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         /* allow the port to be reused */
1321         rc = rdma_set_reuseaddr(cmid, 1);
1322         if (rc != 0) {
1323                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1324                 return rc;
1325         }
1326
1327         /* look for a free privileged port */
1328         for (port = PROT_SOCK-1; port > 0; port--) {
1329                 srcaddr->sin_port = htons(port);
1330                 rc = rdma_resolve_addr(cmid,
1331                                        (struct sockaddr *)srcaddr,
1332                                        (struct sockaddr *)dstaddr,
1333                                        timeout_ms);
1334                 if (rc == 0) {
1335                         CDEBUG(D_NET, "bound to port %hu\n", port);
1336                         return 0;
1337                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1338                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1339                                port, rc);
1340                 } else {
1341                         return rc;
1342                 }
1343         }
1344
1345         CERROR("cannot bind to a free privileged port: rc = %d\n", rc);
1346
1347         return rc;
1348 }
1349
1350 static int
1351 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1352                     struct sockaddr_in *srcaddr,
1353                     struct sockaddr_in *dstaddr,
1354                     int timeout_ms)
1355 {
1356         const struct cred *old_creds = NULL;
1357         struct cred *new_creds;
1358         int rc;
1359
1360         if (!capable(CAP_NET_BIND_SERVICE)) {
1361                 new_creds = prepare_kernel_cred(NULL);
1362                 if (!new_creds)
1363                         return -ENOMEM;
1364
1365                 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1366                 old_creds = override_creds(new_creds);
1367         }
1368
1369         rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1370
1371         if (old_creds)
1372                 revert_creds(old_creds);
1373
1374         return rc;
1375 }
1376
1377 static void
1378 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1379 {
1380         struct rdma_cm_id *cmid;
1381         struct kib_dev *dev;
1382         struct kib_net *net = peer_ni->ibp_ni->ni_data;
1383         struct sockaddr_in srcaddr;
1384         struct sockaddr_in dstaddr;
1385         int rc;
1386
1387         LASSERT (net != NULL);
1388         LASSERT (peer_ni->ibp_connecting > 0);
1389
1390         cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1391                                      kiblnd_cm_callback, peer_ni,
1392                                      RDMA_PS_TCP, IB_QPT_RC);
1393
1394         if (IS_ERR(cmid)) {
1395                 CERROR("Can't create CMID for %s: %ld\n",
1396                        libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1397                 rc = PTR_ERR(cmid);
1398                 goto failed;
1399         }
1400
1401         dev = net->ibn_dev;
1402         memset(&srcaddr, 0, sizeof(srcaddr));
1403         srcaddr.sin_family = AF_INET;
1404         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1405
1406         memset(&dstaddr, 0, sizeof(dstaddr));
1407         dstaddr.sin_family = AF_INET;
1408         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1409         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1410
1411         kiblnd_peer_addref(peer_ni);               /* cmid's ref */
1412
1413         if (*kiblnd_tunables.kib_use_priv_port) {
1414                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1415                                          kiblnd_timeout() * 1000);
1416         } else {
1417                 rc = rdma_resolve_addr(cmid,
1418                                        (struct sockaddr *)&srcaddr,
1419                                        (struct sockaddr *)&dstaddr,
1420                                        kiblnd_timeout() * 1000);
1421         }
1422         if (rc != 0) {
1423                 /* Can't initiate address resolution:  */
1424                 CERROR("Can't resolve addr for %s: %d\n",
1425                        libcfs_nid2str(peer_ni->ibp_nid), rc);
1426                 goto failed2;
1427         }
1428
1429         return;
1430
1431  failed2:
1432         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1433         kiblnd_peer_decref(peer_ni);               /* cmid's ref */
1434         rdma_destroy_id(cmid);
1435         return;
1436  failed:
1437         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1438 }
1439
1440 bool
1441 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1442 {
1443         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1444         char *reason = NULL;
1445         LIST_HEAD(txs);
1446         unsigned long flags;
1447
1448         write_lock_irqsave(glock, flags);
1449         if (peer_ni->ibp_reconnecting == 0) {
1450                 if (peer_ni->ibp_accepting)
1451                         reason = "accepting";
1452                 else if (peer_ni->ibp_connecting)
1453                         reason = "connecting";
1454                 else if (!list_empty(&peer_ni->ibp_conns))
1455                         reason = "connected";
1456                 else /* connected then closed */
1457                         reason = "closed";
1458
1459                 goto no_reconnect;
1460         }
1461
1462         if (peer_ni->ibp_accepting)
1463                 CNETERR("Detecting race between accepting and reconnecting\n");
1464         peer_ni->ibp_reconnecting--;
1465
1466         if (!kiblnd_peer_active(peer_ni)) {
1467                 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1468                 reason = "unlinked";
1469                 goto no_reconnect;
1470         }
1471
1472         peer_ni->ibp_connecting++;
1473         peer_ni->ibp_reconnected++;
1474
1475         write_unlock_irqrestore(glock, flags);
1476
1477         kiblnd_connect_peer(peer_ni);
1478         return true;
1479
1480  no_reconnect:
1481         write_unlock_irqrestore(glock, flags);
1482
1483         CWARN("Abort reconnection of %s: %s\n",
1484               libcfs_nid2str(peer_ni->ibp_nid), reason);
1485         kiblnd_txlist_done(&txs, -ECONNABORTED,
1486                            LNET_MSG_STATUS_LOCAL_ABORTED);
1487         return false;
1488 }
1489
1490 void
1491 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1492 {
1493         struct kib_peer_ni *peer_ni;
1494         struct kib_peer_ni *peer2;
1495         struct kib_conn *conn;
1496         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1497         unsigned long      flags;
1498         int                rc;
1499         int                i;
1500         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1501
1502         /* If I get here, I've committed to send, so I complete the tx with
1503          * failure on any problems */
1504
1505         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1506         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1507
1508         /* First time, just use a read lock since I expect to find my peer_ni
1509          * connected */
1510         read_lock_irqsave(g_lock, flags);
1511
1512         peer_ni = kiblnd_find_peer_locked(ni, nid);
1513         if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1514                 /* Found a peer_ni with an established connection */
1515                 conn = kiblnd_get_conn_locked(peer_ni);
1516                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1517
1518                 read_unlock_irqrestore(g_lock, flags);
1519
1520                 if (tx != NULL)
1521                         kiblnd_queue_tx(tx, conn);
1522                 kiblnd_conn_decref(conn); /* ...to here */
1523                 return;
1524         }
1525
1526         read_unlock(g_lock);
1527         /* Re-try with a write lock */
1528         write_lock(g_lock);
1529
1530         peer_ni = kiblnd_find_peer_locked(ni, nid);
1531         if (peer_ni != NULL) {
1532                 if (list_empty(&peer_ni->ibp_conns)) {
1533                         /* found a peer_ni, but it's still connecting... */
1534                         LASSERT(kiblnd_peer_connecting(peer_ni));
1535                         if (tx != NULL)
1536                                 list_add_tail(&tx->tx_list,
1537                                                   &peer_ni->ibp_tx_queue);
1538                         write_unlock_irqrestore(g_lock, flags);
1539                 } else {
1540                         conn = kiblnd_get_conn_locked(peer_ni);
1541                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1542
1543                         write_unlock_irqrestore(g_lock, flags);
1544
1545                         if (tx != NULL)
1546                                 kiblnd_queue_tx(tx, conn);
1547                         kiblnd_conn_decref(conn); /* ...to here */
1548                 }
1549                 return;
1550         }
1551
1552         write_unlock_irqrestore(g_lock, flags);
1553
1554         /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1555         rc = kiblnd_create_peer(ni, &peer_ni, nid);
1556         if (rc != 0) {
1557                 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1558                 if (tx != NULL) {
1559                         tx->tx_status = -EHOSTUNREACH;
1560                         tx->tx_waiting = 0;
1561                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1562                         kiblnd_tx_done(tx);
1563                 }
1564                 return;
1565         }
1566
1567         write_lock_irqsave(g_lock, flags);
1568
1569         peer2 = kiblnd_find_peer_locked(ni, nid);
1570         if (peer2 != NULL) {
1571                 if (list_empty(&peer2->ibp_conns)) {
1572                         /* found a peer_ni, but it's still connecting... */
1573                         LASSERT(kiblnd_peer_connecting(peer2));
1574                         if (tx != NULL)
1575                                 list_add_tail(&tx->tx_list,
1576                                                   &peer2->ibp_tx_queue);
1577                         write_unlock_irqrestore(g_lock, flags);
1578                 } else {
1579                         conn = kiblnd_get_conn_locked(peer2);
1580                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1581
1582                         write_unlock_irqrestore(g_lock, flags);
1583
1584                         if (tx != NULL)
1585                                 kiblnd_queue_tx(tx, conn);
1586                         kiblnd_conn_decref(conn); /* ...to here */
1587                 }
1588
1589                 kiblnd_peer_decref(peer_ni);
1590                 return;
1591         }
1592
1593         /* Brand new peer_ni */
1594         LASSERT(peer_ni->ibp_connecting == 0);
1595         tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1596         peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1597
1598         /* always called with a ref on ni, which prevents ni being shutdown */
1599         LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1600
1601         if (tx != NULL)
1602                 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1603
1604         kiblnd_peer_addref(peer_ni);
1605         list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1606
1607         write_unlock_irqrestore(g_lock, flags);
1608
1609         for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1610                 kiblnd_connect_peer(peer_ni);
1611         kiblnd_peer_decref(peer_ni);
1612 }
1613
1614 int
1615 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1616 {
1617         struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1618         int               type = lntmsg->msg_type;
1619         struct lnet_process_id target = lntmsg->msg_target;
1620         int               target_is_router = lntmsg->msg_target_is_router;
1621         int               routing = lntmsg->msg_routing;
1622         unsigned int      payload_niov = lntmsg->msg_niov;
1623         struct bio_vec   *payload_kiov = lntmsg->msg_kiov;
1624         unsigned int      payload_offset = lntmsg->msg_offset;
1625         unsigned int      payload_nob = lntmsg->msg_len;
1626         struct kib_msg *ibmsg;
1627         struct kib_rdma_desc *rd;
1628         struct kib_tx *tx;
1629         int               nob;
1630         int               rc;
1631
1632         /* NB 'private' is different depending on what we're sending.... */
1633
1634         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1635                payload_nob, payload_niov, libcfs_id2str(target));
1636
1637         LASSERT (payload_nob == 0 || payload_niov > 0);
1638         LASSERT (payload_niov <= LNET_MAX_IOV);
1639
1640         /* Thread context */
1641         LASSERT (!in_interrupt());
1642
1643         switch (type) {
1644         default:
1645                 LBUG();
1646                 return (-EIO);
1647
1648         case LNET_MSG_ACK:
1649                 LASSERT (payload_nob == 0);
1650                 break;
1651
1652         case LNET_MSG_GET:
1653                 if (routing || target_is_router)
1654                         break;                  /* send IMMEDIATE */
1655
1656                 /* is the REPLY message too small for RDMA? */
1657                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1658                 if (nob <= IBLND_MSG_SIZE)
1659                         break;                  /* send IMMEDIATE */
1660
1661                 tx = kiblnd_get_idle_tx(ni, target.nid);
1662                 if (tx == NULL) {
1663                         CERROR("Can't allocate txd for GET to %s\n",
1664                                libcfs_nid2str(target.nid));
1665                         return -ENOMEM;
1666                 }
1667
1668                 ibmsg = tx->tx_msg;
1669                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1670                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1671                                           lntmsg->msg_md->md_niov,
1672                                           lntmsg->msg_md->md_kiov,
1673                                           0, lntmsg->msg_md->md_length);
1674                 if (rc != 0) {
1675                         CERROR("Can't setup GET sink for %s: %d\n",
1676                                libcfs_nid2str(target.nid), rc);
1677                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1678                         kiblnd_tx_done(tx);
1679                         return -EIO;
1680                 }
1681
1682                 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1683                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1684                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1685
1686                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1687
1688                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1689                 if (tx->tx_lntmsg[1] == NULL) {
1690                         CERROR("Can't create reply for GET -> %s\n",
1691                                libcfs_nid2str(target.nid));
1692                         kiblnd_tx_done(tx);
1693                         return -EIO;
1694                 }
1695
1696                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1697                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1698                 kiblnd_launch_tx(ni, tx, target.nid);
1699                 return 0;
1700
1701         case LNET_MSG_REPLY:
1702         case LNET_MSG_PUT:
1703                 /* Is the payload small enough not to need RDMA? */
1704                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1705                 if (nob <= IBLND_MSG_SIZE)
1706                         break;                  /* send IMMEDIATE */
1707
1708                 tx = kiblnd_get_idle_tx(ni, target.nid);
1709                 if (tx == NULL) {
1710                         CERROR("Can't allocate %s txd for %s\n",
1711                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1712                                libcfs_nid2str(target.nid));
1713                         return -ENOMEM;
1714                 }
1715
1716                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1717                                           payload_niov, payload_kiov,
1718                                           payload_offset, payload_nob);
1719                 if (rc != 0) {
1720                         CERROR("Can't setup PUT src for %s: %d\n",
1721                                libcfs_nid2str(target.nid), rc);
1722                         kiblnd_tx_done(tx);
1723                         return -EIO;
1724                 }
1725
1726                 ibmsg = tx->tx_msg;
1727                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1728                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1729                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1730                                    sizeof(struct kib_putreq_msg));
1731
1732                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1733                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1734                 kiblnd_launch_tx(ni, tx, target.nid);
1735                 return 0;
1736         }
1737
1738         /* send IMMEDIATE */
1739         LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1740                 <= IBLND_MSG_SIZE);
1741
1742         tx = kiblnd_get_idle_tx(ni, target.nid);
1743         if (tx == NULL) {
1744                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1745                         type, libcfs_nid2str(target.nid));
1746                 return -ENOMEM;
1747         }
1748
1749         ibmsg = tx->tx_msg;
1750         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1751
1752         lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1753                             offsetof(struct kib_msg,
1754                                      ibm_u.immediate.ibim_payload),
1755                             payload_niov, payload_kiov,
1756                             payload_offset, payload_nob);
1757
1758         nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1759         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1760
1761         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1762         kiblnd_launch_tx(ni, tx, target.nid);
1763         return 0;
1764 }
1765
1766 static void
1767 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1768 {
1769         struct lnet_process_id target = lntmsg->msg_target;
1770         unsigned int niov = lntmsg->msg_niov;
1771         struct bio_vec *kiov = lntmsg->msg_kiov;
1772         unsigned int offset = lntmsg->msg_offset;
1773         unsigned int nob = lntmsg->msg_len;
1774         struct kib_tx *tx;
1775         int rc;
1776
1777         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1778         if (tx == NULL) {
1779                 CERROR("Can't get tx for REPLY to %s\n",
1780                        libcfs_nid2str(target.nid));
1781                 goto failed_0;
1782         }
1783
1784         if (nob == 0)
1785                 rc = 0;
1786         else
1787                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1788                                           niov, kiov, offset, nob);
1789
1790         if (rc != 0) {
1791                 CERROR("Can't setup GET src for %s: %d\n",
1792                        libcfs_nid2str(target.nid), rc);
1793                 goto failed_1;
1794         }
1795
1796         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1797                               IBLND_MSG_GET_DONE, nob,
1798                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1799                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1800         if (rc < 0) {
1801                 CERROR("Can't setup rdma for GET from %s: %d\n",
1802                        libcfs_nid2str(target.nid), rc);
1803                 goto failed_1;
1804         }
1805
1806         if (nob == 0) {
1807                 /* No RDMA: local completion may happen now! */
1808                 lnet_finalize(lntmsg, 0);
1809         } else {
1810                 /* RDMA: lnet_finalize(lntmsg) when it
1811                  * completes */
1812                 tx->tx_lntmsg[0] = lntmsg;
1813         }
1814
1815         kiblnd_queue_tx(tx, rx->rx_conn);
1816         return;
1817
1818
1819 failed_1:
1820         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1821         kiblnd_tx_done(tx);
1822 failed_0:
1823         lnet_finalize(lntmsg, -EIO);
1824 }
1825
1826 int
1827 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1828             int delayed, unsigned int niov, struct bio_vec *kiov,
1829             unsigned int offset, unsigned int mlen, unsigned int rlen)
1830 {
1831         struct kib_rx *rx = private;
1832         struct kib_msg *rxmsg = rx->rx_msg;
1833         struct kib_conn *conn = rx->rx_conn;
1834         struct kib_tx *tx;
1835         __u64        ibprm_cookie;
1836         int          nob;
1837         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1838         int          rc = 0;
1839
1840         LASSERT (mlen <= rlen);
1841         LASSERT (!in_interrupt());
1842
1843         switch (rxmsg->ibm_type) {
1844         default:
1845                 LBUG();
1846
1847         case IBLND_MSG_IMMEDIATE:
1848                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1849                 if (nob > rx->rx_nob) {
1850                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1851                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1852                                 nob, rx->rx_nob);
1853                         rc = -EPROTO;
1854                         break;
1855                 }
1856
1857                 lnet_copy_flat2kiov(niov, kiov, offset,
1858                                     IBLND_MSG_SIZE, rxmsg,
1859                                     offsetof(struct kib_msg,
1860                                              ibm_u.immediate.ibim_payload),
1861                                     mlen);
1862                 lnet_finalize(lntmsg, 0);
1863                 break;
1864
1865         case IBLND_MSG_PUT_REQ: {
1866                 struct kib_msg  *txmsg;
1867                 struct kib_rdma_desc *rd;
1868                 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1869
1870                 if (mlen == 0) {
1871                         lnet_finalize(lntmsg, 0);
1872                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1873                                                0, ibprm_cookie);
1874                         break;
1875                 }
1876
1877                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1878                 if (tx == NULL) {
1879                         CERROR("Can't allocate tx for %s\n",
1880                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1881                         /* Not replying will break the connection */
1882                         rc = -ENOMEM;
1883                         break;
1884                 }
1885
1886                 txmsg = tx->tx_msg;
1887                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1888                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1889                                           niov, kiov, offset, mlen);
1890                 if (rc != 0) {
1891                         CERROR("Can't setup PUT sink for %s: %d\n",
1892                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1893                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1894                         kiblnd_tx_done(tx);
1895                         /* tell peer_ni it's over */
1896                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1897                                                rc, ibprm_cookie);
1898                         break;
1899                 }
1900
1901                 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1902                 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1903                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1904
1905                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1906
1907                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1908                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1909                 kiblnd_queue_tx(tx, conn);
1910
1911                 /* reposted buffer reserved for PUT_DONE */
1912                 post_credit = IBLND_POSTRX_NO_CREDIT;
1913                 break;
1914                 }
1915
1916         case IBLND_MSG_GET_REQ:
1917                 if (lntmsg != NULL) {
1918                         /* Optimized GET; RDMA lntmsg's payload */
1919                         kiblnd_reply(ni, rx, lntmsg);
1920                 } else {
1921                         /* GET didn't match anything */
1922                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1923                                                -ENODATA,
1924                                                rxmsg->ibm_u.get.ibgm_cookie);
1925                 }
1926                 break;
1927         }
1928
1929         kiblnd_post_rx(rx, post_credit);
1930         return rc;
1931 }
1932
1933 int
1934 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1935 {
1936         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1937
1938         if (IS_ERR(task))
1939                 return PTR_ERR(task);
1940
1941         atomic_inc(&kiblnd_data.kib_nthreads);
1942         return 0;
1943 }
1944
1945 static void
1946 kiblnd_thread_fini (void)
1947 {
1948         atomic_dec (&kiblnd_data.kib_nthreads);
1949 }
1950
1951 static void
1952 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1953 {
1954         /* This is racy, but everyone's only writing ktime_get_seconds() */
1955         peer_ni->ibp_last_alive = ktime_get_seconds();
1956         smp_mb();
1957 }
1958
1959 static void
1960 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1961 {
1962         int           error = 0;
1963         time64_t last_alive = 0;
1964         unsigned long flags;
1965
1966         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1967
1968         if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1969                 error = peer_ni->ibp_error;
1970                 peer_ni->ibp_error = 0;
1971
1972                 last_alive = peer_ni->ibp_last_alive;
1973         }
1974
1975         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1976
1977         if (error != 0)
1978                 lnet_notify(peer_ni->ibp_ni,
1979                             peer_ni->ibp_nid, false, false, last_alive);
1980 }
1981
1982 void
1983 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1984 {
1985         /* This just does the immediate housekeeping.  'error' is zero for a
1986          * normal shutdown which can happen only after the connection has been
1987          * established.  If the connection is established, schedule the
1988          * connection to be finished off by the connd.  Otherwise the connd is
1989          * already dealing with it (either to set it up or tear it down).
1990          * Caller holds kib_global_lock exclusively in irq context */
1991         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1992         struct kib_dev *dev;
1993         unsigned long flags;
1994
1995         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1996
1997         if (error != 0 && conn->ibc_comms_error == 0)
1998                 conn->ibc_comms_error = error;
1999
2000         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
2001                 return; /* already being handled  */
2002
2003         if (error == 0 &&
2004             list_empty(&conn->ibc_tx_noops) &&
2005             list_empty(&conn->ibc_tx_queue) &&
2006             list_empty(&conn->ibc_tx_queue_rsrvd) &&
2007             list_empty(&conn->ibc_tx_queue_nocred) &&
2008             list_empty(&conn->ibc_active_txs)) {
2009                 CDEBUG(D_NET, "closing conn to %s\n",
2010                        libcfs_nid2str(peer_ni->ibp_nid));
2011         } else {
2012                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
2013                        libcfs_nid2str(peer_ni->ibp_nid), error,
2014                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2015                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2016                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
2017                                                 "" : "(sending_rsrvd)",
2018                        list_empty(&conn->ibc_tx_queue_nocred) ?
2019                                                  "" : "(sending_nocred)",
2020                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2021         }
2022
2023         dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2024         if (peer_ni->ibp_next_conn == conn)
2025                 /* clear next_conn so it won't be used */
2026                 peer_ni->ibp_next_conn = NULL;
2027         list_del(&conn->ibc_list);
2028         /* connd (see below) takes over ibc_list's ref */
2029
2030         if (list_empty(&peer_ni->ibp_conns) &&    /* no more conns */
2031             kiblnd_peer_active(peer_ni)) {         /* still in peer_ni table */
2032                 kiblnd_unlink_peer_locked(peer_ni);
2033
2034                 /* set/clear error on last conn */
2035                 peer_ni->ibp_error = conn->ibc_comms_error;
2036         }
2037
2038         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2039
2040         if (error != 0 &&
2041             kiblnd_dev_can_failover(dev)) {
2042                 list_add_tail(&dev->ibd_fail_list,
2043                               &kiblnd_data.kib_failed_devs);
2044                 wake_up(&kiblnd_data.kib_failover_waitq);
2045         }
2046
2047         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2048
2049         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2050         wake_up(&kiblnd_data.kib_connd_waitq);
2051
2052         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2053 }
2054
2055 void
2056 kiblnd_close_conn(struct kib_conn *conn, int error)
2057 {
2058         unsigned long flags;
2059
2060         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2061
2062         kiblnd_close_conn_locked(conn, error);
2063
2064         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2065 }
2066
2067 static void
2068 kiblnd_handle_early_rxs(struct kib_conn *conn)
2069 {
2070         unsigned long flags;
2071         struct kib_rx *rx;
2072
2073         LASSERT(!in_interrupt());
2074         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2075
2076         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2077         while (!list_empty(&conn->ibc_early_rxs)) {
2078                 rx = list_entry(conn->ibc_early_rxs.next,
2079                                 struct kib_rx, rx_list);
2080                 list_del(&rx->rx_list);
2081                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2082
2083                 kiblnd_handle_rx(rx);
2084
2085                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2086         }
2087         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2088 }
2089
2090 void
2091 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2092 {
2093         LIST_HEAD(zombies);
2094         struct kib_tx *nxt;
2095         struct kib_tx *tx;
2096
2097         spin_lock(&conn->ibc_lock);
2098
2099         list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2100                 if (txs == &conn->ibc_active_txs) {
2101                         LASSERT(!tx->tx_queued);
2102                         LASSERT(tx->tx_waiting ||
2103                                 tx->tx_sending != 0);
2104                         if (conn->ibc_comms_error == -ETIMEDOUT) {
2105                                 if (tx->tx_waiting && !tx->tx_sending)
2106                                         tx->tx_hstatus =
2107                                           LNET_MSG_STATUS_REMOTE_TIMEOUT;
2108                                 else if (tx->tx_sending)
2109                                         tx->tx_hstatus =
2110                                           LNET_MSG_STATUS_NETWORK_TIMEOUT;
2111                         }
2112                 } else {
2113                         LASSERT(tx->tx_queued);
2114                         if (conn->ibc_comms_error == -ETIMEDOUT)
2115                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2116                         else
2117                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2118                 }
2119
2120                 tx->tx_status = -ECONNABORTED;
2121                 tx->tx_waiting = 0;
2122
2123                 /*
2124                  * TODO: This makes an assumption that
2125                  * kiblnd_tx_complete() will be called for each tx. If
2126                  * that event is dropped we could end up with stale
2127                  * connections floating around. We'd like to deal with
2128                  * that in a better way.
2129                  *
2130                  * Also that means we can exceed the timeout by many
2131                  * seconds.
2132                  */
2133                 if (tx->tx_sending == 0) {
2134                         tx->tx_queued = 0;
2135                         list_move(&tx->tx_list, &zombies);
2136                 } else {
2137                         /* keep tx until cq destroy */
2138                         list_move(&tx->tx_list, &conn->ibc_zombie_txs);
2139                         conn->ibc_waits ++;
2140                 }
2141         }
2142
2143         spin_unlock(&conn->ibc_lock);
2144
2145         /*
2146          * aborting transmits occurs when finalizing the connection.
2147          * The connection is finalized on error.
2148          * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2149          * override the value already set in tx->tx_hstatus above.
2150          */
2151         kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2152 }
2153
2154 static int
2155 kiblnd_tx_may_discard(struct kib_conn *conn)
2156 {
2157         int rc = 0;
2158         struct kib_tx *nxt;
2159         struct kib_tx *tx;
2160
2161         spin_lock(&conn->ibc_lock);
2162
2163         list_for_each_entry_safe(tx, nxt, &conn->ibc_zombie_txs, tx_list) {
2164                 if (tx->tx_sending > 0 && tx->tx_lntmsg[0] &&
2165                     lnet_md_discarded(tx->tx_lntmsg[0]->msg_md)) {
2166                         tx->tx_sending --;
2167                         if (tx->tx_sending == 0) {
2168                                 kiblnd_conn_decref(tx->tx_conn);
2169                                 tx->tx_conn = NULL;
2170                                 rc = 1;
2171                         }
2172                 }
2173         }
2174
2175         spin_unlock(&conn->ibc_lock);
2176         return rc;
2177 }
2178
2179 static void
2180 kiblnd_finalise_conn(struct kib_conn *conn)
2181 {
2182         LASSERT (!in_interrupt());
2183         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2184
2185         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2186          * for connections that didn't get as far as being connected, because
2187          * rdma_disconnect() does this for free. */
2188         kiblnd_abort_receives(conn);
2189
2190         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2191
2192         /* Complete all tx descs not waiting for sends to complete.
2193          * NB we should be safe from RDMA now that the QP has changed state */
2194
2195         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2196         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2197         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2198         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2199         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2200
2201         kiblnd_handle_early_rxs(conn);
2202 }
2203
2204 static void
2205 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2206                            int error)
2207 {
2208         LIST_HEAD(zombies);
2209         unsigned long   flags;
2210
2211         LASSERT (error != 0);
2212         LASSERT (!in_interrupt());
2213
2214         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2215
2216         if (active) {
2217                 LASSERT(peer_ni->ibp_connecting > 0);
2218                 peer_ni->ibp_connecting--;
2219         } else {
2220                 LASSERT (peer_ni->ibp_accepting > 0);
2221                 peer_ni->ibp_accepting--;
2222         }
2223
2224         if (kiblnd_peer_connecting(peer_ni)) {
2225                 /* another connection attempt under way... */
2226                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2227                                         flags);
2228                 return;
2229         }
2230
2231         peer_ni->ibp_reconnected = 0;
2232         if (list_empty(&peer_ni->ibp_conns)) {
2233                 /* Take peer_ni's blocked transmits to complete with error */
2234                 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2235
2236                 if (kiblnd_peer_active(peer_ni))
2237                         kiblnd_unlink_peer_locked(peer_ni);
2238
2239                 peer_ni->ibp_error = error;
2240         } else {
2241                 /* Can't have blocked transmits if there are connections */
2242                 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2243         }
2244
2245         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2246
2247         kiblnd_peer_notify(peer_ni);
2248
2249         if (list_empty(&zombies))
2250                 return;
2251
2252         CNETERR("Deleting messages for %s: connection failed\n",
2253                 libcfs_nid2str(peer_ni->ibp_nid));
2254
2255         if (error == -EHOSTUNREACH || error == -ETIMEDOUT)
2256                 kiblnd_txlist_done(&zombies, error,
2257                                    LNET_MSG_STATUS_NETWORK_TIMEOUT);
2258         else
2259                 kiblnd_txlist_done(&zombies, error,
2260                                    LNET_MSG_STATUS_LOCAL_DROPPED);
2261 }
2262
2263 static void
2264 kiblnd_connreq_done(struct kib_conn *conn, int status)
2265 {
2266         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2267         struct kib_tx *tx;
2268         LIST_HEAD(txs);
2269         unsigned long    flags;
2270         int              active;
2271
2272         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2273
2274         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2275                libcfs_nid2str(peer_ni->ibp_nid), active,
2276                conn->ibc_version, status);
2277
2278         LASSERT (!in_interrupt());
2279         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2280                   peer_ni->ibp_connecting > 0) ||
2281                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2282                   peer_ni->ibp_accepting > 0));
2283
2284         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2285         conn->ibc_connvars = NULL;
2286
2287         if (status != 0) {
2288                 /* failed to establish connection */
2289                 kiblnd_peer_connect_failed(peer_ni, active, status);
2290                 kiblnd_finalise_conn(conn);
2291                 return;
2292         }
2293
2294         /* connection established */
2295         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2296
2297         /* reset retry count */
2298         peer_ni->ibp_retries = 0;
2299
2300         conn->ibc_last_send = ktime_get();
2301         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2302         kiblnd_peer_alive(peer_ni);
2303
2304         /* Add conn to peer_ni's list and nuke any dangling conns from a different
2305          * peer_ni instance... */
2306         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2307         list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2308         peer_ni->ibp_reconnected = 0;
2309         if (active)
2310                 peer_ni->ibp_connecting--;
2311         else
2312                 peer_ni->ibp_accepting--;
2313
2314         if (peer_ni->ibp_version == 0) {
2315                 peer_ni->ibp_version     = conn->ibc_version;
2316                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2317         }
2318
2319         if (peer_ni->ibp_version     != conn->ibc_version ||
2320             peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2321                 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2322                                                 conn->ibc_incarnation);
2323                 peer_ni->ibp_version     = conn->ibc_version;
2324                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2325         }
2326
2327         /* grab pending txs while I have the lock */
2328         list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2329
2330         if (!kiblnd_peer_active(peer_ni) ||        /* peer_ni has been deleted */
2331             conn->ibc_comms_error != 0) {       /* error has happened already */
2332
2333                 /* start to shut down connection */
2334                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2335                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2336
2337                 kiblnd_txlist_done(&txs, -ECONNABORTED,
2338                                    LNET_MSG_STATUS_LOCAL_ERROR);
2339
2340                 return;
2341         }
2342
2343         /* +1 ref for myself, this connection is visible to other threads
2344          * now, refcount of peer:ibp_conns can be released by connection
2345          * close from either a different thread, or the calling of
2346          * kiblnd_check_sends_locked() below. See bz21911 for details.
2347          */
2348         kiblnd_conn_addref(conn);
2349         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2350
2351         /* Schedule blocked txs
2352          * Note: if we are running with conns_per_peer > 1, these blocked
2353          * txs will all get scheduled to the first connection which gets
2354          * scheduled.  We won't be using round robin on this first batch.
2355          */
2356         spin_lock(&conn->ibc_lock);
2357         while (!list_empty(&txs)) {
2358                 tx = list_entry(txs.next, struct kib_tx, tx_list);
2359                 list_del(&tx->tx_list);
2360
2361                 kiblnd_queue_tx_locked(tx, conn);
2362         }
2363         kiblnd_check_sends_locked(conn);
2364         spin_unlock(&conn->ibc_lock);
2365
2366         /* schedule blocked rxs */
2367         kiblnd_handle_early_rxs(conn);
2368         kiblnd_conn_decref(conn);
2369 }
2370
2371 static void
2372 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2373 {
2374         int          rc;
2375
2376 #ifdef HAVE_RDMA_REJECT_4ARGS
2377         rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2378 #else
2379         rc = rdma_reject(cmid, rej, sizeof(*rej));
2380 #endif
2381
2382         if (rc != 0)
2383                 CWARN("Error %d sending reject\n", rc);
2384 }
2385
2386 static int
2387 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2388 {
2389         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2390         struct kib_msg *reqmsg = priv;
2391         struct kib_msg *ackmsg;
2392         struct kib_dev *ibdev;
2393         struct kib_peer_ni *peer_ni;
2394         struct kib_peer_ni *peer2;
2395         struct kib_conn *conn;
2396         struct lnet_ni *ni = NULL;
2397         struct kib_net *net = NULL;
2398         lnet_nid_t             nid;
2399         struct rdma_conn_param cp;
2400         struct kib_rej rej;
2401         int                    version = IBLND_MSG_VERSION;
2402         unsigned long          flags;
2403         int                    rc;
2404         struct sockaddr_in    *peer_addr;
2405         LASSERT (!in_interrupt());
2406
2407         /* cmid inherits 'context' from the corresponding listener id */
2408         ibdev = cmid->context;
2409         LASSERT(ibdev);
2410
2411         memset(&rej, 0, sizeof(rej));
2412         rej.ibr_magic                = IBLND_MSG_MAGIC;
2413         rej.ibr_why                  = IBLND_REJECT_FATAL;
2414         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2415
2416         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2417         if (*kiblnd_tunables.kib_require_priv_port &&
2418             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2419                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2420                 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2421                        &ip, ntohs(peer_addr->sin_port));
2422                 goto failed;
2423         }
2424
2425         if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2426                 CERROR("Short connection request\n");
2427                 goto failed;
2428         }
2429
2430         /* Future protocol version compatibility support!  If the
2431          * o2iblnd-specific protocol changes, or when LNET unifies
2432          * protocols over all LNDs, the initial connection will
2433          * negotiate a protocol version.  I trap this here to avoid
2434          * console errors; the reject tells the peer_ni which protocol I
2435          * speak. */
2436         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2437             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2438                 goto failed;
2439         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2440             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2441             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2442                 goto failed;
2443         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2444             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2445             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2446                 goto failed;
2447
2448         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2449         if (rc != 0) {
2450                 CERROR("Can't parse connection request: %d\n", rc);
2451                 goto failed;
2452         }
2453
2454         nid = reqmsg->ibm_srcnid;
2455         ni  = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2456
2457         if (ni != NULL) {
2458                 net = (struct kib_net *)ni->ni_data;
2459                 rej.ibr_incarnation = net->ibn_incarnation;
2460         }
2461
2462         if (ni == NULL ||                         /* no matching net */
2463             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2464             net->ibn_dev != ibdev) {              /* wrong device */
2465                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2466                        "bad dst nid %s\n", libcfs_nid2str(nid),
2467                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2468                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2469                         &ibdev->ibd_ifip,
2470                        libcfs_nid2str(reqmsg->ibm_dstnid));
2471
2472                 goto failed;
2473         }
2474
2475        /* check time stamp as soon as possible */
2476         if (reqmsg->ibm_dststamp != 0 &&
2477             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2478                 CWARN("Stale connection request\n");
2479                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2480                 goto failed;
2481         }
2482
2483         /* I can accept peer_ni's version */
2484         version = reqmsg->ibm_version;
2485
2486         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2487                 CERROR("Unexpected connreq msg type: %x from %s\n",
2488                        reqmsg->ibm_type, libcfs_nid2str(nid));
2489                 goto failed;
2490         }
2491
2492         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2493             kiblnd_msg_queue_size(version, ni)) {
2494                 CERROR("Can't accept conn from %s, queue depth too large: "
2495                        " %d (<=%d wanted)\n",
2496                        libcfs_nid2str(nid),
2497                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2498                        kiblnd_msg_queue_size(version, ni));
2499
2500                 if (version == IBLND_MSG_VERSION)
2501                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2502
2503                 goto failed;
2504         }
2505
2506         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2507             IBLND_MAX_RDMA_FRAGS) {
2508                 CWARN("Can't accept conn from %s (version %x): "
2509                       "max_frags %d too large (%d wanted)\n",
2510                       libcfs_nid2str(nid), version,
2511                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2512                       IBLND_MAX_RDMA_FRAGS);
2513
2514                 if (version >= IBLND_MSG_VERSION)
2515                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2516
2517                 goto failed;
2518         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2519                    IBLND_MAX_RDMA_FRAGS &&
2520                    net->ibn_fmr_ps == NULL) {
2521                 CWARN("Can't accept conn from %s (version %x): "
2522                       "max_frags %d incompatible without FMR pool "
2523                       "(%d wanted)\n",
2524                       libcfs_nid2str(nid), version,
2525                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2526                       IBLND_MAX_RDMA_FRAGS);
2527
2528                 if (version == IBLND_MSG_VERSION)
2529                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2530
2531                 goto failed;
2532         }
2533
2534         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2535                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2536                        libcfs_nid2str(nid),
2537                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2538                        IBLND_MSG_SIZE);
2539                 goto failed;
2540         }
2541
2542         /* assume 'nid' is a new peer_ni; create  */
2543         rc = kiblnd_create_peer(ni, &peer_ni, nid);
2544         if (rc != 0) {
2545                 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2546                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2547                 goto failed;
2548         }
2549
2550         /* We have validated the peer's parameters so use those */
2551         peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2552         peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2553
2554         write_lock_irqsave(g_lock, flags);
2555
2556         peer2 = kiblnd_find_peer_locked(ni, nid);
2557         if (peer2 != NULL) {
2558                 if (peer2->ibp_version == 0) {
2559                         peer2->ibp_version     = version;
2560                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2561                 }
2562
2563                 /* not the guy I've talked with */
2564                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2565                     peer2->ibp_version     != version) {
2566                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2567
2568                         if (kiblnd_peer_active(peer2)) {
2569                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2570                                 peer2->ibp_version = version;
2571                         }
2572                         write_unlock_irqrestore(g_lock, flags);
2573
2574                         CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2575                               libcfs_nid2str(nid), peer2->ibp_version, version,
2576                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2577
2578                         kiblnd_peer_decref(peer_ni);
2579                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2580                         goto failed;
2581                 }
2582
2583                 /* Tie-break connection race in favour of the higher NID.
2584                  * If we keep running into a race condition multiple times,
2585                  * we have to assume that the connection attempt with the
2586                  * higher NID is stuck in a connecting state and will never
2587                  * recover.  As such, we pass through this if-block and let
2588                  * the lower NID connection win so we can move forward.
2589                  */
2590                 if (peer2->ibp_connecting != 0 &&
2591                     nid < ni->ni_nid && peer2->ibp_races <
2592                     MAX_CONN_RACES_BEFORE_ABORT) {
2593                         peer2->ibp_races++;
2594                         write_unlock_irqrestore(g_lock, flags);
2595
2596                         CDEBUG(D_NET, "Conn race %s\n",
2597                                libcfs_nid2str(peer2->ibp_nid));
2598
2599                         kiblnd_peer_decref(peer_ni);
2600                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2601                         goto failed;
2602                 }
2603                 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2604                         CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2605                                 libcfs_nid2str(peer2->ibp_nid),
2606                                 MAX_CONN_RACES_BEFORE_ABORT);
2607                 /*
2608                  * passive connection is allowed even this peer_ni is waiting for
2609                  * reconnection.
2610                  */
2611                 peer2->ibp_reconnecting = 0;
2612                 peer2->ibp_races = 0;
2613                 peer2->ibp_accepting++;
2614                 kiblnd_peer_addref(peer2);
2615
2616                 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2617                  * so copy validated parameters since we now know what the
2618                  * peer_ni's limits are */
2619                 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2620                 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2621
2622                 write_unlock_irqrestore(g_lock, flags);
2623                 kiblnd_peer_decref(peer_ni);
2624                 peer_ni = peer2;
2625         } else {
2626                 /* Brand new peer_ni */
2627                 LASSERT (peer_ni->ibp_accepting == 0);
2628                 LASSERT (peer_ni->ibp_version == 0 &&
2629                          peer_ni->ibp_incarnation == 0);
2630
2631                 peer_ni->ibp_accepting   = 1;
2632                 peer_ni->ibp_version     = version;
2633                 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2634
2635                 /* I have a ref on ni that prevents it being shutdown */
2636                 LASSERT (net->ibn_shutdown == 0);
2637
2638                 kiblnd_peer_addref(peer_ni);
2639                 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2640
2641                 write_unlock_irqrestore(g_lock, flags);
2642         }
2643
2644         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2645         if (conn == NULL) {
2646                 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2647                 kiblnd_peer_decref(peer_ni);
2648                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2649                 goto failed;
2650         }
2651
2652         /* conn now "owns" cmid, so I return success from here on to ensure the
2653          * CM callback doesn't destroy cmid. */
2654         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2655         conn->ibc_credits          = conn->ibc_queue_depth;
2656         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2657         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2658                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2659
2660         ackmsg = &conn->ibc_connvars->cv_msg;
2661         memset(ackmsg, 0, sizeof(*ackmsg));
2662
2663         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2664                         sizeof(ackmsg->ibm_u.connparams));
2665         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2666         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2667         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2668
2669         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2670
2671         memset(&cp, 0, sizeof(cp));
2672         cp.private_data        = ackmsg;
2673         cp.private_data_len    = ackmsg->ibm_nob;
2674         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2675         cp.initiator_depth     = 0;
2676         cp.flow_control        = 1;
2677         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2678         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2679
2680         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2681
2682         rc = rdma_accept(cmid, &cp);
2683         if (rc != 0) {
2684                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2685                 rej.ibr_version = version;
2686                 rej.ibr_why     = IBLND_REJECT_FATAL;
2687
2688                 kiblnd_reject(cmid, &rej);
2689                 kiblnd_connreq_done(conn, rc);
2690                 kiblnd_conn_decref(conn);
2691         }
2692
2693         lnet_ni_decref(ni);
2694         return 0;
2695
2696  failed:
2697         if (ni != NULL) {
2698                 rej.ibr_cp.ibcp_queue_depth =
2699                         kiblnd_msg_queue_size(version, ni);
2700                 rej.ibr_cp.ibcp_max_frags   = IBLND_MAX_RDMA_FRAGS;
2701                 lnet_ni_decref(ni);
2702         }
2703
2704         rej.ibr_version = version;
2705         kiblnd_reject(cmid, &rej);
2706
2707         return -ECONNREFUSED;
2708 }
2709
2710 static void
2711 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2712                        u64 incarnation, int why, struct kib_connparams *cp)
2713 {
2714         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
2715         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2716         char            *reason;
2717         int              msg_size = IBLND_MSG_SIZE;
2718         int              frag_num = -1;
2719         int              queue_dep = -1;
2720         bool             reconnect;
2721         unsigned long    flags;
2722
2723         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2724         LASSERT(peer_ni->ibp_connecting > 0);   /* 'conn' at least */
2725
2726         if (cp) {
2727                 msg_size        = cp->ibcp_max_msg_size;
2728                 frag_num        = cp->ibcp_max_frags;
2729                 queue_dep       = cp->ibcp_queue_depth;
2730         }
2731
2732         write_lock_irqsave(glock, flags);
2733         /* retry connection if it's still needed and no other connection
2734          * attempts (active or passive) are in progress
2735          * NB: reconnect is still needed even when ibp_tx_queue is
2736          * empty if ibp_version != version because reconnect may be
2737          * initiated.
2738          */
2739         reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2740                      peer_ni->ibp_version != version) &&
2741                     peer_ni->ibp_connecting &&
2742                     peer_ni->ibp_accepting == 0;
2743         if (!reconnect) {
2744                 reason = "no need";
2745                 goto out;
2746         }
2747
2748         if (peer_ni->ibp_retries > *kiblnd_tunables.kib_retry_count) {
2749                 reason = "retry count exceeded due to no listener";
2750                 goto out;
2751         }
2752
2753         switch (why) {
2754         default:
2755                 reason = "Unknown";
2756                 break;
2757
2758         case IBLND_REJECT_RDMA_FRAGS: {
2759                 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2760
2761                 if (!cp) {
2762                         reason = "can't negotiate max frags";
2763                         goto out;
2764                 }
2765                 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2766 #ifdef HAVE_IB_GET_DMA_MR
2767                 /*
2768                  * This check only makes sense if the kernel supports global
2769                  * memory registration. Otherwise, map_on_demand will never == 0
2770                  */
2771                 if (!tunables->lnd_map_on_demand) {
2772                         reason = "map_on_demand must be enabled";
2773                         goto out;
2774                 }
2775 #endif
2776                 if (conn->ibc_max_frags <= frag_num) {
2777                         reason = "unsupported max frags";
2778                         goto out;
2779                 }
2780
2781                 peer_ni->ibp_max_frags = frag_num;
2782                 reason = "rdma fragments";
2783                 break;
2784         }
2785         case IBLND_REJECT_MSG_QUEUE_SIZE:
2786                 if (!cp) {
2787                         reason = "can't negotiate queue depth";
2788                         goto out;
2789                 }
2790                 if (conn->ibc_queue_depth <= queue_dep) {
2791                         reason = "unsupported queue depth";
2792                         goto out;
2793                 }
2794
2795                 peer_ni->ibp_queue_depth = queue_dep;
2796                 reason = "queue depth";
2797                 break;
2798
2799         case IBLND_REJECT_CONN_STALE:
2800                 reason = "stale";
2801                 break;
2802
2803         case IBLND_REJECT_CONN_RACE:
2804                 reason = "conn race";
2805                 break;
2806
2807         case IBLND_REJECT_CONN_UNCOMPAT:
2808                 reason = "version negotiation";
2809                 break;
2810
2811         case IBLND_REJECT_INVALID_SRV_ID:
2812                 reason = "invalid service id";
2813                 break;
2814         }
2815
2816         conn->ibc_reconnect = 1;
2817         peer_ni->ibp_reconnecting++;
2818         peer_ni->ibp_version = version;
2819         if (incarnation != 0)
2820                 peer_ni->ibp_incarnation = incarnation;
2821  out:
2822         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2823
2824         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2825                 libcfs_nid2str(peer_ni->ibp_nid),
2826                 reconnect ? "reconnect" : "don't reconnect",
2827                 reason, IBLND_MSG_VERSION, version, msg_size,
2828                 conn->ibc_queue_depth, queue_dep,
2829                 conn->ibc_max_frags, frag_num);
2830         /*
2831          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2832          * while destroying the zombie
2833          */
2834 }
2835
2836 static void
2837 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2838 {
2839         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2840
2841         LASSERT (!in_interrupt());
2842         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2843
2844         switch (reason) {
2845         case IB_CM_REJ_STALE_CONN:
2846                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2847                                        IBLND_REJECT_CONN_STALE, NULL);
2848                 break;
2849
2850         case IB_CM_REJ_INVALID_SERVICE_ID:
2851                 peer_ni->ibp_retries++;
2852                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2853                                        IBLND_REJECT_INVALID_SRV_ID, NULL);
2854                 CNETERR("%s rejected: no listener at %d\n",
2855                         libcfs_nid2str(peer_ni->ibp_nid),
2856                         *kiblnd_tunables.kib_service);
2857                 break;
2858
2859         case IB_CM_REJ_CONSUMER_DEFINED:
2860                 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2861                         struct kib_rej *rej = priv;
2862                         struct kib_connparams *cp = NULL;
2863                         int               flip        = 0;
2864                         __u64             incarnation = -1;
2865
2866                         /* NB. default incarnation is -1 because:
2867                          * a) V1 will ignore dst incarnation in connreq.
2868                          * b) V2 will provide incarnation while rejecting me,
2869                          *    -1 will be overwrote.
2870                          *
2871                          * if I try to connect to a V1 peer_ni with V2 protocol,
2872                          * it rejected me then upgrade to V2, I have no idea
2873                          * about the upgrading and try to reconnect with V1,
2874                          * in this case upgraded V2 can find out I'm trying to
2875                          * talk to the old guy and reject me(incarnation is -1).
2876                          */
2877
2878                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2879                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2880                                 __swab32s(&rej->ibr_magic);
2881                                 __swab16s(&rej->ibr_version);
2882                                 flip = 1;
2883                         }
2884
2885                         if (priv_nob >= sizeof(struct kib_rej) &&
2886                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2887                                 /* priv_nob is always 148 in current version
2888                                  * of OFED, so we still need to check version.
2889                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2890                                 cp = &rej->ibr_cp;
2891
2892                                 if (flip) {
2893                                         __swab64s(&rej->ibr_incarnation);
2894                                         __swab16s(&cp->ibcp_queue_depth);
2895                                         __swab16s(&cp->ibcp_max_frags);
2896                                         __swab32s(&cp->ibcp_max_msg_size);
2897                                 }
2898
2899                                 incarnation = rej->ibr_incarnation;
2900                         }
2901
2902                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2903                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2904                                 CERROR("%s rejected: consumer defined fatal error\n",
2905                                        libcfs_nid2str(peer_ni->ibp_nid));
2906                                 break;
2907                         }
2908
2909                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2910                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2911                                 CERROR("%s rejected: o2iblnd version %x error\n",
2912                                        libcfs_nid2str(peer_ni->ibp_nid),
2913                                        rej->ibr_version);
2914                                 break;
2915                         }
2916
2917                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2918                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2919                                 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2920                                        libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2921
2922                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2923                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2924                         }
2925
2926                         switch (rej->ibr_why) {
2927                         case IBLND_REJECT_CONN_RACE:
2928                         case IBLND_REJECT_CONN_STALE:
2929                         case IBLND_REJECT_CONN_UNCOMPAT:
2930                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2931                         case IBLND_REJECT_RDMA_FRAGS:
2932                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2933                                                 incarnation, rej->ibr_why, cp);
2934                                 break;
2935
2936                         case IBLND_REJECT_NO_RESOURCES:
2937                                 CERROR("%s rejected: o2iblnd no resources\n",
2938                                        libcfs_nid2str(peer_ni->ibp_nid));
2939                                 break;
2940
2941                         case IBLND_REJECT_FATAL:
2942                                 CERROR("%s rejected: o2iblnd fatal error\n",
2943                                        libcfs_nid2str(peer_ni->ibp_nid));
2944                                 break;
2945
2946                         default:
2947                                 CERROR("%s rejected: o2iblnd reason %d\n",
2948                                        libcfs_nid2str(peer_ni->ibp_nid),
2949                                        rej->ibr_why);
2950                                 break;
2951                         }
2952                         break;
2953                 }
2954                 /* fall through */
2955         default:
2956                 CNETERR("%s rejected: reason %d, size %d\n",
2957                         libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2958                 break;
2959         }
2960
2961         kiblnd_connreq_done(conn, -ECONNREFUSED);
2962 }
2963
2964 static void
2965 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2966 {
2967         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2968         struct lnet_ni *ni = peer_ni->ibp_ni;
2969         struct kib_net *net = ni->ni_data;
2970         struct kib_msg *msg = priv;
2971         int            ver  = conn->ibc_version;
2972         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2973         unsigned long  flags;
2974
2975         LASSERT (net != NULL);
2976
2977         if (rc != 0) {
2978                 CERROR("Can't unpack connack from %s: %d\n",
2979                        libcfs_nid2str(peer_ni->ibp_nid), rc);
2980                 goto failed;
2981         }
2982
2983         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2984                 CERROR("Unexpected message %d from %s\n",
2985                        msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2986                 rc = -EPROTO;
2987                 goto failed;
2988         }
2989
2990         if (ver != msg->ibm_version) {
2991                 CERROR("%s replied version %x is different with "
2992                        "requested version %x\n",
2993                        libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2994                 rc = -EPROTO;
2995                 goto failed;
2996         }
2997
2998         if (msg->ibm_u.connparams.ibcp_queue_depth >
2999             conn->ibc_queue_depth) {
3000                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
3001                        libcfs_nid2str(peer_ni->ibp_nid),
3002                        msg->ibm_u.connparams.ibcp_queue_depth,
3003                        conn->ibc_queue_depth);
3004                 rc = -EPROTO;
3005                 goto failed;
3006         }
3007
3008         if (msg->ibm_u.connparams.ibcp_max_frags >
3009             conn->ibc_max_frags) {
3010                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
3011                        libcfs_nid2str(peer_ni->ibp_nid),
3012                        msg->ibm_u.connparams.ibcp_max_frags,
3013                        conn->ibc_max_frags);
3014                 rc = -EPROTO;
3015                 goto failed;
3016         }
3017
3018         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
3019                 CERROR("%s max message size %d too big (%d max)\n",
3020                        libcfs_nid2str(peer_ni->ibp_nid),
3021                        msg->ibm_u.connparams.ibcp_max_msg_size,
3022                        IBLND_MSG_SIZE);
3023                 rc = -EPROTO;
3024                 goto failed;
3025         }
3026
3027         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3028         if (msg->ibm_dstnid == ni->ni_nid &&
3029             msg->ibm_dststamp == net->ibn_incarnation)
3030                 rc = 0;
3031         else
3032                 rc = -ESTALE;
3033         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3034
3035         if (rc != 0) {
3036                 CERROR("Bad connection reply from %s, rc = %d, "
3037                        "version: %x max_frags: %d\n",
3038                        libcfs_nid2str(peer_ni->ibp_nid), rc,
3039                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
3040                 goto failed;
3041         }
3042
3043         conn->ibc_incarnation      = msg->ibm_srcstamp;
3044         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
3045         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
3046         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
3047         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
3048         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
3049                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
3050
3051         kiblnd_connreq_done(conn, 0);
3052         return;
3053
3054  failed:
3055         /* NB My QP has already established itself, so I handle anything going
3056          * wrong here by setting ibc_comms_error.
3057          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3058          * immediately tears it down. */
3059
3060         LASSERT (rc != 0);
3061         conn->ibc_comms_error = rc;
3062         kiblnd_connreq_done(conn, 0);
3063 }
3064
3065 static int
3066 kiblnd_active_connect(struct rdma_cm_id *cmid)
3067 {
3068         struct kib_peer_ni *peer_ni = cmid->context;
3069         struct kib_conn *conn;
3070         struct kib_msg *msg;
3071         struct rdma_conn_param cp;
3072         int                      version;
3073         __u64                    incarnation;
3074         unsigned long            flags;
3075         int                      rc;
3076
3077         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3078
3079         incarnation = peer_ni->ibp_incarnation;
3080         version     = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3081                                                  peer_ni->ibp_version;
3082
3083         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3084
3085         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3086                                   version);
3087         if (conn == NULL) {
3088                 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3089                 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3090                 return -ENOMEM;
3091         }
3092
3093         /* conn "owns" cmid now, so I return success from here on to ensure the
3094          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3095          * on peer_ni */
3096
3097         msg = &conn->ibc_connvars->cv_msg;
3098
3099         memset(msg, 0, sizeof(*msg));
3100         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3101         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
3102         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
3103         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3104
3105         kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3106                         0, peer_ni->ibp_nid, incarnation);
3107
3108         memset(&cp, 0, sizeof(cp));
3109         cp.private_data        = msg;
3110         cp.private_data_len    = msg->ibm_nob;
3111         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
3112         cp.initiator_depth     = 0;
3113         cp.flow_control        = 1;
3114         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
3115         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
3116
3117         LASSERT(cmid->context == (void *)conn);
3118         LASSERT(conn->ibc_cmid == cmid);
3119
3120         rc = rdma_connect(cmid, &cp);
3121         if (rc != 0) {
3122                 CERROR("Can't connect to %s: %d\n",
3123                        libcfs_nid2str(peer_ni->ibp_nid), rc);
3124                 kiblnd_connreq_done(conn, rc);
3125                 kiblnd_conn_decref(conn);
3126         }
3127
3128         return 0;
3129 }
3130
3131 int
3132 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3133 {
3134         struct kib_peer_ni *peer_ni;
3135         struct kib_conn *conn;
3136         int rc;
3137
3138         switch (event->event) {
3139         default:
3140                 CERROR("Unexpected event: %d, status: %d\n",
3141                        event->event, event->status);
3142                 LBUG();
3143
3144         case RDMA_CM_EVENT_CONNECT_REQUEST:
3145                 /* destroy cmid on failure */
3146                 rc = kiblnd_passive_connect(cmid,
3147                                             (void *)KIBLND_CONN_PARAM(event),
3148                                             KIBLND_CONN_PARAM_LEN(event));
3149                 CDEBUG(D_NET, "connreq: %d\n", rc);
3150                 return rc;
3151
3152         case RDMA_CM_EVENT_ADDR_ERROR:
3153                 peer_ni = cmid->context;
3154                 CNETERR("%s: ADDR ERROR %d\n",
3155                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3156                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3157                 kiblnd_peer_decref(peer_ni);
3158                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
3159
3160         case RDMA_CM_EVENT_ADDR_RESOLVED:
3161                 peer_ni = cmid->context;
3162
3163                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3164                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3165
3166                 if (event->status != 0) {
3167                         CNETERR("Can't resolve address for %s: %d\n",
3168                                 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3169                         rc = event->status;
3170                 } else {
3171                         rc = rdma_resolve_route(
3172                                 cmid, kiblnd_timeout() * 1000);
3173                         if (rc == 0) {
3174                                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3175                                 struct kib_dev *dev = net->ibn_dev;
3176
3177                                 CDEBUG(D_NET, "%s: connection bound to "\
3178                                        "%s:%pI4h:%s\n",
3179                                        libcfs_nid2str(peer_ni->ibp_nid),
3180                                        dev->ibd_ifname,
3181                                        &dev->ibd_ifip, cmid->device->name);
3182
3183                                 return 0;
3184                         }
3185
3186                         /* Can't initiate route resolution */
3187                         CERROR("Can't resolve route for %s: %d\n",
3188                                libcfs_nid2str(peer_ni->ibp_nid), rc);
3189                 }
3190                 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3191                 kiblnd_peer_decref(peer_ni);
3192                 return rc;                      /* rc != 0 destroys cmid */
3193
3194         case RDMA_CM_EVENT_ROUTE_ERROR:
3195                 peer_ni = cmid->context;
3196                 CNETERR("%s: ROUTE ERROR %d\n",
3197                         libcfs_nid2str(peer_ni->ibp_nid), event->status);
3198                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3199                 kiblnd_peer_decref(peer_ni);
3200                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
3201
3202         case RDMA_CM_EVENT_ROUTE_RESOLVED:
3203                 peer_ni = cmid->context;
3204                 CDEBUG(D_NET,"%s Route resolved: %d\n",
3205                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3206
3207                 if (event->status == 0)
3208                         return kiblnd_active_connect(cmid);
3209
3210                 CNETERR("Can't resolve route for %s: %d\n",
3211                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3212                 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3213                 kiblnd_peer_decref(peer_ni);
3214                 return event->status;           /* rc != 0 destroys cmid */
3215
3216         case RDMA_CM_EVENT_UNREACHABLE:
3217                 conn = cmid->context;
3218                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3219                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3220                 CNETERR("%s: UNREACHABLE %d\n",
3221                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3222                 kiblnd_connreq_done(conn, -ENETDOWN);
3223                 kiblnd_conn_decref(conn);
3224                 return 0;
3225
3226         case RDMA_CM_EVENT_CONNECT_ERROR:
3227                 conn = cmid->context;
3228                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3229                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3230                 CNETERR("%s: CONNECT ERROR %d\n",
3231                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3232                 kiblnd_connreq_done(conn, -ENOTCONN);
3233                 kiblnd_conn_decref(conn);
3234                 return 0;
3235
3236         case RDMA_CM_EVENT_REJECTED:
3237                 conn = cmid->context;
3238                 switch (conn->ibc_state) {
3239                 default:
3240                         LBUG();
3241
3242                 case IBLND_CONN_PASSIVE_WAIT:
3243                         CERROR ("%s: REJECTED %d\n",
3244                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3245                                 event->status);
3246                         kiblnd_connreq_done(conn, -ECONNRESET);
3247                         break;
3248
3249                 case IBLND_CONN_ACTIVE_CONNECT:
3250                         kiblnd_rejected(conn, event->status,
3251                                         (void *)KIBLND_CONN_PARAM(event),
3252                                         KIBLND_CONN_PARAM_LEN(event));
3253                         break;
3254                 }
3255                 kiblnd_conn_decref(conn);
3256                 return 0;
3257
3258         case RDMA_CM_EVENT_ESTABLISHED:
3259                 conn = cmid->context;
3260                 switch (conn->ibc_state) {
3261                 default:
3262                         LBUG();
3263
3264                 case IBLND_CONN_PASSIVE_WAIT:
3265                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3266                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3267                         kiblnd_connreq_done(conn, 0);
3268                         break;
3269
3270                 case IBLND_CONN_ACTIVE_CONNECT:
3271                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3272                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3273                         kiblnd_check_connreply(conn,
3274                                                (void *)KIBLND_CONN_PARAM(event),
3275                                                KIBLND_CONN_PARAM_LEN(event));
3276                         break;
3277                 }
3278                 /* net keeps its ref on conn! */
3279                 return 0;
3280
3281         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3282                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3283                 return 0;
3284
3285         case RDMA_CM_EVENT_DISCONNECTED:
3286                 conn = cmid->context;
3287                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3288                         CERROR("%s DISCONNECTED\n",
3289                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3290                         kiblnd_connreq_done(conn, -ECONNRESET);
3291                 } else {
3292                         kiblnd_close_conn(conn, 0);
3293                 }
3294                 kiblnd_conn_decref(conn);
3295                 cmid->context = NULL;
3296                 return 0;
3297
3298         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3299                 LCONSOLE_ERROR_MSG(0x131,
3300                                    "Received notification of device removal\n"
3301                                    "Please shutdown LNET to allow this to proceed\n");
3302                 /* Can't remove network from underneath LNET for now, so I have
3303                  * to ignore this */
3304                 return 0;
3305
3306         case RDMA_CM_EVENT_ADDR_CHANGE:
3307                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3308                 return 0;
3309         }
3310 }
3311
3312 static int
3313 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3314 {
3315         struct kib_tx *tx;
3316         struct list_head *ttmp;
3317
3318         list_for_each(ttmp, txs) {
3319                 tx = list_entry(ttmp, struct kib_tx, tx_list);
3320
3321                 if (txs != &conn->ibc_active_txs) {
3322                         LASSERT(tx->tx_queued);
3323                 } else {
3324                         LASSERT(!tx->tx_queued);
3325                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3326                 }
3327
3328                 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3329                         CERROR("Timed out tx: %s(WSQ:%d%d%d), %lld seconds\n",
3330                                kiblnd_queue2str(conn, txs),
3331                                tx->tx_waiting, tx->tx_sending, tx->tx_queued,
3332                                kiblnd_timeout() +
3333                                ktime_ms_delta(ktime_get(),
3334                                               tx->tx_deadline) / MSEC_PER_SEC);
3335                         return 1;
3336                 }
3337         }
3338
3339         return 0;
3340 }
3341
3342 static int
3343 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3344 {
3345         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3346                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3347                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3348                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3349                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3350 }
3351
3352 static void
3353 kiblnd_check_conns (int idx)
3354 {
3355         LIST_HEAD(closes);
3356         LIST_HEAD(checksends);
3357         LIST_HEAD(timedout_txs);
3358         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3359         struct list_head *ptmp;
3360         struct kib_peer_ni *peer_ni;
3361         struct kib_conn *conn;
3362         struct kib_tx *tx, *tx_tmp;
3363         struct list_head *ctmp;
3364         unsigned long     flags;
3365
3366         /* NB. We expect to have a look at all the peers and not find any
3367          * RDMAs to time out, so we just use a shared lock while we
3368          * take a look... */
3369         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3370
3371         list_for_each(ptmp, peers) {
3372                 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
3373
3374                 /* Check tx_deadline */
3375                 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3376                         if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3377                                 CWARN("Timed out tx for %s: %lld seconds\n",
3378                                       libcfs_nid2str(peer_ni->ibp_nid),
3379                                       ktime_ms_delta(ktime_get(),
3380                                                      tx->tx_deadline) / MSEC_PER_SEC);
3381                                 list_move(&tx->tx_list, &timedout_txs);
3382                         }
3383                 }
3384
3385                 list_for_each(ctmp, &peer_ni->ibp_conns) {
3386                         int timedout;
3387                         int sendnoop;
3388
3389                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
3390
3391                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3392
3393                         spin_lock(&conn->ibc_lock);
3394
3395                         sendnoop = kiblnd_need_noop(conn);
3396                         timedout = kiblnd_conn_timed_out_locked(conn);
3397                         if (!sendnoop && !timedout) {
3398                                 spin_unlock(&conn->ibc_lock);
3399                                 continue;
3400                         }
3401
3402                         if (timedout) {
3403                                 CERROR("Timed out RDMA with %s (%lld): "
3404                                        "c: %u, oc: %u, rc: %u\n",
3405                                        libcfs_nid2str(peer_ni->ibp_nid),
3406                                        ktime_get_seconds() - peer_ni->ibp_last_alive,
3407                                        conn->ibc_credits,
3408                                        conn->ibc_outstanding_credits,
3409                                        conn->ibc_reserved_credits);
3410                                 list_add(&conn->ibc_connd_list, &closes);
3411                         } else {
3412                                 list_add(&conn->ibc_connd_list, &checksends);
3413                         }
3414                         /* +ref for 'closes' or 'checksends' */
3415                         kiblnd_conn_addref(conn);
3416
3417                         spin_unlock(&conn->ibc_lock);
3418                 }
3419         }
3420
3421         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3422
3423         if (!list_empty(&timedout_txs))
3424                 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3425                                    LNET_MSG_STATUS_NETWORK_TIMEOUT);
3426
3427         /* Handle timeout by closing the whole
3428          * connection. We can only be sure RDMA activity
3429          * has ceased once the QP has been modified. */
3430         while (!list_empty(&closes)) {
3431                 conn = list_entry(closes.next,
3432                                   struct kib_conn, ibc_connd_list);
3433                 list_del(&conn->ibc_connd_list);
3434                 kiblnd_close_conn(conn, -ETIMEDOUT);
3435                 kiblnd_conn_decref(conn);
3436         }
3437
3438         /* In case we have enough credits to return via a
3439          * NOOP, but there were no non-blocking tx descs
3440          * free to do it last time... */
3441         while (!list_empty(&checksends)) {
3442                 conn = list_entry(checksends.next,
3443                                   struct kib_conn, ibc_connd_list);
3444                 list_del(&conn->ibc_connd_list);
3445
3446                 spin_lock(&conn->ibc_lock);
3447                 kiblnd_check_sends_locked(conn);
3448                 spin_unlock(&conn->ibc_lock);
3449
3450                 kiblnd_conn_decref(conn);
3451         }
3452 }
3453
3454 static void
3455 kiblnd_disconnect_conn(struct kib_conn *conn)
3456 {
3457         LASSERT (!in_interrupt());
3458         LASSERT (current == kiblnd_data.kib_connd);
3459         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3460
3461         rdma_disconnect(conn->ibc_cmid);
3462         kiblnd_finalise_conn(conn);
3463
3464         kiblnd_peer_notify(conn->ibc_peer);
3465 }
3466
3467 /*
3468  * High-water for reconnection to the same peer_ni, reconnection attempt should
3469  * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3470  */
3471 #define KIB_RECONN_HIGH_RACE    10
3472 /*
3473  * Allow connd to take a break and handle other things after consecutive
3474  * reconnection attemps.
3475  */
3476 #define KIB_RECONN_BREAK        100
3477
3478 int
3479 kiblnd_connd (void *arg)
3480 {
3481         spinlock_t        *lock= &kiblnd_data.kib_connd_lock;
3482         wait_queue_entry_t wait;
3483         unsigned long      flags;
3484         struct kib_conn *conn;
3485         int                timeout;
3486         int                i;
3487         int                dropped_lock;
3488         int                peer_index = 0;
3489         unsigned long      deadline = jiffies;
3490
3491         init_wait(&wait);
3492         kiblnd_data.kib_connd = current;
3493
3494         spin_lock_irqsave(lock, flags);
3495
3496         while (!kiblnd_data.kib_shutdown) {
3497                 int reconn = 0;
3498
3499                 dropped_lock = 0;
3500
3501                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3502                         struct kib_peer_ni *peer_ni = NULL;
3503
3504                         conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3505                                           struct kib_conn, ibc_list);
3506                         list_del(&conn->ibc_list);
3507                         if (conn->ibc_reconnect) {
3508                                 peer_ni = conn->ibc_peer;
3509                                 kiblnd_peer_addref(peer_ni);
3510                         }
3511
3512                         spin_unlock_irqrestore(lock, flags);
3513                         dropped_lock = 1;
3514
3515                         kiblnd_destroy_conn(conn);
3516
3517                         spin_lock_irqsave(lock, flags);
3518                         if (!peer_ni) {
3519                                 LIBCFS_FREE(conn, sizeof(*conn));
3520                                 continue;
3521                         }
3522
3523                         conn->ibc_peer = peer_ni;
3524                         if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3525                                 list_add_tail(&conn->ibc_list,
3526                                               &kiblnd_data.kib_reconn_list);
3527                         else
3528                                 list_add_tail(&conn->ibc_list,
3529                                               &kiblnd_data.kib_reconn_wait);
3530                 }
3531
3532                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3533                         int wait;
3534                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3535                                           struct kib_conn, ibc_list);
3536                         list_del(&conn->ibc_list);
3537
3538                         spin_unlock_irqrestore(lock, flags);
3539                         dropped_lock = 1;
3540
3541                         kiblnd_disconnect_conn(conn);
3542                         wait = conn->ibc_waits;
3543                         if (wait == 0) /* keep ref for connd_wait, see below */
3544                                 kiblnd_conn_decref(conn);
3545
3546                         spin_lock_irqsave(lock, flags);
3547
3548                         if (wait)
3549                                 list_add_tail(&conn->ibc_list,
3550                                               &kiblnd_data.kib_connd_waits);
3551                 }
3552
3553                 while (reconn < KIB_RECONN_BREAK) {
3554                         if (kiblnd_data.kib_reconn_sec !=
3555                             ktime_get_real_seconds()) {
3556                                 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3557                                 list_splice_init(&kiblnd_data.kib_reconn_wait,
3558                                                  &kiblnd_data.kib_reconn_list);
3559                         }
3560
3561                         if (list_empty(&kiblnd_data.kib_reconn_list))
3562                                 break;
3563
3564                         conn = list_entry(kiblnd_data.kib_reconn_list.next,
3565                                           struct kib_conn, ibc_list);
3566                         list_del(&conn->ibc_list);
3567
3568                         spin_unlock_irqrestore(lock, flags);
3569                         dropped_lock = 1;
3570
3571                         reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3572                         kiblnd_peer_decref(conn->ibc_peer);
3573                         LIBCFS_FREE(conn, sizeof(*conn));
3574
3575                         spin_lock_irqsave(lock, flags);
3576                 }
3577
3578                 if (!list_empty(&kiblnd_data.kib_connd_waits)) {
3579                         conn = list_entry(kiblnd_data.kib_connd_waits.next,
3580                                           struct kib_conn, ibc_list);
3581                         list_del(&conn->ibc_list);
3582                         spin_unlock_irqrestore(lock, flags);
3583
3584                         dropped_lock = kiblnd_tx_may_discard(conn);
3585                         if (dropped_lock)
3586                                 kiblnd_conn_decref(conn);
3587
3588                         spin_lock_irqsave(lock, flags);
3589                         if (dropped_lock == 0)
3590                                 list_add_tail(&conn->ibc_list,
3591                                               &kiblnd_data.kib_connd_waits);
3592                 }
3593
3594                 /* careful with the jiffy wrap... */
3595                 timeout = (int)(deadline - jiffies);
3596                 if (timeout <= 0) {
3597                         const int n = 4;
3598                         const int p = 1;
3599                         int       chunk = kiblnd_data.kib_peer_hash_size;
3600                         unsigned int lnd_timeout;
3601
3602                         spin_unlock_irqrestore(lock, flags);
3603                         dropped_lock = 1;
3604
3605                         /* Time to check for RDMA timeouts on a few more
3606                          * peers: I do checks every 'p' seconds on a
3607                          * proportion of the peer_ni table and I need to check
3608                          * every connection 'n' times within a timeout
3609                          * interval, to ensure I detect a timeout on any
3610                          * connection within (n+1)/n times the timeout
3611                          * interval. */
3612
3613                         lnd_timeout = kiblnd_timeout();
3614                         if (lnd_timeout > n * p)
3615                                 chunk = (chunk * n * p) / lnd_timeout;
3616                         if (chunk == 0)
3617                                 chunk = 1;
3618
3619                         for (i = 0; i < chunk; i++) {
3620                                 kiblnd_check_conns(peer_index);
3621                                 peer_index = (peer_index + 1) %
3622                                              kiblnd_data.kib_peer_hash_size;
3623                         }
3624
3625                         deadline += cfs_time_seconds(p);
3626                         spin_lock_irqsave(lock, flags);
3627                 }
3628
3629                 if (dropped_lock)
3630                         continue;
3631
3632                 /* Nothing to do for 'timeout'  */
3633                 set_current_state(TASK_INTERRUPTIBLE);
3634                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3635                 spin_unlock_irqrestore(lock, flags);
3636
3637                 schedule_timeout(timeout);
3638
3639                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3640                 spin_lock_irqsave(lock, flags);
3641         }
3642
3643         spin_unlock_irqrestore(lock, flags);
3644
3645         kiblnd_thread_fini();
3646         return 0;
3647 }
3648
3649 void
3650 kiblnd_qp_event(struct ib_event *event, void *arg)
3651 {
3652         struct kib_conn *conn = arg;
3653
3654         switch (event->event) {
3655         case IB_EVENT_COMM_EST:
3656                 CDEBUG(D_NET, "%s established\n",
3657                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3658                 /* We received a packet but connection isn't established
3659                  * probably handshake packet was lost, so free to
3660                  * force make connection established */
3661                 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3662                 return;
3663
3664         case IB_EVENT_PORT_ERR:
3665         case IB_EVENT_DEVICE_FATAL:
3666                 CERROR("Fatal device error for NI %s\n",
3667                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3668                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3669                 return;
3670
3671         case IB_EVENT_PORT_ACTIVE:
3672                 CERROR("Port reactivated for NI %s\n",
3673                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3674                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3675                 return;
3676
3677         default:
3678                 CERROR("%s: Async QP event type %d\n",
3679                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3680                 return;
3681         }
3682 }
3683
3684 static void
3685 kiblnd_complete (struct ib_wc *wc)
3686 {
3687         switch (kiblnd_wreqid2type(wc->wr_id)) {
3688         default:
3689                 LBUG();
3690
3691         case IBLND_WID_MR:
3692                 if (wc->status != IB_WC_SUCCESS &&
3693                     wc->status != IB_WC_WR_FLUSH_ERR)
3694                         CNETERR("FastReg failed: %d\n", wc->status);
3695                 return;
3696
3697         case IBLND_WID_RDMA:
3698                 /* We only get RDMA completion notification if it fails.  All
3699                  * subsequent work items, including the final SEND will fail
3700                  * too.  However we can't print out any more info about the
3701                  * failing RDMA because 'tx' might be back on the idle list or
3702                  * even reused already if we didn't manage to post all our work
3703                  * items */
3704                 CNETERR("RDMA (tx: %p) failed: %d\n",
3705                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3706                 return;
3707
3708         case IBLND_WID_TX:
3709                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3710                 return;
3711
3712         case IBLND_WID_RX:
3713                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3714                                    wc->byte_len);
3715                 return;
3716         }
3717 }
3718
3719 void
3720 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3721 {
3722         /* NB I'm not allowed to schedule this conn once its refcount has
3723          * reached 0.  Since fundamentally I'm racing with scheduler threads
3724          * consuming my CQ I could be called after all completions have
3725          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3726          * and this CQ is about to be destroyed so I NOOP. */
3727         struct kib_conn *conn = arg;
3728         struct kib_sched_info *sched = conn->ibc_sched;
3729         unsigned long flags;
3730
3731         LASSERT(cq == conn->ibc_cq);
3732
3733         spin_lock_irqsave(&sched->ibs_lock, flags);
3734
3735         conn->ibc_ready = 1;
3736
3737         if (!conn->ibc_scheduled &&
3738             (conn->ibc_nrx > 0 ||
3739              conn->ibc_nsends_posted > 0)) {
3740                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3741                 conn->ibc_scheduled = 1;
3742                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3743
3744                 if (waitqueue_active(&sched->ibs_waitq))
3745                         wake_up(&sched->ibs_waitq);
3746         }
3747
3748         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3749 }
3750
3751 void
3752 kiblnd_cq_event(struct ib_event *event, void *arg)
3753 {
3754         struct kib_conn *conn = arg;
3755
3756         CERROR("%s: async CQ event type %d\n",
3757                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3758 }
3759
3760 int
3761 kiblnd_scheduler(void *arg)
3762 {
3763         long                    id = (long)arg;
3764         struct kib_sched_info   *sched;
3765         struct kib_conn *conn;
3766         wait_queue_entry_t      wait;
3767         unsigned long           flags;
3768         struct ib_wc            wc;
3769         int                     did_something;
3770         int                     rc;
3771
3772         init_wait(&wait);
3773
3774         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3775
3776         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3777         if (rc != 0) {
3778                 CWARN("Unable to bind on CPU partition %d, please verify "
3779                       "whether all CPUs are healthy and reload modules if "
3780                       "necessary, otherwise your system might under risk of "
3781                       "low performance\n", sched->ibs_cpt);
3782         }
3783
3784         spin_lock_irqsave(&sched->ibs_lock, flags);
3785
3786         while (!kiblnd_data.kib_shutdown) {
3787                 if (need_resched()) {
3788                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3789
3790                         cond_resched();
3791
3792                         spin_lock_irqsave(&sched->ibs_lock, flags);
3793                 }
3794
3795                 did_something = 0;
3796
3797                 if (!list_empty(&sched->ibs_conns)) {
3798                         conn = list_entry(sched->ibs_conns.next,
3799                                           struct kib_conn, ibc_sched_list);
3800                         /* take over kib_sched_conns' ref on conn... */
3801                         LASSERT(conn->ibc_scheduled);
3802                         list_del(&conn->ibc_sched_list);
3803                         conn->ibc_ready = 0;
3804
3805                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3806
3807                         wc.wr_id = IBLND_WID_INVAL;
3808
3809                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3810                         if (rc == 0) {
3811                                 rc = ib_req_notify_cq(conn->ibc_cq,
3812                                                       IB_CQ_NEXT_COMP);
3813                                 if (rc < 0) {
3814                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3815                                               "closing connection\n",
3816                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3817                                         kiblnd_close_conn(conn, -EIO);
3818                                         kiblnd_conn_decref(conn);
3819                                         spin_lock_irqsave(&sched->ibs_lock,
3820                                                               flags);
3821                                         continue;
3822                                 }
3823
3824                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3825                         }
3826
3827                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3828                                 LCONSOLE_ERROR(
3829                                         "ib_poll_cq (rc: %d) returned invalid "
3830                                         "wr_id, opcode %d, status: %d, "
3831                                         "vendor_err: %d, conn: %s status: %d\n"
3832                                         "please upgrade firmware and OFED or "
3833                                         "contact vendor.\n", rc,
3834                                         wc.opcode, wc.status, wc.vendor_err,
3835                                         libcfs_nid2str(conn->ibc_peer->ibp_nid),
3836                                         conn->ibc_state);
3837                                 rc = -EINVAL;
3838                         }
3839
3840                         if (rc < 0) {
3841                                 CWARN("%s: ib_poll_cq failed: %d, "
3842                                       "closing connection\n",
3843                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3844                                       rc);
3845                                 kiblnd_close_conn(conn, -EIO);
3846                                 kiblnd_conn_decref(conn);
3847                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3848                                 continue;
3849                         }
3850
3851                         spin_lock_irqsave(&sched->ibs_lock, flags);
3852
3853                         if (rc != 0 || conn->ibc_ready) {
3854                                 /* There may be another completion waiting; get
3855                                  * another scheduler to check while I handle
3856                                  * this one... */
3857                                 /* +1 ref for sched_conns */
3858                                 kiblnd_conn_addref(conn);
3859                                 list_add_tail(&conn->ibc_sched_list,
3860                                                   &sched->ibs_conns);
3861                                 if (waitqueue_active(&sched->ibs_waitq))
3862                                         wake_up(&sched->ibs_waitq);
3863                         } else {
3864                                 conn->ibc_scheduled = 0;
3865                         }
3866
3867                         if (rc != 0) {
3868                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3869                                 kiblnd_complete(&wc);
3870
3871                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3872                         }
3873
3874                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3875                         did_something = 1;
3876                 }
3877
3878                 if (did_something)
3879                         continue;
3880
3881                 set_current_state(TASK_INTERRUPTIBLE);
3882                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3883                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3884
3885                 schedule();
3886
3887                 remove_wait_queue(&sched->ibs_waitq, &wait);
3888                 set_current_state(TASK_RUNNING);
3889                 spin_lock_irqsave(&sched->ibs_lock, flags);
3890         }
3891
3892         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3893
3894         kiblnd_thread_fini();
3895         return 0;
3896 }
3897
3898 int
3899 kiblnd_failover_thread(void *arg)
3900 {
3901         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
3902         struct kib_dev *dev;
3903         struct net *ns = arg;
3904         wait_queue_entry_t wait;
3905         unsigned long    flags;
3906         int              rc;
3907
3908         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3909
3910         init_wait(&wait);
3911         write_lock_irqsave(glock, flags);
3912
3913         while (!kiblnd_data.kib_shutdown) {
3914                 int     do_failover = 0;
3915                 int     long_sleep;
3916
3917                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3918                                     ibd_fail_list) {
3919                         if (ktime_get_seconds() < dev->ibd_next_failover)
3920                                 continue;
3921                         do_failover = 1;
3922                         break;
3923                 }
3924
3925                 if (do_failover) {
3926                         list_del_init(&dev->ibd_fail_list);
3927                         dev->ibd_failover = 1;
3928                         write_unlock_irqrestore(glock, flags);
3929
3930                         rc = kiblnd_dev_failover(dev, ns);
3931
3932                         write_lock_irqsave(glock, flags);
3933
3934                         LASSERT (dev->ibd_failover);
3935                         dev->ibd_failover = 0;
3936                         if (rc >= 0) { /* Device is OK or failover succeed */
3937                                 dev->ibd_next_failover = ktime_get_seconds() + 3;
3938                                 continue;
3939                         }
3940
3941                         /* failed to failover, retry later */
3942                         dev->ibd_next_failover = ktime_get_seconds() +
3943                                                  min(dev->ibd_failed_failover, 10);
3944                         if (kiblnd_dev_can_failover(dev)) {
3945                                 list_add_tail(&dev->ibd_fail_list,
3946                                               &kiblnd_data.kib_failed_devs);
3947                         }
3948
3949                         continue;
3950                 }
3951
3952                 /* long sleep if no more pending failover */
3953                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3954
3955                 set_current_state(TASK_INTERRUPTIBLE);
3956                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3957                 write_unlock_irqrestore(glock, flags);
3958
3959                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3960                                                    cfs_time_seconds(1));
3961                 set_current_state(TASK_RUNNING);
3962                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3963                 write_lock_irqsave(glock, flags);
3964
3965                 if (!long_sleep || rc != 0)
3966                         continue;
3967
3968                 /* have a long sleep, routine check all active devices,
3969                  * we need checking like this because if there is not active
3970                  * connection on the dev and no SEND from local, we may listen
3971                  * on wrong HCA for ever while there is a bonding failover */
3972                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3973                         if (kiblnd_dev_can_failover(dev)) {
3974                                 list_add_tail(&dev->ibd_fail_list,
3975                                               &kiblnd_data.kib_failed_devs);
3976                         }
3977                 }
3978         }
3979
3980         write_unlock_irqrestore(glock, flags);
3981
3982         kiblnd_thread_fini();
3983         return 0;
3984 }