpumask_weight(mask));
 627         for_each_cpu(i, mask) {
 628                 if (off-- == 0)
 629                         return i % vectors;
 630         }
 631
 632         LBUG();
 633         return 1;
 634 }
 635
 636 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 637                                 int state, int version)
 638 {
 639         /* CAVEAT EMPTOR:
 640          * If the new conn is created successfully it takes over the caller's
 641          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 642          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 643          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 644          * to destroy 'cmid' here since I'm called from the CM which still has
 645          * its ref on 'cmid'). */
 646         rwlock_t *glock = &kiblnd_data.kib_global_lock;
 647         kib_net_t *net = peer->ibp_ni->ni_data;
 648         kib_dev_t *dev;
 649         struct ib_qp_init_attr *init_qp_attr;
 650         struct kib_sched_info *sched;
 651         struct ib_cq_init_attr cq_attr = {};
 652         kib_conn_t *conn;
 653         struct ib_cq *cq;
 654         unsigned long flags;
 655         int cpt;
 656         int rc;
 657         int i;
 658
 659         LASSERT(net != NULL);
 660         LASSERT(!in_interrupt());
 661
 662         dev = net->ibn_dev;
 663
 664         cpt = lnet_cpt_of_nid(peer->ibp_nid);
 665         sched = kiblnd_data.kib_scheds[cpt];
 666
 667         LASSERT(sched->ibs_nthreads > 0);
 668
 669         LIBCFS_CPT_ALLOC(init_qp_attr, lnet_cpt_table(), cpt,
 670                          sizeof(*init_qp_attr));
 671         if (init_qp_attr == NULL) {
 672                 CERROR("Can't allocate qp_attr for %s\n",
 673                        libcfs_nid2str(peer->ibp_nid));
 674                 goto failed_0;
 675         }
 676
 677         LIBCFS_CPT_ALLOC(conn, lnet_cpt_table(), cpt, sizeof(*conn));
 678         if (conn == NULL) {
 679                 CERROR("Can't allocate connection for %s\n",
 680                        libcfs_nid2str(peer->ibp_nid));
 681                 goto failed_1;
 682         }
 683
 684         conn->ibc_state = IBLND_CONN_INIT;
 685         conn->ibc_version = version;
 686         conn->ibc_peer = peer;            /* I take the caller's ref */
 687         cmid->context = conn;              /* for future CM callbacks */
 688         conn->ibc_cmid = cmid;
 689
 690         INIT_LIST_HEAD(&conn->ibc_early_rxs);
 691         INIT_LIST_HEAD(&conn->ibc_tx_noops);
 692         INIT_LIST_HEAD(&conn->ibc_tx_queue);
 693         INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
 694         INIT_LIST_HEAD(&conn->ibc_tx_queue_nocred);
 695         INIT_LIST_HEAD(&conn->ibc_active_txs);
 696         spin_lock_init(&conn->ibc_lock);
 697
 698         LIBCFS_CPT_ALLOC(conn->ibc_connvars, lnet_cpt_table(), cpt,
 699                          sizeof(*conn->ibc_connvars));
 700         if (conn->ibc_connvars == NULL) {
 701                 CERROR("Can't allocate in-progress connection state\n");
 702                 goto failed_2;
 703         }
 704
 705         write_lock_irqsave(glock, flags);
 706         if (dev->ibd_failover) {
 707                 write_unlock_irqrestore(glock, flags);
 708                 CERROR("%s: failover in progress\n", dev->ibd_ifname);
 709                 goto failed_2;
 710         }
 711
 712         if (dev->ibd_hdev->ibh_ibdev != cmid->device) {
 713                 /* wakeup failover thread and teardown connection */
 714                 if (kiblnd_dev_can_failover(dev)) {
 715                         list_add_tail(&dev->ibd_fail_list,
 716                                       &kiblnd_data.kib_failed_devs);
 717                         wake_up(&kiblnd_data.kib_failover_waitq);
 718                 }
 719
 720                 write_unlock_irqrestore(glock, flags);
 721                 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
 722                        cmid->device->name, dev->ibd_ifname);
 723                 goto failed_2;
 724         }
 725
 726         kiblnd_hdev_addref_locked(dev->ibd_hdev);
 727         conn->ibc_hdev = dev->ibd_hdev;
 728
 729         kiblnd_setup_mtu_locked(cmid);
 730
 731         write_unlock_irqrestore(glock, flags);
 732
 733         LIBCFS_CPT_ALLOC(conn->ibc_rxs, lnet_cpt_table(), cpt,
 734                          IBLND_RX_MSGS(version) * sizeof(kib_rx_t));
 735         if (conn->ibc_rxs == NULL) {
 736                 CERROR("Cannot allocate RX buffers\n");
 737                 goto failed_2;
 738         }
 739
 740         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, cpt,
 741                                 IBLND_RX_MSG_PAGES(version));
 742         if (rc != 0)
 743                 goto failed_2;
 744
 745         kiblnd_map_rx_descs(conn);
 746
 747         cq_attr.cqe = IBLND_CQ_ENTRIES(version);
 748         cq_attr.comp_vector = kiblnd_get_completion_vector(conn, cpt);
 749         cq = ib_create_cq(cmid->device,
 750                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 751                           &cq_attr);
 752         if (IS_ERR(cq)) {
 753                 CERROR("Can't create CQ: %ld, cqe: %d\n",
 754                        PTR_ERR(cq), IBLND_CQ_ENTRIES(version));
 755                 goto failed_2;
 756         }
 757
 758         conn->ibc_cq = cq;
 759
 760         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 761         if (rc != 0) {
 762                 CERROR("Can't request completion notificiation: %d\n", rc);
 763                 goto failed_2;
 764         }
 765
 766         init_qp_attr->event_handler = kiblnd_qp_event;
 767         init_qp_attr->qp_context = conn;
 768         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS(version);
 769         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS(version);
 770         init_qp_attr->cap.max_send_sge = 1;
 771         init_qp_attr->cap.max_recv_sge = 1;
 772         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 773         init_qp_attr->qp_type = IB_QPT_RC;
 774         init_qp_attr->send_cq = cq;
 775         init_qp_attr->recv_cq = cq;
 776
 777         conn->ibc_sched = sched;
 778
 779         rc = rdma_create_qp(cmid, conn->ibc_hdev->ibh_pd, init_qp_attr);
 780         if (rc != 0) {
 781                 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
 782                        rc, init_qp_attr->cap.max_send_wr,
 783                        init_qp_attr->cap.max_recv_wr);
 784                 goto failed_2;
 785         }
 786
 787         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 788
 789         /* 1 ref for caller and each rxmsg */
 790         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS(version));
 791         conn->ibc_nrx = IBLND_RX_MSGS(version);
 792
 793         /* post receives */
 794         for (i = 0; i < IBLND_RX_MSGS(version); i++) {
 795                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 796                                     IBLND_POSTRX_NO_CREDIT);
 797                 if (rc != 0) {
 798                         CERROR("Can't post rxmsg: %d\n", rc);
 799
 800                         /* Make posted receives complete */
 801                         kiblnd_abort_receives(conn);
 802
 803                         /* correct # of posted buffers
 804                          * NB locking needed now I'm racing with completion */
 805                         spin_lock_irqsave(&sched->ibs_lock, flags);
 806                         conn->ibc_nrx -= IBLND_RX_MSGS(version) - i;
 807                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 808
 809                         /* cmid will be destroyed by CM(ofed) after cm_callback
 810                          * returned, so we can't refer it anymore
 811                          * (by kiblnd_connd()->kiblnd_destroy_conn) */
 812                         rdma_destroy_qp(conn->ibc_cmid);
 813                         conn->ibc_cmid = NULL;
 814
 815                         /* Drop my own and unused rxbuffer refcounts */
 816                         while (i++ <= IBLND_RX_MSGS(version))
 817                                 kiblnd_conn_decref(conn);
 818
 819                         return NULL;
 820                 }
 821         }
 822
 823         /* Init successful! */
 824         LASSERT(state == IBLND_CONN_ACTIVE_CONNECT ||
 825                  state == IBLND_CONN_PASSIVE_WAIT);
 826         conn->ibc_state = state;
 827
 828         /* 1 more conn */
 829         atomic_inc(&net->ibn_nconns);
 830         return conn;
 831
 832  failed_2:
 833         kiblnd_destroy_conn(conn);
 834  failed_1:
 835         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 836  failed_0:
 837         return NULL;
 838 }
 839
 840 void kiblnd_destroy_conn(kib_conn_t *conn)
 841 {
 842         struct rdma_cm_id *cmid = conn->ibc_cmid;
 843         kib_peer_t *peer = conn->ibc_peer;
 844         int rc;
 845
 846         LASSERT(!in_interrupt());
 847         LASSERT(atomic_read(&conn->ibc_refcount) == 0);
 848         LASSERT(list_empty(&conn->ibc_early_rxs));
 849         LASSERT(list_empty(&conn->ibc_tx_noops));
 850         LASSERT(list_empty(&conn->ibc_tx_queue));
 851         LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));
 852         LASSERT(list_empty(&conn->ibc_tx_queue_nocred));
 853         LASSERT(list_empty(&conn->ibc_active_txs));
 854         LASSERT(conn->ibc_noops_posted == 0);
 855         LASSERT(conn->ibc_nsends_posted == 0);
 856
 857         switch (conn->ibc_state) {
 858         default:
 859                 /* conn must be completely disengaged from the network */
 860                 LBUG();
 861
 862         case IBLND_CONN_DISCONNECTED:
 863                 /* connvars should have been freed already */
 864                 LASSERT(conn->ibc_connvars == NULL);
 865                 break;
 866
 867         case IBLND_CONN_INIT:
 868                 break;
 869         }
 870
 871         /* conn->ibc_cmid might be destroyed by CM already */
 872         if (cmid != NULL && cmid->qp != NULL)
 873                 rdma_destroy_qp(cmid);
 874
 875         if (conn->ibc_cq != NULL) {
 876                 rc = ib_destroy_cq(conn->ibc_cq);
 877                 if (rc != 0)
 878                         CWARN("Error destroying CQ: %d\n", rc);
 879         }
 880
 881         if (conn->ibc_rx_pages != NULL)
 882                 kiblnd_unmap_rx_descs(conn);
 883
 884         if (conn->ibc_rxs != NULL) {
 885                 LIBCFS_FREE(conn->ibc_rxs,
 886                             IBLND_RX_MSGS(conn->ibc_version)
 887                               * sizeof(kib_rx_t));
 888         }
 889
 890         if (conn->ibc_connvars != NULL)
 891                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 892
 893         if (conn->ibc_hdev != NULL)
 894                 kiblnd_hdev_decref(conn->ibc_hdev);
 895
 896         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 897         if (conn->ibc_state != IBLND_CONN_INIT) {
 898                 kib_net_t *net = peer->ibp_ni->ni_data;
 899
 900                 kiblnd_peer_decref(peer);
 901                 rdma_destroy_id(cmid);
 902                 atomic_dec(&net->ibn_nconns);
 903         }
 904
 905         LIBCFS_FREE(conn, sizeof(*conn));
 906 }
 907
 908 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why)
 909 {
 910         kib_conn_t *conn;
 911         struct list_head *ctmp;
 912         struct list_head *cnxt;
 913         int count = 0;
 914
 915         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 916                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 917
 918                 CDEBUG(D_NET, "Closing conn -> %s, version: %x, reason: %d\n",
 919                        libcfs_nid2str(peer->ibp_nid),
 920                        conn->ibc_version, why);
 921
 922                 kiblnd_close_conn_locked(conn, why);
 923                 count++;
 924         }
 925
 926         return count;
 927 }
 928
 929 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
 930                                      int version, __u64 incarnation)
 931 {
 932         kib_conn_t *conn;
 933         struct list_head *ctmp;
 934         struct list_head *cnxt;
 935         int count = 0;
 936
 937         list_for_each_safe(ctmp, cnxt, &peer->ibp_conns) {
 938                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 939
 940                 if (conn->ibc_version     == version &&
 941                     conn->ibc_incarnation == incarnation)
 942                         continue;
 943
 944                 CDEBUG(D_NET,
 945                        "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
 946                        libcfs_nid2str(peer->ibp_nid),
 947                        conn->ibc_version, conn->ibc_incarnation,
 948                        version, incarnation);
 949
 950                 kiblnd_close_conn_locked(conn, -ESTALE);
 951                 count++;
 952         }
 953
 954         return count;
 955 }
 956
 957 static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 958 {
 959         kib_peer_t *peer;
 960         struct list_head *ptmp;
 961         struct list_head *pnxt;
 962         int lo;
 963         int hi;
 964         int i;
 965         unsigned long flags;
 966         int count = 0;
 967
 968         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 969
 970         if (nid != LNET_NID_ANY)
 971                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 972         else {
 973                 lo = 0;
 974                 hi = kiblnd_data.kib_peer_hash_size - 1;
 975         }
 976
 977         for (i = lo; i <= hi; i++) {
 978                 list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 979
 980                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 981                         LASSERT(peer->ibp_connecting > 0 ||
 982                                  peer->ibp_accepting > 0 ||
 983                                  !list_empty(&peer->ibp_conns));
 984
 985                         if (peer->ibp_ni != ni)
 986                                 continue;
 987
 988                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
 989                                 continue;
 990
 991                         count += kiblnd_close_peer_conns_locked(peer, 0);
 992                 }
 993         }
 994
 995         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 996
 997         /* wildcards always succeed */
 998         if (nid == LNET_NID_ANY)
 999                 return 0;
1000
1001         return (count == 0) ? -ENOENT : 0;
1002 }
1003
1004 int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1005 {
1006         struct libcfs_ioctl_data *data = arg;
1007         int rc = -EINVAL;
1008
1009         switch (cmd) {
1010         case IOC_LIBCFS_GET_PEER: {
1011                 lnet_nid_t nid = 0;
1012                 int count = 0;
1013
1014                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1015                                           &nid, &count);
1016                 data->ioc_nid   = nid;
1017                 data->ioc_count = count;
1018                 break;
1019         }
1020
1021         case IOC_LIBCFS_DEL_PEER: {
1022                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1023                 break;
1024         }
1025         case IOC_LIBCFS_GET_CONN: {
1026                 kib_conn_t *conn;
1027
1028                 rc = 0;
1029                 conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1030                 if (conn == NULL) {
1031                         rc = -ENOENT;
1032                         break;
1033                 }
1034
1035                 LASSERT(conn->ibc_cmid != NULL);
1036                 data->ioc_nid = conn->ibc_peer->ibp_nid;
1037                 if (conn->ibc_cmid->route.path_rec == NULL)
1038                         data->ioc_u32[0] = 0; /* iWarp has no path MTU */
1039                 else
1040                         data->ioc_u32[0] =
1041                         ib_mtu_enum_to_int(conn->ibc_cmid->route.path_rec->mtu);
1042                 kiblnd_conn_decref(conn);
1043                 break;
1044         }
1045         case IOC_LIBCFS_CLOSE_CONNECTION: {
1046                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1047                 break;
1048         }
1049
1050         default:
1051                 break;
1052         }
1053
1054         return rc;
1055 }
1056
1057 void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
1058 {
1059         unsigned long last_alive = 0;
1060         unsigned long now = cfs_time_current();
1061         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1062         kib_peer_t *peer;
1063         unsigned long flags;
1064
1065         read_lock_irqsave(glock, flags);
1066
1067         peer = kiblnd_find_peer_locked(nid);
1068         if (peer != NULL) {
1069                 LASSERT(peer->ibp_connecting > 0 || /* creating conns */
1070                          peer->ibp_accepting > 0 ||
1071                          !list_empty(&peer->ibp_conns));  /* active conn */
1072                 last_alive = peer->ibp_last_alive;
1073         }
1074
1075         read_unlock_irqrestore(glock, flags);
1076
1077         if (last_alive != 0)
1078                 *when = last_alive;
1079
1080         /* peer is not persistent in hash, trigger peer creation
1081          * and connection establishment with a NULL tx */
1082         if (peer == NULL)
1083                 kiblnd_launch_tx(ni, NULL, nid);
1084
1085         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago\n",
1086                libcfs_nid2str(nid), peer,
1087                last_alive ? cfs_duration_sec(now - last_alive) : -1);
1088 }
1089
1090 void kiblnd_free_pages(kib_pages_t *p)
1091 {
1092         int npages = p->ibp_npages;
1093         int i;
1094
1095         for (i = 0; i < npages; i++) {
1096                 if (p->ibp_pages[i] != NULL)
1097                         __free_page(p->ibp_pages[i]);
1098         }
1099
1100         LIBCFS_FREE(p, offsetof(kib_pages_t, ibp_pages[npages]));
1101 }
1102
1103 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages)
1104 {
1105         kib_pages_t *p;
1106         int i;
1107
1108         LIBCFS_CPT_ALLOC(p, lnet_cpt_table(), cpt,
1109                          offsetof(kib_pages_t, ibp_pages[npages]));
1110         if (p == NULL) {
1111                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1112                 return -ENOMEM;
1113         }
1114
1115         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1116         p->ibp_npages = npages;
1117
1118         for (i = 0; i < npages; i++) {
1119                 p->ibp_pages[i] = alloc_pages_node(
1120                                     cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1121                                     GFP_NOFS, 0);
1122                 if (p->ibp_pages[i] == NULL) {
1123                         CERROR("Can't allocate page %d of %d\n", i, npages);
1124                         kiblnd_free_pages(p);
1125                         return -ENOMEM;
1126                 }
1127         }
1128
1129         *pp = p;
1130         return 0;
1131 }
1132
1133 void kiblnd_unmap_rx_descs(kib_conn_t *conn)
1134 {
1135         kib_rx_t *rx;
1136         int i;
1137
1138         LASSERT(conn->ibc_rxs != NULL);
1139         LASSERT(conn->ibc_hdev != NULL);
1140
1141         for (i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1142                 rx = &conn->ibc_rxs[i];
1143
1144                 LASSERT(rx->rx_nob >= 0); /* not posted */
1145
1146                 kiblnd_dma_unmap_single(conn->ibc_hdev->ibh_ibdev,
1147                                         KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
1148                                                           rx->rx_msgaddr),
1149                                         IBLND_MSG_SIZE, DMA_FROM_DEVICE);
1150         }
1151
1152         kiblnd_free_pages(conn->ibc_rx_pages);
1153
1154         conn->ibc_rx_pages = NULL;
1155 }
1156
1157 void kiblnd_map_rx_descs(kib_conn_t *conn)
1158 {
1159         kib_rx_t *rx;
1160         struct page *pg;
1161         int pg_off;
1162         int ipg;
1163         int i;
1164
1165         for (pg_off = ipg = i = 0; i < IBLND_RX_MSGS(conn->ibc_version); i++) {
1166                 pg = conn->ibc_rx_pages->ibp_pages[ipg];
1167                 rx = &conn->ibc_rxs[i];
1168
1169                 rx->rx_conn = conn;
1170                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(pg)) + pg_off);
1171
1172                 rx->rx_msgaddr = kiblnd_dma_map_single(conn->ibc_hdev->ibh_ibdev,
1173                                                        rx->rx_msg,
1174                                                        IBLND_MSG_SIZE,
1175                                                        DMA_FROM_DEVICE);
1176                 LASSERT(!kiblnd_dma_mapping_error(conn->ibc_hdev->ibh_ibdev,
1177                                                    rx->rx_msgaddr));
1178                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
1179
1180                 CDEBUG(D_NET, "rx %d: %p %#llx(%#llx)\n",
1181                        i, rx->rx_msg, rx->rx_msgaddr,
1182                        (__u64)(page_to_phys(pg) + pg_off));
1183
1184                 pg_off += IBLND_MSG_SIZE;
1185                 LASSERT(pg_off <= PAGE_SIZE);
1186
1187                 if (pg_off == PAGE_SIZE) {
1188                         pg_off = 0;
1189                         ipg++;
1190                         LASSERT(ipg <= IBLND_RX_MSG_PAGES(conn->ibc_version));
1191                 }
1192         }
1193 }
1194
1195 static void kiblnd_unmap_tx_pool(kib_tx_pool_t *tpo)
1196 {
1197         kib_hca_dev_t *hdev = tpo->tpo_hdev;
1198         kib_tx_t *tx;
1199         int i;
1200
1201         LASSERT(tpo->tpo_pool.po_allocated == 0);
1202
1203         if (hdev == NULL)
1204                 return;
1205
1206         for (i = 0; i < tpo->tpo_pool.po_size; i++) {
1207                 tx = &tpo->tpo_tx_descs[i];
1208                 kiblnd_dma_unmap_single(hdev->ibh_ibdev,
1209                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1210                                                           tx->tx_msgaddr),
1211                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1212         }
1213
1214         kiblnd_hdev_decref(hdev);
1215         tpo->tpo_hdev = NULL;
1216 }
1217
1218 static kib_hca_dev_t *kiblnd_current_hdev(kib_dev_t *dev)
1219 {
1220         kib_hca_dev_t *hdev;
1221         unsigned long flags;
1222         int i = 0;
1223
1224         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1225         while (dev->ibd_failover) {
1226                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1227                 if (i++ % 50 == 0)
1228                         CDEBUG(D_NET, "%s: Wait for failover\n",
1229                                dev->ibd_ifname);
1230                 schedule_timeout(cfs_time_seconds(1) / 100);
1231
1232                 read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1233         }
1234
1235         kiblnd_hdev_addref_locked(dev->ibd_hdev);
1236         hdev = dev->ibd_hdev;
1237
1238         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1239
1240         return hdev;
1241 }
1242
1243 static void kiblnd_map_tx_pool(kib_tx_pool_t *tpo)
1244 {
1245         kib_pages_t *txpgs = tpo->tpo_tx_pages;
1246         kib_pool_t *pool = &tpo->tpo_pool;
1247         kib_net_t *net = pool->po_owner->ps_net;
1248         kib_dev_t *dev;
1249         struct page *page;
1250         kib_tx_t *tx;
1251         int page_offset;
1252         int ipage;
1253         int i;
1254
1255         LASSERT(net != NULL);
1256
1257         dev = net->ibn_dev;
1258
1259         /* pre-mapped messages are not bigger than 1 page */
1260         CLASSERT(IBLND_MSG_SIZE <= PAGE_SIZE);
1261
1262         /* No fancy arithmetic when we do the buffer calculations */
1263         CLASSERT(PAGE_SIZE % IBLND_MSG_SIZE == 0);
1264
1265         tpo->tpo_hdev = kiblnd_current_hdev(dev);
1266
1267         for (ipage = page_offset = i = 0; i < pool->po_size; i++) {
1268                 page = txpgs->ibp_pages[ipage];
1269                 tx = &tpo->tpo_tx_descs[i];
1270
1271                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1272                                            page_offset);
1273
1274                 tx->tx_msgaddr = kiblnd_dma_map_single(
1275                         tpo->tpo_hdev->ibh_ibdev, tx->tx_msg,
1276                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1277                 LASSERT(!kiblnd_dma_mapping_error(tpo->tpo_hdev->ibh_ibdev,
1278                                                    tx->tx_msgaddr));
1279                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1280
1281                 list_add(&tx->tx_list, &pool->po_free_list);
1282
1283                 page_offset += IBLND_MSG_SIZE;
1284                 LASSERT(page_offset <= PAGE_SIZE);
1285
1286                 if (page_offset == PAGE_SIZE) {
1287                         page_offset = 0;
1288                         ipage++;
1289                         LASSERT(ipage <= txpgs->ibp_npages);
1290                 }
1291         }
1292 }
1293
1294 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev, __u64 addr, __u64 size)
1295 {
1296         __u64 index;
1297
1298         LASSERT(hdev->ibh_mrs[0] != NULL);
1299
1300         if (hdev->ibh_nmrs == 1)
1301                 return hdev->ibh_mrs[0];
1302
1303         index = addr >> hdev->ibh_mr_shift;
1304
1305         if (index <  hdev->ibh_nmrs &&
1306             index == ((addr + size - 1) >> hdev->ibh_mr_shift))
1307                 return hdev->ibh_mrs[index];
1308
1309         return NULL;
1310 }
1311
1312 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev, kib_rdma_desc_t *rd)
1313 {
1314         struct ib_mr *prev_mr;
1315         struct ib_mr *mr;
1316         int i;
1317
1318         LASSERT(hdev->ibh_mrs[0] != NULL);
1319
1320         if (*kiblnd_tunables.kib_map_on_demand > 0 &&
1321             *kiblnd_tunables.kib_map_on_demand <= rd->rd_nfrags)
1322                 return NULL;
1323
1324         if (hdev->ibh_nmrs == 1)
1325                 return hdev->ibh_mrs[0];
1326
1327         for (i = 0, mr = prev_mr = NULL;
1328              i < rd->rd_nfrags; i++) {
1329                 mr = kiblnd_find_dma_mr(hdev,
1330                                         rd->rd_frags[i].rf_addr,
1331                                         rd->rd_frags[i].rf_nob);
1332                 if (prev_mr == NULL)
1333                         prev_mr = mr;
1334
1335                 if (mr == NULL || prev_mr != mr) {
1336                         /* Can't covered by one single MR */
1337                         mr = NULL;
1338                         break;
1339                 }
1340         }
1341
1342         return mr;
1343 }
1344
1345 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t *pool)
1346 {
1347         LASSERT(pool->fpo_map_count == 0);
1348
1349         if (pool->fpo_fmr_pool != NULL)
1350                 ib_destroy_fmr_pool(pool->fpo_fmr_pool);
1351
1352         if (pool->fpo_hdev != NULL)
1353                 kiblnd_hdev_decref(pool->fpo_hdev);
1354
1355         LIBCFS_FREE(pool, sizeof(kib_fmr_pool_t));
1356 }
1357
1358 static void kiblnd_destroy_fmr_pool_list(struct list_head *head)
1359 {
1360         kib_fmr_pool_t *pool;
1361
1362         while (!list_empty(head)) {
1363                 pool = list_entry(head->next, kib_fmr_pool_t, fpo_list);
1364                 list_del(&pool->fpo_list);
1365                 kiblnd_destroy_fmr_pool(pool);
1366         }
1367 }
1368
1369 static int kiblnd_fmr_pool_size(int ncpts)
1370 {
1371         int size = *kiblnd_tunables.kib_fmr_pool_size / ncpts;
1372
1373         return max(IBLND_FMR_POOL, size);
1374 }
1375
1376 static int kiblnd_fmr_flush_trigger(int ncpts)
1377 {
1378         int size = *kiblnd_tunables.kib_fmr_flush_trigger / ncpts;
1379
1380         return max(IBLND_FMR_POOL_FLUSH, size);
1381 }
1382
1383 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t *fps,
1384                                   kib_fmr_pool_t **pp_fpo)
1385 {
1386         /* FMR pool for RDMA */
1387         kib_dev_t *dev = fps->fps_net->ibn_dev;
1388         kib_fmr_pool_t *fpo;
1389         struct ib_fmr_pool_param param = {
1390                 .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1391                 .page_shift        = PAGE_SHIFT,
1392                 .access            = (IB_ACCESS_LOCAL_WRITE |
1393                                       IB_ACCESS_REMOTE_WRITE),
1394                 .pool_size         = fps->fps_pool_size,
1395                 .dirty_watermark   = fps->fps_flush_trigger,
1396                 .flush_function    = NULL,
1397                 .flush_arg         = NULL,
1398                 .cache             = !!*kiblnd_tunables.kib_fmr_cache};
1399         int rc;
1400
1401         LIBCFS_CPT_ALLOC(fpo, lnet_cpt_table(), fps->fps_cpt, sizeof(*fpo));
1402         if (fpo == NULL)
1403                 return -ENOMEM;
1404
1405         fpo->fpo_hdev = kiblnd_current_hdev(dev);
1406
1407         fpo->fpo_fmr_pool = ib_create_fmr_pool(fpo->fpo_hdev->ibh_pd, &param);
1408         if (IS_ERR(fpo->fpo_fmr_pool)) {
1409                 rc = PTR_ERR(fpo->fpo_fmr_pool);
1410                 CERROR("Failed to create FMR pool: %d\n", rc);
1411
1412                 kiblnd_hdev_decref(fpo->fpo_hdev);
1413                 LIBCFS_FREE(fpo, sizeof(kib_fmr_pool_t));
1414                 return rc;
1415         }
1416
1417         fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1418         fpo->fpo_owner    = fps;
1419         *pp_fpo = fpo;
1420
1421         return 0;
1422 }
1423
1424 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t *fps,
1425                                     struct list_head *zombies)
1426 {
1427         if (fps->fps_net == NULL) /* intialized? */
1428                 return;
1429
1430         spin_lock(&fps->fps_lock);
1431
1432         while (!list_empty(&fps->fps_pool_list)) {
1433                 kib_fmr_pool_t *fpo = list_entry(fps->fps_pool_list.next,
1434                                                  kib_fmr_pool_t, fpo_list);
1435                 fpo->fpo_failed = 1;
1436                 list_del(&fpo->fpo_list);
1437                 if (fpo->fpo_map_count == 0)
1438                         list_add(&fpo->fpo_list, zombies);
1439                 else
1440                         list_add(&fpo->fpo_list, &fps->fps_failed_pool_list);
1441         }
1442
1443         spin_unlock(&fps->fps_lock);
1444 }
1445
1446 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t *fps)
1447 {
1448         if (fps->fps_net != NULL) { /* initialized? */
1449                 kiblnd_destroy_fmr_pool_list(&fps->fps_failed_pool_list);
1450                 kiblnd_destroy_fmr_pool_list(&fps->fps_pool_list);
1451         }
1452 }
1453
1454 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t *fps, int cpt,
1455                                    kib_net_t *net, int pool_size,
1456                                    int flush_trigger)
1457 {
1458         kib_fmr_pool_t *fpo;
1459         int rc;
1460
1461         memset(fps, 0, sizeof(kib_fmr_poolset_t));
1462
1463         fps->fps_net = net;
1464         fps->fps_cpt = cpt;
1465         fps->fps_pool_size = pool_size;
1466         fps->fps_flush_trigger = flush_trigger;
1467         spin_lock_init(&fps->fps_lock);
1468         INIT_LIST_HEAD(&fps->fps_pool_list);
1469         INIT_LIST_HEAD(&fps->fps_failed_pool_list);
1470
1471         rc = kiblnd_create_fmr_pool(fps, &fpo);
1472         if (rc == 0)
1473                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1474
1475         return rc;
1476 }
1477
1478 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t *fpo, unsigned long now)
1479 {
1480         if (fpo->fpo_map_count != 0) /* still in use */
1481                 return 0;
1482         if (fpo->fpo_failed)
1483                 return 1;
1484         return cfs_time_aftereq(now, fpo->fpo_deadline);
1485 }
1486
1487 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
1488 {
1489         LIST_HEAD(zombies);
1490         kib_fmr_pool_t *fpo = fmr->fmr_pool;
1491         kib_fmr_poolset_t *fps = fpo->fpo_owner;
1492         unsigned long now = cfs_time_current();
1493         kib_fmr_pool_t *tmp;
1494         int rc;
1495
1496         rc = ib_fmr_pool_unmap(fmr->fmr_pfmr);
1497         LASSERT(rc == 0);
1498
1499         if (status != 0) {
1500                 rc = ib_flush_fmr_pool(fpo->fpo_fmr_pool);
1501                 LASSERT(rc == 0);
1502         }
1503
1504         fmr->fmr_pool = NULL;
1505         fmr->fmr_pfmr = NULL;
1506
1507         spin_lock(&fps->fps_lock);
1508         fpo->fpo_map_count--;  /* decref the pool */
1509
1510         list_for_each_entry_safe(fpo, tmp, &fps->fps_pool_list, fpo_list) {
1511                 /* the first pool is persistent */
1512                 if (fps->fps_pool_list.next == &fpo->fpo_list)
1513                         continue;
1514
1515                 if (kiblnd_fmr_pool_is_idle(fpo, now)) {
1516                         list_move(&fpo->fpo_list, &zombies);
1517                         fps->fps_version++;
1518                 }
1519         }
1520         spin_unlock(&fps->fps_lock);
1521
1522         if (!list_empty(&zombies))
1523                 kiblnd_destroy_fmr_pool_list(&zombies);
1524 }
1525
1526 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1527                         __u64 iov, kib_fmr_t *fmr)
1528 {
1529         struct ib_pool_fmr *pfmr;
1530         kib_fmr_pool_t *fpo;
1531         __u64 version;
1532         int rc;
1533
1534  again:
1535         spin_lock(&fps->fps_lock);
1536         version = fps->fps_version;
1537         list_for_each_entry(fpo, &fps->fps_pool_list, fpo_list) {
1538                 fpo->fpo_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1539                 fpo->fpo_map_count++;
1540                 spin_unlock(&fps->fps_lock);
1541
1542                 pfmr = ib_fmr_pool_map_phys(fpo->fpo_fmr_pool,
1543                                             pages, npages, iov);
1544                 if (likely(!IS_ERR(pfmr))) {
1545                         fmr->fmr_pool = fpo;
1546                         fmr->fmr_pfmr = pfmr;
1547                         return 0;
1548                 }
1549
1550                 spin_lock(&fps->fps_lock);
1551                 fpo->fpo_map_count--;
1552                 if (PTR_ERR(pfmr) != -EAGAIN) {
1553                         spin_unlock(&fps->fps_lock);
1554                         return PTR_ERR(pfmr);
1555                 }
1556
1557                 /* EAGAIN and ... */
1558                 if (version != fps->fps_version) {
1559                         spin_unlock(&fps->fps_lock);
1560                         goto again;
1561                 }
1562         }
1563
1564         if (fps->fps_increasing) {
1565                 spin_unlock(&fps->fps_lock);
1566                 CDEBUG(D_NET,
1567                         "Another thread is allocating new FMR pool, waiting for her to complete\n");
1568                 schedule();
1569                 goto again;
1570
1571         }
1572
1573         if (time_before(cfs_time_current(), fps->fps_next_retry)) {
1574                 /* someone failed recently */
1575                 spin_unlock(&fps->fps_lock);
1576                 return -EAGAIN;
1577         }
1578
1579         fps->fps_increasing = 1;
1580         spin_unlock(&fps->fps_lock);
1581
1582         CDEBUG(D_NET, "Allocate new FMR pool\n");
1583         rc = kiblnd_create_fmr_pool(fps, &fpo);
1584         spin_lock(&fps->fps_lock);
1585         fps->fps_increasing = 0;
1586         if (rc == 0) {
1587                 fps->fps_version++;
1588                 list_add_tail(&fpo->fpo_list, &fps->fps_pool_list);
1589         } else {
1590                 fps->fps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1591         }
1592         spin_unlock(&fps->fps_lock);
1593
1594         goto again;
1595 }
1596
1597 static void kiblnd_fini_pool(kib_pool_t *pool)
1598 {
1599         LASSERT(list_empty(&pool->po_free_list));
1600         LASSERT(pool->po_allocated == 0);
1601
1602         CDEBUG(D_NET, "Finalize %s pool\n", pool->po_owner->ps_name);
1603 }
1604
1605 static void kiblnd_init_pool(kib_poolset_t *ps, kib_pool_t *pool, int size)
1606 {
1607         CDEBUG(D_NET, "Initialize %s pool\n", ps->ps_name);
1608
1609         memset(pool, 0, sizeof(kib_pool_t));
1610         INIT_LIST_HEAD(&pool->po_free_list);
1611         pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1612         pool->po_owner    = ps;
1613         pool->po_size     = size;
1614 }
1615
1616 static void kiblnd_destroy_pool_list(struct list_head *head)
1617 {
1618         kib_pool_t *pool;
1619
1620         while (!list_empty(head)) {
1621                 pool = list_entry(head->next, kib_pool_t, po_list);
1622                 list_del(&pool->po_list);
1623
1624                 LASSERT(pool->po_owner != NULL);
1625                 pool->po_owner->ps_pool_destroy(pool);
1626         }
1627 }
1628
1629 static void kiblnd_fail_poolset(kib_poolset_t *ps, struct list_head *zombies)
1630 {
1631         if (ps->ps_net == NULL) /* intialized? */
1632                 return;
1633
1634         spin_lock(&ps->ps_lock);
1635         while (!list_empty(&ps->ps_pool_list)) {
1636                 kib_pool_t *po = list_entry(ps->ps_pool_list.next,
1637                                             kib_pool_t, po_list);
1638                 po->po_failed = 1;
1639                 list_del(&po->po_list);
1640                 if (po->po_allocated == 0)
1641                         list_add(&po->po_list, zombies);
1642                 else
1643                         list_add(&po->po_list, &ps->ps_failed_pool_list);
1644         }
1645         spin_unlock(&ps->ps_lock);
1646 }
1647
1648 static void kiblnd_fini_poolset(kib_poolset_t *ps)
1649 {
1650         if (ps->ps_net != NULL) { /* initialized? */
1651                 kiblnd_destroy_pool_list(&ps->ps_failed_pool_list);
1652                 kiblnd_destroy_pool_list(&ps->ps_pool_list);
1653         }
1654 }
1655
1656 static int kiblnd_init_poolset(kib_poolset_t *ps, int cpt,
1657                                kib_net_t *net, char *name, int size,
1658                                kib_ps_pool_create_t po_create,
1659                                kib_ps_pool_destroy_t po_destroy,
1660                                kib_ps_node_init_t nd_init,
1661                                kib_ps_node_fini_t nd_fini)
1662 {
1663         kib_pool_t *pool;
1664         int rc;
1665
1666         memset(ps, 0, sizeof(kib_poolset_t));
1667
1668         ps->ps_cpt          = cpt;
1669         ps->ps_net          = net;
1670         ps->ps_pool_create  = po_create;
1671         ps->ps_pool_destroy = po_destroy;
1672         ps->ps_node_init    = nd_init;
1673         ps->ps_node_fini    = nd_fini;
1674         ps->ps_pool_size    = size;
1675         if (strlcpy(ps->ps_name, name, sizeof(ps->ps_name))
1676             >= sizeof(ps->ps_name))
1677                 return -E2BIG;
1678         spin_lock_init(&ps->ps_lock);
1679         INIT_LIST_HEAD(&ps->ps_pool_list);
1680         INIT_LIST_HEAD(&ps->ps_failed_pool_list);
1681
1682         rc = ps->ps_pool_create(ps, size, &pool);
1683         if (rc == 0)
1684                 list_add(&pool->po_list, &ps->ps_pool_list);
1685         else
1686                 CERROR("Failed to create the first pool for %s\n", ps->ps_name);
1687
1688         return rc;
1689 }
1690
1691 static int kiblnd_pool_is_idle(kib_pool_t *pool, unsigned long now)
1692 {
1693         if (pool->po_allocated != 0) /* still in use */
1694                 return 0;
1695         if (pool->po_failed)
1696                 return 1;
1697         return cfs_time_aftereq(now, pool->po_deadline);
1698 }
1699
1700 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node)
1701 {
1702         LIST_HEAD(zombies);
1703         kib_poolset_t *ps = pool->po_owner;
1704         kib_pool_t *tmp;
1705         unsigned long now = cfs_time_current();
1706
1707         spin_lock(&ps->ps_lock);
1708
1709         if (ps->ps_node_fini != NULL)
1710                 ps->ps_node_fini(pool, node);
1711
1712         LASSERT(pool->po_allocated > 0);
1713         list_add(node, &pool->po_free_list);
1714         pool->po_allocated--;
1715
1716         list_for_each_entry_safe(pool, tmp, &ps->ps_pool_list, po_list) {
1717                 /* the first pool is persistent */
1718                 if (ps->ps_pool_list.next == &pool->po_list)
1719                         continue;
1720
1721                 if (kiblnd_pool_is_idle(pool, now))
1722                         list_move(&pool->po_list, &zombies);
1723         }
1724         spin_unlock(&ps->ps_lock);
1725
1726         if (!list_empty(&zombies))
1727                 kiblnd_destroy_pool_list(&zombies);
1728 }
1729
1730 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps)
1731 {
1732         struct list_head *node;
1733         kib_pool_t *pool;
1734         int rc;
1735
1736  again:
1737         spin_lock(&ps->ps_lock);
1738         list_for_each_entry(pool, &ps->ps_pool_list, po_list) {
1739                 if (list_empty(&pool->po_free_list))
1740                         continue;
1741
1742                 pool->po_allocated++;
1743                 pool->po_deadline = cfs_time_shift(IBLND_POOL_DEADLINE);
1744                 node = pool->po_free_list.next;
1745                 list_del(node);
1746
1747                 if (ps->ps_node_init != NULL) {
1748                         /* still hold the lock */
1749                         ps->ps_node_init(pool, node);
1750                 }
1751                 spin_unlock(&ps->ps_lock);
1752                 return node;
1753         }
1754
1755         /* no available tx pool and ... */
1756         if (ps->ps_increasing) {
1757                 /* another thread is allocating a new pool */
1758                 spin_unlock(&ps->ps_lock);
1759                 CDEBUG(D_NET, "Another thread is allocating new %s pool, waiting for her to complete\n",
1760                        ps->ps_name);
1761                 schedule();
1762                 goto again;
1763         }
1764
1765         if (time_before(cfs_time_current(), ps->ps_next_retry)) {
1766                 /* someone failed recently */
1767                 spin_unlock(&ps->ps_lock);
1768                 return NULL;
1769         }
1770
1771         ps->ps_increasing = 1;
1772         spin_unlock(&ps->ps_lock);
1773
1774         CDEBUG(D_NET, "%s pool exhausted, allocate new pool\n", ps->ps_name);
1775
1776         rc = ps->ps_pool_create(ps, ps->ps_pool_size, &pool);
1777
1778         spin_lock(&ps->ps_lock);
1779         ps->ps_increasing = 0;
1780         if (rc == 0) {
1781                 list_add_tail(&pool->po_list, &ps->ps_pool_list);
1782         } else {
1783                 ps->ps_next_retry = cfs_time_shift(IBLND_POOL_RETRY);
1784                 CERROR("Can't allocate new %s pool because out of memory\n",
1785                        ps->ps_name);
1786         }
1787         spin_unlock(&ps->ps_lock);
1788
1789         goto again;
1790 }
1791
1792 static void kiblnd_destroy_tx_pool(kib_pool_t *pool)
1793 {
1794         kib_tx_pool_t *tpo = container_of(pool, kib_tx_pool_t, tpo_pool);
1795         int i;
1796
1797         LASSERT(pool->po_allocated == 0);
1798
1799         if (tpo->tpo_tx_pages != NULL) {
1800                 kiblnd_unmap_tx_pool(tpo);
1801                 kiblnd_free_pages(tpo->tpo_tx_pages);
1802         }
1803
1804         if (tpo->tpo_tx_descs == NULL)
1805                 goto out;
1806
1807         for (i = 0; i < pool->po_size; i++) {
1808                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1809
1810                 list_del(&tx->tx_list);
1811                 if (tx->tx_pages != NULL)
1812                         LIBCFS_FREE(tx->tx_pages,
1813                                     LNET_MAX_IOV *
1814                                     sizeof(*tx->tx_pages));
1815                 if (tx->tx_frags != NULL)
1816                         LIBCFS_FREE(tx->tx_frags,
1817                                     IBLND_MAX_RDMA_FRAGS *
1818                                             sizeof(*tx->tx_frags));
1819                 if (tx->tx_wrq != NULL)
1820                         LIBCFS_FREE(tx->tx_wrq,
1821                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1822                                     sizeof(*tx->tx_wrq));
1823                 if (tx->tx_sge != NULL)
1824                         LIBCFS_FREE(tx->tx_sge,
1825                                     (1 + IBLND_MAX_RDMA_FRAGS) *
1826                                     sizeof(*tx->tx_sge));
1827                 if (tx->tx_rd != NULL)
1828                         LIBCFS_FREE(tx->tx_rd,
1829                                     offsetof(kib_rdma_desc_t,
1830                                              rd_frags[IBLND_MAX_RDMA_FRAGS]));
1831         }
1832
1833         LIBCFS_FREE(tpo->tpo_tx_descs,
1834                     pool->po_size * sizeof(kib_tx_t));
1835 out:
1836         kiblnd_fini_pool(pool);
1837         LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1838 }
1839
1840 static int kiblnd_tx_pool_size(int ncpts)
1841 {
1842         int ntx = *kiblnd_tunables.kib_ntx / ncpts;
1843
1844         return max(IBLND_TX_POOL, ntx);
1845 }
1846
1847 static int kiblnd_create_tx_pool(kib_poolset_t *ps, int size,
1848                                  kib_pool_t **pp_po)
1849 {
1850         int i;
1851         int npg;
1852         kib_pool_t *pool;
1853         kib_tx_pool_t *tpo;
1854
1855         LIBCFS_CPT_ALLOC(tpo, lnet_cpt_table(), ps->ps_cpt, sizeof(*tpo));
1856         if (tpo == NULL) {
1857                 CERROR("Failed to allocate TX pool\n");
1858                 return -ENOMEM;
1859         }
1860
1861         pool = &tpo->tpo_pool;
1862         kiblnd_init_pool(ps, pool, size);
1863         tpo->tpo_tx_descs = NULL;
1864         tpo->tpo_tx_pages = NULL;
1865
1866         npg = (size * IBLND_MSG_SIZE + PAGE_SIZE - 1) / PAGE_SIZE;
1867         if (kiblnd_alloc_pages(&tpo->tpo_tx_pages, ps->ps_cpt, npg) != 0) {
1868                 CERROR("Can't allocate tx pages: %d\n", npg);
1869                 LIBCFS_FREE(tpo, sizeof(kib_tx_pool_t));
1870                 return -ENOMEM;
1871         }
1872
1873         LIBCFS_CPT_ALLOC(tpo->tpo_tx_descs, lnet_cpt_table(), ps->ps_cpt,
1874                          size * sizeof(kib_tx_t));
1875         if (tpo->tpo_tx_descs == NULL) {
1876                 CERROR("Can't allocate %d tx descriptors\n", size);
1877                 ps->ps_pool_destroy(pool);
1878                 return -ENOMEM;
1879         }
1880
1881         memset(tpo->tpo_tx_descs, 0, size * sizeof(kib_tx_t));
1882
1883         for (i = 0; i < size; i++) {
1884                 kib_tx_t *tx = &tpo->tpo_tx_descs[i];
1885
1886                 tx->tx_pool = tpo;
1887                 if (ps->ps_net->ibn_fmr_ps != NULL) {
1888                         LIBCFS_CPT_ALLOC(tx->tx_pages,
1889                                          lnet_cpt_table(), ps->ps_cpt,
1890                                          LNET_MAX_IOV * sizeof(*tx->tx_pages));
1891                         if (tx->tx_pages == NULL)
1892                                 break;
1893                 }
1894
1895                 LIBCFS_CPT_ALLOC(tx->tx_frags, lnet_cpt_table(), ps->ps_cpt,
1896                                  IBLND_MAX_RDMA_FRAGS * sizeof(*tx->tx_frags));
1897                 if (tx->tx_frags == NULL)
1898                         break;
1899
1900                 sg_init_table(tx->tx_frags, IBLND_MAX_RDMA_FRAGS);
1901
1902                 LIBCFS_CPT_ALLOC(tx->tx_wrq, lnet_cpt_table(), ps->ps_cpt,
1903                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1904                                  sizeof(*tx->tx_wrq));
1905                 if (tx->tx_wrq == NULL)
1906                         break;
1907
1908                 LIBCFS_CPT_ALLOC(tx->tx_sge, lnet_cpt_table(), ps->ps_cpt,
1909                                  (1 + IBLND_MAX_RDMA_FRAGS) *
1910                                  sizeof(*tx->tx_sge));
1911                 if (tx->tx_sge == NULL)
1912                         break;
1913
1914                 LIBCFS_CPT_ALLOC(tx->tx_rd, lnet_cpt_table(), ps->ps_cpt,
1915                                  offsetof(kib_rdma_desc_t,
1916                                           rd_frags[IBLND_MAX_RDMA_FRAGS]));
1917                 if (tx->tx_rd == NULL)
1918                         break;
1919         }
1920
1921         if (i == size) {
1922                 kiblnd_map_tx_pool(tpo);
1923                 *pp_po = pool;
1924                 return 0;
1925         }
1926
1927         ps->ps_pool_destroy(pool);
1928         return -ENOMEM;
1929 }
1930
1931 static void kiblnd_tx_init(kib_pool_t *pool, struct list_head *node)
1932 {
1933         kib_tx_poolset_t *tps = container_of(pool->po_owner, kib_tx_poolset_t,
1934                                              tps_poolset);
1935         kib_tx_t *tx  = list_entry(node, kib_tx_t, tx_list);
1936
1937         tx->tx_cookie = tps->tps_next_tx_cookie++;
1938 }
1939
1940 static void kiblnd_net_fini_pools(kib_net_t *net)
1941 {
1942         int i;
1943
1944         cfs_cpt_for_each(i, lnet_cpt_table()) {
1945                 kib_tx_poolset_t *tps;
1946                 kib_fmr_poolset_t *fps;
1947
1948                 if (net->ibn_tx_ps != NULL) {
1949                         tps = net->ibn_tx_ps[i];
1950                         kiblnd_fini_poolset(&tps->tps_poolset);
1951                 }
1952
1953                 if (net->ibn_fmr_ps != NULL) {
1954                         fps = net->ibn_fmr_ps[i];
1955                         kiblnd_fini_fmr_poolset(fps);
1956                 }
1957         }
1958
1959         if (net->ibn_tx_ps != NULL) {
1960                 cfs_percpt_free(net->ibn_tx_ps);
1961                 net->ibn_tx_ps = NULL;
1962         }
1963
1964         if (net->ibn_fmr_ps != NULL) {
1965                 cfs_percpt_free(net->ibn_fmr_ps);
1966                 net->ibn_fmr_ps = NULL;
1967         }
1968 }
1969
1970 static int kiblnd_net_init_pools(kib_net_t *net, __u32 *cpts, int ncpts)
1971 {
1972         unsigned long flags;
1973         int cpt;
1974         int             rc = 0;
1975         int i;
1976
1977         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1978         if (*kiblnd_tunables.kib_map_on_demand == 0 &&
1979             net->ibn_dev->ibd_hdev->ibh_nmrs == 1) {
1980                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1981                 goto create_tx_pool;
1982         }
1983
1984         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1985
1986         if (*kiblnd_tunables.kib_fmr_pool_size <
1987             *kiblnd_tunables.kib_ntx / 4) {
1988                 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
1989                        *kiblnd_tunables.kib_fmr_pool_size,
1990                        *kiblnd_tunables.kib_ntx / 4);
1991                 rc = -EINVAL;
1992                 goto failed;
1993         }
1994
1995         /*
1996          * TX pool must be created later than FMR, see LU-2268
1997          * for details
1998          */
1999         LASSERT(net->ibn_tx_ps == NULL);
2000
2001         /*
2002          * premapping can fail if ibd_nmr > 1, so we always create
2003          * FMR pool and map-on-demand if premapping failed
2004          */
2005
2006         net->ibn_fmr_ps = cfs_percpt_alloc(lnet_cpt_table(),
2007                                            sizeof(kib_fmr_poolset_t));
2008         if (net->ibn_fmr_ps == NULL) {
2009                 CERROR("Failed to allocate FMR pool array\n");
2010                 rc = -ENOMEM;
2011                 goto failed;
2012         }
2013
2014         for (i = 0; i < ncpts; i++) {
2015                 cpt = (cpts == NULL) ? i : cpts[i];
2016                 rc = kiblnd_init_fmr_poolset(net->ibn_fmr_ps[cpt], cpt, net,
2017                                              kiblnd_fmr_pool_size(ncpts),
2018                                              kiblnd_fmr_flush_trigger(ncpts));
2019                 if (rc == -ENOSYS && i == 0) /* no FMR */
2020                         break;
2021
2022                 if (rc != 0) { /* a real error */
2023                         CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2024                                cpt, rc);
2025                         goto failed;
2026                 }
2027         }
2028
2029         if (i > 0) {
2030                 LASSERT(i == ncpts);
2031                 goto create_tx_pool;
2032         }
2033
2034         cfs_percpt_free(net->ibn_fmr_ps);
2035         net->ibn_fmr_ps = NULL;
2036
2037         CWARN("Device does not support FMR\n");
2038                         goto failed;
2039
2040  create_tx_pool:
2041         net->ibn_tx_ps = cfs_percpt_alloc(lnet_cpt_table(),
2042                                           sizeof(kib_tx_poolset_t));
2043         if (net->ibn_tx_ps == NULL) {
2044                 CERROR("Failed to allocate tx pool array\n");
2045                 rc = -ENOMEM;
2046                 goto failed;
2047         }
2048
2049         for (i = 0; i < ncpts; i++) {
2050                 cpt = (cpts == NULL) ? i : cpts[i];
2051                 rc = kiblnd_init_poolset(&net->ibn_tx_ps[cpt]->tps_poolset,
2052                                          cpt, net, "TX",
2053                                          kiblnd_tx_pool_size(ncpts),
2054                                          kiblnd_create_tx_pool,
2055                                          kiblnd_destroy_tx_pool,
2056                                          kiblnd_tx_init, NULL);
2057                 if (rc != 0) {
2058                         CERROR("Can't initialize TX pool for CPT %d: %d\n",
2059                                cpt, rc);
2060                         goto failed;
2061                 }
2062         }
2063
2064         return 0;
2065  failed:
2066         kiblnd_net_fini_pools(net);
2067         LASSERT(rc != 0);
2068         return rc;
2069 }
2070
2071 static int kiblnd_hdev_get_attr(kib_hca_dev_t *hdev)
2072 {
2073         struct ib_device_attr *attr;
2074         int rc;
2075
2076         /* It's safe to assume a HCA can handle a page size
2077          * matching that of the native system */
2078         hdev->ibh_page_shift = PAGE_SHIFT;
2079         hdev->ibh_page_size  = 1 << PAGE_SHIFT;
2080         hdev->ibh_page_mask  = ~((__u64)hdev->ibh_page_size - 1);
2081
2082         LIBCFS_ALLOC(attr, sizeof(*attr));
2083         if (attr == NULL) {
2084                 CERROR("Out of memory\n");
2085                 return -ENOMEM;
2086         }
2087
2088         rc = ib_query_device(hdev->ibh_ibdev, attr);
2089         if (rc == 0)
2090                 hdev->ibh_mr_size = attr->max_mr_size;
2091
2092         LIBCFS_FREE(attr, sizeof(*attr));
2093
2094         if (rc != 0) {
2095                 CERROR("Failed to query IB device: %d\n", rc);
2096                 return rc;
2097         }
2098
2099         if (hdev->ibh_mr_size == ~0ULL) {
2100                 hdev->ibh_mr_shift = 64;
2101                 return 0;
2102         }
2103
2104         for (hdev->ibh_mr_shift = 0;
2105              hdev->ibh_mr_shift < 64; hdev->ibh_mr_shift++) {
2106                 if (hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) ||
2107                     hdev->ibh_mr_size == (1ULL << hdev->ibh_mr_shift) - 1)
2108                         return 0;
2109         }
2110
2111         CERROR("Invalid mr size: %#llx\n", hdev->ibh_mr_size);
2112         return -EINVAL;
2113 }
2114
2115 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t *hdev)
2116 {
2117         int i;
2118
2119         if (hdev->ibh_nmrs == 0 || hdev->ibh_mrs == NULL)
2120                 return;
2121
2122         for (i = 0; i < hdev->ibh_nmrs; i++) {
2123                 if (hdev->ibh_mrs[i] == NULL)
2124                         break;
2125
2126                 ib_dereg_mr(hdev->ibh_mrs[i]);
2127         }
2128
2129         LIBCFS_FREE(hdev->ibh_mrs, sizeof(*hdev->ibh_mrs) * hdev->ibh_nmrs);
2130         hdev->ibh_mrs  = NULL;
2131         hdev->ibh_nmrs = 0;
2132 }
2133
2134 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev)
2135 {
2136         kiblnd_hdev_cleanup_mrs(hdev);
2137
2138         if (hdev->ibh_pd != NULL)
2139                 ib_dealloc_pd(hdev->ibh_pd);
2140
2141         if (hdev->ibh_cmid != NULL)
2142                 rdma_destroy_id(hdev->ibh_cmid);
2143
2144         LIBCFS_FREE(hdev, sizeof(*hdev));
2145 }
2146
2147 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t *hdev)
2148 {
2149         struct ib_mr *mr;
2150         int rc;
2151         int acflags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
2152
2153         rc = kiblnd_hdev_get_attr(hdev);
2154         if (rc != 0)
2155                 return rc;
2156
2157         LIBCFS_ALLOC(hdev->ibh_mrs, 1 * sizeof(*hdev->ibh_mrs));
2158         if (hdev->ibh_mrs == NULL) {
2159                 CERROR("Failed to allocate MRs table\n");
2160                 return -ENOMEM;
2161         }
2162
2163         hdev->ibh_mrs[0] = NULL;
2164         hdev->ibh_nmrs   = 1;
2165
2166         mr = ib_get_dma_mr(hdev->ibh_pd, acflags);
2167         if (IS_ERR(mr)) {
2168                 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr));
2169                 kiblnd_hdev_cleanup_mrs(hdev);
2170                 return PTR_ERR(mr);
2171         }
2172
2173         hdev->ibh_mrs[0] = mr;
2174
2175         return 0;
2176 }
2177
2178 /* DUMMY */
2179 static int kiblnd_dummy_callback(struct rdma_cm_id *cmid,
2180                                  struct rdma_cm_event *event)
2181 {
2182         return 0;
2183 }
2184
2185 static int kiblnd_dev_need_failover(kib_dev_t *dev)
2186 {
2187         struct rdma_cm_id *cmid;
2188         struct sockaddr_in srcaddr;
2189         struct sockaddr_in dstaddr;
2190         int rc;
2191
2192         if (dev->ibd_hdev == NULL || /* initializing */
2193             dev->ibd_hdev->ibh_cmid == NULL || /* listener is dead */
2194             *kiblnd_tunables.kib_dev_failover > 1) /* debugging */
2195                 return 1;
2196
2197         /* XXX: it's UGLY, but I don't have better way to find
2198          * ib-bonding HCA failover because:
2199          *
2200          * a. no reliable CM event for HCA failover...
2201          * b. no OFED API to get ib_device for current net_device...
2202          *
2203          * We have only two choices at this point:
2204          *
2205          * a. rdma_bind_addr(), it will conflict with listener cmid
2206          * b. rdma_resolve_addr() to zero addr */
2207         cmid = kiblnd_rdma_create_id(kiblnd_dummy_callback, dev, RDMA_PS_TCP,
2208                                      IB_QPT_RC);
2209         if (IS_ERR(cmid)) {
2210                 rc = PTR_ERR(cmid);
2211                 CERROR("Failed to create cmid for failover: %d\n", rc);
2212                 return rc;
2213         }
2214
2215         memset(&srcaddr, 0, sizeof(srcaddr));
2216         srcaddr.sin_family = AF_INET;
2217         srcaddr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2218
2219         memset(&dstaddr, 0, sizeof(dstaddr));
2220         dstaddr.sin_family = AF_INET;
2221         rc = rdma_resolve_addr(cmid, (struct sockaddr *)&srcaddr,
2222                                (struct sockaddr *)&dstaddr, 1);
2223         if (rc != 0 || cmid->device == NULL) {
2224                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2225                        dev->ibd_ifname, &dev->ibd_ifip,
2226                        cmid->device, rc);
2227                 rdma_destroy_id(cmid);
2228                 return rc;
2229         }
2230
2231         rc = dev->ibd_hdev->ibh_ibdev != cmid->device; /* true for failover */
2232         rdma_destroy_id(cmid);
2233
2234         return rc;
2235 }
2236
2237 int kiblnd_dev_failover(kib_dev_t *dev)
2238 {
2239         LIST_HEAD(zombie_tpo);
2240         LIST_HEAD(zombie_ppo);
2241         LIST_HEAD(zombie_fpo);
2242         struct rdma_cm_id *cmid  = NULL;
2243         kib_hca_dev_t *hdev  = NULL;
2244         struct ib_pd *pd;
2245         kib_net_t *net;
2246         struct sockaddr_in addr;
2247         unsigned long flags;
2248         int rc = 0;
2249         int i;
2250
2251         LASSERT(*kiblnd_tunables.kib_dev_failover > 1 ||
2252                  dev->ibd_can_failover ||
2253                  dev->ibd_hdev == NULL);
2254
2255         rc = kiblnd_dev_need_failover(dev);
2256         if (rc <= 0)
2257                 goto out;
2258
2259         if (dev->ibd_hdev != NULL &&
2260             dev->ibd_hdev->ibh_cmid != NULL) {
2261                 /* XXX it's not good to close old listener at here,
2262                  * because we can fail to create new listener.
2263                  * But we have to close it now, otherwise rdma_bind_addr
2264                  * will return EADDRINUSE... How crap! */
2265                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2266
2267                 cmid = dev->ibd_hdev->ibh_cmid;
2268                 /* make next schedule of kiblnd_dev_need_failover()
2269                  * return 1 for me */
2270                 dev->ibd_hdev->ibh_cmid  = NULL;
2271                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2272
2273                 rdma_destroy_id(cmid);
2274         }
2275
2276         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, dev, RDMA_PS_TCP,
2277                                      IB_QPT_RC);
2278         if (IS_ERR(cmid)) {
2279                 rc = PTR_ERR(cmid);
2280                 CERROR("Failed to create cmid for failover: %d\n", rc);
2281                 goto out;
2282         }
2283
2284         memset(&addr, 0, sizeof(addr));
2285         addr.sin_family      = AF_INET;
2286         addr.sin_addr.s_addr = (__force u32)htonl(dev->ibd_ifip);
2287         addr.sin_port   = htons(*kiblnd_tunables.kib_service);
2288
2289         /* Bind to failover device or port */
2290         rc = rdma_bind_addr(cmid, (struct sockaddr *)&addr);
2291         if (rc != 0 || cmid->device == NULL) {
2292                 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2293                        dev->ibd_ifname, &dev->ibd_ifip,
2294                        cmid->device, rc);
2295                 rdma_destroy_id(cmid);
2296                 goto out;
2297         }
2298
2299         LIBCFS_ALLOC(hdev, sizeof(*hdev));
2300         if (hdev == NULL) {
2301                 CERROR("Failed to allocate kib_hca_dev\n");
2302                 rdma_destroy_id(cmid);
2303                 rc = -ENOMEM;
2304                 goto out;
2305         }
2306
2307         atomic_set(&hdev->ibh_ref, 1);
2308         hdev->ibh_dev   = dev;
2309         hdev->ibh_cmid  = cmid;
2310         hdev->ibh_ibdev = cmid->device;
2311
2312         pd = ib_alloc_pd(cmid->device);
2313         if (IS_ERR(pd)) {
2314                 rc = PTR_ERR(pd);
2315                 CERROR("Can't allocate PD: %d\n", rc);
2316                 goto out;
2317         }
2318
2319         hdev->ibh_pd = pd;
2320
2321         rc = rdma_listen(cmid, 0);
2322         if (rc != 0) {
2323                 CERROR("Can't start new listener: %d\n", rc);
2324                 goto out;
2325         }
2326
2327         rc = kiblnd_hdev_setup_mrs(hdev);
2328         if (rc != 0) {
2329                 CERROR("Can't setup device: %d\n", rc);
2330                 goto out;
2331         }
2332
2333         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2334
2335         swap(dev->ibd_hdev, hdev); /* take over the refcount */
2336
2337         list_for_each_entry(net, &dev->ibd_nets, ibn_list) {
2338                 cfs_cpt_for_each(i, lnet_cpt_table()) {
2339                         kiblnd_fail_poolset(&net->ibn_tx_ps[i]->tps_poolset,
2340                                             &zombie_tpo);
2341
2342                         if (net->ibn_fmr_ps)
2343                                 kiblnd_fail_fmr_poolset(net->ibn_fmr_ps[i],
2344                                                         &zombie_fpo);
2345                 }
2346         }
2347
2348         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2349  out:
2350         if (!list_empty(&zombie_tpo))
2351                 kiblnd_destroy_pool_list(&zombie_tpo);
2352         if (!list_empty(&zombie_ppo))
2353                 kiblnd_destroy_pool_list(&zombie_ppo);
2354         if (!list_empty(&zombie_fpo))
2355                 kiblnd_destroy_fmr_pool_list(&zombie_fpo);
2356         if (hdev != NULL)
2357                 kiblnd_hdev_decref(hdev);
2358
2359         if (rc != 0)
2360                 dev->ibd_failed_failover++;
2361         else
2362                 dev->ibd_failed_failover = 0;
2363
2364         return rc;
2365 }
2366
2367 void kiblnd_destroy_dev(kib_dev_t *dev)
2368 {
2369         LASSERT(dev->ibd_nnets == 0);
2370         LASSERT(list_empty(&dev->ibd_nets));
2371
2372         list_del(&dev->ibd_fail_list);
2373         list_del(&dev->ibd_list);
2374
2375         if (dev->ibd_hdev != NULL)
2376                 kiblnd_hdev_decref(dev->ibd_hdev);
2377
2378         LIBCFS_FREE(dev, sizeof(*dev));
2379 }
2380
2381 static kib_dev_t *kiblnd_create_dev(char *ifname)
2382 {
2383         struct net_device *netdev;
2384         kib_dev_t *dev;
2385         __u32 netmask;
2386         __u32 ip;
2387         int up;
2388         int rc;
2389
2390         rc = lnet_ipif_query(ifname, &up, &ip, &netmask);
2391         if (rc != 0) {
2392                 CERROR("Can't query IPoIB interface %s: %d\n",
2393                        ifname, rc);
2394                 return NULL;
2395         }
2396
2397         if (!up) {
2398                 CERROR("Can't query IPoIB interface %s: it's down\n", ifname);
2399                 return NULL;
2400         }
2401
2402         LIBCFS_ALLOC(dev, sizeof(*dev));
2403         if (dev == NULL)
2404                 return NULL;
2405
2406         netdev = dev_get_by_name(&init_net, ifname);
2407         if (netdev == NULL) {
2408                 dev->ibd_can_failover = 0;
2409         } else {
2410                 dev->ibd_can_failover = !!(netdev->flags & IFF_MASTER);
2411                 dev_put(netdev);
2412         }
2413
2414         INIT_LIST_HEAD(&dev->ibd_nets);
2415         INIT_LIST_HEAD(&dev->ibd_list); /* not yet in kib_devs */
2416         INIT_LIST_HEAD(&dev->ibd_fail_list);
2417         dev->ibd_ifip = ip;
2418         strcpy(&dev->ibd_ifname[0], ifname);
2419
2420         /* initialize the device */
2421         rc = kiblnd_dev_failover(dev);
2422         if (rc != 0) {
2423                 CERROR("Can't initialize device: %d\n", rc);
2424                 LIBCFS_FREE(dev, sizeof(*dev));
2425                 return NULL;
2426         }
2427
2428         list_add_tail(&dev->ibd_list,
2429                           &kiblnd_data.kib_devs);
2430         return dev;
2431 }
2432
2433 static void kiblnd_base_shutdown(void)
2434 {
2435         struct kib_sched_info *sched;
2436         int i;
2437
2438         LASSERT(list_empty(&kiblnd_data.kib_devs));
2439
2440         switch (kiblnd_data.kib_init) {
2441         default:
2442                 LBUG();
2443
2444         case IBLND_INIT_ALL:
2445         case IBLND_INIT_DATA:
2446                 LASSERT(kiblnd_data.kib_peers != NULL);
2447                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2448                         LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
2449                 LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
2450                 LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
2451
2452                 /* flag threads to terminate; wake and wait for them to die */
2453                 kiblnd_data.kib_shutdown = 1;
2454
2455                 /* NB: we really want to stop scheduler threads net by net
2456                  * instead of the whole module, this should be improved
2457                  * with dynamic configuration LNet */
2458                 cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds)
2459                         wake_up_all(&sched->ibs_waitq);
2460
2461                 wake_up_all(&kiblnd_data.kib_connd_waitq);
2462                 wake_up_all(&kiblnd_data.kib_failover_waitq);
2463
2464                 i = 2;
2465                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
2466                         i++;
2467                         /* power of 2 ? */
2468                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2469                                "Waiting for %d threads to terminate\n",
2470                                atomic_read(&kiblnd_data.kib_nthreads));
2471                         set_current_state(TASK_UNINTERRUPTIBLE);
2472                         schedule_timeout(cfs_time_seconds(1));
2473                 }
2474
2475                 /* fall through */
2476
2477         case IBLND_INIT_NOTHING:
2478                 break;
2479         }
2480
2481         if (kiblnd_data.kib_peers != NULL) {
2482                 LIBCFS_FREE(kiblnd_data.kib_peers,
2483                             sizeof(struct list_head) *
2484                             kiblnd_data.kib_peer_hash_size);
2485         }
2486
2487         if (kiblnd_data.kib_scheds != NULL)
2488                 cfs_percpt_free(kiblnd_data.kib_scheds);
2489
2490         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
2491         module_put(THIS_MODULE);
2492 }
2493
2494 void kiblnd_shutdown(lnet_ni_t *ni)
2495 {
2496         kib_net_t *net = ni->ni_data;
2497         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2498         int i;
2499         unsigned long flags;
2500
2501         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
2502
2503         if (net == NULL)
2504                 goto out;
2505
2506         write_lock_irqsave(g_lock, flags);
2507         net->ibn_shutdown = 1;
2508         write_unlock_irqrestore(g_lock, flags);
2509
2510         switch (net->ibn_init) {
2511         default:
2512                 LBUG();
2513
2514         case IBLND_INIT_ALL:
2515                 /* nuke all existing peers within this net */
2516                 kiblnd_del_peer(ni, LNET_NID_ANY);
2517
2518                 /* Wait for all peer state to clean up */
2519                 i = 2;
2520                 while (atomic_read(&net->ibn_npeers) != 0) {
2521                         i++;
2522                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
2523                                "%s: waiting for %d peers to disconnect\n",
2524                                libcfs_nid2str(ni->ni_nid),
2525                                atomic_read(&net->ibn_npeers));
2526                         set_current_state(TASK_UNINTERRUPTIBLE);
2527                         schedule_timeout(cfs_time_seconds(1));
2528                 }
2529
2530                 kiblnd_net_fini_pools(net);
2531
2532                 write_lock_irqsave(g_lock, flags);
2533                 LASSERT(net->ibn_dev->ibd_nnets > 0);
2534                 net->ibn_dev->ibd_nnets--;
2535                 list_del(&net->ibn_list);
2536                 write_unlock_irqrestore(g_lock, flags);
2537
2538                 /* fall through */
2539
2540         case IBLND_INIT_NOTHING:
2541                 LASSERT(atomic_read(&net->ibn_nconns) == 0);
2542
2543                 if (net->ibn_dev != NULL &&
2544                     net->ibn_dev->ibd_nnets == 0)
2545                         kiblnd_destroy_dev(net->ibn_dev);
2546
2547                 break;
2548         }
2549
2550         net->ibn_init = IBLND_INIT_NOTHING;
2551         ni->ni_data = NULL;
2552
2553         LIBCFS_FREE(net, sizeof(*net));
2554
2555 out:
2556         if (list_empty(&kiblnd_data.kib_devs))
2557                 kiblnd_base_shutdown();
2558 }
2559
2560 static int kiblnd_base_startup(void)
2561 {
2562         struct kib_sched_info *sched;
2563         int rc;
2564         int i;
2565
2566         LASSERT(kiblnd_data.kib_init == IBLND_INIT_NOTHING);
2567
2568         try_module_get(THIS_MODULE);
2569         /* zero pointers, flags etc */
2570         memset(&kiblnd_data, 0, sizeof(kiblnd_data));
2571
2572         rwlock_init(&kiblnd_data.kib_global_lock);
2573
2574         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
2575         INIT_LIST_HEAD(&kiblnd_data.kib_failed_devs);
2576
2577         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
2578         LIBCFS_ALLOC(kiblnd_data.kib_peers,
2579                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
2580         if (kiblnd_data.kib_peers == NULL)
2581                 goto failed;
2582         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
2583                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
2584
2585         spin_lock_init(&kiblnd_data.kib_connd_lock);
2586         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
2587         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
2588         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
2589         init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
2590
2591         kiblnd_data.kib_scheds = cfs_percpt_alloc(lnet_cpt_table(),
2592                                                   sizeof(*sched));
2593         if (kiblnd_data.kib_scheds == NULL)
2594                 goto failed;
2595
2596         cfs_percpt_for_each(sched, i, kiblnd_data.kib_scheds) {
2597                 int nthrs;
2598
2599                 spin_lock_init(&sched->ibs_lock);
2600                 INIT_LIST_HEAD(&sched->ibs_conns);
2601                 init_waitqueue_head(&sched->ibs_waitq);
2602
2603                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2604                 if (*kiblnd_tunables.kib_nscheds > 0) {
2605                         nthrs = min(nthrs, *kiblnd_tunables.kib_nscheds);
2606                 } else {
2607                         /* max to half of CPUs, another half is reserved for
2608                          * upper layer modules */
2609                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2610                 }
2611
2612                 sched->ibs_nthreads_max = nthrs;
2613                 sched->ibs_cpt = i;
2614         }
2615
2616         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
2617
2618         /* lists/ptrs/locks initialised */
2619         kiblnd_data.kib_init = IBLND_INIT_DATA;
2620         /*****************************************************/
2621
2622         rc = kiblnd_thread_start(kiblnd_connd, NULL, "kiblnd_connd");
2623         if (rc != 0) {
2624                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
2625                 goto failed;
2626         }
2627
2628         if (*kiblnd_tunables.kib_dev_failover != 0)
2629                 rc = kiblnd_thread_start(kiblnd_failover_thread, NULL,
2630                                          "kiblnd_failover");
2631
2632         if (rc != 0) {
2633                 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc);
2634                 goto failed;
2635         }
2636
2637         /* flag everything initialised */
2638         kiblnd_data.kib_init = IBLND_INIT_ALL;
2639         /*****************************************************/
2640
2641         return 0;
2642
2643  failed:
2644         kiblnd_base_shutdown();
2645         return -ENETDOWN;
2646 }
2647
2648 static int kiblnd_start_schedulers(struct kib_sched_info *sched)
2649 {
2650         int rc = 0;
2651         int nthrs;
2652         int i;
2653
2654         if (sched->ibs_nthreads == 0) {
2655                 if (*kiblnd_tunables.kib_nscheds > 0) {
2656                         nthrs = sched->ibs_nthreads_max;
2657                 } else {
2658                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2659                                                sched->ibs_cpt);
2660                         nthrs = min(max(IBLND_N_SCHED, nthrs >> 1), nthrs);
2661                         nthrs = min(IBLND_N_SCHED_HIGH, nthrs);
2662                 }
2663         } else {
2664                 LASSERT(sched->ibs_nthreads <= sched->ibs_nthreads_max);
2665                 /* increase one thread if there is new interface */
2666                 nthrs = sched->ibs_nthreads < sched->ibs_nthreads_max;
2667         }
2668
2669         for (i = 0; i < nthrs; i++) {
2670                 long id;
2671                 char name[20];
2672
2673                 id = KIB_THREAD_ID(sched->ibs_cpt, sched->ibs_nthreads + i);
2674                 snprintf(name, sizeof(name), "kiblnd_sd_%02ld_%02ld",
2675                          KIB_THREAD_CPT(id), KIB_THREAD_TID(id));
2676                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)id, name);
2677                 if (rc == 0)
2678                         continue;
2679
2680                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2681                        sched->ibs_cpt, sched->ibs_nthreads + i, rc);
2682                 break;
2683         }
2684
2685         sched->ibs_nthreads += i;
2686         return rc;
2687 }
2688
2689 static int kiblnd_dev_start_threads(kib_dev_t *dev, int newdev, __u32 *cpts,
2690                                     int ncpts)
2691 {
2692         int cpt;
2693         int rc;
2694         int i;
2695
2696         for (i = 0; i < ncpts; i++) {
2697                 struct kib_sched_info *sched;
2698
2699                 cpt = (cpts == NULL) ? i : cpts[i];
2700                 sched = kiblnd_data.kib_scheds[cpt];
2701
2702                 if (!newdev && sched->ibs_nthreads > 0)
2703                         continue;
2704
2705                 rc = kiblnd_start_schedulers(kiblnd_data.kib_scheds[cpt]);
2706                 if (rc != 0) {
2707                         CERROR("Failed to start scheduler threads for %s\n",
2708                                dev->ibd_ifname);
2709                         return rc;
2710                 }
2711         }
2712         return 0;
2713 }
2714
2715 static kib_dev_t *kiblnd_dev_search(char *ifname)
2716 {
2717         kib_dev_t *alias = NULL;
2718         kib_dev_t *dev;
2719         char *colon;
2720         char *colon2;
2721
2722         colon = strchr(ifname, ':');
2723         list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
2724                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2725                         return dev;
2726
2727                 if (alias != NULL)
2728                         continue;
2729
2730                 colon2 = strchr(dev->ibd_ifname, ':');
2731                 if (colon != NULL)
2732                         *colon = 0;
2733                 if (colon2 != NULL)
2734                         *colon2 = 0;
2735
2736                 if (strcmp(&dev->ibd_ifname[0], ifname) == 0)
2737                         alias = dev;
2738
2739                 if (colon != NULL)
2740                         *colon = ':';
2741                 if (colon2 != NULL)
2742                         *colon2 = ':';
2743         }
2744         return alias;
2745 }
2746
2747 int kiblnd_startup(lnet_ni_t *ni)
2748 {
2749         char *ifname;
2750         kib_dev_t *ibdev = NULL;
2751         kib_net_t *net;
2752         struct timespec64 tv;
2753         unsigned long flags;
2754         int rc;
2755         int newdev;
2756
2757         LASSERT(ni->ni_lnd == &the_o2iblnd);
2758
2759         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
2760                 rc = kiblnd_base_startup();
2761                 if (rc != 0)
2762                         return rc;
2763         }
2764
2765         LIBCFS_ALLOC(net, sizeof(*net));
2766         ni->ni_data = net;
2767         if (net == NULL)
2768                 goto net_failed;
2769
2770         ktime_get_real_ts64(&tv);
2771         net->ibn_incarnation = tv.tv_sec * USEC_PER_SEC +
2772                                tv.tv_nsec / NSEC_PER_USEC;
2773
2774         ni->ni_peertimeout    = *kiblnd_tunables.kib_peertimeout;
2775         ni->ni_maxtxcredits   = *kiblnd_tunables.kib_credits;
2776         ni->ni_peertxcredits  = *kiblnd_tunables.kib_peertxcredits;
2777         ni->ni_peerrtrcredits = *kiblnd_tunables.kib_peerrtrcredits;
2778
2779         if (ni->ni_interfaces[0] != NULL) {
2780                 /* Use the IPoIB interface specified in 'networks=' */
2781
2782                 CLASSERT(LNET_MAX_INTERFACES > 1);
2783                 if (ni->ni_interfaces[1] != NULL) {
2784                         CERROR("Multiple interfaces not supported\n");
2785                         goto failed;
2786                 }
2787
2788                 ifname = ni->ni_interfaces[0];
2789         } else {
2790                 ifname = *kiblnd_tunables.kib_default_ipif;
2791         }
2792
2793         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
2794                 CERROR("IPoIB interface name too long: %s\n", ifname);
2795                 goto failed;
2796         }
2797
2798         ibdev = kiblnd_dev_search(ifname);
2799
2800         newdev = ibdev == NULL;
2801         /* hmm...create kib_dev even for alias */
2802         if (ibdev == NULL || strcmp(&ibdev->ibd_ifname[0], ifname) != 0)
2803                 ibdev = kiblnd_create_dev(ifname);
2804
2805         if (ibdev == NULL)
2806                 goto failed;
2807
2808         net->ibn_dev = ibdev;
2809         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
2810
2811         rc = kiblnd_dev_start_threads(ibdev, newdev,
2812                                       ni->ni_cpts, ni->ni_ncpts);
2813         if (rc != 0)
2814                 goto failed;
2815
2816         rc = kiblnd_net_init_pools(net, ni->ni_cpts, ni->ni_ncpts);
2817         if (rc != 0) {
2818                 CERROR("Failed to initialize NI pools: %d\n", rc);
2819                 goto failed;
2820         }
2821
2822         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2823         ibdev->ibd_nnets++;
2824         list_add_tail(&net->ibn_list, &ibdev->ibd_nets);
2825         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2826
2827         net->ibn_init = IBLND_INIT_ALL;
2828
2829         return 0;
2830
2831 failed:
2832         if (net->ibn_dev == NULL && ibdev != NULL)
2833                 kiblnd_destroy_dev(ibdev);
2834
2835 net_failed:
2836         kiblnd_shutdown(ni);
2837
2838         CDEBUG(D_NET, "kiblnd_startup failed\n");
2839         return -ENETDOWN;
2840 }
2841
2842 static void __exit kiblnd_module_fini(void)
2843 {
2844         lnet_unregister_lnd(&the_o2iblnd);
2845 }
2846
2847 static int __init kiblnd_module_init(void)
2848 {
2849         int rc;
2850
2851         CLASSERT(sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
2852         CLASSERT(offsetof(kib_msg_t,
2853                 ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2854                 <= IBLND_MSG_SIZE);
2855         CLASSERT(offsetof(kib_msg_t,
2856                 ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
2857                 <= IBLND_MSG_SIZE);
2858
2859         rc = kiblnd_tunables_init();
2860         if (rc != 0)
2861                 return rc;
2862
2863         lnet_register_lnd(&the_o2iblnd);
2864
2865         return 0;
2866 }
2867
2868 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2869 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2870 MODULE_LICENSE("GPL");
2871
2872 module_init(kiblnd_module_init);
2873 module_exit(kiblnd_module_fini);