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