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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/o2iblnd/o2iblnd_cb.c
  33  *
  34  * Author: Eric Barton <eric@bartonsoftware.com>
  35  */
  36
  37 #include "o2iblnd.h"
  38
  39 #define MAX_CONN_RACES_BEFORE_ABORT 20
  40
  41 static void kiblnd_peer_alive(struct kib_peer_ni *peer_ni);
  42 static void kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
  43                                        int error);
  44 static void kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx,
  45                                int type, int body_nob);
  46 static int kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
  47                             int resid, struct kib_rdma_desc *dstrd, u64 dstcookie);
  48 static void kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn);
  49 static void kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn);
  50
  51 static void kiblnd_unmap_tx(struct kib_tx *tx);
  52 static void kiblnd_check_sends_locked(struct kib_conn *conn);
  53
  54 void
  55 kiblnd_tx_done(struct kib_tx *tx)
  56 {
  57         struct lnet_msg *lntmsg[2];
  58         int         rc;
  59         int         i;
  60
  61         LASSERT (!in_interrupt());
  62         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
  63         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
  64         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer_ni response */
  65         LASSERT (tx->tx_pool != NULL);
  66
  67         kiblnd_unmap_tx(tx);
  68
  69         /* tx may have up to 2 lnet msgs to finalise */
  70         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  71         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  72         rc = tx->tx_status;
  73
  74         if (tx->tx_conn != NULL) {
  75                 kiblnd_conn_decref(tx->tx_conn);
  76                 tx->tx_conn = NULL;
  77         }
  78
  79         tx->tx_nwrq = tx->tx_nsge = 0;
  80         tx->tx_status = 0;
  81
  82         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  83
  84         /* delay finalize until my descs have been freed */
  85         for (i = 0; i < 2; i++) {
  86                 if (lntmsg[i] == NULL)
  87                         continue;
  88
  89                 /* propagate health status to LNet for requests */
  90                 if (i == 0 && lntmsg[i])
  91                         lntmsg[i]->msg_health_status = tx->tx_hstatus;
  92
  93                 lnet_finalize(lntmsg[i], rc);
  94         }
  95 }
  96
  97 void
  98 kiblnd_txlist_done(struct list_head *txlist, int status,
  99                    enum lnet_msg_hstatus hstatus)
 100 {
 101         struct kib_tx *tx;
 102
 103         while (!list_empty(txlist)) {
 104                 tx = list_entry(txlist->next, struct kib_tx, tx_list);
 105
 106                 list_del(&tx->tx_list);
 107                 /* complete now */
 108                 tx->tx_waiting = 0;
 109                 tx->tx_status = status;
 110                 if (hstatus != LNET_MSG_STATUS_OK)
 111                         tx->tx_hstatus = hstatus;
 112                 kiblnd_tx_done(tx);
 113         }
 114 }
 115
 116 static struct kib_tx *
 117 kiblnd_get_idle_tx(struct lnet_ni *ni, lnet_nid_t target)
 118 {
 119         struct kib_net *net = ni->ni_data;
 120         struct list_head *node;
 121         struct kib_tx *tx;
 122         struct kib_tx_poolset *tps;
 123
 124         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
 125         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
 126         if (node == NULL)
 127                 return NULL;
 128         tx = container_of(node, struct kib_tx, tx_list);
 129
 130         LASSERT (tx->tx_nwrq == 0);
 131         LASSERT (!tx->tx_queued);
 132         LASSERT (tx->tx_sending == 0);
 133         LASSERT (!tx->tx_waiting);
 134         LASSERT (tx->tx_status == 0);
 135         LASSERT (tx->tx_conn == NULL);
 136         LASSERT (tx->tx_lntmsg[0] == NULL);
 137         LASSERT (tx->tx_lntmsg[1] == NULL);
 138         LASSERT (tx->tx_nfrags == 0);
 139
 140         tx->tx_gaps = false;
 141         tx->tx_hstatus = LNET_MSG_STATUS_OK;
 142
 143         return tx;
 144 }
 145
 146 static void
 147 kiblnd_drop_rx(struct kib_rx *rx)
 148 {
 149         struct kib_conn *conn = rx->rx_conn;
 150         struct kib_sched_info *sched = conn->ibc_sched;
 151         unsigned long flags;
 152
 153         spin_lock_irqsave(&sched->ibs_lock, flags);
 154         LASSERT(conn->ibc_nrx > 0);
 155         conn->ibc_nrx--;
 156         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 157
 158         kiblnd_conn_decref(conn);
 159 }
 160
 161 int
 162 kiblnd_post_rx(struct kib_rx *rx, int credit)
 163 {
 164         struct kib_conn *conn = rx->rx_conn;
 165         struct kib_net *net = conn->ibc_peer->ibp_ni->ni_data;
 166         struct ib_recv_wr *bad_wrq = NULL;
 167 #ifdef HAVE_IB_GET_DMA_MR
 168         struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
 169 #endif
 170         int rc;
 171
 172         LASSERT (net != NULL);
 173         LASSERT (!in_interrupt());
 174         LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
 175                  credit == IBLND_POSTRX_PEER_CREDIT ||
 176                  credit == IBLND_POSTRX_RSRVD_CREDIT);
 177 #ifdef HAVE_IB_GET_DMA_MR
 178         LASSERT(mr != NULL);
 179
 180         rx->rx_sge.lkey   = mr->lkey;
 181 #else
 182         rx->rx_sge.lkey   = conn->ibc_hdev->ibh_pd->local_dma_lkey;
 183 #endif
 184         rx->rx_sge.addr   = rx->rx_msgaddr;
 185         rx->rx_sge.length = IBLND_MSG_SIZE;
 186
 187         rx->rx_wrq.next = NULL;
 188         rx->rx_wrq.sg_list = &rx->rx_sge;
 189         rx->rx_wrq.num_sge = 1;
 190         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 191
 192         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
 193         LASSERT (rx->rx_nob >= 0);              /* not posted */
 194
 195         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 196                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 197                 return 0;
 198         }
 199
 200         rx->rx_nob = -1;                        /* flag posted */
 201
 202         /* NB: need an extra reference after ib_post_recv because we don't
 203          * own this rx (and rx::rx_conn) anymore, LU-5678.
 204          */
 205         kiblnd_conn_addref(conn);
 206 #ifdef HAVE_IB_POST_SEND_RECV_CONST
 207         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq,
 208                           (const struct ib_recv_wr **)&bad_wrq);
 209 #else
 210         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 211 #endif
 212         if (unlikely(rc != 0)) {
 213                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 214                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 215                 rx->rx_nob = 0;
 216         }
 217
 218         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 219                 goto out;
 220
 221         if (unlikely(rc != 0)) {
 222                 kiblnd_close_conn(conn, rc);
 223                 kiblnd_drop_rx(rx);     /* No more posts for this rx */
 224                 goto out;
 225         }
 226
 227         if (credit == IBLND_POSTRX_NO_CREDIT)
 228                 goto out;
 229
 230         spin_lock(&conn->ibc_lock);
 231         if (credit == IBLND_POSTRX_PEER_CREDIT)
 232                 conn->ibc_outstanding_credits++;
 233         else
 234                 conn->ibc_reserved_credits++;
 235         kiblnd_check_sends_locked(conn);
 236         spin_unlock(&conn->ibc_lock);
 237
 238 out:
 239         kiblnd_conn_decref(conn);
 240         return rc;
 241 }
 242
 243 static struct kib_tx *
 244 kiblnd_find_waiting_tx_locked(struct kib_conn *conn, int txtype, u64 cookie)
 245 {
 246         struct list_head *tmp;
 247
 248         list_for_each(tmp, &conn->ibc_active_txs) {
 249                 struct kib_tx *tx = list_entry(tmp, struct kib_tx, tx_list);
 250
 251                 LASSERT(!tx->tx_queued);
 252                 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
 253
 254                 if (tx->tx_cookie != cookie)
 255                         continue;
 256
 257                 if (tx->tx_waiting &&
 258                     tx->tx_msg->ibm_type == txtype)
 259                         return tx;
 260
 261                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 262                       tx->tx_waiting ? "" : "NOT ",
 263                       tx->tx_msg->ibm_type, txtype);
 264         }
 265         return NULL;
 266 }
 267
 268 static void
 269 kiblnd_handle_completion(struct kib_conn *conn, int txtype, int status, u64 cookie)
 270 {
 271         struct kib_tx *tx;
 272         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
 273         int idle;
 274
 275         spin_lock(&conn->ibc_lock);
 276
 277         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 278         if (tx == NULL) {
 279                 spin_unlock(&conn->ibc_lock);
 280
 281                 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
 282                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 283                 kiblnd_close_conn(conn, -EPROTO);
 284                 return;
 285         }
 286
 287         if (tx->tx_status == 0) {               /* success so far */
 288                 if (status < 0) {               /* failed? */
 289                         tx->tx_status = status;
 290                         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 291                 } else if (txtype == IBLND_MSG_GET_REQ) {
 292                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 293                 }
 294         }
 295
 296         tx->tx_waiting = 0;
 297
 298         idle = !tx->tx_queued && (tx->tx_sending == 0);
 299         if (idle)
 300                 list_del(&tx->tx_list);
 301
 302         spin_unlock(&conn->ibc_lock);
 303
 304         if (idle)
 305                 kiblnd_tx_done(tx);
 306 }
 307
 308 static void
 309 kiblnd_send_completion(struct kib_conn *conn, int type, int status, u64 cookie)
 310 {
 311         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
 312         struct kib_tx *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 313
 314         if (tx == NULL) {
 315                 CERROR("Can't get tx for completion %x for %s\n",
 316                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 317                 return;
 318         }
 319
 320         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 321         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 322         kiblnd_init_tx_msg(ni, tx, type, sizeof(struct kib_completion_msg));
 323
 324         kiblnd_queue_tx(tx, conn);
 325 }
 326
 327 static void
 328 kiblnd_handle_rx(struct kib_rx *rx)
 329 {
 330         struct kib_msg *msg = rx->rx_msg;
 331         struct kib_conn   *conn = rx->rx_conn;
 332         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
 333         int           credits = msg->ibm_credits;
 334         struct kib_tx *tx;
 335         int           rc = 0;
 336         int           rc2;
 337         int           post_credit;
 338
 339         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 340
 341         CDEBUG (D_NET, "Received %x[%d] from %s\n",
 342                 msg->ibm_type, credits,
 343                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
 344
 345         if (credits != 0) {
 346                 /* Have I received credits that will let me send? */
 347                 spin_lock(&conn->ibc_lock);
 348
 349                 if (conn->ibc_credits + credits >
 350                     conn->ibc_queue_depth) {
 351                         rc2 = conn->ibc_credits;
 352                         spin_unlock(&conn->ibc_lock);
 353
 354                         CERROR("Bad credits from %s: %d + %d > %d\n",
 355                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 356                                rc2, credits,
 357                                conn->ibc_queue_depth);
 358
 359                         kiblnd_close_conn(conn, -EPROTO);
 360                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 361                         return;
 362                 }
 363
 364                 conn->ibc_credits += credits;
 365
 366                 /* This ensures the credit taken by NOOP can be returned */
 367                 if (msg->ibm_type == IBLND_MSG_NOOP &&
 368                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
 369                         conn->ibc_outstanding_credits++;
 370
 371                 kiblnd_check_sends_locked(conn);
 372                 spin_unlock(&conn->ibc_lock);
 373         }
 374
 375         switch (msg->ibm_type) {
 376         default:
 377                 CERROR("Bad IBLND message type %x from %s\n",
 378                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 379                 post_credit = IBLND_POSTRX_NO_CREDIT;
 380                 rc = -EPROTO;
 381                 break;
 382
 383         case IBLND_MSG_NOOP:
 384                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
 385                         post_credit = IBLND_POSTRX_NO_CREDIT;
 386                         break;
 387                 }
 388
 389                 if (credits != 0) /* credit already posted */
 390                         post_credit = IBLND_POSTRX_NO_CREDIT;
 391                 else              /* a keepalive NOOP */
 392                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 393                 break;
 394
 395         case IBLND_MSG_IMMEDIATE:
 396                 post_credit = IBLND_POSTRX_DONT_POST;
 397                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 398                                 msg->ibm_srcnid, rx, 0);
 399                 if (rc < 0)                     /* repost on error */
 400                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 401                 break;
 402
 403         case IBLND_MSG_PUT_REQ:
 404                 post_credit = IBLND_POSTRX_DONT_POST;
 405                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 406                                 msg->ibm_srcnid, rx, 1);
 407                 if (rc < 0)                     /* repost on error */
 408                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 409                 break;
 410
 411         case IBLND_MSG_PUT_NAK:
 412                 CWARN ("PUT_NACK from %s\n",
 413                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 414                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 415                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 416                                          msg->ibm_u.completion.ibcm_status,
 417                                          msg->ibm_u.completion.ibcm_cookie);
 418                 break;
 419
 420         case IBLND_MSG_PUT_ACK:
 421                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 422
 423                 spin_lock(&conn->ibc_lock);
 424                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 425                                         msg->ibm_u.putack.ibpam_src_cookie);
 426                 if (tx != NULL)
 427                         list_del(&tx->tx_list);
 428                 spin_unlock(&conn->ibc_lock);
 429
 430                 if (tx == NULL) {
 431                         CERROR("Unmatched PUT_ACK from %s\n",
 432                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 433                         rc = -EPROTO;
 434                         break;
 435                 }
 436
 437                 LASSERT (tx->tx_waiting);
 438                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
 439                  * (a) I can overwrite tx_msg since my peer_ni has received it!
 440                  * (b) tx_waiting set tells tx_complete() it's not done. */
 441
 442                 tx->tx_nwrq = tx->tx_nsge = 0;  /* overwrite PUT_REQ */
 443
 444                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 445                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 446                                        &msg->ibm_u.putack.ibpam_rd,
 447                                        msg->ibm_u.putack.ibpam_dst_cookie);
 448                 if (rc2 < 0)
 449                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 450                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 451
 452                 spin_lock(&conn->ibc_lock);
 453                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
 454                 kiblnd_queue_tx_locked(tx, conn);
 455                 spin_unlock(&conn->ibc_lock);
 456                 break;
 457
 458         case IBLND_MSG_PUT_DONE:
 459                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 460                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 461                                          msg->ibm_u.completion.ibcm_status,
 462                                          msg->ibm_u.completion.ibcm_cookie);
 463                 break;
 464
 465         case IBLND_MSG_GET_REQ:
 466                 post_credit = IBLND_POSTRX_DONT_POST;
 467                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 468                                 msg->ibm_srcnid, rx, 1);
 469                 if (rc < 0)                     /* repost on error */
 470                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 471                 break;
 472
 473         case IBLND_MSG_GET_DONE:
 474                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 475                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 476                                          msg->ibm_u.completion.ibcm_status,
 477                                          msg->ibm_u.completion.ibcm_cookie);
 478                 break;
 479         }
 480
 481         if (rc < 0)                             /* protocol error */
 482                 kiblnd_close_conn(conn, rc);
 483
 484         if (post_credit != IBLND_POSTRX_DONT_POST)
 485                 kiblnd_post_rx(rx, post_credit);
 486 }
 487
 488 static void
 489 kiblnd_rx_complete(struct kib_rx *rx, int status, int nob)
 490 {
 491         struct kib_msg *msg = rx->rx_msg;
 492         struct kib_conn   *conn = rx->rx_conn;
 493         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
 494         struct kib_net *net = ni->ni_data;
 495         int rc;
 496         int err = -EIO;
 497
 498         LASSERT (net != NULL);
 499         LASSERT (rx->rx_nob < 0);               /* was posted */
 500         rx->rx_nob = 0;                         /* isn't now */
 501
 502         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 503                 goto ignore;
 504
 505         if (status != IB_WC_SUCCESS) {
 506                 CNETERR("Rx from %s failed: %d\n",
 507                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 508                 goto failed;
 509         }
 510
 511         LASSERT (nob >= 0);
 512         rx->rx_nob = nob;
 513
 514         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 515         if (rc != 0) {
 516                 CERROR ("Error %d unpacking rx from %s\n",
 517                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 518                 goto failed;
 519         }
 520
 521         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 522             msg->ibm_dstnid != ni->ni_nid ||
 523             msg->ibm_srcstamp != conn->ibc_incarnation ||
 524             msg->ibm_dststamp != net->ibn_incarnation) {
 525                 CERROR ("Stale rx from %s\n",
 526                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
 527                 err = -ESTALE;
 528                 goto failed;
 529         }
 530
 531         /* set time last known alive */
 532         kiblnd_peer_alive(conn->ibc_peer);
 533
 534         /* racing with connection establishment/teardown! */
 535
 536         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 537                 rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
 538                 unsigned long  flags;
 539
 540                 write_lock_irqsave(g_lock, flags);
 541                 /* must check holding global lock to eliminate race */
 542                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 543                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 544                         write_unlock_irqrestore(g_lock, flags);
 545                         return;
 546                 }
 547                 write_unlock_irqrestore(g_lock, flags);
 548         }
 549         kiblnd_handle_rx(rx);
 550         return;
 551
 552  failed:
 553         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 554         kiblnd_close_conn(conn, err);
 555  ignore:
 556         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
 557 }
 558
 559 static int
 560 kiblnd_fmr_map_tx(struct kib_net *net, struct kib_tx *tx,
 561                   struct kib_rdma_desc *rd, u32 nob)
 562 {
 563         struct kib_hca_dev *hdev;
 564         struct kib_dev *dev;
 565         struct kib_fmr_poolset *fps;
 566         int                     cpt;
 567         int                     rc;
 568         int i;
 569
 570         LASSERT(tx->tx_pool != NULL);
 571         LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
 572
 573         dev = net->ibn_dev;
 574         hdev = tx->tx_pool->tpo_hdev;
 575         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 576
 577         /*
 578          * If we're dealing with FastReg, but the device doesn't
 579          * support GAPS and the tx has GAPS, then there is no real point
 580          * in trying to map the memory, because it'll just fail. So
 581          * preemptively fail with an appropriate message
 582          */
 583         if ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED) &&
 584             !(dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT) &&
 585             tx->tx_gaps) {
 586                 CERROR("Using FastReg with no GAPS support, but tx has gaps. "
 587                        "Try setting use_fastreg_gaps to 1\n");
 588                 return -EPROTONOSUPPORT;
 589         }
 590
 591 #ifdef HAVE_FMR_POOL_API
 592         /*
 593          * FMR does not support gaps but the tx has gaps then
 594          * we should make sure that the number of fragments we'll be sending
 595          * over fits within the number of fragments negotiated on the
 596          * connection, otherwise, we won't be able to RDMA the data.
 597          * We need to maintain the number of fragments negotiation on the
 598          * connection for backwards compatibility.
 599          */
 600         if (tx->tx_gaps && (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)) {
 601                 if (tx->tx_conn &&
 602                     tx->tx_conn->ibc_max_frags <= rd->rd_nfrags) {
 603                         CERROR("TX number of frags (%d) is <= than connection"
 604                                " number of frags (%d). Consider setting peer's"
 605                                " map_on_demand to 256\n", tx->tx_nfrags,
 606                                tx->tx_conn->ibc_max_frags);
 607                         return -EFBIG;
 608                 }
 609         }
 610 #endif
 611
 612         fps = net->ibn_fmr_ps[cpt];
 613         rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->tx_fmr);
 614         if (rc != 0) {
 615                 CERROR("Can't map %u bytes (%u/%u)s: %d\n", nob,
 616                        tx->tx_nfrags, rd->rd_nfrags, rc);
 617                 return rc;
 618         }
 619
 620         /*
 621          * If rd is not tx_rd, it's going to get sent to a peer_ni, who will
 622          * need the rkey
 623          */
 624         rd->rd_key = tx->tx_fmr.fmr_key;
 625         /*
 626          * for FastReg or FMR with no gaps we can accumulate all
 627          * the fragments in one FastReg or FMR fragment.
 628          */
 629         if (
 630 #ifdef HAVE_FMR_POOL_API
 631             ((dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
 632              && !tx->tx_gaps) ||
 633 #endif
 634             (dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED)) {
 635                 /* FMR requires zero based address */
 636 #ifdef HAVE_FMR_POOL_API
 637                 if (dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED)
 638                         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
 639 #endif
 640                 rd->rd_frags[0].rf_nob = nob;
 641                 rd->rd_nfrags = 1;
 642         } else {
 643                 /*
 644                  * We're transmitting with gaps using FMR.
 645                  * We'll need to use multiple fragments and identify the
 646                  * zero based address of each fragment.
 647                  */
 648                 for (i = 0; i < rd->rd_nfrags; i++) {
 649                         rd->rd_frags[i].rf_addr &= ~hdev->ibh_page_mask;
 650                         rd->rd_frags[i].rf_addr += i << hdev->ibh_page_shift;
 651                 }
 652         }
 653
 654         return 0;
 655 }
 656
 657 static void
 658 kiblnd_unmap_tx(struct kib_tx *tx)
 659 {
 660         if (
 661 #ifdef HAVE_FMR_POOL_API
 662                 tx->tx_fmr.fmr_pfmr ||
 663 #endif
 664                 tx->tx_fmr.fmr_frd)
 665                 kiblnd_fmr_pool_unmap(&tx->tx_fmr, tx->tx_status);
 666
 667         if (tx->tx_nfrags != 0) {
 668                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
 669                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 670                 tx->tx_nfrags = 0;
 671         }
 672 }
 673
 674 #ifdef HAVE_IB_GET_DMA_MR
 675 static struct ib_mr *
 676 kiblnd_find_rd_dma_mr(struct lnet_ni *ni, struct kib_rdma_desc *rd)
 677 {
 678         struct kib_net *net = ni->ni_data;
 679         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
 680         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
 681
 682         tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
 683
 684         /*
 685          * if map-on-demand is turned on and the device supports
 686          * either FMR or FastReg then use that. Otherwise use global
 687          * memory regions. If that's not available either, then you're
 688          * dead in the water and fail the operation.
 689          */
 690         if (tunables->lnd_map_on_demand &&
 691             (net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FASTREG_ENABLED
 692 #ifdef HAVE_FMR_POOL_API
 693              || net->ibn_dev->ibd_dev_caps & IBLND_DEV_CAPS_FMR_ENABLED
 694 #endif
 695         ))
 696                 return NULL;
 697
 698         /*
 699          * hdev->ibh_mrs can be NULL. This case is dealt with gracefully
 700          * in the call chain. The mapping will fail with appropriate error
 701          * message.
 702          */
 703         return hdev->ibh_mrs;
 704 }
 705 #endif
 706
 707 static int kiblnd_map_tx(struct lnet_ni *ni, struct kib_tx *tx,
 708                          struct kib_rdma_desc *rd, int nfrags)
 709 {
 710         struct kib_net *net = ni->ni_data;
 711         struct kib_hca_dev *hdev = net->ibn_dev->ibd_hdev;
 712 #ifdef HAVE_IB_GET_DMA_MR
 713         struct ib_mr *mr = NULL;
 714 #endif
 715         __u32 nob;
 716         int i;
 717
 718         /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
 719          * RDMA sink */
 720         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 721         tx->tx_nfrags = nfrags;
 722
 723         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
 724                                           tx->tx_nfrags, tx->tx_dmadir);
 725
 726         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 727                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 728                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 729                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 730                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 731                 nob += rd->rd_frags[i].rf_nob;
 732         }
 733
 734 #ifdef HAVE_IB_GET_DMA_MR
 735         mr = kiblnd_find_rd_dma_mr(ni, rd);
 736         if (mr != NULL) {
 737                 /* found pre-mapping MR */
 738                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 739                 return 0;
 740         }
 741 #endif
 742
 743         if (net->ibn_fmr_ps != NULL)
 744                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 745
 746         return -EINVAL;
 747 }
 748
 749 static int kiblnd_setup_rd_kiov(struct lnet_ni *ni, struct kib_tx *tx,
 750                                 struct kib_rdma_desc *rd, int nkiov,
 751                                 struct bio_vec *kiov, int offset, int nob)
 752 {
 753         struct kib_net *net = ni->ni_data;
 754         struct scatterlist *sg;
 755         int                 fragnob;
 756         int                 max_nkiov;
 757
 758         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 759
 760         LASSERT(nob > 0);
 761         LASSERT(nkiov > 0);
 762         LASSERT(net != NULL);
 763
 764         while (offset >= kiov->bv_len) {
 765                 offset -= kiov->bv_len;
 766                 nkiov--;
 767                 kiov++;
 768                 LASSERT(nkiov > 0);
 769         }
 770
 771         max_nkiov = nkiov;
 772
 773         sg = tx->tx_frags;
 774         do {
 775                 LASSERT(nkiov > 0);
 776
 777                 fragnob = min((int)(kiov->bv_len - offset), nob);
 778
 779                 /*
 780                  * We're allowed to start at a non-aligned page offset in
 781                  * the first fragment and end at a non-aligned page offset
 782                  * in the last fragment.
 783                  */
 784                 if ((fragnob < (int)(kiov->bv_len - offset)) &&
 785                     nkiov < max_nkiov && nob > fragnob) {
 786                         CDEBUG(D_NET, "fragnob %d < available page %d: with"
 787                                       " remaining %d kiovs with %d nob left\n",
 788                                fragnob, (int)(kiov->bv_len - offset),
 789                                nkiov, nob);
 790                         tx->tx_gaps = true;
 791                 }
 792
 793                 sg_set_page(sg, kiov->bv_page, fragnob,
 794                             kiov->bv_offset + offset);
 795                 sg = sg_next(sg);
 796                 if (!sg) {
 797                         CERROR("lacking enough sg entries to map tx\n");
 798                         return -EFAULT;
 799                 }
 800
 801                 offset = 0;
 802                 kiov++;
 803                 nkiov--;
 804                 nob -= fragnob;
 805         } while (nob > 0);
 806
 807         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 808 }
 809
 810 static int
 811 kiblnd_post_tx_locked(struct kib_conn *conn, struct kib_tx *tx, int credit)
 812 __must_hold(&conn->ibc_lock)
 813 {
 814         struct kib_msg *msg = tx->tx_msg;
 815         struct kib_peer_ni *peer_ni = conn->ibc_peer;
 816         struct lnet_ni *ni = peer_ni->ibp_ni;
 817         int ver = conn->ibc_version;
 818         int rc;
 819         int done;
 820
 821         LASSERT(tx->tx_queued);
 822         /* We rely on this for QP sizing */
 823         LASSERT(tx->tx_nwrq > 0 && tx->tx_nsge >= 0);
 824         LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
 825
 826         LASSERT(credit == 0 || credit == 1);
 827         LASSERT(conn->ibc_outstanding_credits >= 0);
 828         LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
 829         LASSERT(conn->ibc_credits >= 0);
 830         LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
 831
 832         if (conn->ibc_nsends_posted ==
 833             kiblnd_concurrent_sends(ver, ni)) {
 834                 /* tx completions outstanding... */
 835                 CDEBUG(D_NET, "%s: posted enough\n",
 836                        libcfs_nid2str(peer_ni->ibp_nid));
 837                 return -EAGAIN;
 838         }
 839
 840         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 841                 CDEBUG(D_NET, "%s: no credits\n",
 842                        libcfs_nid2str(peer_ni->ibp_nid));
 843                 return -EAGAIN;
 844         }
 845
 846         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 847             conn->ibc_credits == 1 &&   /* last credit reserved */
 848             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
 849                 CDEBUG(D_NET, "%s: not using last credit\n",
 850                        libcfs_nid2str(peer_ni->ibp_nid));
 851                 return -EAGAIN;
 852         }
 853
 854         /* NB don't drop ibc_lock before bumping tx_sending */
 855         list_del(&tx->tx_list);
 856         tx->tx_queued = 0;
 857
 858         if (msg->ibm_type == IBLND_MSG_NOOP &&
 859             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
 860              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 861               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 862                 /* OK to drop when posted enough NOOPs, since
 863                  * kiblnd_check_sends_locked will queue NOOP again when
 864                  * posted NOOPs complete */
 865                 spin_unlock(&conn->ibc_lock);
 866                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 867                 kiblnd_tx_done(tx);
 868                 spin_lock(&conn->ibc_lock);
 869                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 870                        libcfs_nid2str(peer_ni->ibp_nid),
 871                        conn->ibc_noops_posted);
 872                 return 0;
 873         }
 874
 875         kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 876                         peer_ni->ibp_nid, conn->ibc_incarnation);
 877
 878         conn->ibc_credits -= credit;
 879         conn->ibc_outstanding_credits = 0;
 880         conn->ibc_nsends_posted++;
 881         if (msg->ibm_type == IBLND_MSG_NOOP)
 882                 conn->ibc_noops_posted++;
 883
 884         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 885          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 886          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 887          * and then re-queued here.  It's (just) possible that
 888          * tx_sending is non-zero if we've not done the tx_complete()
 889          * from the first send; hence the ++ rather than = below. */
 890         tx->tx_sending++;
 891         list_add(&tx->tx_list, &conn->ibc_active_txs);
 892
 893         /* I'm still holding ibc_lock! */
 894         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
 895                 rc = -ECONNABORTED;
 896         } else if (tx->tx_pool->tpo_pool.po_failed ||
 897                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
 898                 /* close_conn will launch failover */
 899                 rc = -ENETDOWN;
 900         } else {
 901                 struct kib_fast_reg_descriptor *frd = tx->tx_fmr.fmr_frd;
 902                 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
 903                 struct ib_send_wr *wr  = &tx->tx_wrq[0].wr;
 904
 905                 if (frd != NULL) {
 906                         if (!frd->frd_valid) {
 907                                 wr = &frd->frd_inv_wr.wr;
 908                                 wr->next = &frd->frd_fastreg_wr.wr;
 909                         } else {
 910                                 wr = &frd->frd_fastreg_wr.wr;
 911                         }
 912                         frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
 913                 }
 914
 915                 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
 916                          "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
 917                          bad->wr_id, bad->opcode, bad->send_flags,
 918                          libcfs_nid2str(conn->ibc_peer->ibp_nid));
 919
 920                 bad = NULL;
 921                 if (lnet_send_error_simulation(tx->tx_lntmsg[0], &tx->tx_hstatus))
 922                         rc = -EINVAL;
 923                 else
 924 #ifdef HAVE_IB_POST_SEND_RECV_CONST
 925                         rc = ib_post_send(conn->ibc_cmid->qp, wr,
 926                                           (const struct ib_send_wr **)&bad);
 927 #else
 928                         rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
 929 #endif
 930         }
 931
 932         conn->ibc_last_send = ktime_get();
 933
 934         if (rc == 0)
 935                 return 0;
 936
 937         /* NB credits are transferred in the actual
 938          * message, which can only be the last work item */
 939         conn->ibc_credits += credit;
 940         conn->ibc_outstanding_credits += msg->ibm_credits;
 941         conn->ibc_nsends_posted--;
 942         if (msg->ibm_type == IBLND_MSG_NOOP)
 943                 conn->ibc_noops_posted--;
 944
 945         tx->tx_status = rc;
 946         tx->tx_waiting = 0;
 947         tx->tx_sending--;
 948
 949         done = (tx->tx_sending == 0);
 950         if (done)
 951                 list_del(&tx->tx_list);
 952
 953         spin_unlock(&conn->ibc_lock);
 954
 955         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 956                 CERROR("Error %d posting transmit to %s\n",
 957                        rc, libcfs_nid2str(peer_ni->ibp_nid));
 958         else
 959                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 960                        rc, libcfs_nid2str(peer_ni->ibp_nid));
 961
 962         kiblnd_close_conn(conn, rc);
 963
 964         if (done)
 965                 kiblnd_tx_done(tx);
 966
 967         spin_lock(&conn->ibc_lock);
 968
 969         return -EIO;
 970 }
 971
 972 static void
 973 kiblnd_check_sends_locked(struct kib_conn *conn)
 974 {
 975         int ver = conn->ibc_version;
 976         struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
 977         struct kib_tx *tx;
 978
 979         /* Don't send anything until after the connection is established */
 980         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 981                 CDEBUG(D_NET, "%s too soon\n",
 982                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 983                 return;
 984         }
 985
 986         LASSERT(conn->ibc_nsends_posted <=
 987                 kiblnd_concurrent_sends(ver, ni));
 988         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 989                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 990         LASSERT (conn->ibc_reserved_credits >= 0);
 991
 992         while (conn->ibc_reserved_credits > 0 &&
 993                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 994                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 995                                 struct kib_tx, tx_list);
 996                 list_move_tail(&tx->tx_list, &conn->ibc_tx_queue);
 997                 conn->ibc_reserved_credits--;
 998         }
 999
1000         if (kiblnd_need_noop(conn)) {
1001                 spin_unlock(&conn->ibc_lock);
1002
1003                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1004                 if (tx != NULL)
1005                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
1006
1007                 spin_lock(&conn->ibc_lock);
1008                 if (tx != NULL)
1009                         kiblnd_queue_tx_locked(tx, conn);
1010         }
1011
1012         for (;;) {
1013                 int credit;
1014
1015                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
1016                         credit = 0;
1017                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
1018                                         struct kib_tx, tx_list);
1019                 } else if (!list_empty(&conn->ibc_tx_noops)) {
1020                         LASSERT (!IBLND_OOB_CAPABLE(ver));
1021                         credit = 1;
1022                         tx = list_entry(conn->ibc_tx_noops.next,
1023                                         struct kib_tx, tx_list);
1024                 } else if (!list_empty(&conn->ibc_tx_queue)) {
1025                         credit = 1;
1026                         tx = list_entry(conn->ibc_tx_queue.next,
1027                                         struct kib_tx, tx_list);
1028                 } else
1029                         break;
1030
1031                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1032                         break;
1033         }
1034 }
1035
1036 static void
1037 kiblnd_tx_complete(struct kib_tx *tx, int status)
1038 {
1039         int           failed = (status != IB_WC_SUCCESS);
1040         struct kib_conn   *conn = tx->tx_conn;
1041         int           idle;
1042
1043         if (tx->tx_sending <= 0) {
1044                 CERROR("Received an event on a freed tx: %p status %d\n",
1045                        tx, tx->tx_status);
1046                 return;
1047         }
1048
1049         if (failed) {
1050                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1051                         CNETERR("Tx -> %s cookie %#llx"
1052                                 " sending %d waiting %d: failed %d\n",
1053                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1054                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1055                                 status);
1056
1057                 kiblnd_close_conn(conn, -EIO);
1058         } else {
1059                 kiblnd_peer_alive(conn->ibc_peer);
1060         }
1061
1062         spin_lock(&conn->ibc_lock);
1063
1064         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
1065          * gets to free it, which also drops its ref on 'conn'. */
1066
1067         tx->tx_sending--;
1068         conn->ibc_nsends_posted--;
1069         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1070                 conn->ibc_noops_posted--;
1071
1072         if (failed) {
1073                 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
1074                 tx->tx_waiting = 0;             /* don't wait for peer_ni */
1075                 tx->tx_status = -EIO;
1076         }
1077
1078         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1079                !tx->tx_waiting &&               /* Not waiting for peer_ni */
1080                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1081         if (idle)
1082                 list_del(&tx->tx_list);
1083
1084         kiblnd_check_sends_locked(conn);
1085         spin_unlock(&conn->ibc_lock);
1086
1087         if (idle)
1088                 kiblnd_tx_done(tx);
1089 }
1090
1091 static void
1092 kiblnd_init_tx_msg(struct lnet_ni *ni, struct kib_tx *tx, int type,
1093                    int body_nob)
1094 {
1095         struct kib_hca_dev *hdev = tx->tx_pool->tpo_hdev;
1096         struct ib_sge *sge = &tx->tx_msgsge;
1097         struct ib_rdma_wr *wrq;
1098         int nob = offsetof(struct kib_msg, ibm_u) + body_nob;
1099 #ifdef HAVE_IB_GET_DMA_MR
1100         struct ib_mr *mr = hdev->ibh_mrs;
1101 #endif
1102
1103         LASSERT(tx->tx_nwrq >= 0);
1104         LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1105         LASSERT(nob <= IBLND_MSG_SIZE);
1106 #ifdef HAVE_IB_GET_DMA_MR
1107         LASSERT(mr != NULL);
1108 #endif
1109
1110         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1111
1112 #ifdef HAVE_IB_GET_DMA_MR
1113         sge->lkey   = mr->lkey;
1114 #else
1115         sge->lkey   = hdev->ibh_pd->local_dma_lkey;
1116 #endif
1117         sge->addr   = tx->tx_msgaddr;
1118         sge->length = nob;
1119
1120         wrq = &tx->tx_wrq[tx->tx_nwrq];
1121         memset(wrq, 0, sizeof(*wrq));
1122
1123         wrq->wr.next            = NULL;
1124         wrq->wr.wr_id           = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1125         wrq->wr.sg_list         = sge;
1126         wrq->wr.num_sge         = 1;
1127         wrq->wr.opcode          = IB_WR_SEND;
1128         wrq->wr.send_flags      = IB_SEND_SIGNALED;
1129
1130         tx->tx_nwrq++;
1131 }
1132
1133 static int
1134 kiblnd_init_rdma(struct kib_conn *conn, struct kib_tx *tx, int type,
1135                  int resid, struct kib_rdma_desc *dstrd, u64 dstcookie)
1136 {
1137         struct kib_msg *ibmsg = tx->tx_msg;
1138         struct kib_rdma_desc *srcrd = tx->tx_rd;
1139         struct ib_rdma_wr *wrq = NULL;
1140         struct ib_sge     *sge;
1141         int                rc  = resid;
1142         int                srcidx;
1143         int                dstidx;
1144         int                sge_nob;
1145         int                wrq_sge;
1146
1147         LASSERT(!in_interrupt());
1148         LASSERT(tx->tx_nwrq == 0 && tx->tx_nsge == 0);
1149         LASSERT(type == IBLND_MSG_GET_DONE || type == IBLND_MSG_PUT_DONE);
1150
1151         for (srcidx = dstidx = wrq_sge = sge_nob = 0;
1152              resid > 0; resid -= sge_nob) {
1153                 int     prev = dstidx;
1154
1155                 if (srcidx >= srcrd->rd_nfrags) {
1156                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1157                         rc = -EPROTO;
1158                         break;
1159                 }
1160
1161                 if (dstidx >= dstrd->rd_nfrags) {
1162                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1163                         rc = -EPROTO;
1164                         break;
1165                 }
1166
1167                 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1168                         CERROR("RDMA has too many fragments for peer_ni %s (%d), "
1169                                "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1170                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1171                                conn->ibc_max_frags,
1172                                srcidx, srcrd->rd_nfrags,
1173                                dstidx, dstrd->rd_nfrags);
1174                         rc = -EMSGSIZE;
1175                         break;
1176                 }
1177
1178                 sge_nob = min3(kiblnd_rd_frag_size(srcrd, srcidx),
1179                                kiblnd_rd_frag_size(dstrd, dstidx),
1180                                resid);
1181
1182                 sge = &tx->tx_sge[tx->tx_nsge];
1183                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1184                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1185                 sge->length = sge_nob;
1186
1187                 if (wrq_sge == 0) {
1188                         wrq = &tx->tx_wrq[tx->tx_nwrq];
1189
1190                         wrq->wr.next    = &(wrq + 1)->wr;
1191                         wrq->wr.wr_id   = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1192                         wrq->wr.sg_list = sge;
1193                         wrq->wr.opcode  = IB_WR_RDMA_WRITE;
1194                         wrq->wr.send_flags = 0;
1195
1196 #ifdef HAVE_IB_RDMA_WR
1197                         wrq->remote_addr        = kiblnd_rd_frag_addr(dstrd,
1198                                                                       dstidx);
1199                         wrq->rkey               = kiblnd_rd_frag_key(dstrd,
1200                                                                      dstidx);
1201 #else
1202                         wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd,
1203                                                                         dstidx);
1204                         wrq->wr.wr.rdma.rkey    = kiblnd_rd_frag_key(dstrd,
1205                                                                      dstidx);
1206 #endif
1207                 }
1208
1209                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, sge_nob);
1210                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, sge_nob);
1211
1212                 wrq_sge++;
1213                 if (wrq_sge == *kiblnd_tunables.kib_wrq_sge || dstidx != prev) {
1214                         tx->tx_nwrq++;
1215                         wrq->wr.num_sge = wrq_sge;
1216                         wrq_sge = 0;
1217                 }
1218                 tx->tx_nsge++;
1219         }
1220
1221         if (rc < 0)     /* no RDMA if completing with failure */
1222                 tx->tx_nwrq = tx->tx_nsge = 0;
1223
1224         ibmsg->ibm_u.completion.ibcm_status = rc;
1225         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1226         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1227                            type, sizeof(struct kib_completion_msg));
1228
1229         return rc;
1230 }
1231
1232 static void
1233 kiblnd_queue_tx_locked(struct kib_tx *tx, struct kib_conn *conn)
1234 {
1235         struct list_head *q;
1236         s64 timeout_ns;
1237
1238         LASSERT(tx->tx_nwrq > 0);       /* work items set up */
1239         LASSERT(!tx->tx_queued);        /* not queued for sending already */
1240         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1241
1242         if (conn->ibc_state >= IBLND_CONN_DISCONNECTED) {
1243                 tx->tx_status = -ECONNABORTED;
1244                 tx->tx_waiting = 0;
1245                 if (tx->tx_conn != NULL) {
1246                         /* PUT_DONE first attached to conn as a PUT_REQ */
1247                         LASSERT(tx->tx_conn == conn);
1248                         LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1249                         tx->tx_conn = NULL;
1250                         kiblnd_conn_decref(conn);
1251                 }
1252                 list_add(&tx->tx_list, &conn->ibc_zombie_txs);
1253
1254                 return;
1255         }
1256
1257         timeout_ns = kiblnd_timeout() * NSEC_PER_SEC;
1258         tx->tx_queued = 1;
1259         tx->tx_deadline = ktime_add_ns(ktime_get(), timeout_ns);
1260
1261         if (tx->tx_conn == NULL) {
1262                 kiblnd_conn_addref(conn);
1263                 tx->tx_conn = conn;
1264                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1265         } else {
1266                 /* PUT_DONE first attached to conn as a PUT_REQ */
1267                 LASSERT (tx->tx_conn == conn);
1268                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1269         }
1270
1271         switch (tx->tx_msg->ibm_type) {
1272         default:
1273                 LBUG();
1274
1275         case IBLND_MSG_PUT_REQ:
1276         case IBLND_MSG_GET_REQ:
1277                 q = &conn->ibc_tx_queue_rsrvd;
1278                 break;
1279
1280         case IBLND_MSG_PUT_NAK:
1281         case IBLND_MSG_PUT_ACK:
1282         case IBLND_MSG_PUT_DONE:
1283         case IBLND_MSG_GET_DONE:
1284                 q = &conn->ibc_tx_queue_nocred;
1285                 break;
1286
1287         case IBLND_MSG_NOOP:
1288                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1289                         q = &conn->ibc_tx_queue_nocred;
1290                 else
1291                         q = &conn->ibc_tx_noops;
1292                 break;
1293
1294         case IBLND_MSG_IMMEDIATE:
1295                 q = &conn->ibc_tx_queue;
1296                 break;
1297         }
1298
1299         list_add_tail(&tx->tx_list, q);
1300 }
1301
1302 static void
1303 kiblnd_queue_tx(struct kib_tx *tx, struct kib_conn *conn)
1304 {
1305         spin_lock(&conn->ibc_lock);
1306         kiblnd_queue_tx_locked(tx, conn);
1307         kiblnd_check_sends_locked(conn);
1308         spin_unlock(&conn->ibc_lock);
1309 }
1310
1311 static int
1312 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1313                         struct sockaddr_in *srcaddr,
1314                         struct sockaddr_in *dstaddr,
1315                         int timeout_ms)
1316 {
1317         unsigned short port;
1318         int rc;
1319
1320         LASSERT(capable(CAP_NET_BIND_SERVICE));
1321
1322         /* allow the port to be reused */
1323         rc = rdma_set_reuseaddr(cmid, 1);
1324         if (rc != 0) {
1325                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1326                 return rc;
1327         }
1328
1329         /* look for a free privileged port */
1330         for (port = PROT_SOCK-1; port > 0; port--) {
1331                 srcaddr->sin_port = htons(port);
1332                 rc = rdma_resolve_addr(cmid,
1333                                        (struct sockaddr *)srcaddr,
1334                                        (struct sockaddr *)dstaddr,
1335                                        timeout_ms);
1336                 if (rc == 0) {
1337                         CDEBUG(D_NET, "bound to port %hu\n", port);
1338                         return 0;
1339                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1340                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1341                                port, rc);
1342                 } else {
1343                         return rc;
1344                 }
1345         }
1346
1347         CERROR("Failed to bind to a free privileged port\n");
1348         return rc;
1349 }
1350
1351 static int
1352 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1353                     struct sockaddr_in *srcaddr,
1354                     struct sockaddr_in *dstaddr,
1355                     int timeout_ms)
1356 {
1357         const struct cred *old_creds = NULL;
1358         struct cred *new_creds;
1359         int rc;
1360
1361         if (!capable(CAP_NET_BIND_SERVICE)) {
1362                 new_creds = prepare_creds();
1363                 if (!new_creds)
1364                         return -ENOMEM;
1365
1366                 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1367                 old_creds = override_creds(new_creds);
1368         }
1369
1370         rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1371
1372         if (old_creds)
1373                 revert_creds(old_creds);
1374
1375         return rc;
1376 }
1377
1378 static void
1379 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1380 {
1381         struct rdma_cm_id *cmid;
1382         struct kib_dev *dev;
1383         struct kib_net *net = peer_ni->ibp_ni->ni_data;
1384         struct sockaddr_in srcaddr;
1385         struct sockaddr_in dstaddr;
1386         int rc;
1387
1388         LASSERT (net != NULL);
1389         LASSERT (peer_ni->ibp_connecting > 0);
1390
1391         cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1392                                      kiblnd_cm_callback, peer_ni,
1393                                      RDMA_PS_TCP, IB_QPT_RC);
1394
1395         if (IS_ERR(cmid)) {
1396                 CERROR("Can't create CMID for %s: %ld\n",
1397                        libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1398                 rc = PTR_ERR(cmid);
1399                 goto failed;
1400         }
1401
1402         dev = net->ibn_dev;
1403         memset(&srcaddr, 0, sizeof(srcaddr));
1404         srcaddr.sin_family = AF_INET;
1405         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1406
1407         memset(&dstaddr, 0, sizeof(dstaddr));
1408         dstaddr.sin_family = AF_INET;
1409         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1410         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1411
1412         kiblnd_peer_addref(peer_ni);               /* cmid's ref */
1413
1414         if (*kiblnd_tunables.kib_use_priv_port) {
1415                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1416                                          kiblnd_timeout() * 1000);
1417         } else {
1418                 rc = rdma_resolve_addr(cmid,
1419                                        (struct sockaddr *)&srcaddr,
1420                                        (struct sockaddr *)&dstaddr,
1421                                        kiblnd_timeout() * 1000);
1422         }
1423         if (rc != 0) {
1424                 /* Can't initiate address resolution:  */
1425                 CERROR("Can't resolve addr for %s: %d\n",
1426                        libcfs_nid2str(peer_ni->ibp_nid), rc);
1427                 goto failed2;
1428         }
1429
1430         return;
1431
1432  failed2:
1433         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1434         kiblnd_peer_decref(peer_ni);               /* cmid's ref */
1435         rdma_destroy_id(cmid);
1436         return;
1437  failed:
1438         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1439 }
1440
1441 bool
1442 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1443 {
1444         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1445         char *reason = NULL;
1446         LIST_HEAD(txs);
1447         unsigned long flags;
1448
1449         write_lock_irqsave(glock, flags);
1450         if (peer_ni->ibp_reconnecting == 0) {
1451                 if (peer_ni->ibp_accepting)
1452                         reason = "accepting";
1453                 else if (peer_ni->ibp_connecting)
1454                         reason = "connecting";
1455                 else if (!list_empty(&peer_ni->ibp_conns))
1456                         reason = "connected";
1457                 else /* connected then closed */
1458                         reason = "closed";
1459
1460                 goto no_reconnect;
1461         }
1462
1463         if (peer_ni->ibp_accepting)
1464                 CNETERR("Detecting race between accepting and reconnecting\n");
1465         peer_ni->ibp_reconnecting--;
1466
1467         if (!kiblnd_peer_active(peer_ni)) {
1468                 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1469                 reason = "unlinked";
1470                 goto no_reconnect;
1471         }
1472
1473         peer_ni->ibp_connecting++;
1474         peer_ni->ibp_reconnected++;
1475
1476         write_unlock_irqrestore(glock, flags);
1477
1478         kiblnd_connect_peer(peer_ni);
1479         return true;
1480
1481  no_reconnect:
1482         write_unlock_irqrestore(glock, flags);
1483
1484         CWARN("Abort reconnection of %s: %s\n",
1485               libcfs_nid2str(peer_ni->ibp_nid), reason);
1486         kiblnd_txlist_done(&txs, -ECONNABORTED,
1487                            LNET_MSG_STATUS_LOCAL_ABORTED);
1488         return false;
1489 }
1490
1491 void
1492 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1493 {
1494         struct kib_peer_ni *peer_ni;
1495         struct kib_peer_ni *peer2;
1496         struct kib_conn *conn;
1497         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1498         unsigned long      flags;
1499         int                rc;
1500         int                i;
1501         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1502
1503         /* If I get here, I've committed to send, so I complete the tx with
1504          * failure on any problems */
1505
1506         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1507         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1508
1509         /* First time, just use a read lock since I expect to find my peer_ni
1510          * connected */
1511         read_lock_irqsave(g_lock, flags);
1512
1513         peer_ni = kiblnd_find_peer_locked(ni, nid);
1514         if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1515                 /* Found a peer_ni with an established connection */
1516                 conn = kiblnd_get_conn_locked(peer_ni);
1517                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1518
1519                 read_unlock_irqrestore(g_lock, flags);
1520
1521                 if (tx != NULL)
1522                         kiblnd_queue_tx(tx, conn);
1523                 kiblnd_conn_decref(conn); /* ...to here */
1524                 return;
1525         }
1526
1527         read_unlock(g_lock);
1528         /* Re-try with a write lock */
1529         write_lock(g_lock);
1530
1531         peer_ni = kiblnd_find_peer_locked(ni, nid);
1532         if (peer_ni != NULL) {
1533                 if (list_empty(&peer_ni->ibp_conns)) {
1534                         /* found a peer_ni, but it's still connecting... */
1535                         LASSERT(kiblnd_peer_connecting(peer_ni));
1536                         if (tx != NULL)
1537                                 list_add_tail(&tx->tx_list,
1538                                                   &peer_ni->ibp_tx_queue);
1539                         write_unlock_irqrestore(g_lock, flags);
1540                 } else {
1541                         conn = kiblnd_get_conn_locked(peer_ni);
1542                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1543
1544                         write_unlock_irqrestore(g_lock, flags);
1545
1546                         if (tx != NULL)
1547                                 kiblnd_queue_tx(tx, conn);
1548                         kiblnd_conn_decref(conn); /* ...to here */
1549                 }
1550                 return;
1551         }
1552
1553         write_unlock_irqrestore(g_lock, flags);
1554
1555         /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1556         rc = kiblnd_create_peer(ni, &peer_ni, nid);
1557         if (rc != 0) {
1558                 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1559                 if (tx != NULL) {
1560                         tx->tx_status = -EHOSTUNREACH;
1561                         tx->tx_waiting = 0;
1562                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1563                         kiblnd_tx_done(tx);
1564                 }
1565                 return;
1566         }
1567
1568         write_lock_irqsave(g_lock, flags);
1569
1570         peer2 = kiblnd_find_peer_locked(ni, nid);
1571         if (peer2 != NULL) {
1572                 if (list_empty(&peer2->ibp_conns)) {
1573                         /* found a peer_ni, but it's still connecting... */
1574                         LASSERT(kiblnd_peer_connecting(peer2));
1575                         if (tx != NULL)
1576                                 list_add_tail(&tx->tx_list,
1577                                                   &peer2->ibp_tx_queue);
1578                         write_unlock_irqrestore(g_lock, flags);
1579                 } else {
1580                         conn = kiblnd_get_conn_locked(peer2);
1581                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1582
1583                         write_unlock_irqrestore(g_lock, flags);
1584
1585                         if (tx != NULL)
1586                                 kiblnd_queue_tx(tx, conn);
1587                         kiblnd_conn_decref(conn); /* ...to here */
1588                 }
1589
1590                 kiblnd_peer_decref(peer_ni);
1591                 return;
1592         }
1593
1594         /* Brand new peer_ni */
1595         LASSERT(peer_ni->ibp_connecting == 0);
1596         tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1597         peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1598
1599         /* always called with a ref on ni, which prevents ni being shutdown */
1600         LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1601
1602         if (tx != NULL)
1603                 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1604
1605         kiblnd_peer_addref(peer_ni);
1606         list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1607
1608         write_unlock_irqrestore(g_lock, flags);
1609
1610         for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1611                 kiblnd_connect_peer(peer_ni);
1612         kiblnd_peer_decref(peer_ni);
1613 }
1614
1615 int
1616 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1617 {
1618         struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1619         int               type = lntmsg->msg_type;
1620         struct lnet_process_id target = lntmsg->msg_target;
1621         int               target_is_router = lntmsg->msg_target_is_router;
1622         int               routing = lntmsg->msg_routing;
1623         unsigned int      payload_niov = lntmsg->msg_niov;
1624         struct bio_vec   *payload_kiov = lntmsg->msg_kiov;
1625         unsigned int      payload_offset = lntmsg->msg_offset;
1626         unsigned int      payload_nob = lntmsg->msg_len;
1627         struct kib_msg *ibmsg;
1628         struct kib_rdma_desc *rd;
1629         struct kib_tx *tx;
1630         int               nob;
1631         int               rc;
1632
1633         /* NB 'private' is different depending on what we're sending.... */
1634
1635         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1636                payload_nob, payload_niov, libcfs_id2str(target));
1637
1638         LASSERT (payload_nob == 0 || payload_niov > 0);
1639         LASSERT (payload_niov <= LNET_MAX_IOV);
1640
1641         /* Thread context */
1642         LASSERT (!in_interrupt());
1643
1644         switch (type) {
1645         default:
1646                 LBUG();
1647                 return (-EIO);
1648
1649         case LNET_MSG_ACK:
1650                 LASSERT (payload_nob == 0);
1651                 break;
1652
1653         case LNET_MSG_GET:
1654                 if (routing || target_is_router)
1655                         break;                  /* send IMMEDIATE */
1656
1657                 /* is the REPLY message too small for RDMA? */
1658                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1659                 if (nob <= IBLND_MSG_SIZE)
1660                         break;                  /* send IMMEDIATE */
1661
1662                 tx = kiblnd_get_idle_tx(ni, target.nid);
1663                 if (tx == NULL) {
1664                         CERROR("Can't allocate txd for GET to %s\n",
1665                                libcfs_nid2str(target.nid));
1666                         return -ENOMEM;
1667                 }
1668
1669                 ibmsg = tx->tx_msg;
1670                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1671                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1672                                           lntmsg->msg_md->md_niov,
1673                                           lntmsg->msg_md->md_kiov,
1674                                           0, lntmsg->msg_md->md_length);
1675                 if (rc != 0) {
1676                         CERROR("Can't setup GET sink for %s: %d\n",
1677                                libcfs_nid2str(target.nid), rc);
1678                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1679                         kiblnd_tx_done(tx);
1680                         return -EIO;
1681                 }
1682
1683                 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1684                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1685                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1686
1687                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1688
1689                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1690                 if (tx->tx_lntmsg[1] == NULL) {
1691                         CERROR("Can't create reply for GET -> %s\n",
1692                                libcfs_nid2str(target.nid));
1693                         kiblnd_tx_done(tx);
1694                         return -EIO;
1695                 }
1696
1697                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1698                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1699                 kiblnd_launch_tx(ni, tx, target.nid);
1700                 return 0;
1701
1702         case LNET_MSG_REPLY:
1703         case LNET_MSG_PUT:
1704                 /* Is the payload small enough not to need RDMA? */
1705                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1706                 if (nob <= IBLND_MSG_SIZE)
1707                         break;                  /* send IMMEDIATE */
1708
1709                 tx = kiblnd_get_idle_tx(ni, target.nid);
1710                 if (tx == NULL) {
1711                         CERROR("Can't allocate %s txd for %s\n",
1712                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1713                                libcfs_nid2str(target.nid));
1714                         return -ENOMEM;
1715                 }
1716
1717                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1718                                           payload_niov, payload_kiov,
1719                                           payload_offset, payload_nob);
1720                 if (rc != 0) {
1721                         CERROR("Can't setup PUT src for %s: %d\n",
1722                                libcfs_nid2str(target.nid), rc);
1723                         kiblnd_tx_done(tx);
1724                         return -EIO;
1725                 }
1726
1727                 ibmsg = tx->tx_msg;
1728                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1729                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1730                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1731                                    sizeof(struct kib_putreq_msg));
1732
1733                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1734                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1735                 kiblnd_launch_tx(ni, tx, target.nid);
1736                 return 0;
1737         }
1738
1739         /* send IMMEDIATE */
1740         LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1741                 <= IBLND_MSG_SIZE);
1742
1743         tx = kiblnd_get_idle_tx(ni, target.nid);
1744         if (tx == NULL) {
1745                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1746                         type, libcfs_nid2str(target.nid));
1747                 return -ENOMEM;
1748         }
1749
1750         ibmsg = tx->tx_msg;
1751         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1752
1753         lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1754                             offsetof(struct kib_msg,
1755                                      ibm_u.immediate.ibim_payload),
1756                             payload_niov, payload_kiov,
1757                             payload_offset, payload_nob);
1758
1759         nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1760         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1761
1762         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1763         kiblnd_launch_tx(ni, tx, target.nid);
1764         return 0;
1765 }
1766
1767 static void
1768 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1769 {
1770         struct lnet_process_id target = lntmsg->msg_target;
1771         unsigned int niov = lntmsg->msg_niov;
1772         struct bio_vec *kiov = lntmsg->msg_kiov;
1773         unsigned int offset = lntmsg->msg_offset;
1774         unsigned int nob = lntmsg->msg_len;
1775         struct kib_tx *tx;
1776         int rc;
1777
1778         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1779         if (tx == NULL) {
1780                 CERROR("Can't get tx for REPLY to %s\n",
1781                        libcfs_nid2str(target.nid));
1782                 goto failed_0;
1783         }
1784
1785         if (nob == 0)
1786                 rc = 0;
1787         else
1788                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1789                                           niov, kiov, offset, nob);
1790
1791         if (rc != 0) {
1792                 CERROR("Can't setup GET src for %s: %d\n",
1793                        libcfs_nid2str(target.nid), rc);
1794                 goto failed_1;
1795         }
1796
1797         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1798                               IBLND_MSG_GET_DONE, nob,
1799                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1800                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1801         if (rc < 0) {
1802                 CERROR("Can't setup rdma for GET from %s: %d\n",
1803                        libcfs_nid2str(target.nid), rc);
1804                 goto failed_1;
1805         }
1806
1807         if (nob == 0) {
1808                 /* No RDMA: local completion may happen now! */
1809                 lnet_finalize(lntmsg, 0);
1810         } else {
1811                 /* RDMA: lnet_finalize(lntmsg) when it
1812                  * completes */
1813                 tx->tx_lntmsg[0] = lntmsg;
1814         }
1815
1816         kiblnd_queue_tx(tx, rx->rx_conn);
1817         return;
1818
1819
1820 failed_1:
1821         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1822         kiblnd_tx_done(tx);
1823 failed_0:
1824         lnet_finalize(lntmsg, -EIO);
1825 }
1826
1827 int
1828 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1829             int delayed, unsigned int niov, struct bio_vec *kiov,
1830             unsigned int offset, unsigned int mlen, unsigned int rlen)
1831 {
1832         struct kib_rx *rx = private;
1833         struct kib_msg *rxmsg = rx->rx_msg;
1834         struct kib_conn *conn = rx->rx_conn;
1835         struct kib_tx *tx;
1836         __u64        ibprm_cookie;
1837         int          nob;
1838         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1839         int          rc = 0;
1840
1841         LASSERT (mlen <= rlen);
1842         LASSERT (!in_interrupt());
1843
1844         switch (rxmsg->ibm_type) {
1845         default:
1846                 LBUG();
1847
1848         case IBLND_MSG_IMMEDIATE:
1849                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1850                 if (nob > rx->rx_nob) {
1851                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1852                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1853                                 nob, rx->rx_nob);
1854                         rc = -EPROTO;
1855                         break;
1856                 }
1857
1858                 lnet_copy_flat2kiov(niov, kiov, offset,
1859                                     IBLND_MSG_SIZE, rxmsg,
1860                                     offsetof(struct kib_msg,
1861                                              ibm_u.immediate.ibim_payload),
1862                                     mlen);
1863                 lnet_finalize(lntmsg, 0);
1864                 break;
1865
1866         case IBLND_MSG_PUT_REQ: {
1867                 struct kib_msg  *txmsg;
1868                 struct kib_rdma_desc *rd;
1869                 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1870
1871                 if (mlen == 0) {
1872                         lnet_finalize(lntmsg, 0);
1873                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1874                                                0, ibprm_cookie);
1875                         break;
1876                 }
1877
1878                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1879                 if (tx == NULL) {
1880                         CERROR("Can't allocate tx for %s\n",
1881                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1882                         /* Not replying will break the connection */
1883                         rc = -ENOMEM;
1884                         break;
1885                 }
1886
1887                 txmsg = tx->tx_msg;
1888                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1889                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1890                                           niov, kiov, offset, mlen);
1891                 if (rc != 0) {
1892                         CERROR("Can't setup PUT sink for %s: %d\n",
1893                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1894                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1895                         kiblnd_tx_done(tx);
1896                         /* tell peer_ni it's over */
1897                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1898                                                rc, ibprm_cookie);
1899                         break;
1900                 }
1901
1902                 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1903                 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1904                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1905
1906                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1907
1908                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1909                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1910                 kiblnd_queue_tx(tx, conn);
1911
1912                 /* reposted buffer reserved for PUT_DONE */
1913                 post_credit = IBLND_POSTRX_NO_CREDIT;
1914                 break;
1915                 }
1916
1917         case IBLND_MSG_GET_REQ:
1918                 if (lntmsg != NULL) {
1919                         /* Optimized GET; RDMA lntmsg's payload */
1920                         kiblnd_reply(ni, rx, lntmsg);
1921                 } else {
1922                         /* GET didn't match anything */
1923                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1924                                                -ENODATA,
1925                                                rxmsg->ibm_u.get.ibgm_cookie);
1926                 }
1927                 break;
1928         }
1929
1930         kiblnd_post_rx(rx, post_credit);
1931         return rc;
1932 }
1933
1934 int
1935 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1936 {
1937         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1938
1939         if (IS_ERR(task))
1940                 return PTR_ERR(task);
1941
1942         atomic_inc(&kiblnd_data.kib_nthreads);
1943         return 0;
1944 }
1945
1946 static void
1947 kiblnd_thread_fini (void)
1948 {
1949         atomic_dec (&kiblnd_data.kib_nthreads);
1950 }
1951
1952 static void
1953 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1954 {
1955         /* This is racy, but everyone's only writing ktime_get_seconds() */
1956         peer_ni->ibp_last_alive = ktime_get_seconds();
1957         smp_mb();
1958 }
1959
1960 static void
1961 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1962 {
1963         int           error = 0;
1964         time64_t last_alive = 0;
1965         unsigned long flags;
1966
1967         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1968
1969         if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1970                 error = peer_ni->ibp_error;
1971                 peer_ni->ibp_error = 0;
1972
1973                 last_alive = peer_ni->ibp_last_alive;
1974         }
1975
1976         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1977
1978         if (error != 0)
1979                 lnet_notify(peer_ni->ibp_ni,
1980                             peer_ni->ibp_nid, false, false, last_alive);
1981 }
1982
1983 void
1984 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1985 {
1986         /* This just does the immediate housekeeping.  'error' is zero for a
1987          * normal shutdown which can happen only after the connection has been
1988          * established.  If the connection is established, schedule the
1989          * connection to be finished off by the connd.  Otherwise the connd is
1990          * already dealing with it (either to set it up or tear it down).
1991          * Caller holds kib_global_lock exclusively in irq context */
1992         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1993         struct kib_dev *dev;
1994         unsigned long flags;
1995
1996         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1997
1998         if (error != 0 && conn->ibc_comms_error == 0)
1999                 conn->ibc_comms_error = error;
2000
2001         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
2002                 return; /* already being handled  */
2003
2004         if (error == 0 &&
2005             list_empty(&conn->ibc_tx_noops) &&
2006             list_empty(&conn->ibc_tx_queue) &&
2007             list_empty(&conn->ibc_tx_queue_rsrvd) &&
2008             list_empty(&conn->ibc_tx_queue_nocred) &&
2009             list_empty(&conn->ibc_active_txs)) {
2010                 CDEBUG(D_NET, "closing conn to %s\n",
2011                        libcfs_nid2str(peer_ni->ibp_nid));
2012         } else {
2013                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2014                        libcfs_nid2str(peer_ni->ibp_nid), error,
2015                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2016                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2017                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
2018                                                 "" : "(sending_rsrvd)",
2019                        list_empty(&conn->ibc_tx_queue_nocred) ?
2020                                                  "" : "(sending_nocred)",
2021                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2022         }
2023
2024         dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2025         if (peer_ni->ibp_next_conn == conn)
2026                 /* clear next_conn so it won't be used */
2027                 peer_ni->ibp_next_conn = NULL;
2028         list_del(&conn->ibc_list);
2029         /* connd (see below) takes over ibc_list's ref */
2030
2031         if (list_empty(&peer_ni->ibp_conns) &&    /* no more conns */
2032             kiblnd_peer_active(peer_ni)) {         /* still in peer_ni table */
2033                 kiblnd_unlink_peer_locked(peer_ni);
2034
2035                 /* set/clear error on last conn */
2036                 peer_ni->ibp_error = conn->ibc_comms_error;
2037         }
2038
2039         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2040
2041         if (error != 0 &&
2042             kiblnd_dev_can_failover(dev)) {
2043                 list_add_tail(&dev->ibd_fail_list,
2044                               &kiblnd_data.kib_failed_devs);
2045                 wake_up(&kiblnd_data.kib_failover_waitq);
2046         }
2047
2048         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2049
2050         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2051         wake_up(&kiblnd_data.kib_connd_waitq);
2052
2053         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2054 }
2055
2056 void
2057 kiblnd_close_conn(struct kib_conn *conn, int error)
2058 {
2059         unsigned long flags;
2060
2061         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2062
2063         kiblnd_close_conn_locked(conn, error);
2064
2065         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2066 }
2067
2068 static void
2069 kiblnd_handle_early_rxs(struct kib_conn *conn)
2070 {
2071         unsigned long flags;
2072         struct kib_rx *rx;
2073
2074         LASSERT(!in_interrupt());
2075         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2076
2077         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2078         while (!list_empty(&conn->ibc_early_rxs)) {
2079                 rx = list_entry(conn->ibc_early_rxs.next,
2080                                 struct kib_rx, rx_list);
2081                 list_del(&rx->rx_list);
2082                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2083
2084                 kiblnd_handle_rx(rx);
2085
2086                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2087         }
2088         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2089 }
2090
2091 void
2092 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2093 {
2094         LIST_HEAD(zombies);
2095         struct kib_tx *nxt;
2096         struct kib_tx *tx;
2097
2098         spin_lock(&conn->ibc_lock);
2099
2100         list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2101                 if (txs == &conn->ibc_active_txs) {
2102                         LASSERT(!tx->tx_queued);
2103                         LASSERT(tx->tx_waiting ||
2104                                 tx->tx_sending != 0);
2105                         if (conn->ibc_comms_error == -ETIMEDOUT) {
2106                                 if (tx->tx_waiting && !tx->tx_sending)
2107                                         tx->tx_hstatus =
2108                                           LNET_MSG_STATUS_REMOTE_TIMEOUT;
2109                                 else if (tx->tx_sending)
2110                                         tx->tx_hstatus =
2111                                           LNET_MSG_STATUS_NETWORK_TIMEOUT;
2112                         }
2113                 } else {
2114                         LASSERT(tx->tx_queued);
2115                         if (conn->ibc_comms_error == -ETIMEDOUT)
2116                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2117                         else
2118                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2119                 }
2120
2121                 tx->tx_status = -ECONNABORTED;
2122                 tx->tx_waiting = 0;
2123
2124                 /*
2125                  * TODO: This makes an assumption that
2126                  * kiblnd_tx_complete() will be called for each tx. If
2127                  * that event is dropped we could end up with stale
2128                  * connections floating around. We'd like to deal with
2129                  * that in a better way.
2130                  *
2131                  * Also that means we can exceed the timeout by many
2132                  * seconds.
2133                  */
2134                 if (tx->tx_sending == 0) {
2135                         tx->tx_queued = 0;
2136                         list_move(&tx->tx_list, &zombies);
2137                 } else {
2138                         /* keep tx until cq destroy */
2139                         list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2140                         conn->ibc_waits ++;
2141                 }
2142         }
2143
2144         spin_unlock(&conn->ibc_lock);
2145
2146         /*
2147          * aborting transmits occurs when finalizing the connection.
2148          * The connection is finalized on error.
2149          * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2150          * override the value already set in tx->tx_hstatus above.
2151          */
2152         kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2153 }
2154
2155 static int
2156 kiblnd_tx_may_discard(struct kib_conn *conn)
2157 {
2158         int rc = 0;
2159         struct kib_tx *nxt;
2160         struct kib_tx *tx;
2161
2162         spin_lock(&conn->ibc_lock);
2163
2164         list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2165                 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2166                     lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2167                         tx->tx_sending --;
2168                         if (tx->tx_sending == 0) {
2169                                 kiblnd_conn_decref(tx->tx_conn);
2170                                 tx->tx_conn = NULL;
2171                                 rc = 1;
2172                         }
2173                 }
2174         }
2175
2176         spin_unlock(&conn->ibc_lock);
2177         return rc;
2178 }
2179
2180 static void
2181 kiblnd_finalise_conn(struct kib_conn *conn)
2182 {
2183         LASSERT (!in_interrupt());
2184         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2185
2186         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2187          * for connections that didn't get as far as being connected, because
2188          * rdma_disconnect() does this for free. */
2189         kiblnd_abort_receives(conn);
2190
2191         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2192
2193         /* Complete all tx descs not waiting for sends to complete.
2194          * NB we should be safe from RDMA now that the QP has changed state */
2195
2196         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2197         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2198         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2199         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2200         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2201
2202         kiblnd_handle_early_rxs(conn);
2203 }
2204
2205 static void
2206 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2207                            int error)
2208 {
2209         LIST_HEAD(zombies);
2210         unsigned long   flags;
2211
2212         LASSERT (error != 0);
2213         LASSERT (!in_interrupt());
2214
2215         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2216
2217         if (active) {
2218                 LASSERT(peer_ni->ibp_connecting > 0);
2219                 peer_ni->ibp_connecting--;
2220         } else {
2221                 LASSERT (peer_ni->ibp_accepting > 0);
2222                 peer_ni->ibp_accepting--;
2223         }
2224
2225         if (kiblnd_peer_connecting(peer_ni)) {
2226                 /* another connection attempt under way... */
2227                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2228                                         flags);
2229                 return;
2230         }
2231
2232         peer_ni->ibp_reconnected = 0;
2233         if (list_empty(&peer_ni->ibp_conns)) {
2234                 /* Take peer_ni's blocked transmits to complete with error */
2235                 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2236
2237                 if (kiblnd_peer_active(peer_ni))
2238                         kiblnd_unlink_peer_locked(peer_ni);
2239
2240                 peer_ni->ibp_error = error;
2241         } else {
2242                 /* Can't have blocked transmits if there are connections */
2243                 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2244         }
2245
2246         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2247
2248         kiblnd_peer_notify(peer_ni);
2249
2250         if (list_empty(&zombies))
2251                 return;
2252
2253         CNETERR("Deleting messages for %s: connection failed\n",
2254                 libcfs_nid2str(peer_ni->ibp_nid));
2255
2256         if (error == -EHOSTUNREACH || error == -ETIMEDOUT)
2257                 kiblnd_txlist_done(&zombies, error,
2258                                    LNET_MSG_STATUS_NETWORK_TIMEOUT);
2259         else
2260                 kiblnd_txlist_done(&zombies, error,
2261                                    LNET_MSG_STATUS_LOCAL_DROPPED);
2262 }
2263
2264 static void
2265 kiblnd_connreq_done(struct kib_conn *conn, int status)
2266 {
2267         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2268         struct kib_tx *tx;
2269         LIST_HEAD(txs);
2270         unsigned long    flags;
2271         int              active;
2272
2273         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2274
2275         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2276                libcfs_nid2str(peer_ni->ibp_nid), active,
2277                conn->ibc_version, status);
2278
2279         LASSERT (!in_interrupt());
2280         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2281                   peer_ni->ibp_connecting > 0) ||
2282                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2283                   peer_ni->ibp_accepting > 0));
2284
2285         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2286         conn->ibc_connvars = NULL;
2287
2288         if (status != 0) {
2289                 /* failed to establish connection */
2290                 kiblnd_peer_connect_failed(peer_ni, active, status);
2291                 kiblnd_finalise_conn(conn);
2292                 return;
2293         }
2294
2295         /* connection established */
2296         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2297
2298         /* reset retry count */
2299         peer_ni->ibp_retries = 0;
2300
2301         conn->ibc_last_send = ktime_get();
2302         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2303         kiblnd_peer_alive(peer_ni);
2304
2305         /* Add conn to peer_ni's list and nuke any dangling conns from a different
2306          * peer_ni instance... */
2307         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2308         list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2309         peer_ni->ibp_reconnected = 0;
2310         if (active)
2311                 peer_ni->ibp_connecting--;
2312         else
2313                 peer_ni->ibp_accepting--;
2314
2315         if (peer_ni->ibp_version == 0) {
2316                 peer_ni->ibp_version     = conn->ibc_version;
2317                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2318         }
2319
2320         if (peer_ni->ibp_version     != conn->ibc_version ||
2321             peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2322                 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2323                                                 conn->ibc_incarnation);
2324                 peer_ni->ibp_version     = conn->ibc_version;
2325                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2326         }
2327
2328         /* grab pending txs while I have the lock */
2329         list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2330
2331         if (!kiblnd_peer_active(peer_ni) ||        /* peer_ni has been deleted */
2332             conn->ibc_comms_error != 0) {       /* error has happened already */
2333
2334                 /* start to shut down connection */
2335                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2336                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2337
2338                 kiblnd_txlist_done(&txs, -ECONNABORTED,
2339                                    LNET_MSG_STATUS_LOCAL_ERROR);
2340
2341                 return;
2342         }
2343
2344         /* +1 ref for myself, this connection is visible to other threads
2345          * now, refcount of peer:ibp_conns can be released by connection
2346          * close from either a different thread, or the calling of
2347          * kiblnd_check_sends_locked() below. See bz21911 for details.
2348          */
2349         kiblnd_conn_addref(conn);
2350         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2351
2352         /* Schedule blocked txs
2353          * Note: if we are running with conns_per_peer > 1, these blocked
2354          * txs will all get scheduled to the first connection which gets
2355          * scheduled.  We won't be using round robin on this first batch.
2356          */
2357         spin_lock(&conn->ibc_lock);
2358         while (!list_empty(&txs)) {
2359                 tx = list_entry(txs.next, struct kib_tx, tx_list);
2360                 list_del(&tx->tx_list);
2361
2362                 kiblnd_queue_tx_locked(tx, conn);
2363         }
2364         kiblnd_check_sends_locked(conn);
2365         spin_unlock(&conn->ibc_lock);
2366
2367         /* schedule blocked rxs */
2368         kiblnd_handle_early_rxs(conn);
2369         kiblnd_conn_decref(conn);
2370 }
2371
2372 static void
2373 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2374 {
2375         int          rc;
2376
2377 #ifdef HAVE_RDMA_REJECT_4ARGS
2378         rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2379 #else
2380         rc = rdma_reject(cmid, rej, sizeof(*rej));
2381 #endif
2382
2383         if (rc != 0)
2384                 CWARN("Error %d sending reject\n", rc);
2385 }
2386
2387 static int
2388 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2389 {
2390         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2391         struct kib_msg *reqmsg = priv;
2392         struct kib_msg *ackmsg;
2393         struct kib_dev *ibdev;
2394         struct kib_peer_ni *peer_ni;
2395         struct kib_peer_ni *peer2;
2396         struct kib_conn *conn;
2397         struct lnet_ni *ni = NULL;
2398         struct kib_net *net = NULL;
2399         lnet_nid_t             nid;
2400         struct rdma_conn_param cp;
2401         struct kib_rej rej;
2402         int                    version = IBLND_MSG_VERSION;
2403         unsigned long          flags;
2404         int                    rc;
2405         struct sockaddr_in    *peer_addr;
2406         LASSERT (!in_interrupt());
2407
2408         /* cmid inherits 'context' from the corresponding listener id */
2409         ibdev = cmid->context;
2410         LASSERT(ibdev);
2411
2412         memset(&rej, 0, sizeof(rej));
2413         rej.ibr_magic                = IBLND_MSG_MAGIC;
2414         rej.ibr_why                  = IBLND_REJECT_FATAL;
2415         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2416
2417         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2418         if (*kiblnd_tunables.kib_require_priv_port &&
2419             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2420                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2421                 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2422                        &ip, ntohs(peer_addr->sin_port));
2423                 goto failed;
2424         }
2425
2426         if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2427                 CERROR("Short connection request\n");
2428                 goto failed;
2429         }
2430
2431         /* Future protocol version compatibility support!  If the
2432          * o2iblnd-specific protocol changes, or when LNET unifies
2433          * protocols over all LNDs, the initial connection will
2434          * negotiate a protocol version.  I trap this here to avoid
2435          * console errors; the reject tells the peer_ni which protocol I
2436          * speak. */
2437         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2438             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2439                 goto failed;
2440         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2441             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2442             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2443                 goto failed;
2444         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2445             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2446             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2447                 goto failed;
2448
2449         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2450         if (rc != 0) {
2451                 CERROR("Can't parse connection request: %d\n", rc);
2452                 goto failed;
2453         }
2454
2455         nid = reqmsg->ibm_srcnid;
2456         ni  = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2457
2458         if (ni != NULL) {
2459                 net = (struct kib_net *)ni->ni_data;
2460                 rej.ibr_incarnation = net->ibn_incarnation;
2461         }
2462
2463         if (ni == NULL ||                         /* no matching net */
2464             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2465             net->ibn_dev != ibdev) {              /* wrong device */
2466                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2467                        "bad dst nid %s\n", libcfs_nid2str(nid),
2468                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2469                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2470                         &ibdev->ibd_ifip,
2471                        libcfs_nid2str(reqmsg->ibm_dstnid));
2472
2473                 goto failed;
2474         }
2475
2476        /* check time stamp as soon as possible */
2477         if (reqmsg->ibm_dststamp != 0 &&
2478             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2479                 CWARN("Stale connection request\n");
2480                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2481                 goto failed;
2482         }
2483
2484         /* I can accept peer_ni's version */
2485         version = reqmsg->ibm_version;
2486
2487         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2488                 CERROR("Unexpected connreq msg type: %x from %s\n",
2489                        reqmsg->ibm_type, libcfs_nid2str(nid));
2490                 goto failed;
2491         }
2492
2493         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2494             kiblnd_msg_queue_size(version, ni)) {
2495                 CERROR("Can't accept conn from %s, queue depth too large: "
2496                        " %d (<=%d wanted)\n",
2497                        libcfs_nid2str(nid),
2498                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2499                        kiblnd_msg_queue_size(version, ni));
2500
2501                 if (version == IBLND_MSG_VERSION)
2502                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2503
2504                 goto failed;
2505         }
2506
2507         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2508             IBLND_MAX_RDMA_FRAGS) {
2509                 CWARN("Can't accept conn from %s (version %x): "
2510                       "max_frags %d too large (%d wanted)\n",
2511                       libcfs_nid2str(nid), version,
2512                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2513                       IBLND_MAX_RDMA_FRAGS);
2514
2515                 if (version >= IBLND_MSG_VERSION)
2516                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2517
2518                 goto failed;
2519         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2520                    IBLND_MAX_RDMA_FRAGS &&
2521                    net->ibn_fmr_ps == NULL) {
2522                 CWARN("Can't accept conn from %s (version %x): "
2523                       "max_frags %d incompatible without FMR pool "
2524                       "(%d wanted)\n",
2525                       libcfs_nid2str(nid), version,
2526                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2527                       IBLND_MAX_RDMA_FRAGS);
2528
2529                 if (version == IBLND_MSG_VERSION)
2530                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2531
2532                 goto failed;
2533         }
2534
2535         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2536                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2537                        libcfs_nid2str(nid),
2538                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2539                        IBLND_MSG_SIZE);
2540                 goto failed;
2541         }
2542
2543         /* assume 'nid' is a new peer_ni; create  */
2544         rc = kiblnd_create_peer(ni, &peer_ni, nid);
2545         if (rc != 0) {
2546                 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2547                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2548                 goto failed;
2549         }
2550
2551         /* We have validated the peer's parameters so use those */
2552         peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2553         peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2554
2555         write_lock_irqsave(g_lock, flags);
2556
2557         peer2 = kiblnd_find_peer_locked(ni, nid);
2558         if (peer2 != NULL) {
2559                 if (peer2->ibp_version == 0) {
2560                         peer2->ibp_version     = version;
2561                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2562                 }
2563
2564                 /* not the guy I've talked with */
2565                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2566                     peer2->ibp_version     != version) {
2567                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2568
2569                         if (kiblnd_peer_active(peer2)) {
2570                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2571                                 peer2->ibp_version = version;
2572                         }
2573                         write_unlock_irqrestore(g_lock, flags);
2574
2575                         CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2576                               libcfs_nid2str(nid), peer2->ibp_version, version,
2577                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2578
2579                         kiblnd_peer_decref(peer_ni);
2580                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2581                         goto failed;
2582                 }
2583
2584                 /* Tie-break connection race in favour of the higher NID.
2585                  * If we keep running into a race condition multiple times,
2586                  * we have to assume that the connection attempt with the
2587                  * higher NID is stuck in a connecting state and will never
2588                  * recover.  As such, we pass through this if-block and let
2589                  * the lower NID connection win so we can move forward.
2590                  */
2591                 if (peer2->ibp_connecting != 0 &&
2592                     nid < ni->ni_nid && peer2->ibp_races <
2593                     MAX_CONN_RACES_BEFORE_ABORT) {
2594                         peer2->ibp_races++;
2595                         write_unlock_irqrestore(g_lock, flags);
2596
2597                         CDEBUG(D_NET, "Conn race %s\n",
2598                                libcfs_nid2str(peer2->ibp_nid));
2599
2600                         kiblnd_peer_decref(peer_ni);
2601                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2602                         goto failed;
2603                 }
2604                 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2605                         CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2606                                 libcfs_nid2str(peer2->ibp_nid),
2607                                 MAX_CONN_RACES_BEFORE_ABORT);
2608                 /*
2609                  * passive connection is allowed even this peer_ni is waiting for
2610                  * reconnection.
2611                  */
2612                 peer2->ibp_reconnecting = 0;
2613                 peer2->ibp_races = 0;
2614                 peer2->ibp_accepting++;
2615                 kiblnd_peer_addref(peer2);
2616
2617                 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2618                  * so copy validated parameters since we now know what the
2619                  * peer_ni's limits are */
2620                 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2621                 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2622
2623                 write_unlock_irqrestore(g_lock, flags);
2624                 kiblnd_peer_decref(peer_ni);
2625                 peer_ni = peer2;
2626         } else {
2627                 /* Brand new peer_ni */
2628                 LASSERT (peer_ni->ibp_accepting == 0);
2629                 LASSERT (peer_ni->ibp_version == 0 &&
2630                          peer_ni->ibp_incarnation == 0);
2631
2632                 peer_ni->ibp_accepting   = 1;
2633                 peer_ni->ibp_version     = version;
2634                 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2635
2636                 /* I have a ref on ni that prevents it being shutdown */
2637                 LASSERT (net->ibn_shutdown == 0);
2638
2639                 kiblnd_peer_addref(peer_ni);
2640                 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2641
2642                 write_unlock_irqrestore(g_lock, flags);
2643         }
2644
2645         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2646         if (conn == NULL) {
2647                 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2648                 kiblnd_peer_decref(peer_ni);
2649                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2650                 goto failed;
2651         }
2652
2653         /* conn now "owns" cmid, so I return success from here on to ensure the
2654          * CM callback doesn't destroy cmid. */
2655         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2656         conn->ibc_credits          = conn->ibc_queue_depth;
2657         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2658         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2659                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2660
2661         ackmsg = &conn->ibc_connvars->cv_msg;
2662         memset(ackmsg, 0, sizeof(*ackmsg));
2663
2664         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2665                         sizeof(ackmsg->ibm_u.connparams));
2666         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2667         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2668         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2669
2670         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2671
2672         memset(&cp, 0, sizeof(cp));
2673         cp.private_data        = ackmsg;
2674         cp.private_data_len    = ackmsg->ibm_nob;
2675         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2676         cp.initiator_depth     = 0;
2677         cp.flow_control        = 1;
2678         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2679         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2680
2681         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2682
2683         rc = rdma_accept(cmid, &cp);
2684         if (rc != 0) {
2685                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2686                 rej.ibr_version = version;
2687                 rej.ibr_why     = IBLND_REJECT_FATAL;
2688
2689                 kiblnd_reject(cmid, &rej);
2690                 kiblnd_connreq_done(conn, rc);
2691                 kiblnd_conn_decref(conn);
2692         }
2693
2694         lnet_ni_decref(ni);
2695         return 0;
2696
2697  failed:
2698         if (ni != NULL) {
2699                 rej.ibr_cp.ibcp_queue_depth =
2700                         kiblnd_msg_queue_size(version, ni);
2701                 rej.ibr_cp.ibcp_max_frags   = IBLND_MAX_RDMA_FRAGS;
2702                 lnet_ni_decref(ni);
2703         }
2704
2705         rej.ibr_version = version;
2706         kiblnd_reject(cmid, &rej);
2707
2708         return -ECONNREFUSED;
2709 }
2710
2711 static void
2712 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2713                        u64 incarnation, int why, struct kib_connparams *cp)
2714 {
2715         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
2716         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2717         char            *reason;
2718         int              msg_size = IBLND_MSG_SIZE;
2719         int              frag_num = -1;
2720         int              queue_dep = -1;
2721         bool             reconnect;
2722         unsigned long    flags;
2723
2724         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2725         LASSERT(peer_ni->ibp_connecting > 0);   /* 'conn' at least */
2726
2727         if (cp) {
2728                 msg_size        = cp->ibcp_max_msg_size;
2729                 frag_num        = cp->ibcp_max_frags;
2730                 queue_dep       = cp->ibcp_queue_depth;
2731         }
2732
2733         write_lock_irqsave(glock, flags);
2734         /* retry connection if it's still needed and no other connection
2735          * attempts (active or passive) are in progress
2736          * NB: reconnect is still needed even when ibp_tx_queue is
2737          * empty if ibp_version != version because reconnect may be
2738          * initiated.
2739          */
2740         reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2741                      peer_ni->ibp_version != version) &&
2742                     peer_ni->ibp_connecting &&
2743                     peer_ni->ibp_accepting == 0;
2744         if (!reconnect) {
2745                 reason = "no need";
2746                 goto out;
2747         }
2748
2749         if (peer_ni->ibp_retries > *kiblnd_tunables.kib_retry_count) {
2750                 reason = "retry count exceeded due to no listener";
2751                 goto out;
2752         }
2753
2754         switch (why) {
2755         default:
2756                 reason = "Unknown";
2757                 break;
2758
2759         case IBLND_REJECT_RDMA_FRAGS: {
2760                 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2761
2762                 if (!cp) {
2763                         reason = "can't negotiate max frags";
2764                         goto out;
2765                 }
2766                 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2767 #ifdef HAVE_IB_GET_DMA_MR
2768                 /*
2769                  * This check only makes sense if the kernel supports global
2770                  * memory registration. Otherwise, map_on_demand will never == 0
2771                  */
2772                 if (!tunables->lnd_map_on_demand) {
2773                         reason = "map_on_demand must be enabled";
2774                         goto out;
2775                 }
2776 #endif
2777                 if (conn->ibc_max_frags <= frag_num) {
2778                         reason = "unsupported max frags";
2779                         goto out;
2780                 }
2781
2782                 peer_ni->ibp_max_frags = frag_num;
2783                 reason = "rdma fragments";
2784                 break;
2785         }
2786         case IBLND_REJECT_MSG_QUEUE_SIZE:
2787                 if (!cp) {
2788                         reason = "can't negotiate queue depth";
2789                         goto out;
2790                 }
2791                 if (conn->ibc_queue_depth <= queue_dep) {
2792                         reason = "unsupported queue depth";
2793                         goto out;
2794                 }
2795
2796                 peer_ni->ibp_queue_depth = queue_dep;
2797                 reason = "queue depth";
2798                 break;
2799
2800         case IBLND_REJECT_CONN_STALE:
2801                 reason = "stale";
2802                 break;
2803
2804         case IBLND_REJECT_CONN_RACE:
2805                 reason = "conn race";
2806                 break;
2807
2808         case IBLND_REJECT_CONN_UNCOMPAT:
2809                 reason = "version negotiation";
2810                 break;
2811
2812         case IBLND_REJECT_INVALID_SRV_ID:
2813                 reason = "invalid service id";
2814                 break;
2815         }
2816
2817         conn->ibc_reconnect = 1;
2818         peer_ni->ibp_reconnecting++;
2819         peer_ni->ibp_version = version;
2820         if (incarnation != 0)
2821                 peer_ni->ibp_incarnation = incarnation;
2822  out:
2823         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2824
2825         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2826                 libcfs_nid2str(peer_ni->ibp_nid),
2827                 reconnect ? "reconnect" : "don't reconnect",
2828                 reason, IBLND_MSG_VERSION, version, msg_size,
2829                 conn->ibc_queue_depth, queue_dep,
2830                 conn->ibc_max_frags, frag_num);
2831         /*
2832          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2833          * while destroying the zombie
2834          */
2835 }
2836
2837 static void
2838 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2839 {
2840         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2841
2842         LASSERT (!in_interrupt());
2843         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2844
2845         switch (reason) {
2846         case IB_CM_REJ_STALE_CONN:
2847                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2848                                        IBLND_REJECT_CONN_STALE, NULL);
2849                 break;
2850
2851         case IB_CM_REJ_INVALID_SERVICE_ID:
2852                 peer_ni->ibp_retries++;
2853                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2854                                        IBLND_REJECT_INVALID_SRV_ID, NULL);
2855                 CNETERR("%s rejected: no listener at %d\n",
2856                         libcfs_nid2str(peer_ni->ibp_nid),
2857                         *kiblnd_tunables.kib_service);
2858                 break;
2859
2860         case IB_CM_REJ_CONSUMER_DEFINED:
2861                 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2862                         struct kib_rej *rej = priv;
2863                         struct kib_connparams *cp = NULL;
2864                         int               flip        = 0;
2865                         __u64             incarnation = -1;
2866
2867                         /* NB. default incarnation is -1 because:
2868                          * a) V1 will ignore dst incarnation in connreq.
2869                          * b) V2 will provide incarnation while rejecting me,
2870                          *    -1 will be overwrote.
2871                          *
2872                          * if I try to connect to a V1 peer_ni with V2 protocol,
2873                          * it rejected me then upgrade to V2, I have no idea
2874                          * about the upgrading and try to reconnect with V1,
2875                          * in this case upgraded V2 can find out I'm trying to
2876                          * talk to the old guy and reject me(incarnation is -1).
2877                          */
2878
2879                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2880                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2881                                 __swab32s(&rej->ibr_magic);
2882                                 __swab16s(&rej->ibr_version);
2883                                 flip = 1;
2884                         }
2885
2886                         if (priv_nob >= sizeof(struct kib_rej) &&
2887                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2888                                 /* priv_nob is always 148 in current version
2889                                  * of OFED, so we still need to check version.
2890                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2891                                 cp = &rej->ibr_cp;
2892
2893                                 if (flip) {
2894                                         __swab64s(&rej->ibr_incarnation);
2895                                         __swab16s(&cp->ibcp_queue_depth);
2896                                         __swab16s(&cp->ibcp_max_frags);
2897                                         __swab32s(&cp->ibcp_max_msg_size);
2898                                 }
2899
2900                                 incarnation = rej->ibr_incarnation;
2901                         }
2902
2903                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2904                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2905                                 CERROR("%s rejected: consumer defined fatal error\n",
2906                                        libcfs_nid2str(peer_ni->ibp_nid));
2907                                 break;
2908                         }
2909
2910                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2911                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2912                                 CERROR("%s rejected: o2iblnd version %x error\n",
2913                                        libcfs_nid2str(peer_ni->ibp_nid),
2914                                        rej->ibr_version);
2915                                 break;
2916                         }
2917
2918                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2919                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2920                                 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2921                                        libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2922
2923                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2924                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2925                         }
2926
2927                         switch (rej->ibr_why) {
2928                         case IBLND_REJECT_CONN_RACE:
2929                         case IBLND_REJECT_CONN_STALE:
2930                         case IBLND_REJECT_CONN_UNCOMPAT:
2931                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2932                         case IBLND_REJECT_RDMA_FRAGS:
2933                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2934                                                 incarnation, rej->ibr_why, cp);
2935                                 break;
2936
2937                         case IBLND_REJECT_NO_RESOURCES:
2938                                 CERROR("%s rejected: o2iblnd no resources\n",
2939                                        libcfs_nid2str(peer_ni->ibp_nid));
2940                                 break;
2941
2942                         case IBLND_REJECT_FATAL:
2943                                 CERROR("%s rejected: o2iblnd fatal error\n",
2944                                        libcfs_nid2str(peer_ni->ibp_nid));
2945                                 break;
2946
2947                         default:
2948                                 CERROR("%s rejected: o2iblnd reason %d\n",
2949                                        libcfs_nid2str(peer_ni->ibp_nid),
2950                                        rej->ibr_why);
2951                                 break;
2952                         }
2953                         break;
2954                 }
2955                 /* fall through */
2956         default:
2957                 CNETERR("%s rejected: reason %d, size %d\n",
2958                         libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2959                 break;
2960         }
2961
2962         kiblnd_connreq_done(conn, -ECONNREFUSED);
2963 }
2964
2965 static void
2966 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2967 {
2968         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2969         struct lnet_ni *ni = peer_ni->ibp_ni;
2970         struct kib_net *net = ni->ni_data;
2971         struct kib_msg *msg = priv;
2972         int            ver  = conn->ibc_version;
2973         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2974         unsigned long  flags;
2975
2976         LASSERT (net != NULL);
2977
2978         if (rc != 0) {
2979                 CERROR("Can't unpack connack from %s: %d\n",
2980                        libcfs_nid2str(peer_ni->ibp_nid), rc);
2981                 goto failed;
2982         }
2983
2984         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2985                 CERROR("Unexpected message %d from %s\n",
2986                        msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2987                 rc = -EPROTO;
2988                 goto failed;
2989         }
2990
2991         if (ver != msg->ibm_version) {
2992                 CERROR("%s replied version %x is different with "
2993                        "requested version %x\n",
2994                        libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2995                 rc = -EPROTO;
2996                 goto failed;
2997         }
2998
2999         if (msg->ibm_u.connparams.ibcp_queue_depth >
3000             conn->ibc_queue_depth) {
3001                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
3002                        libcfs_nid2str(peer_ni->ibp_nid),
3003                        msg->ibm_u.connparams.ibcp_queue_depth,
3004                        conn->ibc_queue_depth);
3005                 rc = -EPROTO;
3006                 goto failed;
3007         }
3008
3009         if (msg->ibm_u.connparams.ibcp_max_frags >
3010             conn->ibc_max_frags) {
3011                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
3012                        libcfs_nid2str(peer_ni->ibp_nid),
3013                        msg->ibm_u.connparams.ibcp_max_frags,
3014                        conn->ibc_max_frags);
3015                 rc = -EPROTO;
3016                 goto failed;
3017         }
3018
3019         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
3020                 CERROR("%s max message size %d too big (%d max)\n",
3021                        libcfs_nid2str(peer_ni->ibp_nid),
3022                        msg->ibm_u.connparams.ibcp_max_msg_size,
3023                        IBLND_MSG_SIZE);
3024                 rc = -EPROTO;
3025                 goto failed;
3026         }
3027
3028         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3029         if (msg->ibm_dstnid == ni->ni_nid &&
3030             msg->ibm_dststamp == net->ibn_incarnation)
3031                 rc = 0;
3032         else
3033                 rc = -ESTALE;
3034         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3035
3036         if (rc != 0) {
3037                 CERROR("Bad connection reply from %s, rc = %d, "
3038                        "version: %x max_frags: %d\n",
3039                        libcfs_nid2str(peer_ni->ibp_nid), rc,
3040                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3041                 goto failed;
3042         }
3043
3044         conn->ibc_incarnation      = msg->ibm_srcstamp;
3045         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
3046         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3047         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
3048         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
3049         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3050                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3051
3052         kiblnd_connreq_done(conn, 0);
3053         return;
3054
3055  failed:
3056         /* NB My QP has already established itself, so I handle anything going
3057          * wrong here by setting ibc_comms_error.
3058          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3059          * immediately tears it down. */
3060
3061         LASSERT (rc != 0);
3062         conn->ibc_comms_error = rc;
3063         kiblnd_connreq_done(conn, 0);
3064 }
3065
3066 static int
3067 kiblnd_active_connect(struct rdma_cm_id *cmid)
3068 {
3069         struct kib_peer_ni *peer_ni = cmid->context;
3070         struct kib_conn *conn;
3071         struct kib_msg *msg;
3072         struct rdma_conn_param cp;
3073         int                      version;
3074         __u64                    incarnation;
3075         unsigned long            flags;
3076         int                      rc;
3077
3078         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3079
3080         incarnation = peer_ni->ibp_incarnation;
3081         version     = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3082                                                  peer_ni->ibp_version;
3083
3084         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3085
3086         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3087                                   version);
3088         if (conn == NULL) {
3089                 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3090                 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3091                 return -ENOMEM;
3092         }
3093
3094         /* conn "owns" cmid now, so I return success from here on to ensure the
3095          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3096          * on peer_ni */
3097
3098         msg = &conn->ibc_connvars->cv_msg;
3099
3100         memset(msg, 0, sizeof(*msg));
3101         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3102         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
3103         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
3104         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3105
3106         kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3107                         0, peer_ni->ibp_nid, incarnation);
3108
3109         memset(&cp, 0, sizeof(cp));
3110         cp.private_data        = msg;
3111         cp.private_data_len    = msg->ibm_nob;
3112         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
3113         cp.initiator_depth     = 0;
3114         cp.flow_control        = 1;
3115         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
3116         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
3117
3118         LASSERT(cmid->context == (void *)conn);
3119         LASSERT(conn->ibc_cmid == cmid);
3120
3121         rc = rdma_connect(cmid, &cp);
3122         if (rc != 0) {
3123                 CERROR("Can't connect to %s: %d\n",
3124                        libcfs_nid2str(peer_ni->ibp_nid), rc);
3125                 kiblnd_connreq_done(conn, rc);
3126                 kiblnd_conn_decref(conn);
3127         }
3128
3129         return 0;
3130 }
3131
3132 int
3133 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3134 {
3135         struct kib_peer_ni *peer_ni;
3136         struct kib_conn *conn;
3137         int rc;
3138
3139         switch (event->event) {
3140         default:
3141                 CERROR("Unexpected event: %d, status: %d\n",
3142                        event->event, event->status);
3143                 LBUG();
3144
3145         case RDMA_CM_EVENT_CONNECT_REQUEST:
3146                 /* destroy cmid on failure */
3147                 rc = kiblnd_passive_connect(cmid,
3148                                             (void *)KIBLND_CONN_PARAM(event),
3149                                             KIBLND_CONN_PARAM_LEN(event));
3150                 CDEBUG(D_NET, "connreq: %d\n", rc);
3151                 return rc;
3152
3153         case RDMA_CM_EVENT_ADDR_ERROR:
3154                 peer_ni = cmid->context;
3155                 CNETERR("%s: ADDR ERROR %d\n",
3156                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3157                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3158                 kiblnd_peer_decref(peer_ni);
3159                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
3160
3161         case RDMA_CM_EVENT_ADDR_RESOLVED:
3162                 peer_ni = cmid->context;
3163
3164                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3165                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3166
3167                 if (event->status != 0) {
3168                         CNETERR("Can't resolve address for %s: %d\n",
3169                                 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3170                         rc = event->status;
3171                 } else {
3172                         rc = rdma_resolve_route(
3173                                 cmid, kiblnd_timeout() * 1000);
3174                         if (rc == 0) {
3175                                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3176                                 struct kib_dev *dev = net->ibn_dev;
3177
3178                                 CDEBUG(D_NET, "%s: connection bound to "\
3179                                        "%s:%pI4h:%s\n",
3180                                        libcfs_nid2str(peer_ni->ibp_nid),
3181                                        dev->ibd_ifname,
3182                                        &dev->ibd_ifip, cmid->device->name);
3183
3184                                 return 0;
3185                         }
3186
3187                         /* Can't initiate route resolution */
3188                         CERROR("Can't resolve route for %s: %d\n",
3189                                libcfs_nid2str(peer_ni->ibp_nid), rc);
3190                 }
3191                 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3192                 kiblnd_peer_decref(peer_ni);
3193                 return rc;                      /* rc != 0 destroys cmid */
3194
3195         case RDMA_CM_EVENT_ROUTE_ERROR:
3196                 peer_ni = cmid->context;
3197                 CNETERR("%s: ROUTE ERROR %d\n",
3198                         libcfs_nid2str(peer_ni->ibp_nid), event->status);
3199                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3200                 kiblnd_peer_decref(peer_ni);
3201                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
3202
3203         case RDMA_CM_EVENT_ROUTE_RESOLVED:
3204                 peer_ni = cmid->context;
3205                 CDEBUG(D_NET,"%s Route resolved: %d\n",
3206                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3207
3208                 if (event->status == 0)
3209                         return kiblnd_active_connect(cmid);
3210
3211                 CNETERR("Can't resolve route for %s: %d\n",
3212                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3213                 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3214                 kiblnd_peer_decref(peer_ni);
3215                 return event->status;           /* rc != 0 destroys cmid */
3216
3217         case RDMA_CM_EVENT_UNREACHABLE:
3218                 conn = cmid->context;
3219                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3220                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3221                 CNETERR("%s: UNREACHABLE %d\n",
3222                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3223                 kiblnd_connreq_done(conn, -ENETDOWN);
3224                 kiblnd_conn_decref(conn);
3225                 return 0;
3226
3227         case RDMA_CM_EVENT_CONNECT_ERROR:
3228                 conn = cmid->context;
3229                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3230                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3231                 CNETERR("%s: CONNECT ERROR %d\n",
3232                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3233                 kiblnd_connreq_done(conn, -ENOTCONN);
3234                 kiblnd_conn_decref(conn);
3235                 return 0;
3236
3237         case RDMA_CM_EVENT_REJECTED:
3238                 conn = cmid->context;
3239                 switch (conn->ibc_state) {
3240                 default:
3241                         LBUG();
3242
3243                 case IBLND_CONN_PASSIVE_WAIT:
3244                         CERROR ("%s: REJECTED %d\n",
3245                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3246                                 event->status);
3247                         kiblnd_connreq_done(conn, -ECONNRESET);
3248                         break;
3249
3250                 case IBLND_CONN_ACTIVE_CONNECT:
3251                         kiblnd_rejected(conn, event->status,
3252                                         (void *)KIBLND_CONN_PARAM(event),
3253                                         KIBLND_CONN_PARAM_LEN(event));
3254                         break;
3255                 }
3256                 kiblnd_conn_decref(conn);
3257                 return 0;
3258
3259         case RDMA_CM_EVENT_ESTABLISHED:
3260                 conn = cmid->context;
3261                 switch (conn->ibc_state) {
3262                 default:
3263                         LBUG();
3264
3265                 case IBLND_CONN_PASSIVE_WAIT:
3266                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3267                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3268                         kiblnd_connreq_done(conn, 0);
3269                         break;
3270
3271                 case IBLND_CONN_ACTIVE_CONNECT:
3272                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3273                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3274                         kiblnd_check_connreply(conn,
3275                                                (void *)KIBLND_CONN_PARAM(event),
3276                                                KIBLND_CONN_PARAM_LEN(event));
3277                         break;
3278                 }
3279                 /* net keeps its ref on conn! */
3280                 return 0;
3281
3282         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3283                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3284                 return 0;
3285
3286         case RDMA_CM_EVENT_DISCONNECTED:
3287                 conn = cmid->context;
3288                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3289                         CERROR("%s DISCONNECTED\n",
3290                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3291                         kiblnd_connreq_done(conn, -ECONNRESET);
3292                 } else {
3293                         kiblnd_close_conn(conn, 0);
3294                 }
3295                 kiblnd_conn_decref(conn);
3296                 cmid->context = NULL;
3297                 return 0;
3298
3299         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3300                 LCONSOLE_ERROR_MSG(0x131,
3301                                    "Received notification of device removal\n"
3302                                    "Please shutdown LNET to allow this to proceed\n");
3303                 /* Can't remove network from underneath LNET for now, so I have
3304                  * to ignore this */
3305                 return 0;
3306
3307         case RDMA_CM_EVENT_ADDR_CHANGE:
3308                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3309                 return 0;
3310         }
3311 }
3312
3313 static int
3314 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3315 {
3316         struct kib_tx *tx;
3317         struct list_head *ttmp;
3318
3319         list_for_each(ttmp, txs) {
3320                 tx = list_entry(ttmp, struct kib_tx, tx_list);
3321
3322                 if (txs != &conn->ibc_active_txs) {
3323                         LASSERT(tx->tx_queued);
3324                 } else {
3325                         LASSERT(!tx->tx_queued);
3326                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3327                 }
3328
3329                 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3330                         CERROR("Timed out tx: %s(WSQ:%d%d%d), %lld seconds\n",
3331                                kiblnd_queue2str(conn, txs),
3332                                tx->tx_waiting, tx->tx_sending, tx->tx_queued,
3333                                kiblnd_timeout() +
3334                                ktime_ms_delta(ktime_get(),
3335                                               tx->tx_deadline) / MSEC_PER_SEC);
3336                         return 1;
3337                 }
3338         }
3339
3340         return 0;
3341 }
3342
3343 static int
3344 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3345 {
3346         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3347                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3348                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3349                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3350                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3351 }
3352
3353 static void
3354 kiblnd_check_conns (int idx)
3355 {
3356         LIST_HEAD(closes);
3357         LIST_HEAD(checksends);
3358         LIST_HEAD(timedout_txs);
3359         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3360         struct list_head *ptmp;
3361         struct kib_peer_ni *peer_ni;
3362         struct kib_conn *conn;
3363         struct kib_tx *tx, *tx_tmp;
3364         struct list_head *ctmp;
3365         unsigned long     flags;
3366
3367         /* NB. We expect to have a look at all the peers and not find any
3368          * RDMAs to time out, so we just use a shared lock while we
3369          * take a look... */
3370         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3371
3372         list_for_each(ptmp, peers) {
3373                 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
3374
3375                 /* Check tx_deadline */
3376                 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3377                         if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3378                                 CWARN("Timed out tx for %s: %lld seconds\n",
3379                                       libcfs_nid2str(peer_ni->ibp_nid),
3380                                       ktime_ms_delta(ktime_get(),
3381                                                      tx->tx_deadline) / MSEC_PER_SEC);
3382                                 list_move(&tx->tx_list, &timedout_txs);
3383                         }
3384                 }
3385
3386                 list_for_each(ctmp, &peer_ni->ibp_conns) {
3387                         int timedout;
3388                         int sendnoop;
3389
3390                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
3391
3392                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3393
3394                         spin_lock(&conn->ibc_lock);
3395
3396                         sendnoop = kiblnd_need_noop(conn);
3397                         timedout = kiblnd_conn_timed_out_locked(conn);
3398                         if (!sendnoop && !timedout) {
3399                                 spin_unlock(&conn->ibc_lock);
3400                                 continue;
3401                         }
3402
3403                         if (timedout) {
3404                                 CERROR("Timed out RDMA with %s (%lld): "
3405                                        "c: %u, oc: %u, rc: %u\n",
3406                                        libcfs_nid2str(peer_ni->ibp_nid),
3407                                        ktime_get_seconds() - peer_ni->ibp_last_alive,
3408                                        conn->ibc_credits,
3409                                        conn->ibc_outstanding_credits,
3410                                        conn->ibc_reserved_credits);
3411                                 list_add(&conn->ibc_connd_list, &closes);
3412                         } else {
3413                                 list_add(&conn->ibc_connd_list, &checksends);
3414                         }
3415                         /* +ref for 'closes' or 'checksends' */
3416                         kiblnd_conn_addref(conn);
3417
3418                         spin_unlock(&conn->ibc_lock);
3419                 }
3420         }
3421
3422         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3423
3424         if (!list_empty(&timedout_txs))
3425                 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3426                                    LNET_MSG_STATUS_NETWORK_TIMEOUT);
3427
3428         /* Handle timeout by closing the whole
3429          * connection. We can only be sure RDMA activity
3430          * has ceased once the QP has been modified. */
3431         while (!list_empty(&closes)) {
3432                 conn = list_entry(closes.next,
3433                                   struct kib_conn, ibc_connd_list);
3434                 list_del(&conn->ibc_connd_list);
3435                 kiblnd_close_conn(conn, -ETIMEDOUT);
3436                 kiblnd_conn_decref(conn);
3437         }
3438
3439         /* In case we have enough credits to return via a
3440          * NOOP, but there were no non-blocking tx descs
3441          * free to do it last time... */
3442         while (!list_empty(&checksends)) {
3443                 conn = list_entry(checksends.next,
3444                                   struct kib_conn, ibc_connd_list);
3445                 list_del(&conn->ibc_connd_list);
3446
3447                 spin_lock(&conn->ibc_lock);
3448                 kiblnd_check_sends_locked(conn);
3449                 spin_unlock(&conn->ibc_lock);
3450
3451                 kiblnd_conn_decref(conn);
3452         }
3453 }
3454
3455 static void
3456 kiblnd_disconnect_conn(struct kib_conn *conn)
3457 {
3458         LASSERT (!in_interrupt());
3459         LASSERT (current == kiblnd_data.kib_connd);
3460         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3461
3462         rdma_disconnect(conn->ibc_cmid);
3463         kiblnd_finalise_conn(conn);
3464
3465         kiblnd_peer_notify(conn->ibc_peer);
3466 }
3467
3468 /*
3469  * High-water for reconnection to the same peer_ni, reconnection attempt should
3470  * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3471  */
3472 #define KIB_RECONN_HIGH_RACE    10
3473 /*
3474  * Allow connd to take a break and handle other things after consecutive
3475  * reconnection attemps.
3476  */
3477 #define KIB_RECONN_BREAK        100
3478
3479 int
3480 kiblnd_connd (void *arg)
3481 {
3482         spinlock_t        *lock= &kiblnd_data.kib_connd_lock;
3483         wait_queue_entry_t wait;
3484         unsigned long      flags;
3485         struct kib_conn *conn;
3486         int                timeout;
3487         int                i;
3488         int                dropped_lock;
3489         int                peer_index = 0;
3490         unsigned long      deadline = jiffies;
3491
3492         init_waitqueue_entry(&wait, current);
3493         kiblnd_data.kib_connd = current;
3494
3495         spin_lock_irqsave(lock, flags);
3496
3497         while (!kiblnd_data.kib_shutdown) {
3498                 int reconn = 0;
3499
3500                 dropped_lock = 0;
3501
3502                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3503                         struct kib_peer_ni *peer_ni = NULL;
3504
3505                         conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3506                                           struct kib_conn, ibc_list);
3507                         list_del(&conn->ibc_list);
3508                         if (conn->ibc_reconnect) {
3509                                 peer_ni = conn->ibc_peer;
3510                                 kiblnd_peer_addref(peer_ni);
3511                         }
3512
3513                         spin_unlock_irqrestore(lock, flags);
3514                         dropped_lock = 1;
3515
3516                         kiblnd_destroy_conn(conn);
3517
3518                         spin_lock_irqsave(lock, flags);
3519                         if (!peer_ni) {
3520                                 LIBCFS_FREE(conn, sizeof(*conn));
3521                                 continue;
3522                         }
3523
3524                         conn->ibc_peer = peer_ni;
3525                         if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3526                                 list_add_tail(&conn->ibc_list,
3527                                               &kiblnd_data.kib_reconn_list);
3528                         else
3529                                 list_add_tail(&conn->ibc_list,
3530                                               &kiblnd_data.kib_reconn_wait);
3531                 }
3532
3533                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3534                         int wait;
3535                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3536                                           struct kib_conn, ibc_list);
3537                         list_del(&conn->ibc_list);
3538
3539                         spin_unlock_irqrestore(lock, flags);
3540                         dropped_lock = 1;
3541
3542                         kiblnd_disconnect_conn(conn);
3543                         wait = conn->ibc_waits;
3544                         if (wait == 0) /* keep ref for connd_wait, see below */
3545                                 kiblnd_conn_decref(conn);
3546
3547                         spin_lock_irqsave(lock, flags);
3548
3549                         if (wait)
3550                                 list_add_tail(&conn->ibc_list,
3551                                               &kiblnd_data.kib_connd_waits);
3552                 }
3553
3554                 while (reconn < KIB_RECONN_BREAK) {
3555                         if (kiblnd_data.kib_reconn_sec !=
3556                             ktime_get_real_seconds()) {
3557                                 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3558                                 list_splice_init(&kiblnd_data.kib_reconn_wait,
3559                                                  &kiblnd_data.kib_reconn_list);
3560                         }
3561
3562                         if (list_empty(&kiblnd_data.kib_reconn_list))
3563                                 break;
3564
3565                         conn = list_entry(kiblnd_data.kib_reconn_list.next,
3566                                           struct kib_conn, ibc_list);
3567                         list_del(&conn->ibc_list);
3568
3569                         spin_unlock_irqrestore(lock, flags);
3570                         dropped_lock = 1;
3571
3572                         reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3573                         kiblnd_peer_decref(conn->ibc_peer);
3574                         LIBCFS_FREE(conn, sizeof(*conn));
3575
3576                         spin_lock_irqsave(lock, flags);
3577                 }
3578
3579                 if (!list_empty(&kiblnd_data.kib_connd_waits)) {
3580                         conn = list_entry(kiblnd_data.kib_connd_waits.next,
3581                                           struct kib_conn, ibc_list);
3582                         list_del(&conn->ibc_list);
3583                         spin_unlock_irqrestore(lock, flags);
3584
3585                         dropped_lock = kiblnd_tx_may_discard(conn);
3586                         if (dropped_lock)
3587                                 kiblnd_conn_decref(conn);
3588
3589                         spin_lock_irqsave(lock, flags);
3590                         if (dropped_lock == 0)
3591                                 list_add_tail(&conn->ibc_list,
3592                                               &kiblnd_data.kib_connd_waits);
3593                 }
3594
3595                 /* careful with the jiffy wrap... */
3596                 timeout = (int)(deadline - jiffies);
3597                 if (timeout <= 0) {
3598                         const int n = 4;
3599                         const int p = 1;
3600                         int       chunk = kiblnd_data.kib_peer_hash_size;
3601                         unsigned int lnd_timeout;
3602
3603                         spin_unlock_irqrestore(lock, flags);
3604                         dropped_lock = 1;
3605
3606                         /* Time to check for RDMA timeouts on a few more
3607                          * peers: I do checks every 'p' seconds on a
3608                          * proportion of the peer_ni table and I need to check
3609                          * every connection 'n' times within a timeout
3610                          * interval, to ensure I detect a timeout on any
3611                          * connection within (n+1)/n times the timeout
3612                          * interval. */
3613
3614                         lnd_timeout = kiblnd_timeout();
3615                         if (lnd_timeout > n * p)
3616                                 chunk = (chunk * n * p) / lnd_timeout;
3617                         if (chunk == 0)
3618                                 chunk = 1;
3619
3620                         for (i = 0; i < chunk; i++) {
3621                                 kiblnd_check_conns(peer_index);
3622                                 peer_index = (peer_index + 1) %
3623                                              kiblnd_data.kib_peer_hash_size;
3624                         }
3625
3626                         deadline += cfs_time_seconds(p);
3627                         spin_lock_irqsave(lock, flags);
3628                 }
3629
3630                 if (dropped_lock)
3631                         continue;
3632
3633                 /* Nothing to do for 'timeout'  */
3634                 set_current_state(TASK_INTERRUPTIBLE);
3635                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3636                 spin_unlock_irqrestore(lock, flags);
3637
3638                 schedule_timeout(timeout);
3639
3640                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3641                 spin_lock_irqsave(lock, flags);
3642         }
3643
3644         spin_unlock_irqrestore(lock, flags);
3645
3646         kiblnd_thread_fini();
3647         return 0;
3648 }
3649
3650 void
3651 kiblnd_qp_event(struct ib_event *event, void *arg)
3652 {
3653         struct kib_conn *conn = arg;
3654
3655         switch (event->event) {
3656         case IB_EVENT_COMM_EST:
3657                 CDEBUG(D_NET, "%s established\n",
3658                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3659                 /* We received a packet but connection isn't established
3660                  * probably handshake packet was lost, so free to
3661                  * force make connection established */
3662                 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3663                 return;
3664
3665         case IB_EVENT_PORT_ERR:
3666         case IB_EVENT_DEVICE_FATAL:
3667                 CERROR("Fatal device error for NI %s\n",
3668                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3669                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3670                 return;
3671
3672         case IB_EVENT_PORT_ACTIVE:
3673                 CERROR("Port reactivated for NI %s\n",
3674                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3675                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3676                 return;
3677
3678         default:
3679                 CERROR("%s: Async QP event type %d\n",
3680                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3681                 return;
3682         }
3683 }
3684
3685 static void
3686 kiblnd_complete (struct ib_wc *wc)
3687 {
3688         switch (kiblnd_wreqid2type(wc->wr_id)) {
3689         default:
3690                 LBUG();
3691
3692         case IBLND_WID_MR:
3693                 if (wc->status != IB_WC_SUCCESS &&
3694                     wc->status != IB_WC_WR_FLUSH_ERR)
3695                         CNETERR("FastReg failed: %d\n", wc->status);
3696                 return;
3697
3698         case IBLND_WID_RDMA:
3699                 /* We only get RDMA completion notification if it fails.  All
3700                  * subsequent work items, including the final SEND will fail
3701                  * too.  However we can't print out any more info about the
3702                  * failing RDMA because 'tx' might be back on the idle list or
3703                  * even reused already if we didn't manage to post all our work
3704                  * items */
3705                 CNETERR("RDMA (tx: %p) failed: %d\n",
3706                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3707                 return;
3708
3709         case IBLND_WID_TX:
3710                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3711                 return;
3712
3713         case IBLND_WID_RX:
3714                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3715                                    wc->byte_len);
3716                 return;
3717         }
3718 }
3719
3720 void
3721 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3722 {
3723         /* NB I'm not allowed to schedule this conn once its refcount has
3724          * reached 0.  Since fundamentally I'm racing with scheduler threads
3725          * consuming my CQ I could be called after all completions have
3726          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3727          * and this CQ is about to be destroyed so I NOOP. */
3728         struct kib_conn *conn = arg;
3729         struct kib_sched_info *sched = conn->ibc_sched;
3730         unsigned long flags;
3731
3732         LASSERT(cq == conn->ibc_cq);
3733
3734         spin_lock_irqsave(&sched->ibs_lock, flags);
3735
3736         conn->ibc_ready = 1;
3737
3738         if (!conn->ibc_scheduled &&
3739             (conn->ibc_nrx > 0 ||
3740              conn->ibc_nsends_posted > 0)) {
3741                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3742                 conn->ibc_scheduled = 1;
3743                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3744
3745                 if (waitqueue_active(&sched->ibs_waitq))
3746                         wake_up(&sched->ibs_waitq);
3747         }
3748
3749         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3750 }
3751
3752 void
3753 kiblnd_cq_event(struct ib_event *event, void *arg)
3754 {
3755         struct kib_conn *conn = arg;
3756
3757         CERROR("%s: async CQ event type %d\n",
3758                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3759 }
3760
3761 int
3762 kiblnd_scheduler(void *arg)
3763 {
3764         long                    id = (long)arg;
3765         struct kib_sched_info   *sched;
3766         struct kib_conn *conn;
3767         wait_queue_entry_t      wait;
3768         unsigned long           flags;
3769         struct ib_wc            wc;
3770         int                     did_something;
3771         int                     rc;
3772
3773         init_waitqueue_entry(&wait, current);
3774
3775         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3776
3777         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3778         if (rc != 0) {
3779                 CWARN("Unable to bind on CPU partition %d, please verify "
3780                       "whether all CPUs are healthy and reload modules if "
3781                       "necessary, otherwise your system might under risk of "
3782                       "low performance\n", sched->ibs_cpt);
3783         }
3784
3785         spin_lock_irqsave(&sched->ibs_lock, flags);
3786
3787         while (!kiblnd_data.kib_shutdown) {
3788                 if (need_resched()) {
3789                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3790
3791                         cond_resched();
3792
3793                         spin_lock_irqsave(&sched->ibs_lock, flags);
3794                 }
3795
3796                 did_something = 0;
3797
3798                 if (!list_empty(&sched->ibs_conns)) {
3799                         conn = list_entry(sched->ibs_conns.next,
3800                                           struct kib_conn, ibc_sched_list);
3801                         /* take over kib_sched_conns' ref on conn... */
3802                         LASSERT(conn->ibc_scheduled);
3803                         list_del(&conn->ibc_sched_list);
3804                         conn->ibc_ready = 0;
3805
3806                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3807
3808                         wc.wr_id = IBLND_WID_INVAL;
3809
3810                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3811                         if (rc == 0) {
3812                                 rc = ib_req_notify_cq(conn->ibc_cq,
3813                                                       IB_CQ_NEXT_COMP);
3814                                 if (rc < 0) {
3815                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3816                                               "closing connection\n",
3817                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3818                                         kiblnd_close_conn(conn, -EIO);
3819                                         kiblnd_conn_decref(conn);
3820                                         spin_lock_irqsave(&sched->ibs_lock,
3821                                                               flags);
3822                                         continue;
3823                                 }
3824
3825                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3826                         }
3827
3828                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3829                                 LCONSOLE_ERROR(
3830                                         "ib_poll_cq (rc: %d) returned invalid "
3831                                         "wr_id, opcode %d, status: %d, "
3832                                         "vendor_err: %d, conn: %s status: %d\n"
3833                                         "please upgrade firmware and OFED or "
3834                                         "contact vendor.\n", rc,
3835                                         wc.opcode, wc.status, wc.vendor_err,
3836                                         libcfs_nid2str(conn->ibc_peer->ibp_nid),
3837                                         conn->ibc_state);
3838                                 rc = -EINVAL;
3839                         }
3840
3841                         if (rc < 0) {
3842                                 CWARN("%s: ib_poll_cq failed: %d, "
3843                                       "closing connection\n",
3844                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3845                                       rc);
3846                                 kiblnd_close_conn(conn, -EIO);
3847                                 kiblnd_conn_decref(conn);
3848                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3849                                 continue;
3850                         }
3851
3852                         spin_lock_irqsave(&sched->ibs_lock, flags);
3853
3854                         if (rc != 0 || conn->ibc_ready) {
3855                                 /* There may be another completion waiting; get
3856                                  * another scheduler to check while I handle
3857                                  * this one... */
3858                                 /* +1 ref for sched_conns */
3859                                 kiblnd_conn_addref(conn);
3860                                 list_add_tail(&conn->ibc_sched_list,
3861                                                   &sched->ibs_conns);
3862                                 if (waitqueue_active(&sched->ibs_waitq))
3863                                         wake_up(&sched->ibs_waitq);
3864                         } else {
3865                                 conn->ibc_scheduled = 0;
3866                         }
3867
3868                         if (rc != 0) {
3869                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3870                                 kiblnd_complete(&wc);
3871
3872                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3873                         }
3874
3875                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3876                         did_something = 1;
3877                 }
3878
3879                 if (did_something)
3880                         continue;
3881
3882                 set_current_state(TASK_INTERRUPTIBLE);
3883                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3884                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3885
3886                 schedule();
3887
3888                 remove_wait_queue(&sched->ibs_waitq, &wait);
3889                 set_current_state(TASK_RUNNING);
3890                 spin_lock_irqsave(&sched->ibs_lock, flags);
3891         }
3892
3893         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3894
3895         kiblnd_thread_fini();
3896         return 0;
3897 }
3898
3899 int
3900 kiblnd_failover_thread(void *arg)
3901 {
3902         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
3903         struct kib_dev *dev;
3904         struct net *ns = arg;
3905         wait_queue_entry_t wait;
3906         unsigned long    flags;
3907         int              rc;
3908
3909         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3910
3911         init_waitqueue_entry(&wait, current);
3912         write_lock_irqsave(glock, flags);
3913
3914         while (!kiblnd_data.kib_shutdown) {
3915                 int     do_failover = 0;
3916                 int     long_sleep;
3917
3918                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3919                                     ibd_fail_list) {
3920                         if (ktime_get_seconds() < dev->ibd_next_failover)
3921                                 continue;
3922                         do_failover = 1;
3923                         break;
3924                 }
3925
3926                 if (do_failover) {
3927                         list_del_init(&dev->ibd_fail_list);
3928                         dev->ibd_failover = 1;
3929                         write_unlock_irqrestore(glock, flags);
3930
3931                         rc = kiblnd_dev_failover(dev, ns);
3932
3933                         write_lock_irqsave(glock, flags);
3934
3935                         LASSERT (dev->ibd_failover);
3936                         dev->ibd_failover = 0;
3937                         if (rc >= 0) { /* Device is OK or failover succeed */
3938                                 dev->ibd_next_failover = ktime_get_seconds() + 3;
3939                                 continue;
3940                         }
3941
3942                         /* failed to failover, retry later */
3943                         dev->ibd_next_failover = ktime_get_seconds() +
3944                                                  min(dev->ibd_failed_failover, 10);
3945                         if (kiblnd_dev_can_failover(dev)) {
3946                                 list_add_tail(&dev->ibd_fail_list,
3947                                               &kiblnd_data.kib_failed_devs);
3948                         }
3949
3950                         continue;
3951                 }
3952
3953                 /* long sleep if no more pending failover */
3954                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3955
3956                 set_current_state(TASK_INTERRUPTIBLE);
3957                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3958                 write_unlock_irqrestore(glock, flags);
3959
3960                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3961                                                    cfs_time_seconds(1));
3962                 set_current_state(TASK_RUNNING);
3963                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3964                 write_lock_irqsave(glock, flags);
3965
3966                 if (!long_sleep || rc != 0)
3967                         continue;
3968
3969                 /* have a long sleep, routine check all active devices,
3970                  * we need checking like this because if there is not active
3971                  * connection on the dev and no SEND from local, we may listen
3972                  * on wrong HCA for ever while there is a bonding failover */
3973                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3974                         if (kiblnd_dev_can_failover(dev)) {
3975                                 list_add_tail(&dev->ibd_fail_list,
3976                                               &kiblnd_data.kib_failed_devs);
3977                         }
3978                 }
3979         }
3980
3981         write_unlock_irqrestore(glock, flags);
3982
3983         kiblnd_thread_fini();
3984         return 0;
3985 }