lnet/klnds/o2iblnd/o2iblnd.c

   1 /*
   2  * -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   3  * vim:expandtab:shiftwidth=8:tabstop=8:
   4  *
   5  * GPL HEADER START
   6  *
   7  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 only,
  11  * as published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License version 2 for more details (a copy is included
  17  * in the LICENSE file that accompanied this code).
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * version 2 along with this program; If not, see [sun.com URL with a
  21  * copy of GPLv2].
  22  *
  23  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  24  * CA 95054 USA or visit www.sun.com if you need additional information or
  25  * have any questions.
  26  *
  27  * GPL HEADER END
  28  */
  29 /*
  30  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  31  * Use is subject to license terms.
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * lnet/klnds/o2iblnd/o2iblnd.c
  38  *
  39  * Author: Eric Barton <eric@bartonsoftware.com>
  40  */
  41
  42 #include "o2iblnd.h"
  43
  44 lnd_t the_kiblnd = {
  45         .lnd_type       = O2IBLND,
  46         .lnd_startup    = kiblnd_startup,
  47         .lnd_shutdown   = kiblnd_shutdown,
  48         .lnd_ctl        = kiblnd_ctl,
  49         .lnd_send       = kiblnd_send,
  50         .lnd_recv       = kiblnd_recv,
  51 };
  52
  53 kib_data_t              kiblnd_data;
  54
  55 __u32
  56 kiblnd_cksum (void *ptr, int nob)
  57 {
  58         char  *c  = ptr;
  59         __u32  sum = 0;
  60
  61         while (nob-- > 0)
  62                 sum = ((sum << 1) | (sum >> 31)) + *c++;
  63
  64         /* ensure I don't return 0 (== no checksum) */
  65         return (sum == 0) ? 1 : sum;
  66 }
  67
  68 void
  69 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
  70 {
  71         msg->ibm_type = type;
  72         msg->ibm_nob  = offsetof(kib_msg_t, ibm_u) + body_nob;
  73 }
  74
  75 void
  76 kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg,
  77                  int credits, lnet_nid_t dstnid, __u64 dststamp)
  78 {
  79         kib_net_t *net = ni->ni_data;
  80
  81         /* CAVEAT EMPTOR! all message fields not set here should have been
  82          * initialised previously. */
  83         msg->ibm_magic    = IBLND_MSG_MAGIC;
  84         msg->ibm_version  = IBLND_MSG_VERSION;
  85         /*   ibm_type */
  86         msg->ibm_credits  = credits;
  87         /*   ibm_nob */
  88         msg->ibm_cksum    = 0;
  89         msg->ibm_srcnid   = ni->ni_nid;
  90         msg->ibm_srcstamp = net->ibn_incarnation;
  91         msg->ibm_dstnid   = dstnid;
  92         msg->ibm_dststamp = dststamp;
  93
  94         if (*kiblnd_tunables.kib_cksum) {
  95                 /* NB ibm_cksum zero while computing cksum */
  96                 msg->ibm_cksum = kiblnd_cksum(msg, msg->ibm_nob);
  97         }
  98 }
  99
 100 int
 101 kiblnd_unpack_msg(kib_msg_t *msg, int nob)
 102 {
 103         const int hdr_size = offsetof(kib_msg_t, ibm_u);
 104         __u32     msg_cksum;
 105         int       flip;
 106         int       msg_nob;
 107 #if !IBLND_MAP_ON_DEMAND
 108         int       i;
 109         int       n;
 110 #endif
 111         /* 6 bytes are enough to have received magic + version */
 112         if (nob < 6) {
 113                 CERROR("Short message: %d\n", nob);
 114                 return -EPROTO;
 115         }
 116
 117         if (msg->ibm_magic == IBLND_MSG_MAGIC) {
 118                 flip = 0;
 119         } else if (msg->ibm_magic == __swab32(IBLND_MSG_MAGIC)) {
 120                 flip = 1;
 121         } else {
 122                 CERROR("Bad magic: %08x\n", msg->ibm_magic);
 123                 return -EPROTO;
 124         }
 125
 126         if (msg->ibm_version !=
 127             (flip ? __swab16(IBLND_MSG_VERSION) : IBLND_MSG_VERSION)) {
 128                 CERROR("Bad version: %d\n", msg->ibm_version);
 129                 return -EPROTO;
 130         }
 131
 132         if (nob < hdr_size) {
 133                 CERROR("Short message: %d\n", nob);
 134                 return -EPROTO;
 135         }
 136
 137         msg_nob = flip ? __swab32(msg->ibm_nob) : msg->ibm_nob;
 138         if (msg_nob > nob) {
 139                 CERROR("Short message: got %d, wanted %d\n", nob, msg_nob);
 140                 return -EPROTO;
 141         }
 142
 143         /* checksum must be computed with ibm_cksum zero and BEFORE anything
 144          * gets flipped */
 145         msg_cksum = flip ? __swab32(msg->ibm_cksum) : msg->ibm_cksum;
 146         msg->ibm_cksum = 0;
 147         if (msg_cksum != 0 &&
 148             msg_cksum != kiblnd_cksum(msg, msg_nob)) {
 149                 CERROR("Bad checksum\n");
 150                 return -EPROTO;
 151         }
 152         msg->ibm_cksum = msg_cksum;
 153
 154         if (flip) {
 155                 /* leave magic unflipped as a clue to peer endianness */
 156                 __swab16s(&msg->ibm_version);
 157                 CLASSERT (sizeof(msg->ibm_type) == 1);
 158                 CLASSERT (sizeof(msg->ibm_credits) == 1);
 159                 msg->ibm_nob = msg_nob;
 160                 __swab64s(&msg->ibm_srcnid);
 161                 __swab64s(&msg->ibm_srcstamp);
 162                 __swab64s(&msg->ibm_dstnid);
 163                 __swab64s(&msg->ibm_dststamp);
 164         }
 165
 166         if (msg->ibm_srcnid == LNET_NID_ANY) {
 167                 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg->ibm_srcnid));
 168                 return -EPROTO;
 169         }
 170
 171         switch (msg->ibm_type) {
 172         default:
 173                 CERROR("Unknown message type %x\n", msg->ibm_type);
 174                 return -EPROTO;
 175
 176         case IBLND_MSG_NOOP:
 177                 break;
 178
 179         case IBLND_MSG_IMMEDIATE:
 180                 if (msg_nob < offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0])) {
 181                         CERROR("Short IMMEDIATE: %d(%d)\n", msg_nob,
 182                                (int)offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[0]));
 183                         return -EPROTO;
 184                 }
 185                 break;
 186
 187         case IBLND_MSG_PUT_REQ:
 188                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.putreq)) {
 189                         CERROR("Short PUT_REQ: %d(%d)\n", msg_nob,
 190                                (int)(hdr_size + sizeof(msg->ibm_u.putreq)));
 191                         return -EPROTO;
 192                 }
 193                 break;
 194
 195         case IBLND_MSG_PUT_ACK:
 196                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.putack)) {
 197                         CERROR("Short PUT_ACK: %d(%d)\n", msg_nob,
 198                                (int)(hdr_size + sizeof(msg->ibm_u.putack)));
 199                         return -EPROTO;
 200                 }
 201 #if IBLND_MAP_ON_DEMAND
 202                 if (flip) {
 203                         __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_addr);
 204                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nob);
 205                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key);
 206                 }
 207 #else
 208                 if (flip) {
 209                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_key);
 210                         __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_nfrags);
 211                 }
 212
 213                 n = msg->ibm_u.putack.ibpam_rd.rd_nfrags;
 214                 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 215                         CERROR("Bad PUT_ACK nfrags: %d, should be 0 < n <= %d\n",
 216                                n, IBLND_MAX_RDMA_FRAGS);
 217                         return -EPROTO;
 218                 }
 219
 220                 if (msg_nob < offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n])) {
 221                         CERROR("Short PUT_ACK: %d(%d)\n", msg_nob,
 222                                (int)offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[n]));
 223                         return -EPROTO;
 224                 }
 225
 226                 if (flip) {
 227                         for (i = 0; i < n; i++) {
 228                                 __swab32s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_nob);
 229                                 __swab64s(&msg->ibm_u.putack.ibpam_rd.rd_frags[i].rf_addr);
 230                         }
 231                 }
 232 #endif
 233                 break;
 234
 235         case IBLND_MSG_GET_REQ:
 236                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.get)) {
 237                         CERROR("Short GET_REQ: %d(%d)\n", msg_nob,
 238                                (int)(hdr_size + sizeof(msg->ibm_u.get)));
 239                         return -EPROTO;
 240                 }
 241 #if IBLND_MAP_ON_DEMAND
 242                 if (flip) {
 243                         __swab64s(&msg->ibm_u.get.ibgm_rd.rd_addr);
 244                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nob);
 245                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key);
 246                 }
 247 #else
 248                 if (flip) {
 249                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_key);
 250                         __swab32s(&msg->ibm_u.get.ibgm_rd.rd_nfrags);
 251                 }
 252
 253                 n = msg->ibm_u.get.ibgm_rd.rd_nfrags;
 254                 if (n <= 0 || n > IBLND_MAX_RDMA_FRAGS) {
 255                         CERROR("Bad GET_REQ nfrags: %d, should be 0 < n <= %d\n",
 256                                n, IBLND_MAX_RDMA_FRAGS);
 257                         return -EPROTO;
 258                 }
 259
 260                 if (msg_nob < offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n])) {
 261                         CERROR("Short GET_REQ: %d(%d)\n", msg_nob,
 262                                (int)offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[n]));
 263                         return -EPROTO;
 264                 }
 265
 266                 if (flip)
 267                         for (i = 0; i < msg->ibm_u.get.ibgm_rd.rd_nfrags; i++) {
 268                                 __swab32s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_nob);
 269                                 __swab64s(&msg->ibm_u.get.ibgm_rd.rd_frags[i].rf_addr);
 270                         }
 271 #endif
 272                 break;
 273
 274         case IBLND_MSG_PUT_NAK:
 275         case IBLND_MSG_PUT_DONE:
 276         case IBLND_MSG_GET_DONE:
 277                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.completion)) {
 278                         CERROR("Short RDMA completion: %d(%d)\n", msg_nob,
 279                                (int)(hdr_size + sizeof(msg->ibm_u.completion)));
 280                         return -EPROTO;
 281                 }
 282                 if (flip)
 283                         __swab32s(&msg->ibm_u.completion.ibcm_status);
 284                 break;
 285
 286         case IBLND_MSG_CONNREQ:
 287         case IBLND_MSG_CONNACK:
 288                 if (msg_nob < hdr_size + sizeof(msg->ibm_u.connparams)) {
 289                         CERROR("Short connreq/ack: %d(%d)\n", msg_nob,
 290                                (int)(hdr_size + sizeof(msg->ibm_u.connparams)));
 291                         return -EPROTO;
 292                 }
 293                 if (flip) {
 294                         __swab16s(&msg->ibm_u.connparams.ibcp_queue_depth);
 295                         __swab16s(&msg->ibm_u.connparams.ibcp_max_frags);
 296                         __swab32s(&msg->ibm_u.connparams.ibcp_max_msg_size);
 297                 }
 298                 break;
 299         }
 300         return 0;
 301 }
 302
 303 int
 304 kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
 305 {
 306         kib_peer_t     *peer;
 307         kib_net_t      *net = ni->ni_data;
 308         unsigned long   flags;
 309
 310         LASSERT (net != NULL);
 311         LASSERT (nid != LNET_NID_ANY);
 312
 313         LIBCFS_ALLOC(peer, sizeof(*peer));
 314         if (peer == NULL) {
 315                 CERROR("Cannot allocate peer\n");
 316                 return -ENOMEM;
 317         }
 318
 319         memset(peer, 0, sizeof(*peer));         /* zero flags etc */
 320
 321         peer->ibp_ni = ni;
 322         peer->ibp_nid = nid;
 323         peer->ibp_error = 0;
 324         peer->ibp_last_alive = cfs_time_current();
 325         atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 326
 327         INIT_LIST_HEAD(&peer->ibp_list);       /* not in the peer table yet */
 328         INIT_LIST_HEAD(&peer->ibp_conns);
 329         INIT_LIST_HEAD(&peer->ibp_tx_queue);
 330
 331         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 332
 333         /* always called with a ref on ni, which prevents ni being shutdown */
 334         LASSERT (net->ibn_shutdown == 0);
 335
 336         /* npeers only grows with the global lock held */
 337         atomic_inc(&net->ibn_npeers);
 338
 339         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 340
 341         *peerp = peer;
 342         return 0;
 343 }
 344
 345 void
 346 kiblnd_destroy_peer (kib_peer_t *peer)
 347 {
 348         kib_net_t *net = peer->ibp_ni->ni_data;
 349
 350         LASSERT (net != NULL);
 351         LASSERT (atomic_read(&peer->ibp_refcount) == 0);
 352         LASSERT (!kiblnd_peer_active(peer));
 353         LASSERT (peer->ibp_connecting == 0);
 354         LASSERT (peer->ibp_accepting == 0);
 355         LASSERT (list_empty(&peer->ibp_conns));
 356         LASSERT (list_empty(&peer->ibp_tx_queue));
 357
 358         LIBCFS_FREE(peer, sizeof(*peer));
 359
 360         /* NB a peer's connections keep a reference on their peer until
 361          * they are destroyed, so we can be assured that _all_ state to do
 362          * with this peer has been cleaned up when its refcount drops to
 363          * zero. */
 364         atomic_dec(&net->ibn_npeers);
 365 }
 366
 367 void
 368 kiblnd_destroy_dev (kib_dev_t *dev)
 369 {
 370         LASSERT (dev->ibd_nnets == 0);
 371
 372         if (!list_empty(&dev->ibd_list)) /* on kib_devs? */
 373                 list_del_init(&dev->ibd_list);
 374
 375         if (dev->ibd_mr != NULL)
 376                 ib_dereg_mr(dev->ibd_mr);
 377
 378         if (dev->ibd_pd != NULL)
 379                 ib_dealloc_pd(dev->ibd_pd);
 380
 381         if (dev->ibd_cmid != NULL)
 382                 rdma_destroy_id(dev->ibd_cmid);
 383
 384         LIBCFS_FREE(dev, sizeof(*dev));
 385 }
 386
 387 kib_peer_t *
 388 kiblnd_find_peer_locked (lnet_nid_t nid)
 389 {
 390         /* the caller is responsible for accounting the additional reference
 391          * that this creates */
 392         struct list_head *peer_list = kiblnd_nid2peerlist(nid);
 393         struct list_head *tmp;
 394         kib_peer_t       *peer;
 395
 396         list_for_each (tmp, peer_list) {
 397
 398                 peer = list_entry(tmp, kib_peer_t, ibp_list);
 399
 400                 LASSERT (peer->ibp_connecting > 0 || /* creating conns */
 401                          peer->ibp_accepting > 0 ||
 402                          !list_empty(&peer->ibp_conns));  /* active conn */
 403
 404                 if (peer->ibp_nid != nid)
 405                         continue;
 406
 407                 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
 408                        peer, libcfs_nid2str(nid),
 409                        atomic_read(&peer->ibp_refcount));
 410                 return peer;
 411         }
 412         return NULL;
 413 }
 414
 415 void
 416 kiblnd_unlink_peer_locked (kib_peer_t *peer)
 417 {
 418         LASSERT (list_empty(&peer->ibp_conns));
 419
 420         LASSERT (kiblnd_peer_active(peer));
 421         list_del_init(&peer->ibp_list);
 422         /* lose peerlist's ref */
 423         kiblnd_peer_decref(peer);
 424 }
 425
 426 int
 427 kiblnd_get_peer_info (lnet_ni_t *ni, int index,
 428                       lnet_nid_t *nidp, int *count)
 429 {
 430         kib_peer_t        *peer;
 431         struct list_head  *ptmp;
 432         int                i;
 433         unsigned long      flags;
 434
 435         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 436
 437         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 438
 439                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 440
 441                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 442                         LASSERT (peer->ibp_connecting > 0 ||
 443                                  peer->ibp_accepting > 0 ||
 444                                  !list_empty(&peer->ibp_conns));
 445
 446                         if (peer->ibp_ni != ni)
 447                                 continue;
 448
 449                         if (index-- > 0)
 450                                 continue;
 451
 452                         *nidp = peer->ibp_nid;
 453                         *count = atomic_read(&peer->ibp_refcount);
 454
 455                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 456                                                flags);
 457                         return 0;
 458                 }
 459         }
 460
 461         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 462         return -ENOENT;
 463 }
 464
 465 void
 466 kiblnd_del_peer_locked (kib_peer_t *peer)
 467 {
 468         struct list_head *ctmp;
 469         struct list_head *cnxt;
 470         kib_conn_t       *conn;
 471
 472         if (list_empty(&peer->ibp_conns)) {
 473                 kiblnd_unlink_peer_locked(peer);
 474         } else {
 475                 list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 476                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
 477
 478                         kiblnd_close_conn_locked(conn, 0);
 479                 }
 480                 /* NB closing peer's last conn unlinked it. */
 481         }
 482         /* NB peer now unlinked; might even be freed if the peer table had the
 483          * last ref on it. */
 484 }
 485
 486 int
 487 kiblnd_del_peer (lnet_ni_t *ni, lnet_nid_t nid)
 488 {
 489         CFS_LIST_HEAD     (zombies);
 490         struct list_head  *ptmp;
 491         struct list_head  *pnxt;
 492         kib_peer_t        *peer;
 493         int                lo;
 494         int                hi;
 495         int                i;
 496         unsigned long      flags;
 497         int                rc = -ENOENT;
 498
 499         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 500
 501         if (nid != LNET_NID_ANY) {
 502                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
 503         } else {
 504                 lo = 0;
 505                 hi = kiblnd_data.kib_peer_hash_size - 1;
 506         }
 507
 508         for (i = lo; i <= hi; i++) {
 509                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 510                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 511                         LASSERT (peer->ibp_connecting > 0 ||
 512                                  peer->ibp_accepting > 0 ||
 513                                  !list_empty(&peer->ibp_conns));
 514
 515                         if (peer->ibp_ni != ni)
 516                                 continue;
 517
 518                         if (!(nid == LNET_NID_ANY || peer->ibp_nid == nid))
 519                                 continue;
 520
 521                         if (!list_empty(&peer->ibp_tx_queue)) {
 522                                 LASSERT (list_empty(&peer->ibp_conns));
 523
 524                                 list_splice_init(&peer->ibp_tx_queue, &zombies);
 525                         }
 526
 527                         kiblnd_del_peer_locked(peer);
 528                         rc = 0;         /* matched something */
 529                 }
 530         }
 531
 532         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 533
 534         kiblnd_txlist_done(ni, &zombies, -EIO);
 535
 536         return rc;
 537 }
 538
 539 kib_conn_t *
 540 kiblnd_get_conn_by_idx (lnet_ni_t *ni, int index)
 541 {
 542         kib_peer_t        *peer;
 543         struct list_head  *ptmp;
 544         kib_conn_t        *conn;
 545         struct list_head  *ctmp;
 546         int                i;
 547         unsigned long      flags;
 548
 549         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 550
 551         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 552                 list_for_each (ptmp, &kiblnd_data.kib_peers[i]) {
 553
 554                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
 555                         LASSERT (peer->ibp_connecting > 0 ||
 556                                  peer->ibp_accepting > 0 ||
 557                                  !list_empty(&peer->ibp_conns));
 558
 559                         if (peer->ibp_ni != ni)
 560                                 continue;
 561
 562                         list_for_each (ctmp, &peer->ibp_conns) {
 563                                 if (index-- > 0)
 564                                         continue;
 565
 566                                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 567                                 kiblnd_conn_addref(conn);
 568                                 read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
 569                                                        flags);
 570                                 return conn;
 571                         }
 572                 }
 573         }
 574
 575         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 576         return NULL;
 577 }
 578
 579 void
 580 kiblnd_debug_rx (kib_rx_t *rx)
 581 {
 582         CDEBUG(D_CONSOLE, "      %p status %d msg_type %x cred %d\n",
 583                rx, rx->rx_status, rx->rx_msg->ibm_type,
 584                rx->rx_msg->ibm_credits);
 585 }
 586
 587 void
 588 kiblnd_debug_tx (kib_tx_t *tx)
 589 {
 590         CDEBUG(D_CONSOLE, "      %p snd %d q %d w %d rc %d dl %lx "
 591                "cookie "LPX64" msg %s%s type %x cred %d\n",
 592                tx, tx->tx_sending, tx->tx_queued, tx->tx_waiting,
 593                tx->tx_status, tx->tx_deadline, tx->tx_cookie,
 594                tx->tx_lntmsg[0] == NULL ? "-" : "!",
 595                tx->tx_lntmsg[1] == NULL ? "-" : "!",
 596                tx->tx_msg->ibm_type, tx->tx_msg->ibm_credits);
 597 }
 598
 599 void
 600 kiblnd_debug_conn (kib_conn_t *conn)
 601 {
 602         struct list_head *tmp;
 603         int               i;
 604
 605         spin_lock(&conn->ibc_lock);
 606
 607         CDEBUG(D_CONSOLE, "conn[%d] %p -> %s: \n",
 608                atomic_read(&conn->ibc_refcount), conn,
 609                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 610         CDEBUG(D_CONSOLE, "   state %d nposted %d cred %d o_cred %d r_cred %d\n",
 611                conn->ibc_state, conn->ibc_nsends_posted, conn->ibc_credits,
 612                conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
 613         CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 614
 615         CDEBUG(D_CONSOLE, "   early_rxs:\n");
 616         list_for_each(tmp, &conn->ibc_early_rxs)
 617                 kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
 618
 619         CDEBUG(D_CONSOLE, "   tx_noops:\n");
 620         list_for_each(tmp, &conn->ibc_tx_noops)
 621                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 622
 623         CDEBUG(D_CONSOLE, "   tx_queue_nocred:\n");
 624         list_for_each(tmp, &conn->ibc_tx_queue_nocred)
 625                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 626
 627         CDEBUG(D_CONSOLE, "   tx_queue_rsrvd:\n");
 628         list_for_each(tmp, &conn->ibc_tx_queue_rsrvd)
 629                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 630
 631         CDEBUG(D_CONSOLE, "   tx_queue:\n");
 632         list_for_each(tmp, &conn->ibc_tx_queue)
 633                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 634
 635         CDEBUG(D_CONSOLE, "   active_txs:\n");
 636         list_for_each(tmp, &conn->ibc_active_txs)
 637                 kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
 638
 639         CDEBUG(D_CONSOLE, "   rxs:\n");
 640         for (i = 0; i < IBLND_RX_MSGS; i++)
 641                 kiblnd_debug_rx(&conn->ibc_rxs[i]);
 642
 643         spin_unlock(&conn->ibc_lock);
 644 }
 645
 646 kib_conn_t *
 647 kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid, int state)
 648 {
 649         /* CAVEAT EMPTOR:
 650          * If the new conn is created successfully it takes over the caller's
 651          * ref on 'peer'.  It also "owns" 'cmid' and destroys it when it itself
 652          * is destroyed.  On failure, the caller's ref on 'peer' remains and
 653          * she must dispose of 'cmid'.  (Actually I'd block forever if I tried
 654          * to destroy 'cmid' here since I'm called from the CM which still has
 655          * its ref on 'cmid'). */
 656         kib_conn_t             *conn;
 657         kib_net_t              *net = peer->ibp_ni->ni_data;
 658         int                     i;
 659         int                     page_offset;
 660         int                     ipage;
 661         int                     rc;
 662         struct ib_cq           *cq;
 663         struct ib_qp_init_attr *init_qp_attr;
 664         unsigned long           flags;
 665
 666         LASSERT (net != NULL);
 667         LASSERT (!in_interrupt());
 668
 669         LIBCFS_ALLOC(init_qp_attr, sizeof(*init_qp_attr));
 670         if (init_qp_attr == NULL) {
 671                 CERROR("Can't allocate qp_attr for %s\n",
 672                        libcfs_nid2str(peer->ibp_nid));
 673                 goto failed_0;
 674         }
 675
 676         LIBCFS_ALLOC(conn, sizeof(*conn));
 677         if (conn == NULL) {
 678                 CERROR("Can't allocate connection for %s\n",
 679                        libcfs_nid2str(peer->ibp_nid));
 680                 goto failed_1;
 681         }
 682
 683         memset(conn, 0, sizeof(*conn)); /* zero flags, NULL pointers etc... */
 684
 685         conn->ibc_state = IBLND_CONN_INIT;
 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_ALLOC(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 699         if (conn->ibc_connvars == NULL) {
 700                 CERROR("Can't allocate in-progress connection state\n");
 701                 goto failed_2;
 702         }
 703         memset(conn->ibc_connvars, 0, sizeof(*conn->ibc_connvars));
 704
 705         LIBCFS_ALLOC(conn->ibc_rxs, IBLND_RX_MSGS * sizeof(kib_rx_t));
 706         if (conn->ibc_rxs == NULL) {
 707                 CERROR("Cannot allocate RX buffers\n");
 708                 goto failed_2;
 709         }
 710         memset(conn->ibc_rxs, 0, IBLND_RX_MSGS * sizeof(kib_rx_t));
 711
 712         rc = kiblnd_alloc_pages(&conn->ibc_rx_pages, IBLND_RX_MSG_PAGES);
 713         if (rc != 0)
 714                 goto failed_2;
 715
 716         for (i = ipage = page_offset = 0; i < IBLND_RX_MSGS; i++) {
 717                 struct page *page = conn->ibc_rx_pages->ibp_pages[ipage];
 718                 kib_rx_t    *rx = &conn->ibc_rxs[i];
 719
 720                 rx->rx_conn = conn;
 721                 rx->rx_msg = (kib_msg_t *)(((char *)page_address(page)) +
 722                                            page_offset);
 723                 rx->rx_msgaddr = kiblnd_dma_map_single(cmid->device,
 724                                                        rx->rx_msg, IBLND_MSG_SIZE,
 725                                                        DMA_FROM_DEVICE);
 726                 KIBLND_UNMAP_ADDR_SET(rx, rx_msgunmap, rx->rx_msgaddr);
 727
 728                 CDEBUG(D_NET,"rx %d: %p "LPX64"("LPX64")\n",
 729                        i, rx->rx_msg, rx->rx_msgaddr,
 730                        lnet_page2phys(page) + page_offset);
 731
 732                 page_offset += IBLND_MSG_SIZE;
 733                 LASSERT (page_offset <= PAGE_SIZE);
 734
 735                 if (page_offset == PAGE_SIZE) {
 736                         page_offset = 0;
 737                         ipage++;
 738                         LASSERT (ipage <= IBLND_RX_MSG_PAGES);
 739                 }
 740         }
 741
 742 #ifdef HAVE_OFED_IB_COMP_VECTOR
 743         cq = ib_create_cq(cmid->device,
 744                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 745                           IBLND_CQ_ENTRIES(), 0);
 746 #else
 747         cq = ib_create_cq(cmid->device,
 748                           kiblnd_cq_completion, kiblnd_cq_event, conn,
 749                           IBLND_CQ_ENTRIES());
 750 #endif
 751         if (!IS_ERR(cq)) {
 752                 conn->ibc_cq = cq;
 753         } else {
 754                 CERROR("Can't create CQ: %ld\n", PTR_ERR(cq));
 755                 goto failed_2;
 756         }
 757
 758         rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
 759         if (rc != 0) {
 760                 CERROR("Can't request completion notificiation: %d\n", rc);
 761                 goto failed_2;
 762         }
 763
 764         memset(init_qp_attr, 0, sizeof(*init_qp_attr));
 765         init_qp_attr->event_handler = kiblnd_qp_event;
 766         init_qp_attr->qp_context = conn;
 767         init_qp_attr->cap.max_send_wr = IBLND_SEND_WRS;
 768         init_qp_attr->cap.max_recv_wr = IBLND_RECV_WRS;
 769         init_qp_attr->cap.max_send_sge = 1;
 770         init_qp_attr->cap.max_recv_sge = 1;
 771         init_qp_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 772         init_qp_attr->qp_type = IB_QPT_RC;
 773         init_qp_attr->send_cq = cq;
 774         init_qp_attr->recv_cq = cq;
 775
 776         rc = 0;
 777         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 778         switch (*kiblnd_tunables.kib_ib_mtu) {
 779         default:
 780                 rc = *kiblnd_tunables.kib_ib_mtu;
 781                 /* fall through to... */
 782         case 0: /* set tunable to the default
 783                  * CAVEAT EMPTOR! this assumes the default is one of the MTUs
 784                  * below, otherwise we'll WARN on the next QP create */
 785                 *kiblnd_tunables.kib_ib_mtu =
 786                         ib_mtu_enum_to_int(cmid->route.path_rec->mtu);
 787                 break;
 788         case 256:
 789                 cmid->route.path_rec->mtu = IB_MTU_256;
 790                 break;
 791         case 512:
 792                 cmid->route.path_rec->mtu = IB_MTU_512;
 793                 break;
 794         case 1024:
 795                 cmid->route.path_rec->mtu = IB_MTU_1024;
 796                 break;
 797         case 2048:
 798                 cmid->route.path_rec->mtu = IB_MTU_2048;
 799                 break;
 800         case 4096:
 801                 cmid->route.path_rec->mtu = IB_MTU_4096;
 802                 break;
 803         }
 804         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 805
 806         if (rc != 0)
 807                 CWARN("Invalid IB MTU value %d, using default value %d\n",
 808                       rc, *kiblnd_tunables.kib_ib_mtu);
 809
 810         rc = rdma_create_qp(cmid, net->ibn_dev->ibd_pd, init_qp_attr);
 811         if (rc != 0) {
 812                 CERROR("Can't create QP: %d\n", rc);
 813                 goto failed_2;
 814         }
 815
 816         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 817
 818         /* 1 ref for caller and each rxmsg */
 819         atomic_set(&conn->ibc_refcount, 1 + IBLND_RX_MSGS);
 820         conn->ibc_nrx = IBLND_RX_MSGS;
 821
 822         /* post receives */
 823         for (i = 0; i < IBLND_RX_MSGS; i++) {
 824                 rc = kiblnd_post_rx(&conn->ibc_rxs[i],
 825                                     IBLND_POSTRX_NO_CREDIT);
 826                 if (rc != 0) {
 827                         CERROR("Can't post rxmsg: %d\n", rc);
 828
 829                         /* Make posted receives complete */
 830                         kiblnd_abort_receives(conn);
 831
 832                         /* correct # of posted buffers
 833                          * NB locking needed now I'm racing with completion */
 834                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
 835                         conn->ibc_nrx -= IBLND_RX_MSGS - i;
 836                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
 837                                                flags);
 838
 839                         /* Drop my own and unused rxbuffer refcounts */
 840                         while (i++ <= IBLND_RX_MSGS)
 841                                 kiblnd_conn_decref(conn);
 842
 843                         return NULL;
 844                 }
 845         }
 846
 847         /* Init successful! */
 848         LASSERT (state == IBLND_CONN_ACTIVE_CONNECT ||
 849                  state == IBLND_CONN_PASSIVE_WAIT);
 850         conn->ibc_state = state;
 851
 852         /* 1 more conn */
 853         atomic_inc(&net->ibn_nconns);
 854         return conn;
 855
 856  failed_2:
 857         kiblnd_destroy_conn(conn);
 858  failed_1:
 859         LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
 860  failed_0:
 861         return NULL;
 862 }
 863
 864 void
 865 kiblnd_destroy_conn (kib_conn_t *conn)
 866 {
 867         struct rdma_cm_id *cmid = conn->ibc_cmid;
 868         kib_peer_t        *peer = conn->ibc_peer;
 869         int                rc;
 870         int                i;
 871
 872         LASSERT (!in_interrupt());
 873         LASSERT (atomic_read(&conn->ibc_refcount) == 0);
 874         LASSERT (list_empty(&conn->ibc_early_rxs));
 875         LASSERT (list_empty(&conn->ibc_tx_noops));
 876         LASSERT (list_empty(&conn->ibc_tx_queue));
 877         LASSERT (list_empty(&conn->ibc_tx_queue_rsrvd));
 878         LASSERT (list_empty(&conn->ibc_tx_queue_nocred));
 879         LASSERT (list_empty(&conn->ibc_active_txs));
 880         LASSERT (conn->ibc_nsends_posted == 0);
 881
 882         switch (conn->ibc_state) {
 883         default:
 884                 /* conn must be completely disengaged from the network */
 885                 LBUG();
 886
 887         case IBLND_CONN_DISCONNECTED:
 888                 /* connvars should have been freed already */
 889                 LASSERT (conn->ibc_connvars == NULL);
 890                 break;
 891
 892         case IBLND_CONN_INIT:
 893                 break;
 894         }
 895
 896         if (conn->ibc_cmid->qp != NULL)
 897                 rdma_destroy_qp(conn->ibc_cmid);
 898
 899         if (conn->ibc_cq != NULL) {
 900                 rc = ib_destroy_cq(conn->ibc_cq);
 901                 if (rc != 0)
 902                         CWARN("Error destroying CQ: %d\n", rc);
 903         }
 904
 905         if (conn->ibc_rx_pages != NULL) {
 906                 LASSERT (conn->ibc_rxs != NULL);
 907
 908                 for (i = 0; i < IBLND_RX_MSGS; i++) {
 909                         kib_rx_t *rx = &conn->ibc_rxs[i];
 910
 911                         LASSERT (rx->rx_nob >= 0); /* not posted */
 912
 913                         kiblnd_dma_unmap_single(conn->ibc_cmid->device,
 914                                                 KIBLND_UNMAP_ADDR(rx, rx_msgunmap,
 915                                                                   rx->rx_msgaddr),
 916                                                 IBLND_MSG_SIZE, DMA_FROM_DEVICE);
 917                 }
 918
 919                 kiblnd_free_pages(conn->ibc_rx_pages);
 920         }
 921
 922         if (conn->ibc_rxs != NULL) {
 923                 LIBCFS_FREE(conn->ibc_rxs,
 924                             IBLND_RX_MSGS * sizeof(kib_rx_t));
 925         }
 926
 927         if (conn->ibc_connvars != NULL)
 928                 LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
 929
 930         /* See CAVEAT EMPTOR above in kiblnd_create_conn */
 931         if (conn->ibc_state != IBLND_CONN_INIT) {
 932                 kib_net_t *net = peer->ibp_ni->ni_data;
 933
 934                 kiblnd_peer_decref(peer);
 935                 rdma_destroy_id(cmid);
 936                 atomic_dec(&net->ibn_nconns);
 937         }
 938
 939         LIBCFS_FREE(conn, sizeof(*conn));
 940 }
 941
 942 int
 943 kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why)
 944 {
 945         kib_conn_t         *conn;
 946         struct list_head   *ctmp;
 947         struct list_head   *cnxt;
 948         int                 count = 0;
 949
 950         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 951                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 952
 953                 count++;
 954                 kiblnd_close_conn_locked(conn, why);
 955         }
 956
 957         return count;
 958 }
 959
 960 int
 961 kiblnd_close_stale_conns_locked (kib_peer_t *peer, __u64 incarnation)
 962 {
 963         kib_conn_t         *conn;
 964         struct list_head   *ctmp;
 965         struct list_head   *cnxt;
 966         int                 count = 0;
 967
 968         list_for_each_safe (ctmp, cnxt, &peer->ibp_conns) {
 969                 conn = list_entry(ctmp, kib_conn_t, ibc_list);
 970
 971                 if (conn->ibc_incarnation == incarnation)
 972                         continue;
 973
 974                 CDEBUG(D_NET, "Closing stale conn -> %s incarnation:"LPX64"("LPX64")\n",
 975                        libcfs_nid2str(peer->ibp_nid),
 976                        conn->ibc_incarnation, incarnation);
 977
 978                 count++;
 979                 kiblnd_close_conn_locked(conn, -ESTALE);
 980         }
 981
 982         return count;
 983 }
 984
 985 int
 986 kiblnd_close_matching_conns (lnet_ni_t *ni, lnet_nid_t nid)
 987 {
 988         kib_peer_t         *peer;
 989         struct list_head   *ptmp;
 990         struct list_head   *pnxt;
 991         int                 lo;
 992         int                 hi;
 993         int                 i;
 994         unsigned long       flags;
 995         int                 count = 0;
 996
 997         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 998
 999         if (nid != LNET_NID_ANY)
1000                 lo = hi = kiblnd_nid2peerlist(nid) - kiblnd_data.kib_peers;
1001         else {
1002                 lo = 0;
1003                 hi = kiblnd_data.kib_peer_hash_size - 1;
1004         }
1005
1006         for (i = lo; i <= hi; i++) {
1007                 list_for_each_safe (ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
1008
1009                         peer = list_entry(ptmp, kib_peer_t, ibp_list);
1010                         LASSERT (peer->ibp_connecting > 0 ||
1011                                  peer->ibp_accepting > 0 ||
1012                                  !list_empty(&peer->ibp_conns));
1013
1014                         if (peer->ibp_ni != ni)
1015                                 continue;
1016
1017                         if (!(nid == LNET_NID_ANY || nid == peer->ibp_nid))
1018                                 continue;
1019
1020                         count += kiblnd_close_peer_conns_locked(peer, 0);
1021                 }
1022         }
1023
1024         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1025
1026         /* wildcards always succeed */
1027         if (nid == LNET_NID_ANY)
1028                 return 0;
1029
1030         return (count == 0) ? -ENOENT : 0;
1031 }
1032
1033 int
1034 kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1035 {
1036         struct libcfs_ioctl_data *data = arg;
1037         int                       rc = -EINVAL;
1038
1039         switch(cmd) {
1040         case IOC_LIBCFS_GET_PEER: {
1041                 lnet_nid_t   nid = 0;
1042                 int          count = 0;
1043
1044                 rc = kiblnd_get_peer_info(ni, data->ioc_count,
1045                                           &nid, &count);
1046                 data->ioc_nid    = nid;
1047                 data->ioc_count  = count;
1048                 break;
1049         }
1050
1051         case IOC_LIBCFS_DEL_PEER: {
1052                 rc = kiblnd_del_peer(ni, data->ioc_nid);
1053                 break;
1054         }
1055         case IOC_LIBCFS_GET_CONN: {
1056                 kib_conn_t *conn = kiblnd_get_conn_by_idx(ni, data->ioc_count);
1057
1058                 if (conn == NULL) {
1059                         rc = -ENOENT;
1060                 } else {
1061                         // kiblnd_debug_conn(conn);
1062                         rc = 0;
1063                         data->ioc_nid = conn->ibc_peer->ibp_nid;
1064                         kiblnd_conn_decref(conn);
1065                 }
1066                 break;
1067         }
1068         case IOC_LIBCFS_CLOSE_CONNECTION: {
1069                 rc = kiblnd_close_matching_conns(ni, data->ioc_nid);
1070                 break;
1071         }
1072
1073         default:
1074                 break;
1075         }
1076
1077         return rc;
1078 }
1079
1080 void
1081 kiblnd_free_pages (kib_pages_t *p)
1082 {
1083         int         npages = p->ibp_npages;
1084         int         i;
1085
1086         for (i = 0; i < npages; i++)
1087                 if (p->ibp_pages[i] != NULL)
1088                         __free_page(p->ibp_pages[i]);
1089
1090         LIBCFS_FREE (p, offsetof(kib_pages_t, ibp_pages[npages]));
1091 }
1092
1093 int
1094 kiblnd_alloc_pages (kib_pages_t **pp, int npages)
1095 {
1096         kib_pages_t   *p;
1097         int            i;
1098
1099         LIBCFS_ALLOC(p, offsetof(kib_pages_t, ibp_pages[npages]));
1100         if (p == NULL) {
1101                 CERROR("Can't allocate descriptor for %d pages\n", npages);
1102                 return -ENOMEM;
1103         }
1104
1105         memset(p, 0, offsetof(kib_pages_t, ibp_pages[npages]));
1106         p->ibp_npages = npages;
1107
1108         for (i = 0; i < npages; i++) {
1109                 p->ibp_pages[i] = alloc_page(GFP_KERNEL);
1110                 if (p->ibp_pages[i] == NULL) {
1111                         CERROR("Can't allocate page %d of %d\n", i, npages);
1112                         kiblnd_free_pages(p);
1113                         return -ENOMEM;
1114                 }
1115         }
1116
1117         *pp = p;
1118         return 0;
1119 }
1120
1121 void
1122 kiblnd_free_tx_descs (lnet_ni_t *ni)
1123 {
1124         int        i;
1125         kib_net_t *net = ni->ni_data;
1126
1127         LASSERT (net != NULL);
1128
1129         if (net->ibn_tx_descs != NULL) {
1130                 for (i = 0; i < IBLND_TX_MSGS(); i++) {
1131                         kib_tx_t *tx = &net->ibn_tx_descs[i];
1132
1133 #if IBLND_MAP_ON_DEMAND
1134                         if (tx->tx_pages != NULL)
1135                                 LIBCFS_FREE(tx->tx_pages, LNET_MAX_IOV *
1136                                             sizeof(*tx->tx_pages));
1137 #else
1138                         if (tx->tx_wrq != NULL)
1139                                 LIBCFS_FREE(tx->tx_wrq,
1140                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1141                                             sizeof(*tx->tx_wrq));
1142
1143                         if (tx->tx_sge != NULL)
1144                                 LIBCFS_FREE(tx->tx_sge,
1145                                             (1 + IBLND_MAX_RDMA_FRAGS) *
1146                                             sizeof(*tx->tx_sge));
1147
1148                         if (tx->tx_rd != NULL)
1149                                 LIBCFS_FREE(tx->tx_rd,
1150                                             offsetof(kib_rdma_desc_t,
1151                                                rd_frags[IBLND_MAX_RDMA_FRAGS]));
1152
1153                         if (tx->tx_frags != NULL)
1154                                 LIBCFS_FREE(tx->tx_frags,
1155                                             IBLND_MAX_RDMA_FRAGS *
1156                                             sizeof(*tx->tx_frags));
1157 #endif
1158                 }
1159
1160                 LIBCFS_FREE(net->ibn_tx_descs,
1161                             IBLND_TX_MSGS() * sizeof(kib_tx_t));
1162         }
1163
1164         if (net->ibn_tx_pages != NULL)
1165                 kiblnd_free_pages(net->ibn_tx_pages);
1166 }
1167
1168 int
1169 kiblnd_alloc_tx_descs (lnet_ni_t *ni)
1170 {
1171         int        i;
1172         int        rc;
1173         kib_net_t *net = ni->ni_data;
1174
1175         LASSERT (net != NULL);
1176
1177         rc = kiblnd_alloc_pages(&net->ibn_tx_pages, IBLND_TX_MSG_PAGES());
1178
1179         if (rc != 0) {
1180                 CERROR("Can't allocate tx pages\n");
1181                 return rc;
1182         }
1183
1184         LIBCFS_ALLOC (net->ibn_tx_descs,
1185                       IBLND_TX_MSGS() * sizeof(kib_tx_t));
1186         if (net->ibn_tx_descs == NULL) {
1187                 CERROR("Can't allocate %d tx descriptors\n", IBLND_TX_MSGS());
1188                 return -ENOMEM;
1189         }
1190
1191         memset(net->ibn_tx_descs, 0,
1192                IBLND_TX_MSGS() * sizeof(kib_tx_t));
1193
1194         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1195                 kib_tx_t *tx = &net->ibn_tx_descs[i];
1196
1197 #if IBLND_MAP_ON_DEMAND
1198                 LIBCFS_ALLOC(tx->tx_pages, LNET_MAX_IOV *
1199                              sizeof(*tx->tx_pages));
1200                 if (tx->tx_pages == NULL) {
1201                         CERROR("Can't allocate phys page vector[%d]\n",
1202                                LNET_MAX_IOV);
1203                         return -ENOMEM;
1204                 }
1205 #else
1206                 LIBCFS_ALLOC(tx->tx_wrq,
1207                              (1 + IBLND_MAX_RDMA_FRAGS) *
1208                              sizeof(*tx->tx_wrq));
1209                 if (tx->tx_wrq == NULL)
1210                         return -ENOMEM;
1211
1212                 LIBCFS_ALLOC(tx->tx_sge,
1213                              (1 + IBLND_MAX_RDMA_FRAGS) *
1214                              sizeof(*tx->tx_sge));
1215                 if (tx->tx_sge == NULL)
1216                         return -ENOMEM;
1217
1218                 LIBCFS_ALLOC(tx->tx_rd,
1219                              offsetof(kib_rdma_desc_t,
1220                                       rd_frags[IBLND_MAX_RDMA_FRAGS]));
1221                 if (tx->tx_rd == NULL)
1222                         return -ENOMEM;
1223
1224                 LIBCFS_ALLOC(tx->tx_frags,
1225                              IBLND_MAX_RDMA_FRAGS *
1226                              sizeof(*tx->tx_frags));
1227                 if (tx->tx_frags == NULL)
1228                         return -ENOMEM;
1229 #endif
1230         }
1231
1232         return 0;
1233 }
1234
1235 void
1236 kiblnd_unmap_tx_descs (lnet_ni_t *ni)
1237 {
1238         int             i;
1239         kib_tx_t       *tx;
1240         kib_net_t      *net = ni->ni_data;
1241
1242         LASSERT (net != NULL);
1243
1244         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1245                 tx = &net->ibn_tx_descs[i];
1246
1247                 kiblnd_dma_unmap_single(net->ibn_dev->ibd_cmid->device,
1248                                         KIBLND_UNMAP_ADDR(tx, tx_msgunmap,
1249                                                           tx->tx_msgaddr),
1250                                         IBLND_MSG_SIZE, DMA_TO_DEVICE);
1251         }
1252 }
1253
1254 void
1255 kiblnd_map_tx_descs (lnet_ni_t *ni)
1256 {
1257         int             ipage = 0;
1258         int             page_offset = 0;
1259         int             i;
1260         struct page    *page;
1261         kib_tx_t       *tx;
1262         kib_net_t      *net = ni->ni_data;
1263
1264         LASSERT (net != NULL);
1265
1266         /* pre-mapped messages are not bigger than 1 page */
1267         CLASSERT (IBLND_MSG_SIZE <= PAGE_SIZE);
1268
1269         /* No fancy arithmetic when we do the buffer calculations */
1270         CLASSERT (PAGE_SIZE % IBLND_MSG_SIZE == 0);
1271
1272         for (i = 0; i < IBLND_TX_MSGS(); i++) {
1273                 page = net->ibn_tx_pages->ibp_pages[ipage];
1274                 tx = &net->ibn_tx_descs[i];
1275
1276                 tx->tx_msg = (kib_msg_t *)(((char *)page_address(page)) +
1277                                            page_offset);
1278
1279                 tx->tx_msgaddr = kiblnd_dma_map_single(
1280                         net->ibn_dev->ibd_cmid->device,
1281                         tx->tx_msg, IBLND_MSG_SIZE, DMA_TO_DEVICE);
1282                 KIBLND_UNMAP_ADDR_SET(tx, tx_msgunmap, tx->tx_msgaddr);
1283
1284                 list_add(&tx->tx_list, &net->ibn_idle_txs);
1285
1286                 page_offset += IBLND_MSG_SIZE;
1287                 LASSERT (page_offset <= PAGE_SIZE);
1288
1289                 if (page_offset == PAGE_SIZE) {
1290                         page_offset = 0;
1291                         ipage++;
1292                         LASSERT (ipage <= IBLND_TX_MSG_PAGES());
1293                 }
1294         }
1295 }
1296
1297 void
1298 kiblnd_base_shutdown (void)
1299 {
1300         int i;
1301
1302         LASSERT (list_empty(&kiblnd_data.kib_devs));
1303
1304         CDEBUG(D_MALLOC, "before LND base cleanup: kmem %d\n",
1305                atomic_read(&libcfs_kmemory));
1306
1307         switch (kiblnd_data.kib_init) {
1308         default:
1309                 LBUG();
1310
1311         case IBLND_INIT_ALL:
1312         case IBLND_INIT_DATA:
1313                 LASSERT (kiblnd_data.kib_peers != NULL);
1314                 for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
1315                         LASSERT (list_empty(&kiblnd_data.kib_peers[i]));
1316                 }
1317                 LASSERT (list_empty(&kiblnd_data.kib_connd_zombies));
1318                 LASSERT (list_empty(&kiblnd_data.kib_connd_conns));
1319
1320                 /* flag threads to terminate; wake and wait for them to die */
1321                 kiblnd_data.kib_shutdown = 1;
1322                 wake_up_all(&kiblnd_data.kib_sched_waitq);
1323                 wake_up_all(&kiblnd_data.kib_connd_waitq);
1324
1325                 i = 2;
1326                 while (atomic_read(&kiblnd_data.kib_nthreads) != 0) {
1327                         i++;
1328                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
1329                                "Waiting for %d threads to terminate\n",
1330                                atomic_read(&kiblnd_data.kib_nthreads));
1331                         cfs_pause(cfs_time_seconds(1));
1332                 }
1333
1334                 /* fall through */
1335
1336         case IBLND_INIT_NOTHING:
1337                 break;
1338         }
1339
1340         if (kiblnd_data.kib_peers != NULL)
1341                 LIBCFS_FREE(kiblnd_data.kib_peers,
1342                             sizeof(struct list_head) *
1343                             kiblnd_data.kib_peer_hash_size);
1344
1345         CDEBUG(D_MALLOC, "after LND base cleanup: kmem %d\n",
1346                atomic_read(&libcfs_kmemory));
1347
1348         kiblnd_data.kib_init = IBLND_INIT_NOTHING;
1349         PORTAL_MODULE_UNUSE;
1350 }
1351
1352 void
1353 kiblnd_shutdown (lnet_ni_t *ni)
1354 {
1355         kib_net_t        *net = ni->ni_data;
1356         rwlock_t         *g_lock = &kiblnd_data.kib_global_lock;
1357         int               i;
1358         unsigned long     flags;
1359
1360         LASSERT(kiblnd_data.kib_init == IBLND_INIT_ALL);
1361
1362         if (net == NULL)
1363                 goto out;
1364
1365         CDEBUG(D_MALLOC, "before LND net cleanup: kmem %d\n",
1366                atomic_read(&libcfs_kmemory));
1367
1368         write_lock_irqsave(g_lock, flags);
1369         net->ibn_shutdown = 1;
1370         write_unlock_irqrestore(g_lock, flags);
1371
1372         switch (net->ibn_init) {
1373         default:
1374                 LBUG();
1375
1376         case IBLND_INIT_ALL:
1377                 /* nuke all existing peers within this net */
1378                 kiblnd_del_peer(ni, LNET_NID_ANY);
1379
1380                 /* Wait for all peer state to clean up */
1381                 i = 2;
1382                 while (atomic_read(&net->ibn_npeers) != 0) {
1383                         i++;
1384                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* 2**n? */
1385                                "%s: waiting for %d peers to disconnect\n",
1386                                libcfs_nid2str(ni->ni_nid),
1387                                atomic_read(&net->ibn_npeers));
1388                         cfs_pause(cfs_time_seconds(1));
1389                 }
1390
1391                 kiblnd_unmap_tx_descs(ni);
1392
1393                 LASSERT (net->ibn_dev->ibd_nnets > 0);
1394                 net->ibn_dev->ibd_nnets--;
1395
1396                 /* fall through */
1397
1398         case IBLND_INIT_NOTHING:
1399                 LASSERT (atomic_read(&net->ibn_nconns) == 0);
1400
1401 #if IBLND_MAP_ON_DEMAND
1402                 if (net->ibn_fmrpool != NULL)
1403                         ib_destroy_fmr_pool(net->ibn_fmrpool);
1404 #endif
1405                 if (net->ibn_dev != NULL &&
1406                     net->ibn_dev->ibd_nnets == 0)
1407                         kiblnd_destroy_dev(net->ibn_dev);
1408
1409                 break;
1410         }
1411
1412         kiblnd_free_tx_descs(ni);
1413
1414         CDEBUG(D_MALLOC, "after LND net cleanup: kmem %d\n",
1415                atomic_read(&libcfs_kmemory));
1416
1417         net->ibn_init = IBLND_INIT_NOTHING;
1418         ni->ni_data = NULL;
1419
1420         LIBCFS_FREE(net, sizeof(*net));
1421
1422 out:
1423         if (list_empty(&kiblnd_data.kib_devs))
1424                 kiblnd_base_shutdown();
1425         return;
1426 }
1427
1428 int
1429 kiblnd_base_startup (void)
1430 {
1431         int               rc;
1432         int               i;
1433
1434         LASSERT (kiblnd_data.kib_init == IBLND_INIT_NOTHING);
1435
1436         if (*kiblnd_tunables.kib_credits > *kiblnd_tunables.kib_ntx) {
1437                 CERROR("Can't set credits(%d) > ntx(%d)\n",
1438                        *kiblnd_tunables.kib_credits,
1439                        *kiblnd_tunables.kib_ntx);
1440                 return -EINVAL;
1441         }
1442
1443         PORTAL_MODULE_USE;
1444         memset(&kiblnd_data, 0, sizeof(kiblnd_data)); /* zero pointers, flags etc */
1445
1446         rwlock_init(&kiblnd_data.kib_global_lock);
1447
1448         INIT_LIST_HEAD(&kiblnd_data.kib_devs);
1449
1450         kiblnd_data.kib_peer_hash_size = IBLND_PEER_HASH_SIZE;
1451         LIBCFS_ALLOC(kiblnd_data.kib_peers,
1452                      sizeof(struct list_head) * kiblnd_data.kib_peer_hash_size);
1453         if (kiblnd_data.kib_peers == NULL) {
1454                 goto failed;
1455         }
1456         for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++)
1457                 INIT_LIST_HEAD(&kiblnd_data.kib_peers[i]);
1458
1459         spin_lock_init(&kiblnd_data.kib_connd_lock);
1460         INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
1461         INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
1462         init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
1463
1464         spin_lock_init(&kiblnd_data.kib_sched_lock);
1465         INIT_LIST_HEAD(&kiblnd_data.kib_sched_conns);
1466         init_waitqueue_head(&kiblnd_data.kib_sched_waitq);
1467
1468         kiblnd_data.kib_error_qpa.qp_state = IB_QPS_ERR;
1469
1470         /* lists/ptrs/locks initialised */
1471         kiblnd_data.kib_init = IBLND_INIT_DATA;
1472         /*****************************************************/
1473
1474         for (i = 0; i < IBLND_N_SCHED; i++) {
1475                 rc = kiblnd_thread_start(kiblnd_scheduler, (void *)((long)i));
1476                 if (rc != 0) {
1477                         CERROR("Can't spawn o2iblnd scheduler[%d]: %d\n",
1478                                i, rc);
1479                         goto failed;
1480                 }
1481         }
1482
1483         rc = kiblnd_thread_start(kiblnd_connd, NULL);
1484         if (rc != 0) {
1485                 CERROR("Can't spawn o2iblnd connd: %d\n", rc);
1486                 goto failed;
1487         }
1488
1489         /* flag everything initialised */
1490         kiblnd_data.kib_init = IBLND_INIT_ALL;
1491         /*****************************************************/
1492
1493         return 0;
1494
1495  failed:
1496         kiblnd_base_shutdown();
1497         return -ENETDOWN;
1498 }
1499
1500 int
1501 kiblnd_startup (lnet_ni_t *ni)
1502 {
1503         char                     *ifname;
1504         kib_net_t                *net;
1505         kib_dev_t                *ibdev;
1506         struct list_head         *tmp;
1507         struct timeval            tv;
1508         int                       rc;
1509
1510         LASSERT (ni->ni_lnd == &the_kiblnd);
1511
1512         if (kiblnd_data.kib_init == IBLND_INIT_NOTHING) {
1513                 rc = kiblnd_base_startup();
1514                 if (rc != 0)
1515                         return rc;
1516         }
1517
1518         LIBCFS_ALLOC(net, sizeof(*net));
1519         ni->ni_data = net;
1520         if (net == NULL)
1521                 goto failed;
1522
1523         memset(net, 0, sizeof(*net));
1524
1525         do_gettimeofday(&tv);
1526         net->ibn_incarnation = (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
1527
1528         ni->ni_maxtxcredits = *kiblnd_tunables.kib_credits;
1529         ni->ni_peertxcredits = *kiblnd_tunables.kib_peercredits;
1530
1531         spin_lock_init(&net->ibn_tx_lock);
1532         INIT_LIST_HEAD(&net->ibn_idle_txs);
1533
1534         rc = kiblnd_alloc_tx_descs(ni);
1535         if (rc != 0) {
1536                 CERROR("Can't allocate tx descs\n");
1537                 goto failed;
1538         }
1539
1540         if (ni->ni_interfaces[0] != NULL) {
1541                 /* Use the IPoIB interface specified in 'networks=' */
1542
1543                 CLASSERT (LNET_MAX_INTERFACES > 1);
1544                 if (ni->ni_interfaces[1] != NULL) {
1545                         CERROR("Multiple interfaces not supported\n");
1546                         goto failed;
1547                 }
1548
1549                 ifname = ni->ni_interfaces[0];
1550         } else {
1551                 ifname = *kiblnd_tunables.kib_default_ipif;
1552         }
1553
1554         if (strlen(ifname) >= sizeof(ibdev->ibd_ifname)) {
1555                 CERROR("IPoIB interface name too long: %s\n", ifname);
1556                 goto failed;
1557         }
1558
1559         ibdev = NULL;
1560         list_for_each (tmp, &kiblnd_data.kib_devs) {
1561                 ibdev = list_entry(tmp, kib_dev_t, ibd_list);
1562
1563                 if (!strcmp(&ibdev->ibd_ifname[0], ifname))
1564                         break;
1565
1566                 ibdev = NULL;
1567         }
1568
1569         if (ibdev == NULL) {
1570                 __u32                     ip;
1571                 __u32                     netmask;
1572                 int                       up;
1573                 struct rdma_cm_id        *id;
1574                 struct ib_pd             *pd;
1575                 struct ib_mr             *mr;
1576                 struct sockaddr_in        addr;
1577
1578                 rc = libcfs_ipif_query(ifname, &up, &ip, &netmask);
1579                 if (rc != 0) {
1580                         CERROR("Can't query IPoIB interface %s: %d\n",
1581                                ifname, rc);
1582                         goto failed;
1583                 }
1584
1585                 if (!up) {
1586                         CERROR("Can't query IPoIB interface %s: it's down\n",
1587                                ifname);
1588                         goto failed;
1589                 }
1590
1591                 LIBCFS_ALLOC(ibdev, sizeof(*ibdev));
1592                 if (ibdev == NULL)
1593                         goto failed;
1594
1595                 memset(ibdev, 0, sizeof(*ibdev));
1596
1597                 INIT_LIST_HEAD(&ibdev->ibd_list); /* not yet in kib_devs */
1598                 ibdev->ibd_ifip = ip;
1599                 strcpy(&ibdev->ibd_ifname[0], ifname);
1600
1601                 id = rdma_create_id(kiblnd_cm_callback, ibdev, RDMA_PS_TCP);
1602                 if (!IS_ERR(id)) {
1603                         ibdev->ibd_cmid = id;
1604                 } else {
1605                         CERROR("Can't create listen ID: %ld\n", PTR_ERR(id));
1606                         goto failed;
1607                 }
1608
1609                 memset(&addr, 0, sizeof(addr));
1610                 addr.sin_family      = AF_INET;
1611                 addr.sin_port        = htons(*kiblnd_tunables.kib_service);
1612                 addr.sin_addr.s_addr = htonl(ip);
1613
1614                 rc = rdma_bind_addr(id, (struct sockaddr *)&addr);
1615                 if (rc != 0) {
1616                         CERROR("Can't bind to %s: %d\n", ifname, rc);
1617                         goto failed;
1618                 }
1619
1620                 /* Binding should have assigned me an IB device */
1621                 LASSERT (id->device != NULL);
1622
1623                 pd = ib_alloc_pd(id->device);
1624                 if (!IS_ERR(pd)) {
1625                         ibdev->ibd_pd = pd;
1626                 } else {
1627                         CERROR("Can't allocate PD: %ld\n", PTR_ERR(pd));
1628                         goto failed;
1629                 }
1630
1631 #if IBLND_MAP_ON_DEMAND
1632                 /* MR for sends and receives */
1633                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE);
1634 #else
1635                 /* MR for sends, recieves _and_ RDMA...........v */
1636                 mr = ib_get_dma_mr(pd, IB_ACCESS_LOCAL_WRITE |
1637                                        IB_ACCESS_REMOTE_WRITE);
1638 #endif
1639                 if (!IS_ERR(mr)) {
1640                         ibdev->ibd_mr = mr;
1641                 } else {
1642                         CERROR("Can't get MR: %ld\n", PTR_ERR(mr));
1643                         goto failed;
1644                 }
1645
1646                 rc = rdma_listen(id, 0);
1647                 if (rc != 0) {
1648                         CERROR("Can't start listener: %d\n", rc);
1649                         goto failed;
1650                 }
1651
1652                 list_add_tail(&ibdev->ibd_list,
1653                               &kiblnd_data.kib_devs);
1654         }
1655
1656         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ibdev->ibd_ifip);
1657         net->ibn_dev = ibdev;
1658
1659 #if IBLND_MAP_ON_DEMAND
1660         /* FMR pool for RDMA */
1661         {
1662                 struct ib_fmr_pool      *fmrpool;
1663                 struct ib_fmr_pool_param param = {
1664                         .max_pages_per_fmr = LNET_MAX_PAYLOAD/PAGE_SIZE,
1665                         .page_shift        = PAGE_SHIFT,
1666                         .access            = (IB_ACCESS_LOCAL_WRITE |
1667                                               IB_ACCESS_REMOTE_WRITE),
1668                         .pool_size         = *kiblnd_tunables.kib_fmr_pool_size,
1669                         .dirty_watermark   = *kiblnd_tunables.kib_fmr_flush_trigger,
1670                         .flush_function    = NULL,
1671                         .flush_arg         = NULL,
1672                         .cache             = *kiblnd_tunables.kib_fmr_cache};
1673
1674                 if (*kiblnd_tunables.kib_fmr_pool_size <
1675                     *kiblnd_tunables.kib_ntx) {
1676                         CERROR("Can't set fmr pool size (%d) < ntx(%d)\n",
1677                                *kiblnd_tunables.kib_fmr_pool_size,
1678                                *kiblnd_tunables.kib_ntx);
1679                         goto failed;
1680                 }
1681
1682                 fmrpool = ib_create_fmr_pool(ibdev->ibd_pd, &param);
1683                 if (!IS_ERR(fmrpool)) {
1684                         net->ibn_fmrpool = fmrpool;
1685                 } else {
1686                         CERROR("Can't create FMR pool: %ld\n",
1687                                PTR_ERR(fmrpool));
1688                         goto failed;
1689                 }
1690         }
1691 #endif
1692
1693         kiblnd_map_tx_descs(ni);
1694
1695         ibdev->ibd_nnets++;
1696         net->ibn_init = IBLND_INIT_ALL;
1697
1698         return 0;
1699
1700 failed:
1701         kiblnd_shutdown(ni);
1702
1703         CDEBUG(D_NET, "kiblnd_startup failed\n");
1704         return -ENETDOWN;
1705 }
1706
1707 void __exit
1708 kiblnd_module_fini (void)
1709 {
1710         lnet_unregister_lnd(&the_kiblnd);
1711         kiblnd_tunables_fini();
1712 }
1713
1714 int __init
1715 kiblnd_module_init (void)
1716 {
1717         int    rc;
1718
1719         CLASSERT (sizeof(kib_msg_t) <= IBLND_MSG_SIZE);
1720 #if !IBLND_MAP_ON_DEMAND
1721         CLASSERT (offsetof(kib_msg_t, ibm_u.get.ibgm_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1722                   <= IBLND_MSG_SIZE);
1723         CLASSERT (offsetof(kib_msg_t, ibm_u.putack.ibpam_rd.rd_frags[IBLND_MAX_RDMA_FRAGS])
1724                   <= IBLND_MSG_SIZE);
1725 #endif
1726         rc = kiblnd_tunables_init();
1727         if (rc != 0)
1728                 return rc;
1729
1730         lnet_register_lnd(&the_kiblnd);
1731
1732         return 0;
1733 }
1734
1735 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1736 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v1.00");
1737 MODULE_LICENSE("GPL");
1738
1739 module_init(kiblnd_module_init);
1740 module_exit(kiblnd_module_fini);