lnet/klnds/o2iblnd/o2iblnd_cb.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) 2012, 2015, 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_cb.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 static void kiblnd_peer_alive(kib_peer_t *peer);
  44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
  45 static void kiblnd_check_sends(kib_conn_t *conn);
  46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
  47                                int type, int body_nob);
  48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
  49                             int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
  50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
  51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
  52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
  53
  54 static void
  55 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
  56 {
  57         lnet_msg_t *lntmsg[2];
  58         kib_net_t  *net = ni->ni_data;
  59         int         rc;
  60         int         i;
  61
  62         LASSERT (net != NULL);
  63         LASSERT (!in_interrupt());
  64         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
  65         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
  66         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer response */
  67         LASSERT (tx->tx_pool != NULL);
  68
  69         kiblnd_unmap_tx(ni, tx);
  70
  71         /* tx may have up to 2 lnet msgs to finalise */
  72         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  73         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  74         rc = tx->tx_status;
  75
  76         if (tx->tx_conn != NULL) {
  77                 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
  78
  79                 kiblnd_conn_decref(tx->tx_conn);
  80                 tx->tx_conn = NULL;
  81         }
  82
  83         tx->tx_nwrq = 0;
  84         tx->tx_status = 0;
  85
  86         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  87
  88         /* delay finalize until my descs have been freed */
  89         for (i = 0; i < 2; i++) {
  90                 if (lntmsg[i] == NULL)
  91                         continue;
  92
  93                 lnet_finalize(ni, lntmsg[i], rc);
  94         }
  95 }
  96
  97 void
  98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
  99 {
 100         kib_tx_t *tx;
 101
 102         while (!list_empty(txlist)) {
 103                 tx = list_entry(txlist->next, kib_tx_t, tx_list);
 104
 105                 list_del(&tx->tx_list);
 106                 /* complete now */
 107                 tx->tx_waiting = 0;
 108                 tx->tx_status = status;
 109                 kiblnd_tx_done(ni, tx);
 110         }
 111 }
 112
 113 static kib_tx_t *
 114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
 115 {
 116         kib_net_t               *net = (kib_net_t *)ni->ni_data;
 117         struct list_head        *node;
 118         kib_tx_t                *tx;
 119         kib_tx_poolset_t        *tps;
 120
 121         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
 122         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
 123         if (node == NULL)
 124                 return NULL;
 125         tx = container_of(node, kib_tx_t, tx_list);
 126
 127         LASSERT (tx->tx_nwrq == 0);
 128         LASSERT (!tx->tx_queued);
 129         LASSERT (tx->tx_sending == 0);
 130         LASSERT (!tx->tx_waiting);
 131         LASSERT (tx->tx_status == 0);
 132         LASSERT (tx->tx_conn == NULL);
 133         LASSERT (tx->tx_lntmsg[0] == NULL);
 134         LASSERT (tx->tx_lntmsg[1] == NULL);
 135         LASSERT (tx->tx_nfrags == 0);
 136
 137         return tx;
 138 }
 139
 140 static void
 141 kiblnd_drop_rx(kib_rx_t *rx)
 142 {
 143         kib_conn_t              *conn   = rx->rx_conn;
 144         struct kib_sched_info   *sched  = conn->ibc_sched;
 145         unsigned long           flags;
 146
 147         spin_lock_irqsave(&sched->ibs_lock, flags);
 148         LASSERT(conn->ibc_nrx > 0);
 149         conn->ibc_nrx--;
 150         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 151
 152         kiblnd_conn_decref(conn);
 153 }
 154
 155 int
 156 kiblnd_post_rx (kib_rx_t *rx, int credit)
 157 {
 158         kib_conn_t         *conn = rx->rx_conn;
 159         kib_net_t          *net = conn->ibc_peer->ibp_ni->ni_data;
 160         struct ib_recv_wr  *bad_wrq = NULL;
 161         struct ib_mr       *mr = conn->ibc_hdev->ibh_mrs;
 162         int                 rc;
 163
 164         LASSERT (net != NULL);
 165         LASSERT (!in_interrupt());
 166         LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
 167                  credit == IBLND_POSTRX_PEER_CREDIT ||
 168                  credit == IBLND_POSTRX_RSRVD_CREDIT);
 169         LASSERT(mr != NULL);
 170
 171         rx->rx_sge.lkey   = mr->lkey;
 172         rx->rx_sge.addr   = rx->rx_msgaddr;
 173         rx->rx_sge.length = IBLND_MSG_SIZE;
 174
 175         rx->rx_wrq.next = NULL;
 176         rx->rx_wrq.sg_list = &rx->rx_sge;
 177         rx->rx_wrq.num_sge = 1;
 178         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 179
 180         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
 181         LASSERT (rx->rx_nob >= 0);              /* not posted */
 182
 183         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 184                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 185                 return 0;
 186         }
 187
 188         rx->rx_nob = -1;                        /* flag posted */
 189
 190         /* NB: need an extra reference after ib_post_recv because we don't
 191          * own this rx (and rx::rx_conn) anymore, LU-5678.
 192          */
 193         kiblnd_conn_addref(conn);
 194         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 195         if (unlikely(rc != 0)) {
 196                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 197                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 198                 rx->rx_nob = 0;
 199         }
 200
 201         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 202                 goto out;
 203
 204         if (unlikely(rc != 0)) {
 205                 kiblnd_close_conn(conn, rc);
 206                 kiblnd_drop_rx(rx);     /* No more posts for this rx */
 207                 goto out;
 208         }
 209
 210         if (credit == IBLND_POSTRX_NO_CREDIT)
 211                 goto out;
 212
 213         spin_lock(&conn->ibc_lock);
 214         if (credit == IBLND_POSTRX_PEER_CREDIT)
 215                 conn->ibc_outstanding_credits++;
 216         else
 217                 conn->ibc_reserved_credits++;
 218         spin_unlock(&conn->ibc_lock);
 219
 220         kiblnd_check_sends(conn);
 221 out:
 222         kiblnd_conn_decref(conn);
 223         return rc;
 224 }
 225
 226 static kib_tx_t *
 227 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
 228 {
 229         struct list_head *tmp;
 230
 231         list_for_each(tmp, &conn->ibc_active_txs) {
 232                 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
 233
 234                 LASSERT(!tx->tx_queued);
 235                 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
 236
 237                 if (tx->tx_cookie != cookie)
 238                         continue;
 239
 240                 if (tx->tx_waiting &&
 241                     tx->tx_msg->ibm_type == txtype)
 242                         return tx;
 243
 244                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 245                       tx->tx_waiting ? "" : "NOT ",
 246                       tx->tx_msg->ibm_type, txtype);
 247         }
 248         return NULL;
 249 }
 250
 251 static void
 252 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
 253 {
 254         kib_tx_t    *tx;
 255         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 256         int          idle;
 257
 258         spin_lock(&conn->ibc_lock);
 259
 260         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 261         if (tx == NULL) {
 262                 spin_unlock(&conn->ibc_lock);
 263
 264                 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
 265                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 266                 kiblnd_close_conn(conn, -EPROTO);
 267                 return;
 268         }
 269
 270         if (tx->tx_status == 0) {               /* success so far */
 271                 if (status < 0) {               /* failed? */
 272                         tx->tx_status = status;
 273                 } else if (txtype == IBLND_MSG_GET_REQ) {
 274                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 275                 }
 276         }
 277
 278         tx->tx_waiting = 0;
 279
 280         idle = !tx->tx_queued && (tx->tx_sending == 0);
 281         if (idle)
 282                 list_del(&tx->tx_list);
 283
 284         spin_unlock(&conn->ibc_lock);
 285
 286         if (idle)
 287                 kiblnd_tx_done(ni, tx);
 288 }
 289
 290 static void
 291 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
 292 {
 293         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 294         kib_tx_t    *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 295
 296         if (tx == NULL) {
 297                 CERROR("Can't get tx for completion %x for %s\n",
 298                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 299                 return;
 300         }
 301
 302         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 303         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 304         kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
 305
 306         kiblnd_queue_tx(tx, conn);
 307 }
 308
 309 static void
 310 kiblnd_handle_rx (kib_rx_t *rx)
 311 {
 312         kib_msg_t    *msg = rx->rx_msg;
 313         kib_conn_t   *conn = rx->rx_conn;
 314         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 315         int           credits = msg->ibm_credits;
 316         kib_tx_t     *tx;
 317         int           rc = 0;
 318         int           rc2;
 319         int           post_credit;
 320
 321         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 322
 323         CDEBUG (D_NET, "Received %x[%d] from %s\n",
 324                 msg->ibm_type, credits,
 325                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
 326
 327         if (credits != 0) {
 328                 /* Have I received credits that will let me send? */
 329                 spin_lock(&conn->ibc_lock);
 330
 331                 if (conn->ibc_credits + credits >
 332                     conn->ibc_queue_depth) {
 333                         rc2 = conn->ibc_credits;
 334                         spin_unlock(&conn->ibc_lock);
 335
 336                         CERROR("Bad credits from %s: %d + %d > %d\n",
 337                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 338                                rc2, credits,
 339                                conn->ibc_queue_depth);
 340
 341                         kiblnd_close_conn(conn, -EPROTO);
 342                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 343                         return;
 344                 }
 345
 346                 conn->ibc_credits += credits;
 347
 348                 /* This ensures the credit taken by NOOP can be returned */
 349                 if (msg->ibm_type == IBLND_MSG_NOOP &&
 350                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
 351                         conn->ibc_outstanding_credits++;
 352
 353                 spin_unlock(&conn->ibc_lock);
 354                 kiblnd_check_sends(conn);
 355         }
 356
 357         switch (msg->ibm_type) {
 358         default:
 359                 CERROR("Bad IBLND message type %x from %s\n",
 360                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 361                 post_credit = IBLND_POSTRX_NO_CREDIT;
 362                 rc = -EPROTO;
 363                 break;
 364
 365         case IBLND_MSG_NOOP:
 366                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
 367                         post_credit = IBLND_POSTRX_NO_CREDIT;
 368                         break;
 369                 }
 370
 371                 if (credits != 0) /* credit already posted */
 372                         post_credit = IBLND_POSTRX_NO_CREDIT;
 373                 else              /* a keepalive NOOP */
 374                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 375                 break;
 376
 377         case IBLND_MSG_IMMEDIATE:
 378                 post_credit = IBLND_POSTRX_DONT_POST;
 379                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 380                                 msg->ibm_srcnid, rx, 0);
 381                 if (rc < 0)                     /* repost on error */
 382                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 383                 break;
 384
 385         case IBLND_MSG_PUT_REQ:
 386                 post_credit = IBLND_POSTRX_DONT_POST;
 387                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 388                                 msg->ibm_srcnid, rx, 1);
 389                 if (rc < 0)                     /* repost on error */
 390                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 391                 break;
 392
 393         case IBLND_MSG_PUT_NAK:
 394                 CWARN ("PUT_NACK from %s\n",
 395                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 396                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 397                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 398                                          msg->ibm_u.completion.ibcm_status,
 399                                          msg->ibm_u.completion.ibcm_cookie);
 400                 break;
 401
 402         case IBLND_MSG_PUT_ACK:
 403                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 404
 405                 spin_lock(&conn->ibc_lock);
 406                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 407                                         msg->ibm_u.putack.ibpam_src_cookie);
 408                 if (tx != NULL)
 409                         list_del(&tx->tx_list);
 410                 spin_unlock(&conn->ibc_lock);
 411
 412                 if (tx == NULL) {
 413                         CERROR("Unmatched PUT_ACK from %s\n",
 414                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 415                         rc = -EPROTO;
 416                         break;
 417                 }
 418
 419                 LASSERT (tx->tx_waiting);
 420                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
 421                  * (a) I can overwrite tx_msg since my peer has received it!
 422                  * (b) tx_waiting set tells tx_complete() it's not done. */
 423
 424                 tx->tx_nwrq = 0;                /* overwrite PUT_REQ */
 425
 426                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 427                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 428                                        &msg->ibm_u.putack.ibpam_rd,
 429                                        msg->ibm_u.putack.ibpam_dst_cookie);
 430                 if (rc2 < 0)
 431                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 432                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 433
 434                 spin_lock(&conn->ibc_lock);
 435                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
 436                 kiblnd_queue_tx_locked(tx, conn);
 437                 spin_unlock(&conn->ibc_lock);
 438                 break;
 439
 440         case IBLND_MSG_PUT_DONE:
 441                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 442                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 443                                          msg->ibm_u.completion.ibcm_status,
 444                                          msg->ibm_u.completion.ibcm_cookie);
 445                 break;
 446
 447         case IBLND_MSG_GET_REQ:
 448                 post_credit = IBLND_POSTRX_DONT_POST;
 449                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 450                                 msg->ibm_srcnid, rx, 1);
 451                 if (rc < 0)                     /* repost on error */
 452                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 453                 break;
 454
 455         case IBLND_MSG_GET_DONE:
 456                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 457                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 458                                          msg->ibm_u.completion.ibcm_status,
 459                                          msg->ibm_u.completion.ibcm_cookie);
 460                 break;
 461         }
 462
 463         if (rc < 0)                             /* protocol error */
 464                 kiblnd_close_conn(conn, rc);
 465
 466         if (post_credit != IBLND_POSTRX_DONT_POST)
 467                 kiblnd_post_rx(rx, post_credit);
 468 }
 469
 470 static void
 471 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
 472 {
 473         kib_msg_t    *msg = rx->rx_msg;
 474         kib_conn_t   *conn = rx->rx_conn;
 475         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 476         kib_net_t    *net = ni->ni_data;
 477         int           rc;
 478         int           err = -EIO;
 479
 480         LASSERT (net != NULL);
 481         LASSERT (rx->rx_nob < 0);               /* was posted */
 482         rx->rx_nob = 0;                         /* isn't now */
 483
 484         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 485                 goto ignore;
 486
 487         if (status != IB_WC_SUCCESS) {
 488                 CNETERR("Rx from %s failed: %d\n",
 489                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 490                 goto failed;
 491         }
 492
 493         LASSERT (nob >= 0);
 494         rx->rx_nob = nob;
 495
 496         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 497         if (rc != 0) {
 498                 CERROR ("Error %d unpacking rx from %s\n",
 499                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 500                 goto failed;
 501         }
 502
 503         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 504             msg->ibm_dstnid != ni->ni_nid ||
 505             msg->ibm_srcstamp != conn->ibc_incarnation ||
 506             msg->ibm_dststamp != net->ibn_incarnation) {
 507                 CERROR ("Stale rx from %s\n",
 508                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
 509                 err = -ESTALE;
 510                 goto failed;
 511         }
 512
 513         /* set time last known alive */
 514         kiblnd_peer_alive(conn->ibc_peer);
 515
 516         /* racing with connection establishment/teardown! */
 517
 518         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 519                 rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
 520                 unsigned long  flags;
 521
 522                 write_lock_irqsave(g_lock, flags);
 523                 /* must check holding global lock to eliminate race */
 524                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 525                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 526                         write_unlock_irqrestore(g_lock, flags);
 527                         return;
 528                 }
 529                 write_unlock_irqrestore(g_lock, flags);
 530         }
 531         kiblnd_handle_rx(rx);
 532         return;
 533
 534  failed:
 535         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 536         kiblnd_close_conn(conn, err);
 537  ignore:
 538         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
 539 }
 540
 541 static struct page *
 542 kiblnd_kvaddr_to_page (unsigned long vaddr)
 543 {
 544         struct page *page;
 545
 546         if (is_vmalloc_addr((void *)vaddr)) {
 547                 page = vmalloc_to_page ((void *)vaddr);
 548                 LASSERT (page != NULL);
 549                 return page;
 550         }
 551 #ifdef CONFIG_HIGHMEM
 552         if (vaddr >= PKMAP_BASE &&
 553             vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
 554                 /* No highmem pages only used for bulk (kiov) I/O */
 555                 CERROR("find page for address in highmem\n");
 556                 LBUG();
 557         }
 558 #endif
 559         page = virt_to_page (vaddr);
 560         LASSERT (page != NULL);
 561         return page;
 562 }
 563
 564 static int
 565 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 566 {
 567         kib_hca_dev_t           *hdev;
 568         __u64                   *pages = tx->tx_pages;
 569         kib_fmr_poolset_t       *fps;
 570         int                     npages;
 571         int                     size;
 572         int                     cpt;
 573         int                     rc;
 574         int                     i;
 575
 576         LASSERT(tx->tx_pool != NULL);
 577         LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
 578
 579         hdev  = tx->tx_pool->tpo_hdev;
 580
 581         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
 582                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
 583                                size += hdev->ibh_page_size) {
 584                         pages[npages ++] = (rd->rd_frags[i].rf_addr &
 585                                             hdev->ibh_page_mask) + size;
 586                 }
 587         }
 588
 589         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 590
 591         fps = net->ibn_fmr_ps[cpt];
 592         rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
 593         if (rc != 0) {
 594                 CERROR ("Can't map %d pages: %d\n", npages, rc);
 595                 return rc;
 596         }
 597
 598         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 599          * the rkey */
 600         rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
 601                                          tx->fmr.fmr_pfmr->fmr->lkey;
 602         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
 603         rd->rd_frags[0].rf_nob   = nob;
 604         rd->rd_nfrags = 1;
 605
 606         return 0;
 607 }
 608
 609 static void
 610 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
 611 {
 612         kib_net_t  *net = ni->ni_data;
 613
 614         LASSERT(net != NULL);
 615
 616         if (net->ibn_fmr_ps != NULL && tx->fmr.fmr_pfmr != NULL) {
 617                 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
 618                 tx->fmr.fmr_pfmr = NULL;
 619         }
 620
 621         if (tx->tx_nfrags != 0) {
 622                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
 623                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 624                 tx->tx_nfrags = 0;
 625         }
 626 }
 627
 628 static int
 629 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
 630 {
 631         kib_hca_dev_t *hdev  = tx->tx_pool->tpo_hdev;
 632         kib_net_t     *net   = ni->ni_data;
 633         struct ib_mr  *mr    = NULL;
 634         __u32          nob;
 635         int            i;
 636
 637         /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
 638          * RDMA sink */
 639         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 640         tx->tx_nfrags = nfrags;
 641
 642         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
 643                                           tx->tx_nfrags, tx->tx_dmadir);
 644
 645         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 646                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 647                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 648                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 649                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 650                 nob += rd->rd_frags[i].rf_nob;
 651         }
 652
 653         mr = kiblnd_find_rd_dma_mr(hdev, rd,
 654                                    (tx->tx_conn != NULL) ?
 655                                    tx->tx_conn->ibc_max_frags : -1);
 656         if (mr != NULL) {
 657                 /* found pre-mapping MR */
 658                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 659                 return 0;
 660         }
 661
 662         if (net->ibn_fmr_ps != NULL)
 663                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 664
 665         return -EINVAL;
 666 }
 667
 668
 669 static int
 670 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 671                     unsigned int niov, struct kvec *iov, int offset, int nob)
 672 {
 673         kib_net_t          *net = ni->ni_data;
 674         struct page        *page;
 675         struct scatterlist *sg;
 676         unsigned long       vaddr;
 677         int                 fragnob;
 678         int                 page_offset;
 679
 680         LASSERT (nob > 0);
 681         LASSERT (niov > 0);
 682         LASSERT (net != NULL);
 683
 684         while (offset >= iov->iov_len) {
 685                 offset -= iov->iov_len;
 686                 niov--;
 687                 iov++;
 688                 LASSERT (niov > 0);
 689         }
 690
 691         sg = tx->tx_frags;
 692         do {
 693                 LASSERT (niov > 0);
 694
 695                 vaddr = ((unsigned long)iov->iov_base) + offset;
 696                 page_offset = vaddr & (PAGE_SIZE - 1);
 697                 page = kiblnd_kvaddr_to_page(vaddr);
 698                 if (page == NULL) {
 699                         CERROR ("Can't find page\n");
 700                         return -EFAULT;
 701                 }
 702
 703                 fragnob = min((int)(iov->iov_len - offset), nob);
 704                 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
 705
 706                 sg_set_page(sg, page, fragnob, page_offset);
 707                 sg++;
 708
 709                 if (offset + fragnob < iov->iov_len) {
 710                         offset += fragnob;
 711                 } else {
 712                         offset = 0;
 713                         iov++;
 714                         niov--;
 715                 }
 716                 nob -= fragnob;
 717         } while (nob > 0);
 718
 719         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 720 }
 721
 722 static int
 723 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 724                       int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 725 {
 726         kib_net_t          *net = ni->ni_data;
 727         struct scatterlist *sg;
 728         int                 fragnob;
 729
 730         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 731
 732         LASSERT (nob > 0);
 733         LASSERT (nkiov > 0);
 734         LASSERT (net != NULL);
 735
 736         while (offset >= kiov->kiov_len) {
 737                 offset -= kiov->kiov_len;
 738                 nkiov--;
 739                 kiov++;
 740                 LASSERT (nkiov > 0);
 741         }
 742
 743         sg = tx->tx_frags;
 744         do {
 745                 LASSERT (nkiov > 0);
 746
 747                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 748
 749                 sg_set_page(sg, kiov->kiov_page, fragnob,
 750                             kiov->kiov_offset + offset);
 751                 sg++;
 752
 753                 offset = 0;
 754                 kiov++;
 755                 nkiov--;
 756                 nob -= fragnob;
 757         } while (nob > 0);
 758
 759         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 760 }
 761
 762 static int
 763 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
 764 __must_hold(&conn->ibc_lock)
 765 {
 766         kib_msg_t         *msg = tx->tx_msg;
 767         kib_peer_t        *peer = conn->ibc_peer;
 768         int                ver = conn->ibc_version;
 769         int                rc;
 770         int                done;
 771
 772         LASSERT(tx->tx_queued);
 773         /* We rely on this for QP sizing */
 774         LASSERT(tx->tx_nwrq > 0);
 775         LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
 776
 777         LASSERT(credit == 0 || credit == 1);
 778         LASSERT(conn->ibc_outstanding_credits >= 0);
 779         LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
 780         LASSERT(conn->ibc_credits >= 0);
 781         LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
 782
 783         if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
 784                 /* tx completions outstanding... */
 785                 CDEBUG(D_NET, "%s: posted enough\n",
 786                        libcfs_nid2str(peer->ibp_nid));
 787                 return -EAGAIN;
 788         }
 789
 790         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 791                 CDEBUG(D_NET, "%s: no credits\n",
 792                        libcfs_nid2str(peer->ibp_nid));
 793                 return -EAGAIN;
 794         }
 795
 796         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 797             conn->ibc_credits == 1 &&   /* last credit reserved */
 798             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
 799                 CDEBUG(D_NET, "%s: not using last credit\n",
 800                        libcfs_nid2str(peer->ibp_nid));
 801                 return -EAGAIN;
 802         }
 803
 804         /* NB don't drop ibc_lock before bumping tx_sending */
 805         list_del(&tx->tx_list);
 806         tx->tx_queued = 0;
 807
 808         if (msg->ibm_type == IBLND_MSG_NOOP &&
 809             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
 810              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 811               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 812                 /* OK to drop when posted enough NOOPs, since
 813                  * kiblnd_check_sends will queue NOOP again when
 814                  * posted NOOPs complete */
 815                 spin_unlock(&conn->ibc_lock);
 816                 kiblnd_tx_done(peer->ibp_ni, tx);
 817                 spin_lock(&conn->ibc_lock);
 818                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 819                        libcfs_nid2str(peer->ibp_nid),
 820                        conn->ibc_noops_posted);
 821                 return 0;
 822         }
 823
 824         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 825                         peer->ibp_nid, conn->ibc_incarnation);
 826
 827         conn->ibc_credits -= credit;
 828         conn->ibc_outstanding_credits = 0;
 829         conn->ibc_nsends_posted++;
 830         if (msg->ibm_type == IBLND_MSG_NOOP)
 831                 conn->ibc_noops_posted++;
 832
 833         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 834          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 835          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 836          * and then re-queued here.  It's (just) possible that
 837          * tx_sending is non-zero if we've not done the tx_complete()
 838          * from the first send; hence the ++ rather than = below. */
 839         tx->tx_sending++;
 840         list_add(&tx->tx_list, &conn->ibc_active_txs);
 841
 842         /* I'm still holding ibc_lock! */
 843         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
 844                 rc = -ECONNABORTED;
 845         } else if (tx->tx_pool->tpo_pool.po_failed ||
 846                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
 847                 /* close_conn will launch failover */
 848                 rc = -ENETDOWN;
 849         } else {
 850                 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq - 1];
 851
 852                 LASSERTF(wrq->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
 853                          "bad wr_id "LPX64", opc %d, flags %d, peer: %s\n",
 854                          wrq->wr_id, wrq->opcode, wrq->send_flags,
 855                          libcfs_nid2str(conn->ibc_peer->ibp_nid));
 856                 wrq = NULL;
 857                 rc = ib_post_send(conn->ibc_cmid->qp, tx->tx_wrq, &wrq);
 858         }
 859
 860         conn->ibc_last_send = jiffies;
 861
 862         if (rc == 0)
 863                 return 0;
 864
 865         /* NB credits are transferred in the actual
 866          * message, which can only be the last work item */
 867         conn->ibc_credits += credit;
 868         conn->ibc_outstanding_credits += msg->ibm_credits;
 869         conn->ibc_nsends_posted--;
 870         if (msg->ibm_type == IBLND_MSG_NOOP)
 871                 conn->ibc_noops_posted--;
 872
 873         tx->tx_status = rc;
 874         tx->tx_waiting = 0;
 875         tx->tx_sending--;
 876
 877         done = (tx->tx_sending == 0);
 878         if (done)
 879                 list_del(&tx->tx_list);
 880
 881         spin_unlock(&conn->ibc_lock);
 882
 883         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 884                 CERROR("Error %d posting transmit to %s\n",
 885                        rc, libcfs_nid2str(peer->ibp_nid));
 886         else
 887                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 888                        rc, libcfs_nid2str(peer->ibp_nid));
 889
 890         kiblnd_close_conn(conn, rc);
 891
 892         if (done)
 893                 kiblnd_tx_done(peer->ibp_ni, tx);
 894
 895         spin_lock(&conn->ibc_lock);
 896
 897         return -EIO;
 898 }
 899
 900 static void
 901 kiblnd_check_sends (kib_conn_t *conn)
 902 {
 903         int        ver = conn->ibc_version;
 904         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 905         kib_tx_t  *tx;
 906
 907         /* Don't send anything until after the connection is established */
 908         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 909                 CDEBUG(D_NET, "%s too soon\n",
 910                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 911                 return;
 912         }
 913
 914         spin_lock(&conn->ibc_lock);
 915
 916         LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
 917         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 918                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 919         LASSERT (conn->ibc_reserved_credits >= 0);
 920
 921         while (conn->ibc_reserved_credits > 0 &&
 922                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 923                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 924                                     kib_tx_t, tx_list);
 925                 list_del(&tx->tx_list);
 926                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 927                 conn->ibc_reserved_credits--;
 928         }
 929
 930         if (kiblnd_need_noop(conn)) {
 931                 spin_unlock(&conn->ibc_lock);
 932
 933                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 934                 if (tx != NULL)
 935                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 936
 937                 spin_lock(&conn->ibc_lock);
 938                 if (tx != NULL)
 939                         kiblnd_queue_tx_locked(tx, conn);
 940         }
 941
 942         for (;;) {
 943                 int credit;
 944
 945                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 946                         credit = 0;
 947                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
 948                                             kib_tx_t, tx_list);
 949                 } else if (!list_empty(&conn->ibc_tx_noops)) {
 950                         LASSERT (!IBLND_OOB_CAPABLE(ver));
 951                         credit = 1;
 952                         tx = list_entry(conn->ibc_tx_noops.next,
 953                                         kib_tx_t, tx_list);
 954                 } else if (!list_empty(&conn->ibc_tx_queue)) {
 955                         credit = 1;
 956                         tx = list_entry(conn->ibc_tx_queue.next,
 957                                             kib_tx_t, tx_list);
 958                 } else
 959                         break;
 960
 961                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
 962                         break;
 963         }
 964
 965         spin_unlock(&conn->ibc_lock);
 966 }
 967
 968 static void
 969 kiblnd_tx_complete (kib_tx_t *tx, int status)
 970 {
 971         int           failed = (status != IB_WC_SUCCESS);
 972         kib_conn_t   *conn = tx->tx_conn;
 973         int           idle;
 974
 975         LASSERT (tx->tx_sending > 0);
 976
 977         if (failed) {
 978                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 979                         CNETERR("Tx -> %s cookie "LPX64
 980                                 " sending %d waiting %d: failed %d\n",
 981                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
 982                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
 983                                 status);
 984
 985                 kiblnd_close_conn(conn, -EIO);
 986         } else {
 987                 kiblnd_peer_alive(conn->ibc_peer);
 988         }
 989
 990         spin_lock(&conn->ibc_lock);
 991
 992         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
 993          * gets to free it, which also drops its ref on 'conn'. */
 994
 995         tx->tx_sending--;
 996         conn->ibc_nsends_posted--;
 997         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
 998                 conn->ibc_noops_posted--;
 999
1000         if (failed) {
1001                 tx->tx_waiting = 0;             /* don't wait for peer */
1002                 tx->tx_status = -EIO;
1003         }
1004
1005         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1006                !tx->tx_waiting &&               /* Not waiting for peer */
1007                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1008         if (idle)
1009                 list_del(&tx->tx_list);
1010
1011         kiblnd_conn_addref(conn);               /* 1 ref for me.... */
1012
1013         spin_unlock(&conn->ibc_lock);
1014
1015         if (idle)
1016                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1017
1018         kiblnd_check_sends(conn);
1019
1020         kiblnd_conn_decref(conn);               /* ...until here */
1021 }
1022
1023 static void
1024 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1025 {
1026         kib_hca_dev_t     *hdev = tx->tx_pool->tpo_hdev;
1027         struct ib_sge     *sge = &tx->tx_sge[tx->tx_nwrq];
1028         struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1029         int                nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1030         struct ib_mr      *mr = hdev->ibh_mrs;
1031
1032         LASSERT(tx->tx_nwrq >= 0);
1033         LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1034         LASSERT(nob <= IBLND_MSG_SIZE);
1035         LASSERT(mr != NULL);
1036
1037         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1038
1039         sge->lkey   = mr->lkey;
1040         sge->addr   = tx->tx_msgaddr;
1041         sge->length = nob;
1042
1043         memset(wrq, 0, sizeof(*wrq));
1044
1045         wrq->next       = NULL;
1046         wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1047         wrq->sg_list    = sge;
1048         wrq->num_sge    = 1;
1049         wrq->opcode     = IB_WR_SEND;
1050         wrq->send_flags = IB_SEND_SIGNALED;
1051
1052         tx->tx_nwrq++;
1053 }
1054
1055 static int
1056 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1057                  int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1058 {
1059         kib_msg_t         *ibmsg = tx->tx_msg;
1060         kib_rdma_desc_t   *srcrd = tx->tx_rd;
1061         struct ib_sge     *sge = &tx->tx_sge[0];
1062         struct ib_send_wr *wrq = &tx->tx_wrq[0];
1063         int                rc  = resid;
1064         int                srcidx;
1065         int                dstidx;
1066         int                wrknob;
1067
1068         LASSERT (!in_interrupt());
1069         LASSERT (tx->tx_nwrq == 0);
1070         LASSERT (type == IBLND_MSG_GET_DONE ||
1071                  type == IBLND_MSG_PUT_DONE);
1072
1073         srcidx = dstidx = 0;
1074
1075         while (resid > 0) {
1076                 if (srcidx >= srcrd->rd_nfrags) {
1077                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1078                         rc = -EPROTO;
1079                         break;
1080                 }
1081
1082                 if (dstidx == dstrd->rd_nfrags) {
1083                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1084                         rc = -EPROTO;
1085                         break;
1086                 }
1087
1088                 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1089                         CERROR("RDMA has too many fragments for peer %s (%d), "
1090                                "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1091                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1092                                conn->ibc_max_frags,
1093                                srcidx, srcrd->rd_nfrags,
1094                                dstidx, dstrd->rd_nfrags);
1095                         rc = -EMSGSIZE;
1096                         break;
1097                 }
1098
1099                 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1100                                  kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1101
1102                 sge = &tx->tx_sge[tx->tx_nwrq];
1103                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1104                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1105                 sge->length = wrknob;
1106
1107                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1108
1109                 wrq->next       = wrq + 1;
1110                 wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1111                 wrq->sg_list    = sge;
1112                 wrq->num_sge    = 1;
1113                 wrq->opcode     = IB_WR_RDMA_WRITE;
1114                 wrq->send_flags = 0;
1115
1116                 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1117                 wrq->wr.rdma.rkey        = kiblnd_rd_frag_key(dstrd, dstidx);
1118
1119                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1120                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1121
1122                 resid -= wrknob;
1123
1124                 tx->tx_nwrq++;
1125                 wrq++;
1126                 sge++;
1127         }
1128
1129         if (rc < 0)                             /* no RDMA if completing with failure */
1130                 tx->tx_nwrq = 0;
1131
1132         ibmsg->ibm_u.completion.ibcm_status = rc;
1133         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1134         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1135                            type, sizeof (kib_completion_msg_t));
1136
1137         return rc;
1138 }
1139
1140 static void
1141 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1142 {
1143         struct list_head *q;
1144
1145         LASSERT(tx->tx_nwrq > 0);       /* work items set up */
1146         LASSERT(!tx->tx_queued);        /* not queued for sending already */
1147         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1148
1149         tx->tx_queued = 1;
1150         tx->tx_deadline = jiffies +
1151                           msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1152                                            MSEC_PER_SEC);
1153
1154         if (tx->tx_conn == NULL) {
1155                 kiblnd_conn_addref(conn);
1156                 tx->tx_conn = conn;
1157                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1158         } else {
1159                 /* PUT_DONE first attached to conn as a PUT_REQ */
1160                 LASSERT (tx->tx_conn == conn);
1161                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1162         }
1163
1164         switch (tx->tx_msg->ibm_type) {
1165         default:
1166                 LBUG();
1167
1168         case IBLND_MSG_PUT_REQ:
1169         case IBLND_MSG_GET_REQ:
1170                 q = &conn->ibc_tx_queue_rsrvd;
1171                 break;
1172
1173         case IBLND_MSG_PUT_NAK:
1174         case IBLND_MSG_PUT_ACK:
1175         case IBLND_MSG_PUT_DONE:
1176         case IBLND_MSG_GET_DONE:
1177                 q = &conn->ibc_tx_queue_nocred;
1178                 break;
1179
1180         case IBLND_MSG_NOOP:
1181                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1182                         q = &conn->ibc_tx_queue_nocred;
1183                 else
1184                         q = &conn->ibc_tx_noops;
1185                 break;
1186
1187         case IBLND_MSG_IMMEDIATE:
1188                 q = &conn->ibc_tx_queue;
1189                 break;
1190         }
1191
1192         list_add_tail(&tx->tx_list, q);
1193 }
1194
1195 static void
1196 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1197 {
1198         spin_lock(&conn->ibc_lock);
1199         kiblnd_queue_tx_locked(tx, conn);
1200         spin_unlock(&conn->ibc_lock);
1201
1202         kiblnd_check_sends(conn);
1203 }
1204
1205 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1206                                struct sockaddr_in *srcaddr,
1207                                struct sockaddr_in *dstaddr,
1208                                int timeout_ms)
1209 {
1210         unsigned short port;
1211         int rc;
1212
1213         /* allow the port to be reused */
1214         rc = rdma_set_reuseaddr(cmid, 1);
1215         if (rc != 0) {
1216                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1217                 return rc;
1218         }
1219
1220         /* look for a free privileged port */
1221         for (port = PROT_SOCK-1; port > 0; port--) {
1222                 srcaddr->sin_port = htons(port);
1223                 rc = rdma_resolve_addr(cmid,
1224                                        (struct sockaddr *)srcaddr,
1225                                        (struct sockaddr *)dstaddr,
1226                                        timeout_ms);
1227                 if (rc == 0) {
1228                         CDEBUG(D_NET, "bound to port %hu\n", port);
1229                         return 0;
1230                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1231                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1232                                port, rc);
1233                 } else {
1234                         return rc;
1235                 }
1236         }
1237
1238         CERROR("Failed to bind to a free privileged port\n");
1239         return rc;
1240 }
1241
1242 static void
1243 kiblnd_connect_peer (kib_peer_t *peer)
1244 {
1245         struct rdma_cm_id *cmid;
1246         kib_dev_t         *dev;
1247         kib_net_t         *net = peer->ibp_ni->ni_data;
1248         struct sockaddr_in srcaddr;
1249         struct sockaddr_in dstaddr;
1250         int                rc;
1251
1252         LASSERT (net != NULL);
1253         LASSERT (peer->ibp_connecting > 0);
1254
1255         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1256                                      IB_QPT_RC);
1257
1258         if (IS_ERR(cmid)) {
1259                 CERROR("Can't create CMID for %s: %ld\n",
1260                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1261                 rc = PTR_ERR(cmid);
1262                 goto failed;
1263         }
1264
1265         dev = net->ibn_dev;
1266         memset(&srcaddr, 0, sizeof(srcaddr));
1267         srcaddr.sin_family = AF_INET;
1268         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1269
1270         memset(&dstaddr, 0, sizeof(dstaddr));
1271         dstaddr.sin_family = AF_INET;
1272         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1273         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1274
1275         kiblnd_peer_addref(peer);               /* cmid's ref */
1276
1277         if (*kiblnd_tunables.kib_use_priv_port) {
1278                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1279                                          *kiblnd_tunables.kib_timeout * 1000);
1280         } else {
1281                 rc = rdma_resolve_addr(cmid,
1282                                        (struct sockaddr *)&srcaddr,
1283                                        (struct sockaddr *)&dstaddr,
1284                                        *kiblnd_tunables.kib_timeout * 1000);
1285         }
1286         if (rc != 0) {
1287                 /* Can't initiate address resolution:  */
1288                 CERROR("Can't resolve addr for %s: %d\n",
1289                        libcfs_nid2str(peer->ibp_nid), rc);
1290                 goto failed2;
1291         }
1292
1293         LASSERT (cmid->device != NULL);
1294         CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1295                libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1296                &dev->ibd_ifip, cmid->device->name);
1297
1298         return;
1299
1300  failed2:
1301         kiblnd_peer_connect_failed(peer, 1, rc);
1302         kiblnd_peer_decref(peer);               /* cmid's ref */
1303         rdma_destroy_id(cmid);
1304         return;
1305  failed:
1306         kiblnd_peer_connect_failed(peer, 1, rc);
1307 }
1308
1309 void
1310 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1311 {
1312         kib_peer_t        *peer;
1313         kib_peer_t        *peer2;
1314         kib_conn_t        *conn;
1315         rwlock_t        *g_lock = &kiblnd_data.kib_global_lock;
1316         unsigned long      flags;
1317         int                rc;
1318
1319         /* If I get here, I've committed to send, so I complete the tx with
1320          * failure on any problems */
1321
1322         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1323         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1324
1325         /* First time, just use a read lock since I expect to find my peer
1326          * connected */
1327         read_lock_irqsave(g_lock, flags);
1328
1329         peer = kiblnd_find_peer_locked(nid);
1330         if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1331                 /* Found a peer with an established connection */
1332                 conn = kiblnd_get_conn_locked(peer);
1333                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1334
1335                 read_unlock_irqrestore(g_lock, flags);
1336
1337                 if (tx != NULL)
1338                         kiblnd_queue_tx(tx, conn);
1339                 kiblnd_conn_decref(conn); /* ...to here */
1340                 return;
1341         }
1342
1343         read_unlock(g_lock);
1344         /* Re-try with a write lock */
1345         write_lock(g_lock);
1346
1347         peer = kiblnd_find_peer_locked(nid);
1348         if (peer != NULL) {
1349                 if (list_empty(&peer->ibp_conns)) {
1350                         /* found a peer, but it's still connecting... */
1351                         LASSERT (peer->ibp_connecting != 0 ||
1352                                  peer->ibp_accepting != 0);
1353                         if (tx != NULL)
1354                                 list_add_tail(&tx->tx_list,
1355                                                   &peer->ibp_tx_queue);
1356                         write_unlock_irqrestore(g_lock, flags);
1357                 } else {
1358                         conn = kiblnd_get_conn_locked(peer);
1359                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1360
1361                         write_unlock_irqrestore(g_lock, flags);
1362
1363                         if (tx != NULL)
1364                                 kiblnd_queue_tx(tx, conn);
1365                         kiblnd_conn_decref(conn); /* ...to here */
1366                 }
1367                 return;
1368         }
1369
1370         write_unlock_irqrestore(g_lock, flags);
1371
1372         /* Allocate a peer ready to add to the peer table and retry */
1373         rc = kiblnd_create_peer(ni, &peer, nid);
1374         if (rc != 0) {
1375                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1376                 if (tx != NULL) {
1377                         tx->tx_status = -EHOSTUNREACH;
1378                         tx->tx_waiting = 0;
1379                         kiblnd_tx_done(ni, tx);
1380                 }
1381                 return;
1382         }
1383
1384         write_lock_irqsave(g_lock, flags);
1385
1386         peer2 = kiblnd_find_peer_locked(nid);
1387         if (peer2 != NULL) {
1388                 if (list_empty(&peer2->ibp_conns)) {
1389                         /* found a peer, but it's still connecting... */
1390                         LASSERT (peer2->ibp_connecting != 0 ||
1391                                  peer2->ibp_accepting != 0);
1392                         if (tx != NULL)
1393                                 list_add_tail(&tx->tx_list,
1394                                                   &peer2->ibp_tx_queue);
1395                         write_unlock_irqrestore(g_lock, flags);
1396                 } else {
1397                         conn = kiblnd_get_conn_locked(peer2);
1398                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1399
1400                         write_unlock_irqrestore(g_lock, flags);
1401
1402                         if (tx != NULL)
1403                                 kiblnd_queue_tx(tx, conn);
1404                         kiblnd_conn_decref(conn); /* ...to here */
1405                 }
1406
1407                 kiblnd_peer_decref(peer);
1408                 return;
1409         }
1410
1411         /* Brand new peer */
1412         LASSERT (peer->ibp_connecting == 0);
1413         peer->ibp_connecting = 1;
1414
1415         /* always called with a ref on ni, which prevents ni being shutdown */
1416         LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1417
1418         if (tx != NULL)
1419                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1420
1421         kiblnd_peer_addref(peer);
1422         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1423
1424         write_unlock_irqrestore(g_lock, flags);
1425
1426         kiblnd_connect_peer(peer);
1427         kiblnd_peer_decref(peer);
1428 }
1429
1430 int
1431 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1432 {
1433         lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
1434         int               type = lntmsg->msg_type;
1435         lnet_process_id_t target = lntmsg->msg_target;
1436         int               target_is_router = lntmsg->msg_target_is_router;
1437         int               routing = lntmsg->msg_routing;
1438         unsigned int      payload_niov = lntmsg->msg_niov;
1439         struct kvec      *payload_iov = lntmsg->msg_iov;
1440         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
1441         unsigned int      payload_offset = lntmsg->msg_offset;
1442         unsigned int      payload_nob = lntmsg->msg_len;
1443         kib_msg_t        *ibmsg;
1444         kib_rdma_desc_t  *rd;
1445         kib_tx_t         *tx;
1446         int               nob;
1447         int               rc;
1448
1449         /* NB 'private' is different depending on what we're sending.... */
1450
1451         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1452                payload_nob, payload_niov, libcfs_id2str(target));
1453
1454         LASSERT (payload_nob == 0 || payload_niov > 0);
1455         LASSERT (payload_niov <= LNET_MAX_IOV);
1456
1457         /* Thread context */
1458         LASSERT (!in_interrupt());
1459         /* payload is either all vaddrs or all pages */
1460         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1461
1462         switch (type) {
1463         default:
1464                 LBUG();
1465                 return (-EIO);
1466
1467         case LNET_MSG_ACK:
1468                 LASSERT (payload_nob == 0);
1469                 break;
1470
1471         case LNET_MSG_GET:
1472                 if (routing || target_is_router)
1473                         break;                  /* send IMMEDIATE */
1474
1475                 /* is the REPLY message too small for RDMA? */
1476                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1477                 if (nob <= IBLND_MSG_SIZE)
1478                         break;                  /* send IMMEDIATE */
1479
1480                 tx = kiblnd_get_idle_tx(ni, target.nid);
1481                 if (tx == NULL) {
1482                         CERROR("Can't allocate txd for GET to %s\n",
1483                                libcfs_nid2str(target.nid));
1484                         return -ENOMEM;
1485                 }
1486
1487                 ibmsg = tx->tx_msg;
1488                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1489                 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1490                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1491                                                  lntmsg->msg_md->md_niov,
1492                                                  lntmsg->msg_md->md_iov.iov,
1493                                                  0, lntmsg->msg_md->md_length);
1494                 else
1495                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1496                                                   lntmsg->msg_md->md_niov,
1497                                                   lntmsg->msg_md->md_iov.kiov,
1498                                                   0, lntmsg->msg_md->md_length);
1499                 if (rc != 0) {
1500                         CERROR("Can't setup GET sink for %s: %d\n",
1501                                libcfs_nid2str(target.nid), rc);
1502                         kiblnd_tx_done(ni, tx);
1503                         return -EIO;
1504                 }
1505
1506                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1507                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1508                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1509
1510                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1511
1512                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1513                 if (tx->tx_lntmsg[1] == NULL) {
1514                         CERROR("Can't create reply for GET -> %s\n",
1515                                libcfs_nid2str(target.nid));
1516                         kiblnd_tx_done(ni, tx);
1517                         return -EIO;
1518                 }
1519
1520                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1521                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1522                 kiblnd_launch_tx(ni, tx, target.nid);
1523                 return 0;
1524
1525         case LNET_MSG_REPLY:
1526         case LNET_MSG_PUT:
1527                 /* Is the payload small enough not to need RDMA? */
1528                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1529                 if (nob <= IBLND_MSG_SIZE)
1530                         break;                  /* send IMMEDIATE */
1531
1532                 tx = kiblnd_get_idle_tx(ni, target.nid);
1533                 if (tx == NULL) {
1534                         CERROR("Can't allocate %s txd for %s\n",
1535                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1536                                libcfs_nid2str(target.nid));
1537                         return -ENOMEM;
1538                 }
1539
1540                 if (payload_kiov == NULL)
1541                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1542                                                  payload_niov, payload_iov,
1543                                                  payload_offset, payload_nob);
1544                 else
1545                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1546                                                   payload_niov, payload_kiov,
1547                                                   payload_offset, payload_nob);
1548                 if (rc != 0) {
1549                         CERROR("Can't setup PUT src for %s: %d\n",
1550                                libcfs_nid2str(target.nid), rc);
1551                         kiblnd_tx_done(ni, tx);
1552                         return -EIO;
1553                 }
1554
1555                 ibmsg = tx->tx_msg;
1556                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1557                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1558                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1559
1560                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1561                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1562                 kiblnd_launch_tx(ni, tx, target.nid);
1563                 return 0;
1564         }
1565
1566         /* send IMMEDIATE */
1567
1568         LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1569                  <= IBLND_MSG_SIZE);
1570
1571         tx = kiblnd_get_idle_tx(ni, target.nid);
1572         if (tx == NULL) {
1573                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1574                         type, libcfs_nid2str(target.nid));
1575                 return -ENOMEM;
1576         }
1577
1578         ibmsg = tx->tx_msg;
1579         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1580
1581         if (payload_kiov != NULL)
1582                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1583                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1584                                     payload_niov, payload_kiov,
1585                                     payload_offset, payload_nob);
1586         else
1587                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1588                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1589                                    payload_niov, payload_iov,
1590                                    payload_offset, payload_nob);
1591
1592         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1593         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1594
1595         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1596         kiblnd_launch_tx(ni, tx, target.nid);
1597         return 0;
1598 }
1599
1600 static void
1601 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1602 {
1603         lnet_process_id_t target = lntmsg->msg_target;
1604         unsigned int      niov = lntmsg->msg_niov;
1605         struct kvec      *iov = lntmsg->msg_iov;
1606         lnet_kiov_t      *kiov = lntmsg->msg_kiov;
1607         unsigned int      offset = lntmsg->msg_offset;
1608         unsigned int      nob = lntmsg->msg_len;
1609         kib_tx_t         *tx;
1610         int               rc;
1611
1612         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1613         if (tx == NULL) {
1614                 CERROR("Can't get tx for REPLY to %s\n",
1615                        libcfs_nid2str(target.nid));
1616                 goto failed_0;
1617         }
1618
1619         if (nob == 0)
1620                 rc = 0;
1621         else if (kiov == NULL)
1622                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1623                                          niov, iov, offset, nob);
1624         else
1625                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1626                                           niov, kiov, offset, nob);
1627
1628         if (rc != 0) {
1629                 CERROR("Can't setup GET src for %s: %d\n",
1630                        libcfs_nid2str(target.nid), rc);
1631                 goto failed_1;
1632         }
1633
1634         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1635                               IBLND_MSG_GET_DONE, nob,
1636                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1637                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1638         if (rc < 0) {
1639                 CERROR("Can't setup rdma for GET from %s: %d\n",
1640                        libcfs_nid2str(target.nid), rc);
1641                 goto failed_1;
1642         }
1643
1644         if (nob == 0) {
1645                 /* No RDMA: local completion may happen now! */
1646                 lnet_finalize(ni, lntmsg, 0);
1647         } else {
1648                 /* RDMA: lnet_finalize(lntmsg) when it
1649                  * completes */
1650                 tx->tx_lntmsg[0] = lntmsg;
1651         }
1652
1653         kiblnd_queue_tx(tx, rx->rx_conn);
1654         return;
1655
1656  failed_1:
1657         kiblnd_tx_done(ni, tx);
1658  failed_0:
1659         lnet_finalize(ni, lntmsg, -EIO);
1660 }
1661
1662 int
1663 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1664             unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1665             unsigned int offset, unsigned int mlen, unsigned int rlen)
1666 {
1667         kib_rx_t    *rx = private;
1668         kib_msg_t   *rxmsg = rx->rx_msg;
1669         kib_conn_t  *conn = rx->rx_conn;
1670         kib_tx_t    *tx;
1671         int          nob;
1672         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1673         int          rc = 0;
1674
1675         LASSERT (mlen <= rlen);
1676         LASSERT (!in_interrupt());
1677         /* Either all pages or all vaddrs */
1678         LASSERT (!(kiov != NULL && iov != NULL));
1679
1680         switch (rxmsg->ibm_type) {
1681         default:
1682                 LBUG();
1683
1684         case IBLND_MSG_IMMEDIATE:
1685                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1686                 if (nob > rx->rx_nob) {
1687                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1688                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1689                                 nob, rx->rx_nob);
1690                         rc = -EPROTO;
1691                         break;
1692                 }
1693
1694                 if (kiov != NULL)
1695                         lnet_copy_flat2kiov(niov, kiov, offset,
1696                                             IBLND_MSG_SIZE, rxmsg,
1697                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1698                                             mlen);
1699                 else
1700                         lnet_copy_flat2iov(niov, iov, offset,
1701                                            IBLND_MSG_SIZE, rxmsg,
1702                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1703                                            mlen);
1704                 lnet_finalize (ni, lntmsg, 0);
1705                 break;
1706
1707         case IBLND_MSG_PUT_REQ: {
1708                 kib_msg_t       *txmsg;
1709                 kib_rdma_desc_t *rd;
1710
1711                 if (mlen == 0) {
1712                         lnet_finalize(ni, lntmsg, 0);
1713                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1714                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1715                         break;
1716                 }
1717
1718                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1719                 if (tx == NULL) {
1720                         CERROR("Can't allocate tx for %s\n",
1721                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1722                         /* Not replying will break the connection */
1723                         rc = -ENOMEM;
1724                         break;
1725                 }
1726
1727                 txmsg = tx->tx_msg;
1728                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1729                 if (kiov == NULL)
1730                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1731                                                  niov, iov, offset, mlen);
1732                 else
1733                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1734                                                   niov, kiov, offset, mlen);
1735                 if (rc != 0) {
1736                         CERROR("Can't setup PUT sink for %s: %d\n",
1737                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1738                         kiblnd_tx_done(ni, tx);
1739                         /* tell peer it's over */
1740                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1741                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1742                         break;
1743                 }
1744
1745                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1746                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1747                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1748
1749                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1750
1751                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1752                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1753                 kiblnd_queue_tx(tx, conn);
1754
1755                 /* reposted buffer reserved for PUT_DONE */
1756                 post_credit = IBLND_POSTRX_NO_CREDIT;
1757                 break;
1758                 }
1759
1760         case IBLND_MSG_GET_REQ:
1761                 if (lntmsg != NULL) {
1762                         /* Optimized GET; RDMA lntmsg's payload */
1763                         kiblnd_reply(ni, rx, lntmsg);
1764                 } else {
1765                         /* GET didn't match anything */
1766                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1767                                                -ENODATA,
1768                                                rxmsg->ibm_u.get.ibgm_cookie);
1769                 }
1770                 break;
1771         }
1772
1773         kiblnd_post_rx(rx, post_credit);
1774         return rc;
1775 }
1776
1777 int
1778 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1779 {
1780         struct task_struct *task = kthread_run(fn, arg, name);
1781
1782         if (IS_ERR(task))
1783                 return PTR_ERR(task);
1784
1785         atomic_inc(&kiblnd_data.kib_nthreads);
1786         return 0;
1787 }
1788
1789 static void
1790 kiblnd_thread_fini (void)
1791 {
1792         atomic_dec (&kiblnd_data.kib_nthreads);
1793 }
1794
1795 static void
1796 kiblnd_peer_alive (kib_peer_t *peer)
1797 {
1798         /* This is racy, but everyone's only writing cfs_time_current() */
1799         peer->ibp_last_alive = cfs_time_current();
1800         smp_mb();
1801 }
1802
1803 static void
1804 kiblnd_peer_notify (kib_peer_t *peer)
1805 {
1806         int           error = 0;
1807         cfs_time_t    last_alive = 0;
1808         unsigned long flags;
1809
1810         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1811
1812         if (list_empty(&peer->ibp_conns) &&
1813             peer->ibp_accepting == 0 &&
1814             peer->ibp_connecting == 0 &&
1815             peer->ibp_error != 0) {
1816                 error = peer->ibp_error;
1817                 peer->ibp_error = 0;
1818
1819                 last_alive = peer->ibp_last_alive;
1820         }
1821
1822         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1823
1824         if (error != 0)
1825                 lnet_notify(peer->ibp_ni,
1826                             peer->ibp_nid, 0, last_alive);
1827 }
1828
1829 void
1830 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1831 {
1832         /* This just does the immediate housekeeping.  'error' is zero for a
1833          * normal shutdown which can happen only after the connection has been
1834          * established.  If the connection is established, schedule the
1835          * connection to be finished off by the connd.  Otherwise the connd is
1836          * already dealing with it (either to set it up or tear it down).
1837          * Caller holds kib_global_lock exclusively in irq context */
1838         kib_peer_t       *peer = conn->ibc_peer;
1839         kib_dev_t        *dev;
1840         unsigned long     flags;
1841
1842         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1843
1844         if (error != 0 && conn->ibc_comms_error == 0)
1845                 conn->ibc_comms_error = error;
1846
1847         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1848                 return; /* already being handled  */
1849
1850         if (error == 0 &&
1851             list_empty(&conn->ibc_tx_noops) &&
1852             list_empty(&conn->ibc_tx_queue) &&
1853             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1854             list_empty(&conn->ibc_tx_queue_nocred) &&
1855             list_empty(&conn->ibc_active_txs)) {
1856                 CDEBUG(D_NET, "closing conn to %s\n",
1857                        libcfs_nid2str(peer->ibp_nid));
1858         } else {
1859                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1860                        libcfs_nid2str(peer->ibp_nid), error,
1861                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1862                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1863                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
1864                                                 "" : "(sending_rsrvd)",
1865                        list_empty(&conn->ibc_tx_queue_nocred) ?
1866                                                  "" : "(sending_nocred)",
1867                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1868         }
1869
1870         dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1871         list_del(&conn->ibc_list);
1872         /* connd (see below) takes over ibc_list's ref */
1873
1874         if (list_empty(&peer->ibp_conns) &&    /* no more conns */
1875             kiblnd_peer_active(peer)) {         /* still in peer table */
1876                 kiblnd_unlink_peer_locked(peer);
1877
1878                 /* set/clear error on last conn */
1879                 peer->ibp_error = conn->ibc_comms_error;
1880         }
1881
1882         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1883
1884         if (error != 0 &&
1885             kiblnd_dev_can_failover(dev)) {
1886                 list_add_tail(&dev->ibd_fail_list,
1887                               &kiblnd_data.kib_failed_devs);
1888                 wake_up(&kiblnd_data.kib_failover_waitq);
1889         }
1890
1891         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1892
1893         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1894         wake_up(&kiblnd_data.kib_connd_waitq);
1895
1896         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1897 }
1898
1899 void
1900 kiblnd_close_conn(kib_conn_t *conn, int error)
1901 {
1902         unsigned long flags;
1903
1904         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1905
1906         kiblnd_close_conn_locked(conn, error);
1907
1908         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1909 }
1910
1911 static void
1912 kiblnd_handle_early_rxs(kib_conn_t *conn)
1913 {
1914         unsigned long    flags;
1915         kib_rx_t        *rx;
1916
1917         LASSERT(!in_interrupt());
1918         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1919
1920         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1921         while (!list_empty(&conn->ibc_early_rxs)) {
1922                 rx = list_entry(conn->ibc_early_rxs.next,
1923                                     kib_rx_t, rx_list);
1924                 list_del(&rx->rx_list);
1925                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1926
1927                 kiblnd_handle_rx(rx);
1928
1929                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1930         }
1931         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1932 }
1933
1934 static void
1935 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1936 {
1937         struct list_head         zombies = LIST_HEAD_INIT(zombies);
1938         struct list_head        *tmp;
1939         struct list_head        *nxt;
1940         kib_tx_t                *tx;
1941
1942         spin_lock(&conn->ibc_lock);
1943
1944         list_for_each_safe(tmp, nxt, txs) {
1945                 tx = list_entry(tmp, kib_tx_t, tx_list);
1946
1947                 if (txs == &conn->ibc_active_txs) {
1948                         LASSERT(!tx->tx_queued);
1949                         LASSERT(tx->tx_waiting ||
1950                                 tx->tx_sending != 0);
1951                 } else {
1952                         LASSERT(tx->tx_queued);
1953                 }
1954
1955                 tx->tx_status = -ECONNABORTED;
1956                 tx->tx_waiting = 0;
1957
1958                 if (tx->tx_sending == 0) {
1959                         tx->tx_queued = 0;
1960                         list_del(&tx->tx_list);
1961                         list_add(&tx->tx_list, &zombies);
1962                 }
1963         }
1964
1965         spin_unlock(&conn->ibc_lock);
1966
1967         kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
1968 }
1969
1970 static void
1971 kiblnd_finalise_conn (kib_conn_t *conn)
1972 {
1973         LASSERT (!in_interrupt());
1974         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1975
1976         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1977
1978         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
1979          * for connections that didn't get as far as being connected, because
1980          * rdma_disconnect() does this for free. */
1981         kiblnd_abort_receives(conn);
1982
1983         /* Complete all tx descs not waiting for sends to complete.
1984          * NB we should be safe from RDMA now that the QP has changed state */
1985
1986         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
1987         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1988         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1989         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1990         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1991
1992         kiblnd_handle_early_rxs(conn);
1993 }
1994
1995 static void
1996 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
1997 {
1998         struct list_head zombies = LIST_HEAD_INIT(zombies);
1999         unsigned long   flags;
2000
2001         LASSERT (error != 0);
2002         LASSERT (!in_interrupt());
2003
2004         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2005
2006         if (active) {
2007                 LASSERT (peer->ibp_connecting > 0);
2008                 peer->ibp_connecting--;
2009         } else {
2010                 LASSERT (peer->ibp_accepting > 0);
2011                 peer->ibp_accepting--;
2012         }
2013
2014         if (peer->ibp_connecting != 0 ||
2015             peer->ibp_accepting != 0) {
2016                 /* another connection attempt under way... */
2017                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2018                                         flags);
2019                 return;
2020         }
2021
2022         if (list_empty(&peer->ibp_conns)) {
2023                 /* Take peer's blocked transmits to complete with error */
2024                 list_add(&zombies, &peer->ibp_tx_queue);
2025                 list_del_init(&peer->ibp_tx_queue);
2026
2027                 if (kiblnd_peer_active(peer))
2028                         kiblnd_unlink_peer_locked(peer);
2029
2030                 peer->ibp_error = error;
2031         } else {
2032                 /* Can't have blocked transmits if there are connections */
2033                 LASSERT(list_empty(&peer->ibp_tx_queue));
2034         }
2035
2036         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2037
2038         kiblnd_peer_notify(peer);
2039
2040         if (list_empty(&zombies))
2041                 return;
2042
2043         CNETERR("Deleting messages for %s: connection failed\n",
2044                 libcfs_nid2str(peer->ibp_nid));
2045
2046         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2047 }
2048
2049 static void
2050 kiblnd_connreq_done(kib_conn_t *conn, int status)
2051 {
2052         kib_peer_t       *peer = conn->ibc_peer;
2053         kib_tx_t         *tx;
2054         struct list_head txs;
2055         unsigned long    flags;
2056         int              active;
2057
2058         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2059
2060         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2061                libcfs_nid2str(peer->ibp_nid), active,
2062                conn->ibc_version, status);
2063
2064         LASSERT (!in_interrupt());
2065         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2066                   peer->ibp_connecting > 0) ||
2067                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2068                   peer->ibp_accepting > 0));
2069
2070         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2071         conn->ibc_connvars = NULL;
2072
2073         if (status != 0) {
2074                 /* failed to establish connection */
2075                 kiblnd_peer_connect_failed(peer, active, status);
2076                 kiblnd_finalise_conn(conn);
2077                 return;
2078         }
2079
2080         /* connection established */
2081         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2082
2083         conn->ibc_last_send = jiffies;
2084         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2085         kiblnd_peer_alive(peer);
2086
2087         /* Add conn to peer's list and nuke any dangling conns from a different
2088          * peer instance... */
2089         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2090         list_add(&conn->ibc_list, &peer->ibp_conns);
2091         if (active)
2092                 peer->ibp_connecting--;
2093         else
2094                 peer->ibp_accepting--;
2095
2096         if (peer->ibp_version == 0) {
2097                 peer->ibp_version     = conn->ibc_version;
2098                 peer->ibp_incarnation = conn->ibc_incarnation;
2099         }
2100
2101         if (peer->ibp_version     != conn->ibc_version ||
2102             peer->ibp_incarnation != conn->ibc_incarnation) {
2103                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2104                                                 conn->ibc_incarnation);
2105                 peer->ibp_version     = conn->ibc_version;
2106                 peer->ibp_incarnation = conn->ibc_incarnation;
2107         }
2108
2109         /* grab pending txs while I have the lock */
2110         list_add(&txs, &peer->ibp_tx_queue);
2111         list_del_init(&peer->ibp_tx_queue);
2112
2113         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2114             conn->ibc_comms_error != 0) {       /* error has happened already */
2115                 lnet_ni_t *ni = peer->ibp_ni;
2116
2117                 /* start to shut down connection */
2118                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2119                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2120
2121                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2122
2123                 return;
2124         }
2125
2126         /* refcount taken by cmid is not reliable after I released the glock
2127          * because this connection is visible to other threads now, another
2128          * thread can find and close this connection right after I released
2129          * the glock, if kiblnd_cm_callback for RDMA_CM_EVENT_DISCONNECTED is
2130          * called, it can release the connection refcount taken by cmid.
2131          * It means the connection could be destroyed before I finish my
2132          * operations on it.
2133          */
2134         kiblnd_conn_addref(conn);
2135         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2136
2137         /* Schedule blocked txs */
2138         spin_lock(&conn->ibc_lock);
2139         while (!list_empty(&txs)) {
2140                 tx = list_entry(txs.next, kib_tx_t, tx_list);
2141                 list_del(&tx->tx_list);
2142
2143                 kiblnd_queue_tx_locked(tx, conn);
2144         }
2145         spin_unlock(&conn->ibc_lock);
2146
2147         kiblnd_check_sends(conn);
2148
2149         /* schedule blocked rxs */
2150         kiblnd_handle_early_rxs(conn);
2151
2152         kiblnd_conn_decref(conn);
2153 }
2154
2155 static void
2156 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2157 {
2158         int          rc;
2159
2160         rc = rdma_reject(cmid, rej, sizeof(*rej));
2161
2162         if (rc != 0)
2163                 CWARN("Error %d sending reject\n", rc);
2164 }
2165
2166 static int
2167 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2168 {
2169         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2170         kib_msg_t             *reqmsg = priv;
2171         kib_msg_t             *ackmsg;
2172         kib_dev_t             *ibdev;
2173         kib_peer_t            *peer;
2174         kib_peer_t            *peer2;
2175         kib_conn_t            *conn;
2176         lnet_ni_t             *ni  = NULL;
2177         kib_net_t             *net = NULL;
2178         lnet_nid_t             nid;
2179         struct rdma_conn_param cp;
2180         kib_rej_t              rej;
2181         int                    version = IBLND_MSG_VERSION;
2182         unsigned long          flags;
2183         int                    rc;
2184         struct sockaddr_in    *peer_addr;
2185         LASSERT (!in_interrupt());
2186
2187         /* cmid inherits 'context' from the corresponding listener id */
2188         ibdev = (kib_dev_t *)cmid->context;
2189         LASSERT (ibdev != NULL);
2190
2191         memset(&rej, 0, sizeof(rej));
2192         rej.ibr_magic                = IBLND_MSG_MAGIC;
2193         rej.ibr_why                  = IBLND_REJECT_FATAL;
2194         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2195
2196         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2197         if (*kiblnd_tunables.kib_require_priv_port &&
2198             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2199                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2200                 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2201                        &ip, ntohs(peer_addr->sin_port));
2202                 goto failed;
2203         }
2204
2205         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2206                 CERROR("Short connection request\n");
2207                 goto failed;
2208         }
2209
2210         /* Future protocol version compatibility support!  If the
2211          * o2iblnd-specific protocol changes, or when LNET unifies
2212          * protocols over all LNDs, the initial connection will
2213          * negotiate a protocol version.  I trap this here to avoid
2214          * console errors; the reject tells the peer which protocol I
2215          * speak. */
2216         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2217             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2218                 goto failed;
2219         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2220             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2221             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2222                 goto failed;
2223         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2224             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2225             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2226                 goto failed;
2227
2228         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2229         if (rc != 0) {
2230                 CERROR("Can't parse connection request: %d\n", rc);
2231                 goto failed;
2232         }
2233
2234         nid = reqmsg->ibm_srcnid;
2235         ni  = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2236
2237         if (ni != NULL) {
2238                 net = (kib_net_t *)ni->ni_data;
2239                 rej.ibr_incarnation = net->ibn_incarnation;
2240         }
2241
2242         if (ni == NULL ||                         /* no matching net */
2243             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2244             net->ibn_dev != ibdev) {              /* wrong device */
2245                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2246                        "bad dst nid %s\n", libcfs_nid2str(nid),
2247                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2248                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2249                         &ibdev->ibd_ifip,
2250                        libcfs_nid2str(reqmsg->ibm_dstnid));
2251
2252                 goto failed;
2253         }
2254
2255        /* check time stamp as soon as possible */
2256         if (reqmsg->ibm_dststamp != 0 &&
2257             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2258                 CWARN("Stale connection request\n");
2259                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2260                 goto failed;
2261         }
2262
2263         /* I can accept peer's version */
2264         version = reqmsg->ibm_version;
2265
2266         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2267                 CERROR("Unexpected connreq msg type: %x from %s\n",
2268                        reqmsg->ibm_type, libcfs_nid2str(nid));
2269                 goto failed;
2270         }
2271
2272         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2273             IBLND_MSG_QUEUE_SIZE(version)) {
2274                 CERROR("Can't accept conn from %s, queue depth too large: "
2275                        " %d (<=%d wanted)\n",
2276                        libcfs_nid2str(nid),
2277                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2278                        IBLND_MSG_QUEUE_SIZE(version));
2279
2280                 if (version == IBLND_MSG_VERSION)
2281                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2282
2283                 goto failed;
2284         }
2285
2286         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2287             IBLND_RDMA_FRAGS(version)) {
2288                 CWARN("Can't accept conn from %s (version %x): "
2289                       "max_frags %d too large (%d wanted)\n",
2290                        libcfs_nid2str(nid), version,
2291                        reqmsg->ibm_u.connparams.ibcp_max_frags,
2292                        IBLND_RDMA_FRAGS(version));
2293
2294                 if (version >= IBLND_MSG_VERSION)
2295                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2296
2297                 goto failed;
2298         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2299                    IBLND_RDMA_FRAGS(version) && net->ibn_fmr_ps == NULL) {
2300                 CWARN("Can't accept conn from %s (version %x): "
2301                       "max_frags %d incompatible without FMR pool "
2302                       "(%d wanted)\n",
2303                       libcfs_nid2str(nid), version,
2304                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2305                       IBLND_RDMA_FRAGS(version));
2306
2307                 if (version >= IBLND_MSG_VERSION)
2308                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2309
2310                 goto failed;
2311         }
2312
2313         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2314                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2315                        libcfs_nid2str(nid),
2316                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2317                        IBLND_MSG_SIZE);
2318                 goto failed;
2319         }
2320
2321         /* assume 'nid' is a new peer; create  */
2322         rc = kiblnd_create_peer(ni, &peer, nid);
2323         if (rc != 0) {
2324                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2325                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2326                 goto failed;
2327         }
2328
2329         /* We have validated the peer's parameters so use those */
2330         peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2331         peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2332
2333         write_lock_irqsave(g_lock, flags);
2334
2335         peer2 = kiblnd_find_peer_locked(nid);
2336         if (peer2 != NULL) {
2337                 if (peer2->ibp_version == 0) {
2338                         peer2->ibp_version     = version;
2339                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2340                 }
2341
2342                 /* not the guy I've talked with */
2343                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2344                     peer2->ibp_version     != version) {
2345                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2346                         write_unlock_irqrestore(g_lock, flags);
2347
2348                         CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2349                               libcfs_nid2str(nid), peer2->ibp_version, version);
2350
2351                         kiblnd_peer_decref(peer);
2352                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2353                         goto failed;
2354                 }
2355
2356                 /* tie-break connection race in favour of the higher NID */
2357                 if (peer2->ibp_connecting != 0 &&
2358                     nid < ni->ni_nid) {
2359                         write_unlock_irqrestore(g_lock, flags);
2360
2361                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2362
2363                         kiblnd_peer_decref(peer);
2364                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2365                         goto failed;
2366                 }
2367
2368                 peer2->ibp_accepting++;
2369                 kiblnd_peer_addref(peer2);
2370
2371                 /* Race with kiblnd_launch_tx (active connect) to create peer
2372                  * so copy validated parameters since we now know what the
2373                  * peer's limits are */
2374                 peer2->ibp_max_frags = peer->ibp_max_frags;
2375                 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2376
2377                 write_unlock_irqrestore(g_lock, flags);
2378                 kiblnd_peer_decref(peer);
2379                 peer = peer2;
2380         } else {
2381                 /* Brand new peer */
2382                 LASSERT (peer->ibp_accepting == 0);
2383                 LASSERT (peer->ibp_version == 0 &&
2384                          peer->ibp_incarnation == 0);
2385
2386                 peer->ibp_accepting   = 1;
2387                 peer->ibp_version     = version;
2388                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2389
2390                 /* I have a ref on ni that prevents it being shutdown */
2391                 LASSERT (net->ibn_shutdown == 0);
2392
2393                 kiblnd_peer_addref(peer);
2394                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2395
2396                 write_unlock_irqrestore(g_lock, flags);
2397         }
2398
2399         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2400         if (conn == NULL) {
2401                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2402                 kiblnd_peer_decref(peer);
2403                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2404                 goto failed;
2405         }
2406
2407         /* conn now "owns" cmid, so I return success from here on to ensure the
2408          * CM callback doesn't destroy cmid. */
2409         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2410         conn->ibc_credits          = conn->ibc_queue_depth;
2411         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2412         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2413                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2414
2415         ackmsg = &conn->ibc_connvars->cv_msg;
2416         memset(ackmsg, 0, sizeof(*ackmsg));
2417
2418         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2419                         sizeof(ackmsg->ibm_u.connparams));
2420         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2421         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2422         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2423
2424         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2425
2426         memset(&cp, 0, sizeof(cp));
2427         cp.private_data        = ackmsg;
2428         cp.private_data_len    = ackmsg->ibm_nob;
2429         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2430         cp.initiator_depth     = 0;
2431         cp.flow_control        = 1;
2432         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2433         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2434
2435         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2436
2437         rc = rdma_accept(cmid, &cp);
2438         if (rc != 0) {
2439                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2440                 rej.ibr_version = version;
2441                 rej.ibr_why     = IBLND_REJECT_FATAL;
2442
2443                 kiblnd_reject(cmid, &rej);
2444                 kiblnd_connreq_done(conn, rc);
2445                 kiblnd_conn_decref(conn);
2446         }
2447
2448         lnet_ni_decref(ni);
2449         return 0;
2450
2451  failed:
2452         if (ni != NULL)
2453                 lnet_ni_decref(ni);
2454
2455         rej.ibr_version = version;
2456         rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2457         rej.ibr_cp.ibcp_max_frags   = IBLND_RDMA_FRAGS(version);
2458         kiblnd_reject(cmid, &rej);
2459
2460         return -ECONNREFUSED;
2461 }
2462
2463 static void
2464 kiblnd_reconnect (kib_conn_t *conn, int version,
2465                   __u64 incarnation, int why, kib_connparams_t *cp)
2466 {
2467         kib_peer_t      *peer = conn->ibc_peer;
2468         char            *reason;
2469         int              retry = 0;
2470         unsigned long    flags;
2471
2472         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2473         LASSERT (peer->ibp_connecting > 0);     /* 'conn' at least */
2474
2475         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2476
2477         /* retry connection if it's still needed and no other connection
2478          * attempts (active or passive) are in progress
2479          * NB: reconnect is still needed even when ibp_tx_queue is
2480          * empty if ibp_version != version because reconnect may be
2481          * initiated by kiblnd_query() */
2482         if ((!list_empty(&peer->ibp_tx_queue) ||
2483              peer->ibp_version != version) &&
2484             peer->ibp_connecting == 1 &&
2485             peer->ibp_accepting == 0) {
2486                 retry = 1;
2487                 peer->ibp_connecting++;
2488                 peer->ibp_version     = version;
2489                 peer->ibp_incarnation = incarnation;
2490         }
2491
2492         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2493
2494         if (!retry)
2495                 return;
2496
2497         switch (why) {
2498         default:
2499                 reason = "Unknown";
2500                 break;
2501
2502         case IBLND_REJECT_RDMA_FRAGS:
2503                 if (!cp)
2504                         goto failed;
2505
2506                 if (conn->ibc_max_frags <= cp->ibcp_max_frags) {
2507                         CNETERR("Unsupported max frags, peer supports %d\n",
2508                                 cp->ibcp_max_frags);
2509                         goto failed;
2510                 } else if (*kiblnd_tunables.kib_map_on_demand == 0) {
2511                         CNETERR("map_on_demand must be enabled to support "
2512                                 "map_on_demand peers\n");
2513                         goto failed;
2514                 }
2515
2516                 peer->ibp_max_frags = cp->ibcp_max_frags;
2517                 reason = "rdma fragments";
2518                 break;
2519
2520         case IBLND_REJECT_MSG_QUEUE_SIZE:
2521                 if (!cp)
2522                         goto failed;
2523
2524                 if (conn->ibc_queue_depth <= cp->ibcp_queue_depth) {
2525                         CNETERR("Unsupported queue depth, peer supports %d\n",
2526                                 cp->ibcp_queue_depth);
2527                         goto failed;
2528                 }
2529
2530                 peer->ibp_queue_depth = cp->ibcp_queue_depth;
2531                 reason = "queue depth";
2532                 break;
2533
2534         case IBLND_REJECT_CONN_STALE:
2535                 reason = "stale";
2536                 break;
2537
2538         case IBLND_REJECT_CONN_RACE:
2539                 reason = "conn race";
2540                 break;
2541
2542         case IBLND_REJECT_CONN_UNCOMPAT:
2543                 reason = "version negotiation";
2544                 break;
2545         }
2546
2547         CNETERR("%s: retrying (%s), %x, %x, "
2548                 "queue_depth: %d, max_frags: %d, msg_size: %d\n",
2549                 libcfs_nid2str(peer->ibp_nid),
2550                 reason, IBLND_MSG_VERSION, version,
2551                 conn->ibc_queue_depth, conn->ibc_max_frags,
2552                 cp != NULL ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
2553
2554         kiblnd_connect_peer(peer);
2555         return;
2556
2557  failed:
2558         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2559         peer->ibp_connecting--;
2560         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2561
2562         return;
2563 }
2564
2565 static void
2566 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2567 {
2568         kib_peer_t    *peer = conn->ibc_peer;
2569
2570         LASSERT (!in_interrupt());
2571         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2572
2573         switch (reason) {
2574         case IB_CM_REJ_STALE_CONN:
2575                 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2576                                  IBLND_REJECT_CONN_STALE, NULL);
2577                 break;
2578
2579         case IB_CM_REJ_INVALID_SERVICE_ID:
2580                 CNETERR("%s rejected: no listener at %d\n",
2581                         libcfs_nid2str(peer->ibp_nid),
2582                         *kiblnd_tunables.kib_service);
2583                 break;
2584
2585         case IB_CM_REJ_CONSUMER_DEFINED:
2586                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2587                         kib_rej_t        *rej         = priv;
2588                         kib_connparams_t *cp          = NULL;
2589                         int               flip        = 0;
2590                         __u64             incarnation = -1;
2591
2592                         /* NB. default incarnation is -1 because:
2593                          * a) V1 will ignore dst incarnation in connreq.
2594                          * b) V2 will provide incarnation while rejecting me,
2595                          *    -1 will be overwrote.
2596                          *
2597                          * if I try to connect to a V1 peer with V2 protocol,
2598                          * it rejected me then upgrade to V2, I have no idea
2599                          * about the upgrading and try to reconnect with V1,
2600                          * in this case upgraded V2 can find out I'm trying to
2601                          * talk to the old guy and reject me(incarnation is -1).
2602                          */
2603
2604                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2605                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2606                                 __swab32s(&rej->ibr_magic);
2607                                 __swab16s(&rej->ibr_version);
2608                                 flip = 1;
2609                         }
2610
2611                         if (priv_nob >= sizeof(kib_rej_t) &&
2612                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2613                                 /* priv_nob is always 148 in current version
2614                                  * of OFED, so we still need to check version.
2615                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2616                                 cp = &rej->ibr_cp;
2617
2618                                 if (flip) {
2619                                         __swab64s(&rej->ibr_incarnation);
2620                                         __swab16s(&cp->ibcp_queue_depth);
2621                                         __swab16s(&cp->ibcp_max_frags);
2622                                         __swab32s(&cp->ibcp_max_msg_size);
2623                                 }
2624
2625                                 incarnation = rej->ibr_incarnation;
2626                         }
2627
2628                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2629                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2630                                 CERROR("%s rejected: consumer defined fatal error\n",
2631                                        libcfs_nid2str(peer->ibp_nid));
2632                                 break;
2633                         }
2634
2635                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2636                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2637                                 CERROR("%s rejected: o2iblnd version %x error\n",
2638                                        libcfs_nid2str(peer->ibp_nid),
2639                                        rej->ibr_version);
2640                                 break;
2641                         }
2642
2643                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2644                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2645                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2646                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2647
2648                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2649                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2650                         }
2651
2652                         switch (rej->ibr_why) {
2653                         case IBLND_REJECT_CONN_RACE:
2654                         case IBLND_REJECT_CONN_STALE:
2655                         case IBLND_REJECT_CONN_UNCOMPAT:
2656                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2657                         case IBLND_REJECT_RDMA_FRAGS:
2658                                 kiblnd_reconnect(conn, rej->ibr_version,
2659                                                  incarnation, rej->ibr_why, cp);
2660                                 break;
2661
2662                         case IBLND_REJECT_NO_RESOURCES:
2663                                 CERROR("%s rejected: o2iblnd no resources\n",
2664                                        libcfs_nid2str(peer->ibp_nid));
2665                                 break;
2666
2667                         case IBLND_REJECT_FATAL:
2668                                 CERROR("%s rejected: o2iblnd fatal error\n",
2669                                        libcfs_nid2str(peer->ibp_nid));
2670                                 break;
2671
2672                         default:
2673                                 CERROR("%s rejected: o2iblnd reason %d\n",
2674                                        libcfs_nid2str(peer->ibp_nid),
2675                                        rej->ibr_why);
2676                                 break;
2677                         }
2678                         break;
2679                 }
2680                 /* fall through */
2681         default:
2682                 CNETERR("%s rejected: reason %d, size %d\n",
2683                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2684                 break;
2685         }
2686
2687         kiblnd_connreq_done(conn, -ECONNREFUSED);
2688 }
2689
2690 static void
2691 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2692 {
2693         kib_peer_t    *peer = conn->ibc_peer;
2694         lnet_ni_t     *ni   = peer->ibp_ni;
2695         kib_net_t     *net  = ni->ni_data;
2696         kib_msg_t     *msg  = priv;
2697         int            ver  = conn->ibc_version;
2698         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2699         unsigned long  flags;
2700
2701         LASSERT (net != NULL);
2702
2703         if (rc != 0) {
2704                 CERROR("Can't unpack connack from %s: %d\n",
2705                        libcfs_nid2str(peer->ibp_nid), rc);
2706                 goto failed;
2707         }
2708
2709         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2710                 CERROR("Unexpected message %d from %s\n",
2711                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2712                 rc = -EPROTO;
2713                 goto failed;
2714         }
2715
2716         if (ver != msg->ibm_version) {
2717                 CERROR("%s replied version %x is different with "
2718                        "requested version %x\n",
2719                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2720                 rc = -EPROTO;
2721                 goto failed;
2722         }
2723
2724         if (msg->ibm_u.connparams.ibcp_queue_depth >
2725             conn->ibc_queue_depth) {
2726                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2727                        libcfs_nid2str(peer->ibp_nid),
2728                        msg->ibm_u.connparams.ibcp_queue_depth,
2729                        conn->ibc_queue_depth);
2730                 rc = -EPROTO;
2731                 goto failed;
2732         }
2733
2734         if (msg->ibm_u.connparams.ibcp_max_frags >
2735             conn->ibc_max_frags) {
2736                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2737                        libcfs_nid2str(peer->ibp_nid),
2738                        msg->ibm_u.connparams.ibcp_max_frags,
2739                        conn->ibc_max_frags);
2740                 rc = -EPROTO;
2741                 goto failed;
2742         }
2743
2744         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2745                 CERROR("%s max message size %d too big (%d max)\n",
2746                        libcfs_nid2str(peer->ibp_nid),
2747                        msg->ibm_u.connparams.ibcp_max_msg_size,
2748                        IBLND_MSG_SIZE);
2749                 rc = -EPROTO;
2750                 goto failed;
2751         }
2752
2753         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2754         if (msg->ibm_dstnid == ni->ni_nid &&
2755             msg->ibm_dststamp == net->ibn_incarnation)
2756                 rc = 0;
2757         else
2758                 rc = -ESTALE;
2759         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2760
2761         if (rc != 0) {
2762                 CERROR("Bad connection reply from %s, rc = %d, "
2763                        "version: %x max_frags: %d\n",
2764                        libcfs_nid2str(peer->ibp_nid), rc,
2765                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2766                 goto failed;
2767         }
2768
2769         conn->ibc_incarnation      = msg->ibm_srcstamp;
2770         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
2771         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2772         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
2773         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
2774         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2775                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2776
2777         kiblnd_connreq_done(conn, 0);
2778         return;
2779
2780  failed:
2781         /* NB My QP has already established itself, so I handle anything going
2782          * wrong here by setting ibc_comms_error.
2783          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2784          * immediately tears it down. */
2785
2786         LASSERT (rc != 0);
2787         conn->ibc_comms_error = rc;
2788         kiblnd_connreq_done(conn, 0);
2789 }
2790
2791 static int
2792 kiblnd_active_connect (struct rdma_cm_id *cmid)
2793 {
2794         kib_peer_t              *peer = (kib_peer_t *)cmid->context;
2795         kib_conn_t              *conn;
2796         kib_msg_t               *msg;
2797         struct rdma_conn_param   cp;
2798         int                      version;
2799         __u64                    incarnation;
2800         unsigned long            flags;
2801         int                      rc;
2802
2803         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2804
2805         incarnation = peer->ibp_incarnation;
2806         version     = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2807                                                  peer->ibp_version;
2808
2809         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2810
2811         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2812                                   version);
2813         if (conn == NULL) {
2814                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2815                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2816                 return -ENOMEM;
2817         }
2818
2819         /* conn "owns" cmid now, so I return success from here on to ensure the
2820          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2821          * on peer */
2822
2823         msg = &conn->ibc_connvars->cv_msg;
2824
2825         memset(msg, 0, sizeof(*msg));
2826         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2827         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2828         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2829         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2830
2831         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2832                         0, peer->ibp_nid, incarnation);
2833
2834         memset(&cp, 0, sizeof(cp));
2835         cp.private_data        = msg;
2836         cp.private_data_len    = msg->ibm_nob;
2837         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2838         cp.initiator_depth     = 0;
2839         cp.flow_control        = 1;
2840         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2841         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2842
2843         LASSERT(cmid->context == (void *)conn);
2844         LASSERT(conn->ibc_cmid == cmid);
2845
2846         rc = rdma_connect(cmid, &cp);
2847         if (rc != 0) {
2848                 CERROR("Can't connect to %s: %d\n",
2849                        libcfs_nid2str(peer->ibp_nid), rc);
2850                 kiblnd_connreq_done(conn, rc);
2851                 kiblnd_conn_decref(conn);
2852         }
2853
2854         return 0;
2855 }
2856
2857 int
2858 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2859 {
2860         kib_peer_t  *peer;
2861         kib_conn_t  *conn;
2862         int          rc;
2863
2864         switch (event->event) {
2865         default:
2866                 CERROR("Unexpected event: %d, status: %d\n",
2867                        event->event, event->status);
2868                 LBUG();
2869
2870         case RDMA_CM_EVENT_CONNECT_REQUEST:
2871                 /* destroy cmid on failure */
2872                 rc = kiblnd_passive_connect(cmid,
2873                                             (void *)KIBLND_CONN_PARAM(event),
2874                                             KIBLND_CONN_PARAM_LEN(event));
2875                 CDEBUG(D_NET, "connreq: %d\n", rc);
2876                 return rc;
2877
2878         case RDMA_CM_EVENT_ADDR_ERROR:
2879                 peer = (kib_peer_t *)cmid->context;
2880                 CNETERR("%s: ADDR ERROR %d\n",
2881                        libcfs_nid2str(peer->ibp_nid), event->status);
2882                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2883                 kiblnd_peer_decref(peer);
2884                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
2885
2886         case RDMA_CM_EVENT_ADDR_RESOLVED:
2887                 peer = (kib_peer_t *)cmid->context;
2888
2889                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2890                        libcfs_nid2str(peer->ibp_nid), event->status);
2891
2892                 if (event->status != 0) {
2893                         CNETERR("Can't resolve address for %s: %d\n",
2894                                 libcfs_nid2str(peer->ibp_nid), event->status);
2895                         rc = event->status;
2896                 } else {
2897                         rc = rdma_resolve_route(
2898                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2899                         if (rc == 0)
2900                                 return 0;
2901                         /* Can't initiate route resolution */
2902                         CERROR("Can't resolve route for %s: %d\n",
2903                                libcfs_nid2str(peer->ibp_nid), rc);
2904                 }
2905                 kiblnd_peer_connect_failed(peer, 1, rc);
2906                 kiblnd_peer_decref(peer);
2907                 return rc;                      /* rc != 0 destroys cmid */
2908
2909         case RDMA_CM_EVENT_ROUTE_ERROR:
2910                 peer = (kib_peer_t *)cmid->context;
2911                 CNETERR("%s: ROUTE ERROR %d\n",
2912                         libcfs_nid2str(peer->ibp_nid), event->status);
2913                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2914                 kiblnd_peer_decref(peer);
2915                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
2916
2917         case RDMA_CM_EVENT_ROUTE_RESOLVED:
2918                 peer = (kib_peer_t *)cmid->context;
2919                 CDEBUG(D_NET,"%s Route resolved: %d\n",
2920                        libcfs_nid2str(peer->ibp_nid), event->status);
2921
2922                 if (event->status == 0)
2923                         return kiblnd_active_connect(cmid);
2924
2925                 CNETERR("Can't resolve route for %s: %d\n",
2926                        libcfs_nid2str(peer->ibp_nid), event->status);
2927                 kiblnd_peer_connect_failed(peer, 1, event->status);
2928                 kiblnd_peer_decref(peer);
2929                 return event->status;           /* rc != 0 destroys cmid */
2930
2931         case RDMA_CM_EVENT_UNREACHABLE:
2932                 conn = (kib_conn_t *)cmid->context;
2933                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2934                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2935                 CNETERR("%s: UNREACHABLE %d\n",
2936                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2937                 kiblnd_connreq_done(conn, -ENETDOWN);
2938                 kiblnd_conn_decref(conn);
2939                 return 0;
2940
2941         case RDMA_CM_EVENT_CONNECT_ERROR:
2942                 conn = (kib_conn_t *)cmid->context;
2943                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2944                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2945                 CNETERR("%s: CONNECT ERROR %d\n",
2946                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2947                 kiblnd_connreq_done(conn, -ENOTCONN);
2948                 kiblnd_conn_decref(conn);
2949                 return 0;
2950
2951         case RDMA_CM_EVENT_REJECTED:
2952                 conn = (kib_conn_t *)cmid->context;
2953                 switch (conn->ibc_state) {
2954                 default:
2955                         LBUG();
2956
2957                 case IBLND_CONN_PASSIVE_WAIT:
2958                         CERROR ("%s: REJECTED %d\n",
2959                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2960                                 event->status);
2961                         kiblnd_connreq_done(conn, -ECONNRESET);
2962                         break;
2963
2964                 case IBLND_CONN_ACTIVE_CONNECT:
2965                         kiblnd_rejected(conn, event->status,
2966                                         (void *)KIBLND_CONN_PARAM(event),
2967                                         KIBLND_CONN_PARAM_LEN(event));
2968                         break;
2969                 }
2970                 kiblnd_conn_decref(conn);
2971                 return 0;
2972
2973         case RDMA_CM_EVENT_ESTABLISHED:
2974                 conn = (kib_conn_t *)cmid->context;
2975                 switch (conn->ibc_state) {
2976                 default:
2977                         LBUG();
2978
2979                 case IBLND_CONN_PASSIVE_WAIT:
2980                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2981                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2982                         kiblnd_connreq_done(conn, 0);
2983                         break;
2984
2985                 case IBLND_CONN_ACTIVE_CONNECT:
2986                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2987                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2988                         kiblnd_check_connreply(conn,
2989                                                (void *)KIBLND_CONN_PARAM(event),
2990                                                KIBLND_CONN_PARAM_LEN(event));
2991                         break;
2992                 }
2993                 /* net keeps its ref on conn! */
2994                 return 0;
2995
2996         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2997                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2998                 return 0;
2999
3000         case RDMA_CM_EVENT_DISCONNECTED:
3001                 conn = (kib_conn_t *)cmid->context;
3002                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3003                         CERROR("%s DISCONNECTED\n",
3004                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3005                         kiblnd_connreq_done(conn, -ECONNRESET);
3006                 } else {
3007                         kiblnd_close_conn(conn, 0);
3008                 }
3009                 kiblnd_conn_decref(conn);
3010                 cmid->context = NULL;
3011                 return 0;
3012
3013         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3014                 LCONSOLE_ERROR_MSG(0x131,
3015                                    "Received notification of device removal\n"
3016                                    "Please shutdown LNET to allow this to proceed\n");
3017                 /* Can't remove network from underneath LNET for now, so I have
3018                  * to ignore this */
3019                 return 0;
3020
3021         case RDMA_CM_EVENT_ADDR_CHANGE:
3022                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3023                 return 0;
3024         }
3025 }
3026
3027 static int
3028 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3029 {
3030         kib_tx_t         *tx;
3031         struct list_head *ttmp;
3032
3033         list_for_each(ttmp, txs) {
3034                 tx = list_entry(ttmp, kib_tx_t, tx_list);
3035
3036                 if (txs != &conn->ibc_active_txs) {
3037                         LASSERT(tx->tx_queued);
3038                 } else {
3039                         LASSERT(!tx->tx_queued);
3040                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3041                 }
3042
3043                 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3044                         CERROR("Timed out tx: %s, %lu seconds\n",
3045                                kiblnd_queue2str(conn, txs),
3046                                cfs_duration_sec(jiffies - tx->tx_deadline));
3047                         return 1;
3048                 }
3049         }
3050
3051         return 0;
3052 }
3053
3054 static int
3055 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3056 {
3057         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3058                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3059                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3060                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3061                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3062 }
3063
3064 static void
3065 kiblnd_check_conns (int idx)
3066 {
3067         struct list_head  closes = LIST_HEAD_INIT(closes);
3068         struct list_head  checksends = LIST_HEAD_INIT(checksends);
3069         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3070         struct list_head *ptmp;
3071         kib_peer_t       *peer;
3072         kib_conn_t       *conn;
3073         struct list_head *ctmp;
3074         unsigned long     flags;
3075
3076         /* NB. We expect to have a look at all the peers and not find any
3077          * RDMAs to time out, so we just use a shared lock while we
3078          * take a look... */
3079         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3080
3081         list_for_each(ptmp, peers) {
3082                 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3083
3084                 list_for_each(ctmp, &peer->ibp_conns) {
3085                         int timedout;
3086                         int sendnoop;
3087
3088                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
3089
3090                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3091
3092                         spin_lock(&conn->ibc_lock);
3093
3094                         sendnoop = kiblnd_need_noop(conn);
3095                         timedout = kiblnd_conn_timed_out_locked(conn);
3096                         if (!sendnoop && !timedout) {
3097                                 spin_unlock(&conn->ibc_lock);
3098                                 continue;
3099                         }
3100
3101                         if (timedout) {
3102                                 CERROR("Timed out RDMA with %s (%lu): "
3103                                        "c: %u, oc: %u, rc: %u\n",
3104                                        libcfs_nid2str(peer->ibp_nid),
3105                                        cfs_duration_sec(cfs_time_current() -
3106                                                         peer->ibp_last_alive),
3107                                        conn->ibc_credits,
3108                                        conn->ibc_outstanding_credits,
3109                                        conn->ibc_reserved_credits);
3110                                 list_add(&conn->ibc_connd_list, &closes);
3111                         } else {
3112                                 list_add(&conn->ibc_connd_list, &checksends);
3113                         }
3114                         /* +ref for 'closes' or 'checksends' */
3115                         kiblnd_conn_addref(conn);
3116
3117                         spin_unlock(&conn->ibc_lock);
3118                 }
3119         }
3120
3121         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3122
3123         /* Handle timeout by closing the whole
3124          * connection. We can only be sure RDMA activity
3125          * has ceased once the QP has been modified. */
3126         while (!list_empty(&closes)) {
3127                 conn = list_entry(closes.next,
3128                                   kib_conn_t, ibc_connd_list);
3129                 list_del(&conn->ibc_connd_list);
3130                 kiblnd_close_conn(conn, -ETIMEDOUT);
3131                 kiblnd_conn_decref(conn);
3132         }
3133
3134         /* In case we have enough credits to return via a
3135          * NOOP, but there were no non-blocking tx descs
3136          * free to do it last time... */
3137         while (!list_empty(&checksends)) {
3138                 conn = list_entry(checksends.next,
3139                                   kib_conn_t, ibc_connd_list);
3140                 list_del(&conn->ibc_connd_list);
3141                 kiblnd_check_sends(conn);
3142                 kiblnd_conn_decref(conn);
3143         }
3144 }
3145
3146 static void
3147 kiblnd_disconnect_conn (kib_conn_t *conn)
3148 {
3149         LASSERT (!in_interrupt());
3150         LASSERT (current == kiblnd_data.kib_connd);
3151         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3152
3153         rdma_disconnect(conn->ibc_cmid);
3154         kiblnd_finalise_conn(conn);
3155
3156         kiblnd_peer_notify(conn->ibc_peer);
3157 }
3158
3159 int
3160 kiblnd_connd (void *arg)
3161 {
3162         wait_queue_t     wait;
3163         unsigned long      flags;
3164         kib_conn_t        *conn;
3165         int                timeout;
3166         int                i;
3167         int                dropped_lock;
3168         int                peer_index = 0;
3169         unsigned long      deadline = jiffies;
3170
3171         cfs_block_allsigs();
3172
3173         init_waitqueue_entry(&wait, current);
3174         kiblnd_data.kib_connd = current;
3175
3176         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3177
3178         while (!kiblnd_data.kib_shutdown) {
3179
3180                 dropped_lock = 0;
3181
3182                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3183                         conn = list_entry(kiblnd_data. \
3184                                               kib_connd_zombies.next,
3185                                               kib_conn_t, ibc_list);
3186                         list_del(&conn->ibc_list);
3187
3188                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3189                                                flags);
3190                         dropped_lock = 1;
3191
3192                         kiblnd_destroy_conn(conn);
3193
3194                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3195                 }
3196
3197                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3198                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3199                                               kib_conn_t, ibc_list);
3200                         list_del(&conn->ibc_list);
3201
3202                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3203                                                flags);
3204                         dropped_lock = 1;
3205
3206                         kiblnd_disconnect_conn(conn);
3207                         kiblnd_conn_decref(conn);
3208
3209                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3210                 }
3211
3212                 /* careful with the jiffy wrap... */
3213                 timeout = (int)(deadline - jiffies);
3214                 if (timeout <= 0) {
3215                         const int n = 4;
3216                         const int p = 1;
3217                         int       chunk = kiblnd_data.kib_peer_hash_size;
3218
3219                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3220                         dropped_lock = 1;
3221
3222                         /* Time to check for RDMA timeouts on a few more
3223                          * peers: I do checks every 'p' seconds on a
3224                          * proportion of the peer table and I need to check
3225                          * every connection 'n' times within a timeout
3226                          * interval, to ensure I detect a timeout on any
3227                          * connection within (n+1)/n times the timeout
3228                          * interval. */
3229
3230                         if (*kiblnd_tunables.kib_timeout > n * p)
3231                                 chunk = (chunk * n * p) /
3232                                         *kiblnd_tunables.kib_timeout;
3233                         if (chunk == 0)
3234                                 chunk = 1;
3235
3236                         for (i = 0; i < chunk; i++) {
3237                                 kiblnd_check_conns(peer_index);
3238                                 peer_index = (peer_index + 1) %
3239                                              kiblnd_data.kib_peer_hash_size;
3240                         }
3241
3242                         deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3243                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3244                 }
3245
3246                 if (dropped_lock)
3247                         continue;
3248
3249                 /* Nothing to do for 'timeout'  */
3250                 set_current_state(TASK_INTERRUPTIBLE);
3251                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3252                 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3253
3254                 schedule_timeout(timeout);
3255
3256                 set_current_state(TASK_RUNNING);
3257                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3258                 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3259         }
3260
3261         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3262
3263         kiblnd_thread_fini();
3264         return 0;
3265 }
3266
3267 void
3268 kiblnd_qp_event(struct ib_event *event, void *arg)
3269 {
3270         kib_conn_t *conn = arg;
3271
3272         switch (event->event) {
3273         case IB_EVENT_COMM_EST:
3274                 CDEBUG(D_NET, "%s established\n",
3275                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3276                 return;
3277
3278         default:
3279                 CERROR("%s: Async QP event type %d\n",
3280                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3281                 return;
3282         }
3283 }
3284
3285 static void
3286 kiblnd_complete (struct ib_wc *wc)
3287 {
3288         switch (kiblnd_wreqid2type(wc->wr_id)) {
3289         default:
3290                 LBUG();
3291
3292         case IBLND_WID_RDMA:
3293                 /* We only get RDMA completion notification if it fails.  All
3294                  * subsequent work items, including the final SEND will fail
3295                  * too.  However we can't print out any more info about the
3296                  * failing RDMA because 'tx' might be back on the idle list or
3297                  * even reused already if we didn't manage to post all our work
3298                  * items */
3299                 CNETERR("RDMA (tx: %p) failed: %d\n",
3300                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3301                 return;
3302
3303         case IBLND_WID_TX:
3304                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3305                 return;
3306
3307         case IBLND_WID_RX:
3308                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3309                                    wc->byte_len);
3310                 return;
3311         }
3312 }
3313
3314 void
3315 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3316 {
3317         /* NB I'm not allowed to schedule this conn once its refcount has
3318          * reached 0.  Since fundamentally I'm racing with scheduler threads
3319          * consuming my CQ I could be called after all completions have
3320          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3321          * and this CQ is about to be destroyed so I NOOP. */
3322         kib_conn_t              *conn = (kib_conn_t *)arg;
3323         struct kib_sched_info   *sched = conn->ibc_sched;
3324         unsigned long           flags;
3325
3326         LASSERT(cq == conn->ibc_cq);
3327
3328         spin_lock_irqsave(&sched->ibs_lock, flags);
3329
3330         conn->ibc_ready = 1;
3331
3332         if (!conn->ibc_scheduled &&
3333             (conn->ibc_nrx > 0 ||
3334              conn->ibc_nsends_posted > 0)) {
3335                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3336                 conn->ibc_scheduled = 1;
3337                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3338
3339                 if (waitqueue_active(&sched->ibs_waitq))
3340                         wake_up(&sched->ibs_waitq);
3341         }
3342
3343         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3344 }
3345
3346 void
3347 kiblnd_cq_event(struct ib_event *event, void *arg)
3348 {
3349         kib_conn_t *conn = arg;
3350
3351         CERROR("%s: async CQ event type %d\n",
3352                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3353 }
3354
3355 int
3356 kiblnd_scheduler(void *arg)
3357 {
3358         long                    id = (long)arg;
3359         struct kib_sched_info   *sched;
3360         kib_conn_t              *conn;
3361         wait_queue_t            wait;
3362         unsigned long           flags;
3363         struct ib_wc            wc;
3364         int                     did_something;
3365         int                     busy_loops = 0;
3366         int                     rc;
3367
3368         cfs_block_allsigs();
3369
3370         init_waitqueue_entry(&wait, current);
3371
3372         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3373
3374         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3375         if (rc != 0) {
3376                 CWARN("Failed to bind on CPT %d, please verify whether "
3377                       "all CPUs are healthy and reload modules if necessary, "
3378                       "otherwise your system might under risk of low "
3379                       "performance\n", sched->ibs_cpt);
3380         }
3381
3382         spin_lock_irqsave(&sched->ibs_lock, flags);
3383
3384         while (!kiblnd_data.kib_shutdown) {
3385                 if (busy_loops++ >= IBLND_RESCHED) {
3386                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3387
3388                         cond_resched();
3389                         busy_loops = 0;
3390
3391                         spin_lock_irqsave(&sched->ibs_lock, flags);
3392                 }
3393
3394                 did_something = 0;
3395
3396                 if (!list_empty(&sched->ibs_conns)) {
3397                         conn = list_entry(sched->ibs_conns.next,
3398                                               kib_conn_t, ibc_sched_list);
3399                         /* take over kib_sched_conns' ref on conn... */
3400                         LASSERT(conn->ibc_scheduled);
3401                         list_del(&conn->ibc_sched_list);
3402                         conn->ibc_ready = 0;
3403
3404                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3405
3406                         wc.wr_id = IBLND_WID_INVAL;
3407
3408                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3409                         if (rc == 0) {
3410                                 rc = ib_req_notify_cq(conn->ibc_cq,
3411                                                       IB_CQ_NEXT_COMP);
3412                                 if (rc < 0) {
3413                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3414                                               "closing connection\n",
3415                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3416                                         kiblnd_close_conn(conn, -EIO);
3417                                         kiblnd_conn_decref(conn);
3418                                         spin_lock_irqsave(&sched->ibs_lock,
3419                                                               flags);
3420                                         continue;
3421                                 }
3422
3423                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3424                         }
3425
3426                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3427                                 LCONSOLE_ERROR(
3428                                         "ib_poll_cq (rc: %d) returned invalid "
3429                                         "wr_id, opcode %d, status: %d, "
3430                                         "vendor_err: %d, conn: %s status: %d\n"
3431                                         "please upgrade firmware and OFED or "
3432                                         "contact vendor.\n", rc,
3433                                         wc.opcode, wc.status, wc.vendor_err,
3434                                         libcfs_nid2str(conn->ibc_peer->ibp_nid),
3435                                         conn->ibc_state);
3436                                 rc = -EINVAL;
3437                         }
3438
3439                         if (rc < 0) {
3440                                 CWARN("%s: ib_poll_cq failed: %d, "
3441                                       "closing connection\n",
3442                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3443                                       rc);
3444                                 kiblnd_close_conn(conn, -EIO);
3445                                 kiblnd_conn_decref(conn);
3446                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3447                                 continue;
3448                         }
3449
3450                         spin_lock_irqsave(&sched->ibs_lock, flags);
3451
3452                         if (rc != 0 || conn->ibc_ready) {
3453                                 /* There may be another completion waiting; get
3454                                  * another scheduler to check while I handle
3455                                  * this one... */
3456                                 /* +1 ref for sched_conns */
3457                                 kiblnd_conn_addref(conn);
3458                                 list_add_tail(&conn->ibc_sched_list,
3459                                                   &sched->ibs_conns);
3460                                 if (waitqueue_active(&sched->ibs_waitq))
3461                                         wake_up(&sched->ibs_waitq);
3462                         } else {
3463                                 conn->ibc_scheduled = 0;
3464                         }
3465
3466                         if (rc != 0) {
3467                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3468                                 kiblnd_complete(&wc);
3469
3470                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3471                         }
3472
3473                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3474                         did_something = 1;
3475                 }
3476
3477                 if (did_something)
3478                         continue;
3479
3480                 set_current_state(TASK_INTERRUPTIBLE);
3481                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3482                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3483
3484                 schedule();
3485                 busy_loops = 0;
3486
3487                 remove_wait_queue(&sched->ibs_waitq, &wait);
3488                 set_current_state(TASK_RUNNING);
3489                 spin_lock_irqsave(&sched->ibs_lock, flags);
3490         }
3491
3492         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3493
3494         kiblnd_thread_fini();
3495         return 0;
3496 }
3497
3498 int
3499 kiblnd_failover_thread(void *arg)
3500 {
3501         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
3502         kib_dev_t               *dev;
3503         wait_queue_t            wait;
3504         unsigned long           flags;
3505         int                     rc;
3506
3507         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3508
3509         cfs_block_allsigs();
3510
3511         init_waitqueue_entry(&wait, current);
3512         write_lock_irqsave(glock, flags);
3513
3514         while (!kiblnd_data.kib_shutdown) {
3515                 int     do_failover = 0;
3516                 int     long_sleep;
3517
3518                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3519                                     ibd_fail_list) {
3520                         if (cfs_time_before(cfs_time_current(),
3521                                             dev->ibd_next_failover))
3522                                 continue;
3523                         do_failover = 1;
3524                         break;
3525                 }
3526
3527                 if (do_failover) {
3528                         list_del_init(&dev->ibd_fail_list);
3529                         dev->ibd_failover = 1;
3530                         write_unlock_irqrestore(glock, flags);
3531
3532                         rc = kiblnd_dev_failover(dev);
3533
3534                         write_lock_irqsave(glock, flags);
3535
3536                         LASSERT (dev->ibd_failover);
3537                         dev->ibd_failover = 0;
3538                         if (rc >= 0) { /* Device is OK or failover succeed */
3539                                 dev->ibd_next_failover = cfs_time_shift(3);
3540                                 continue;
3541                         }
3542
3543                         /* failed to failover, retry later */
3544                         dev->ibd_next_failover =
3545                                 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3546                         if (kiblnd_dev_can_failover(dev)) {
3547                                 list_add_tail(&dev->ibd_fail_list,
3548                                               &kiblnd_data.kib_failed_devs);
3549                         }
3550
3551                         continue;
3552                 }
3553
3554                 /* long sleep if no more pending failover */
3555                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3556
3557                 set_current_state(TASK_INTERRUPTIBLE);
3558                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3559                 write_unlock_irqrestore(glock, flags);
3560
3561                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3562                                                    cfs_time_seconds(1));
3563                 set_current_state(TASK_RUNNING);
3564                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3565                 write_lock_irqsave(glock, flags);
3566
3567                 if (!long_sleep || rc != 0)
3568                         continue;
3569
3570                 /* have a long sleep, routine check all active devices,
3571                  * we need checking like this because if there is not active
3572                  * connection on the dev and no SEND from local, we may listen
3573                  * on wrong HCA for ever while there is a bonding failover */
3574                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3575                         if (kiblnd_dev_can_failover(dev)) {
3576                                 list_add_tail(&dev->ibd_fail_list,
3577                                               &kiblnd_data.kib_failed_devs);
3578                         }
3579                 }
3580         }
3581
3582         write_unlock_irqrestore(glock, flags);
3583
3584         kiblnd_thread_fini();
3585         return 0;
3586 }