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
1297 kiblnd_resolve_addr_cap(struct rdma_cm_id *cmid,
1298                         struct sockaddr_in *srcaddr,
1299                         struct sockaddr_in *dstaddr,
1300                         int timeout_ms)
1301 {
1302         unsigned short port;
1303         int rc;
1304
1305         LASSERT(capable(CAP_NET_BIND_SERVICE));
1306
1307         /* allow the port to be reused */
1308         rc = rdma_set_reuseaddr(cmid, 1);
1309         if (rc != 0) {
1310                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1311                 return rc;
1312         }
1313
1314         /* look for a free privileged port */
1315         for (port = PROT_SOCK-1; port > 0; port--) {
1316                 srcaddr->sin_port = htons(port);
1317                 rc = rdma_resolve_addr(cmid,
1318                                        (struct sockaddr *)srcaddr,
1319                                        (struct sockaddr *)dstaddr,
1320                                        timeout_ms);
1321                 if (rc == 0) {
1322                         CDEBUG(D_NET, "bound to port %hu\n", port);
1323                         return 0;
1324                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1325                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1326                                port, rc);
1327                 } else {
1328                         return rc;
1329                 }
1330         }
1331
1332         CERROR("Failed to bind to a free privileged port\n");
1333         return rc;
1334 }
1335
1336 static int
1337 kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1338                     struct sockaddr_in *srcaddr,
1339                     struct sockaddr_in *dstaddr,
1340                     int timeout_ms)
1341 {
1342         const struct cred *old_creds = NULL;
1343         struct cred *new_creds;
1344         int rc;
1345
1346         if (!capable(CAP_NET_BIND_SERVICE)) {
1347                 new_creds = prepare_creds();
1348                 if (!new_creds)
1349                         return -ENOMEM;
1350
1351                 cap_raise(new_creds->cap_effective, CAP_NET_BIND_SERVICE);
1352                 old_creds = override_creds(new_creds);
1353         }
1354
1355         rc = kiblnd_resolve_addr_cap(cmid, srcaddr, dstaddr, timeout_ms);
1356
1357         if (old_creds)
1358                 revert_creds(old_creds);
1359
1360         return rc;
1361 }
1362
1363 static void
1364 kiblnd_connect_peer(struct kib_peer_ni *peer_ni)
1365 {
1366         struct rdma_cm_id *cmid;
1367         struct kib_dev *dev;
1368         struct kib_net *net = peer_ni->ibp_ni->ni_data;
1369         struct sockaddr_in srcaddr;
1370         struct sockaddr_in dstaddr;
1371         int rc;
1372
1373         LASSERT (net != NULL);
1374         LASSERT (peer_ni->ibp_connecting > 0);
1375
1376         cmid = kiblnd_rdma_create_id(peer_ni->ibp_ni->ni_net_ns,
1377                                      kiblnd_cm_callback, peer_ni,
1378                                      RDMA_PS_TCP, IB_QPT_RC);
1379
1380         if (IS_ERR(cmid)) {
1381                 CERROR("Can't create CMID for %s: %ld\n",
1382                        libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
1383                 rc = PTR_ERR(cmid);
1384                 goto failed;
1385         }
1386
1387         dev = net->ibn_dev;
1388         memset(&srcaddr, 0, sizeof(srcaddr));
1389         srcaddr.sin_family = AF_INET;
1390         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1391
1392         memset(&dstaddr, 0, sizeof(dstaddr));
1393         dstaddr.sin_family = AF_INET;
1394         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1395         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
1396
1397         kiblnd_peer_addref(peer_ni);               /* cmid's ref */
1398
1399         if (*kiblnd_tunables.kib_use_priv_port) {
1400                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1401                                          kiblnd_timeout() * 1000);
1402         } else {
1403                 rc = rdma_resolve_addr(cmid,
1404                                        (struct sockaddr *)&srcaddr,
1405                                        (struct sockaddr *)&dstaddr,
1406                                        kiblnd_timeout() * 1000);
1407         }
1408         if (rc != 0) {
1409                 /* Can't initiate address resolution:  */
1410                 CERROR("Can't resolve addr for %s: %d\n",
1411                        libcfs_nid2str(peer_ni->ibp_nid), rc);
1412                 goto failed2;
1413         }
1414
1415         return;
1416
1417  failed2:
1418         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1419         kiblnd_peer_decref(peer_ni);               /* cmid's ref */
1420         rdma_destroy_id(cmid);
1421         return;
1422  failed:
1423         kiblnd_peer_connect_failed(peer_ni, 1, rc);
1424 }
1425
1426 bool
1427 kiblnd_reconnect_peer(struct kib_peer_ni *peer_ni)
1428 {
1429         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1430         char *reason = NULL;
1431         LIST_HEAD(txs);
1432         unsigned long flags;
1433
1434         write_lock_irqsave(glock, flags);
1435         if (peer_ni->ibp_reconnecting == 0) {
1436                 if (peer_ni->ibp_accepting)
1437                         reason = "accepting";
1438                 else if (peer_ni->ibp_connecting)
1439                         reason = "connecting";
1440                 else if (!list_empty(&peer_ni->ibp_conns))
1441                         reason = "connected";
1442                 else /* connected then closed */
1443                         reason = "closed";
1444
1445                 goto no_reconnect;
1446         }
1447
1448         if (peer_ni->ibp_accepting)
1449                 CNETERR("Detecting race between accepting and reconnecting\n");
1450         peer_ni->ibp_reconnecting--;
1451
1452         if (!kiblnd_peer_active(peer_ni)) {
1453                 list_splice_init(&peer_ni->ibp_tx_queue, &txs);
1454                 reason = "unlinked";
1455                 goto no_reconnect;
1456         }
1457
1458         peer_ni->ibp_connecting++;
1459         peer_ni->ibp_reconnected++;
1460
1461         write_unlock_irqrestore(glock, flags);
1462
1463         kiblnd_connect_peer(peer_ni);
1464         return true;
1465
1466  no_reconnect:
1467         write_unlock_irqrestore(glock, flags);
1468
1469         CWARN("Abort reconnection of %s: %s\n",
1470               libcfs_nid2str(peer_ni->ibp_nid), reason);
1471         kiblnd_txlist_done(&txs, -ECONNABORTED,
1472                            LNET_MSG_STATUS_LOCAL_ABORTED);
1473         return false;
1474 }
1475
1476 void
1477 kiblnd_launch_tx(struct lnet_ni *ni, struct kib_tx *tx, lnet_nid_t nid)
1478 {
1479         struct kib_peer_ni *peer_ni;
1480         struct kib_peer_ni *peer2;
1481         struct kib_conn *conn;
1482         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1483         unsigned long      flags;
1484         int                rc;
1485         int                i;
1486         struct lnet_ioctl_config_o2iblnd_tunables *tunables;
1487
1488         /* If I get here, I've committed to send, so I complete the tx with
1489          * failure on any problems */
1490
1491         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1492         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1493
1494         /* First time, just use a read lock since I expect to find my peer_ni
1495          * connected */
1496         read_lock_irqsave(g_lock, flags);
1497
1498         peer_ni = kiblnd_find_peer_locked(ni, nid);
1499         if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
1500                 /* Found a peer_ni with an established connection */
1501                 conn = kiblnd_get_conn_locked(peer_ni);
1502                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1503
1504                 read_unlock_irqrestore(g_lock, flags);
1505
1506                 if (tx != NULL)
1507                         kiblnd_queue_tx(tx, conn);
1508                 kiblnd_conn_decref(conn); /* ...to here */
1509                 return;
1510         }
1511
1512         read_unlock(g_lock);
1513         /* Re-try with a write lock */
1514         write_lock(g_lock);
1515
1516         peer_ni = kiblnd_find_peer_locked(ni, nid);
1517         if (peer_ni != NULL) {
1518                 if (list_empty(&peer_ni->ibp_conns)) {
1519                         /* found a peer_ni, but it's still connecting... */
1520                         LASSERT(kiblnd_peer_connecting(peer_ni));
1521                         if (tx != NULL)
1522                                 list_add_tail(&tx->tx_list,
1523                                                   &peer_ni->ibp_tx_queue);
1524                         write_unlock_irqrestore(g_lock, flags);
1525                 } else {
1526                         conn = kiblnd_get_conn_locked(peer_ni);
1527                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1528
1529                         write_unlock_irqrestore(g_lock, flags);
1530
1531                         if (tx != NULL)
1532                                 kiblnd_queue_tx(tx, conn);
1533                         kiblnd_conn_decref(conn); /* ...to here */
1534                 }
1535                 return;
1536         }
1537
1538         write_unlock_irqrestore(g_lock, flags);
1539
1540         /* Allocate a peer_ni ready to add to the peer_ni table and retry */
1541         rc = kiblnd_create_peer(ni, &peer_ni, nid);
1542         if (rc != 0) {
1543                 CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
1544                 if (tx != NULL) {
1545                         tx->tx_status = -EHOSTUNREACH;
1546                         tx->tx_waiting = 0;
1547                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1548                         kiblnd_tx_done(tx);
1549                 }
1550                 return;
1551         }
1552
1553         write_lock_irqsave(g_lock, flags);
1554
1555         peer2 = kiblnd_find_peer_locked(ni, nid);
1556         if (peer2 != NULL) {
1557                 if (list_empty(&peer2->ibp_conns)) {
1558                         /* found a peer_ni, but it's still connecting... */
1559                         LASSERT(kiblnd_peer_connecting(peer2));
1560                         if (tx != NULL)
1561                                 list_add_tail(&tx->tx_list,
1562                                                   &peer2->ibp_tx_queue);
1563                         write_unlock_irqrestore(g_lock, flags);
1564                 } else {
1565                         conn = kiblnd_get_conn_locked(peer2);
1566                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1567
1568                         write_unlock_irqrestore(g_lock, flags);
1569
1570                         if (tx != NULL)
1571                                 kiblnd_queue_tx(tx, conn);
1572                         kiblnd_conn_decref(conn); /* ...to here */
1573                 }
1574
1575                 kiblnd_peer_decref(peer_ni);
1576                 return;
1577         }
1578
1579         /* Brand new peer_ni */
1580         LASSERT(peer_ni->ibp_connecting == 0);
1581         tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
1582         peer_ni->ibp_connecting = tunables->lnd_conns_per_peer;
1583
1584         /* always called with a ref on ni, which prevents ni being shutdown */
1585         LASSERT(((struct kib_net *)ni->ni_data)->ibn_shutdown == 0);
1586
1587         if (tx != NULL)
1588                 list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
1589
1590         kiblnd_peer_addref(peer_ni);
1591         list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
1592
1593         write_unlock_irqrestore(g_lock, flags);
1594
1595         for (i = 0; i < tunables->lnd_conns_per_peer; i++)
1596                 kiblnd_connect_peer(peer_ni);
1597         kiblnd_peer_decref(peer_ni);
1598 }
1599
1600 int
1601 kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
1602 {
1603         struct lnet_hdr *hdr = &lntmsg->msg_hdr;
1604         int               type = lntmsg->msg_type;
1605         struct lnet_process_id target = lntmsg->msg_target;
1606         int               target_is_router = lntmsg->msg_target_is_router;
1607         int               routing = lntmsg->msg_routing;
1608         unsigned int      payload_niov = lntmsg->msg_niov;
1609         struct bio_vec   *payload_kiov = lntmsg->msg_kiov;
1610         unsigned int      payload_offset = lntmsg->msg_offset;
1611         unsigned int      payload_nob = lntmsg->msg_len;
1612         struct kib_msg *ibmsg;
1613         struct kib_rdma_desc *rd;
1614         struct kib_tx *tx;
1615         int               nob;
1616         int               rc;
1617
1618         /* NB 'private' is different depending on what we're sending.... */
1619
1620         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1621                payload_nob, payload_niov, libcfs_id2str(target));
1622
1623         LASSERT (payload_nob == 0 || payload_niov > 0);
1624         LASSERT (payload_niov <= LNET_MAX_IOV);
1625
1626         /* Thread context */
1627         LASSERT (!in_interrupt());
1628
1629         switch (type) {
1630         default:
1631                 LBUG();
1632                 return (-EIO);
1633
1634         case LNET_MSG_ACK:
1635                 LASSERT (payload_nob == 0);
1636                 break;
1637
1638         case LNET_MSG_GET:
1639                 if (routing || target_is_router)
1640                         break;                  /* send IMMEDIATE */
1641
1642                 /* is the REPLY message too small for RDMA? */
1643                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1644                 if (nob <= IBLND_MSG_SIZE)
1645                         break;                  /* send IMMEDIATE */
1646
1647                 tx = kiblnd_get_idle_tx(ni, target.nid);
1648                 if (tx == NULL) {
1649                         CERROR("Can't allocate txd for GET to %s\n",
1650                                libcfs_nid2str(target.nid));
1651                         return -ENOMEM;
1652                 }
1653
1654                 ibmsg = tx->tx_msg;
1655                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1656                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1657                                           lntmsg->msg_md->md_niov,
1658                                           lntmsg->msg_md->md_kiov,
1659                                           0, lntmsg->msg_md->md_length);
1660                 if (rc != 0) {
1661                         CERROR("Can't setup GET sink for %s: %d\n",
1662                                libcfs_nid2str(target.nid), rc);
1663                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1664                         kiblnd_tx_done(tx);
1665                         return -EIO;
1666                 }
1667
1668                 nob = offsetof(struct kib_get_msg, ibgm_rd.rd_frags[rd->rd_nfrags]);
1669                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1670                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1671
1672                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1673
1674                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1675                 if (tx->tx_lntmsg[1] == NULL) {
1676                         CERROR("Can't create reply for GET -> %s\n",
1677                                libcfs_nid2str(target.nid));
1678                         kiblnd_tx_done(tx);
1679                         return -EIO;
1680                 }
1681
1682                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1683                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1684                 kiblnd_launch_tx(ni, tx, target.nid);
1685                 return 0;
1686
1687         case LNET_MSG_REPLY:
1688         case LNET_MSG_PUT:
1689                 /* Is the payload small enough not to need RDMA? */
1690                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob]);
1691                 if (nob <= IBLND_MSG_SIZE)
1692                         break;                  /* send IMMEDIATE */
1693
1694                 tx = kiblnd_get_idle_tx(ni, target.nid);
1695                 if (tx == NULL) {
1696                         CERROR("Can't allocate %s txd for %s\n",
1697                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1698                                libcfs_nid2str(target.nid));
1699                         return -ENOMEM;
1700                 }
1701
1702                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1703                                           payload_niov, payload_kiov,
1704                                           payload_offset, payload_nob);
1705                 if (rc != 0) {
1706                         CERROR("Can't setup PUT src for %s: %d\n",
1707                                libcfs_nid2str(target.nid), rc);
1708                         kiblnd_tx_done(tx);
1709                         return -EIO;
1710                 }
1711
1712                 ibmsg = tx->tx_msg;
1713                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1714                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1715                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ,
1716                                    sizeof(struct kib_putreq_msg));
1717
1718                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1719                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1720                 kiblnd_launch_tx(ni, tx, target.nid);
1721                 return 0;
1722         }
1723
1724         /* send IMMEDIATE */
1725         LASSERT(offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[payload_nob])
1726                 <= IBLND_MSG_SIZE);
1727
1728         tx = kiblnd_get_idle_tx(ni, target.nid);
1729         if (tx == NULL) {
1730                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1731                         type, libcfs_nid2str(target.nid));
1732                 return -ENOMEM;
1733         }
1734
1735         ibmsg = tx->tx_msg;
1736         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1737
1738         lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1739                             offsetof(struct kib_msg,
1740                                      ibm_u.immediate.ibim_payload),
1741                             payload_niov, payload_kiov,
1742                             payload_offset, payload_nob);
1743
1744         nob = offsetof(struct kib_immediate_msg, ibim_payload[payload_nob]);
1745         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1746
1747         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1748         kiblnd_launch_tx(ni, tx, target.nid);
1749         return 0;
1750 }
1751
1752 static void
1753 kiblnd_reply(struct lnet_ni *ni, struct kib_rx *rx, struct lnet_msg *lntmsg)
1754 {
1755         struct lnet_process_id target = lntmsg->msg_target;
1756         unsigned int niov = lntmsg->msg_niov;
1757         struct bio_vec *kiov = lntmsg->msg_kiov;
1758         unsigned int offset = lntmsg->msg_offset;
1759         unsigned int nob = lntmsg->msg_len;
1760         struct kib_tx *tx;
1761         int rc;
1762
1763         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1764         if (tx == NULL) {
1765                 CERROR("Can't get tx for REPLY to %s\n",
1766                        libcfs_nid2str(target.nid));
1767                 goto failed_0;
1768         }
1769
1770         if (nob == 0)
1771                 rc = 0;
1772         else
1773                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1774                                           niov, kiov, offset, nob);
1775
1776         if (rc != 0) {
1777                 CERROR("Can't setup GET src for %s: %d\n",
1778                        libcfs_nid2str(target.nid), rc);
1779                 goto failed_1;
1780         }
1781
1782         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1783                               IBLND_MSG_GET_DONE, nob,
1784                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1785                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1786         if (rc < 0) {
1787                 CERROR("Can't setup rdma for GET from %s: %d\n",
1788                        libcfs_nid2str(target.nid), rc);
1789                 goto failed_1;
1790         }
1791
1792         if (nob == 0) {
1793                 /* No RDMA: local completion may happen now! */
1794                 lnet_finalize(lntmsg, 0);
1795         } else {
1796                 /* RDMA: lnet_finalize(lntmsg) when it
1797                  * completes */
1798                 tx->tx_lntmsg[0] = lntmsg;
1799         }
1800
1801         kiblnd_queue_tx(tx, rx->rx_conn);
1802         return;
1803
1804
1805 failed_1:
1806         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1807         kiblnd_tx_done(tx);
1808 failed_0:
1809         lnet_finalize(lntmsg, -EIO);
1810 }
1811
1812 int
1813 kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1814             int delayed, unsigned int niov, struct bio_vec *kiov,
1815             unsigned int offset, unsigned int mlen, unsigned int rlen)
1816 {
1817         struct kib_rx *rx = private;
1818         struct kib_msg *rxmsg = rx->rx_msg;
1819         struct kib_conn *conn = rx->rx_conn;
1820         struct kib_tx *tx;
1821         __u64        ibprm_cookie;
1822         int          nob;
1823         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1824         int          rc = 0;
1825
1826         LASSERT (mlen <= rlen);
1827         LASSERT (!in_interrupt());
1828
1829         switch (rxmsg->ibm_type) {
1830         default:
1831                 LBUG();
1832
1833         case IBLND_MSG_IMMEDIATE:
1834                 nob = offsetof(struct kib_msg, ibm_u.immediate.ibim_payload[rlen]);
1835                 if (nob > rx->rx_nob) {
1836                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1837                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1838                                 nob, rx->rx_nob);
1839                         rc = -EPROTO;
1840                         break;
1841                 }
1842
1843                 lnet_copy_flat2kiov(niov, kiov, offset,
1844                                     IBLND_MSG_SIZE, rxmsg,
1845                                     offsetof(struct kib_msg,
1846                                              ibm_u.immediate.ibim_payload),
1847                                     mlen);
1848                 lnet_finalize(lntmsg, 0);
1849                 break;
1850
1851         case IBLND_MSG_PUT_REQ: {
1852                 struct kib_msg  *txmsg;
1853                 struct kib_rdma_desc *rd;
1854                 ibprm_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1855
1856                 if (mlen == 0) {
1857                         lnet_finalize(lntmsg, 0);
1858                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1859                                                0, ibprm_cookie);
1860                         break;
1861                 }
1862
1863                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1864                 if (tx == NULL) {
1865                         CERROR("Can't allocate tx for %s\n",
1866                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1867                         /* Not replying will break the connection */
1868                         rc = -ENOMEM;
1869                         break;
1870                 }
1871
1872                 txmsg = tx->tx_msg;
1873                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1874                 rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1875                                           niov, kiov, offset, mlen);
1876                 if (rc != 0) {
1877                         CERROR("Can't setup PUT sink for %s: %d\n",
1878                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1879                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
1880                         kiblnd_tx_done(tx);
1881                         /* tell peer_ni it's over */
1882                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK,
1883                                                rc, ibprm_cookie);
1884                         break;
1885                 }
1886
1887                 nob = offsetof(struct kib_putack_msg, ibpam_rd.rd_frags[rd->rd_nfrags]);
1888                 txmsg->ibm_u.putack.ibpam_src_cookie = ibprm_cookie;
1889                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1890
1891                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1892
1893                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1894                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1895                 kiblnd_queue_tx(tx, conn);
1896
1897                 /* reposted buffer reserved for PUT_DONE */
1898                 post_credit = IBLND_POSTRX_NO_CREDIT;
1899                 break;
1900                 }
1901
1902         case IBLND_MSG_GET_REQ:
1903                 if (lntmsg != NULL) {
1904                         /* Optimized GET; RDMA lntmsg's payload */
1905                         kiblnd_reply(ni, rx, lntmsg);
1906                 } else {
1907                         /* GET didn't match anything */
1908                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1909                                                -ENODATA,
1910                                                rxmsg->ibm_u.get.ibgm_cookie);
1911                 }
1912                 break;
1913         }
1914
1915         kiblnd_post_rx(rx, post_credit);
1916         return rc;
1917 }
1918
1919 int
1920 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1921 {
1922         struct task_struct *task = kthread_run(fn, arg, name);
1923
1924         if (IS_ERR(task))
1925                 return PTR_ERR(task);
1926
1927         atomic_inc(&kiblnd_data.kib_nthreads);
1928         return 0;
1929 }
1930
1931 static void
1932 kiblnd_thread_fini (void)
1933 {
1934         atomic_dec (&kiblnd_data.kib_nthreads);
1935 }
1936
1937 static void
1938 kiblnd_peer_alive(struct kib_peer_ni *peer_ni)
1939 {
1940         /* This is racy, but everyone's only writing ktime_get_seconds() */
1941         peer_ni->ibp_last_alive = ktime_get_seconds();
1942         smp_mb();
1943 }
1944
1945 static void
1946 kiblnd_peer_notify(struct kib_peer_ni *peer_ni)
1947 {
1948         int           error = 0;
1949         time64_t last_alive = 0;
1950         unsigned long flags;
1951
1952         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1953
1954         if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
1955                 error = peer_ni->ibp_error;
1956                 peer_ni->ibp_error = 0;
1957
1958                 last_alive = peer_ni->ibp_last_alive;
1959         }
1960
1961         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1962
1963         if (error != 0)
1964                 lnet_notify(peer_ni->ibp_ni,
1965                             peer_ni->ibp_nid, false, false, last_alive);
1966 }
1967
1968 void
1969 kiblnd_close_conn_locked(struct kib_conn *conn, int error)
1970 {
1971         /* This just does the immediate housekeeping.  'error' is zero for a
1972          * normal shutdown which can happen only after the connection has been
1973          * established.  If the connection is established, schedule the
1974          * connection to be finished off by the connd.  Otherwise the connd is
1975          * already dealing with it (either to set it up or tear it down).
1976          * Caller holds kib_global_lock exclusively in irq context */
1977         struct kib_peer_ni *peer_ni = conn->ibc_peer;
1978         struct kib_dev *dev;
1979         unsigned long flags;
1980
1981         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1982
1983         if (error != 0 && conn->ibc_comms_error == 0)
1984                 conn->ibc_comms_error = error;
1985
1986         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1987                 return; /* already being handled  */
1988
1989         if (error == 0 &&
1990             list_empty(&conn->ibc_tx_noops) &&
1991             list_empty(&conn->ibc_tx_queue) &&
1992             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1993             list_empty(&conn->ibc_tx_queue_nocred) &&
1994             list_empty(&conn->ibc_active_txs)) {
1995                 CDEBUG(D_NET, "closing conn to %s\n",
1996                        libcfs_nid2str(peer_ni->ibp_nid));
1997         } else {
1998                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1999                        libcfs_nid2str(peer_ni->ibp_nid), error,
2000                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
2001                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
2002                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
2003                                                 "" : "(sending_rsrvd)",
2004                        list_empty(&conn->ibc_tx_queue_nocred) ?
2005                                                  "" : "(sending_nocred)",
2006                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
2007         }
2008
2009         dev = ((struct kib_net *)peer_ni->ibp_ni->ni_data)->ibn_dev;
2010         if (peer_ni->ibp_next_conn == conn)
2011                 /* clear next_conn so it won't be used */
2012                 peer_ni->ibp_next_conn = NULL;
2013         list_del(&conn->ibc_list);
2014         /* connd (see below) takes over ibc_list's ref */
2015
2016         if (list_empty(&peer_ni->ibp_conns) &&    /* no more conns */
2017             kiblnd_peer_active(peer_ni)) {         /* still in peer_ni table */
2018                 kiblnd_unlink_peer_locked(peer_ni);
2019
2020                 /* set/clear error on last conn */
2021                 peer_ni->ibp_error = conn->ibc_comms_error;
2022         }
2023
2024         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
2025
2026         if (error != 0 &&
2027             kiblnd_dev_can_failover(dev)) {
2028                 list_add_tail(&dev->ibd_fail_list,
2029                               &kiblnd_data.kib_failed_devs);
2030                 wake_up(&kiblnd_data.kib_failover_waitq);
2031         }
2032
2033         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
2034
2035         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
2036         wake_up(&kiblnd_data.kib_connd_waitq);
2037
2038         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
2039 }
2040
2041 void
2042 kiblnd_close_conn(struct kib_conn *conn, int error)
2043 {
2044         unsigned long flags;
2045
2046         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2047
2048         kiblnd_close_conn_locked(conn, error);
2049
2050         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2051 }
2052
2053 static void
2054 kiblnd_handle_early_rxs(struct kib_conn *conn)
2055 {
2056         unsigned long flags;
2057         struct kib_rx *rx;
2058
2059         LASSERT(!in_interrupt());
2060         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
2061
2062         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2063         while (!list_empty(&conn->ibc_early_rxs)) {
2064                 rx = list_entry(conn->ibc_early_rxs.next,
2065                                 struct kib_rx, rx_list);
2066                 list_del(&rx->rx_list);
2067                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2068
2069                 kiblnd_handle_rx(rx);
2070
2071                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2072         }
2073         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2074 }
2075
2076 void
2077 kiblnd_abort_txs(struct kib_conn *conn, struct list_head *txs)
2078 {
2079         LIST_HEAD(zombies);
2080         struct kib_tx *nxt;
2081         struct kib_tx *tx;
2082
2083         spin_lock(&conn->ibc_lock);
2084
2085         list_for_each_entry_safe(tx, nxt, txs, tx_list) {
2086                 if (txs == &conn->ibc_active_txs) {
2087                         LASSERT(!tx->tx_queued);
2088                         LASSERT(tx->tx_waiting ||
2089                                 tx->tx_sending != 0);
2090                         if (conn->ibc_comms_error == -ETIMEDOUT) {
2091                                 if (tx->tx_waiting && !tx->tx_sending)
2092                                         tx->tx_hstatus =
2093                                           LNET_MSG_STATUS_REMOTE_TIMEOUT;
2094                                 else if (tx->tx_sending)
2095                                         tx->tx_hstatus =
2096                                           LNET_MSG_STATUS_NETWORK_TIMEOUT;
2097                         }
2098                 } else {
2099                         LASSERT(tx->tx_queued);
2100                         if (conn->ibc_comms_error == -ETIMEDOUT)
2101                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2102                         else
2103                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
2104                 }
2105
2106                 tx->tx_status = -ECONNABORTED;
2107                 tx->tx_waiting = 0;
2108
2109                 /*
2110                  * TODO: This makes an assumption that
2111                  * kiblnd_tx_complete() will be called for each tx. If
2112                  * that event is dropped we could end up with stale
2113                  * connections floating around. We'd like to deal with
2114                  * that in a better way.
2115                  *
2116                  * Also that means we can exceed the timeout by many
2117                  * seconds.
2118                  */
2119                 if (tx->tx_sending == 0) {
2120                         tx->tx_queued = 0;
2121                         list_move(&tx->tx_list, &zombies);
2122                 }
2123         }
2124
2125         spin_unlock(&conn->ibc_lock);
2126
2127         /*
2128          * aborting transmits occurs when finalizing the connection.
2129          * The connection is finalized on error.
2130          * Passing LNET_MSG_STATUS_OK to txlist_done() will not
2131          * override the value already set in tx->tx_hstatus above.
2132          */
2133         kiblnd_txlist_done(&zombies, -ECONNABORTED, LNET_MSG_STATUS_OK);
2134 }
2135
2136 static void
2137 kiblnd_finalise_conn(struct kib_conn *conn)
2138 {
2139         LASSERT (!in_interrupt());
2140         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2141
2142         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2143          * for connections that didn't get as far as being connected, because
2144          * rdma_disconnect() does this for free. */
2145         kiblnd_abort_receives(conn);
2146
2147         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2148
2149         /* Complete all tx descs not waiting for sends to complete.
2150          * NB we should be safe from RDMA now that the QP has changed state */
2151
2152         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2153         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2154         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2155         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2156         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2157
2158         kiblnd_handle_early_rxs(conn);
2159 }
2160
2161 static void
2162 kiblnd_peer_connect_failed(struct kib_peer_ni *peer_ni, int active,
2163                            int error)
2164 {
2165         LIST_HEAD(zombies);
2166         unsigned long   flags;
2167
2168         LASSERT (error != 0);
2169         LASSERT (!in_interrupt());
2170
2171         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2172
2173         if (active) {
2174                 LASSERT(peer_ni->ibp_connecting > 0);
2175                 peer_ni->ibp_connecting--;
2176         } else {
2177                 LASSERT (peer_ni->ibp_accepting > 0);
2178                 peer_ni->ibp_accepting--;
2179         }
2180
2181         if (kiblnd_peer_connecting(peer_ni)) {
2182                 /* another connection attempt under way... */
2183                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2184                                         flags);
2185                 return;
2186         }
2187
2188         peer_ni->ibp_reconnected = 0;
2189         if (list_empty(&peer_ni->ibp_conns)) {
2190                 /* Take peer_ni's blocked transmits to complete with error */
2191                 list_splice_init(&peer_ni->ibp_tx_queue, &zombies);
2192
2193                 if (kiblnd_peer_active(peer_ni))
2194                         kiblnd_unlink_peer_locked(peer_ni);
2195
2196                 peer_ni->ibp_error = error;
2197         } else {
2198                 /* Can't have blocked transmits if there are connections */
2199                 LASSERT(list_empty(&peer_ni->ibp_tx_queue));
2200         }
2201
2202         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2203
2204         kiblnd_peer_notify(peer_ni);
2205
2206         if (list_empty(&zombies))
2207                 return;
2208
2209         CNETERR("Deleting messages for %s: connection failed\n",
2210                 libcfs_nid2str(peer_ni->ibp_nid));
2211
2212         kiblnd_txlist_done(&zombies, error,
2213                            LNET_MSG_STATUS_LOCAL_DROPPED);
2214 }
2215
2216 static void
2217 kiblnd_connreq_done(struct kib_conn *conn, int status)
2218 {
2219         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2220         struct kib_tx *tx;
2221         LIST_HEAD(txs);
2222         unsigned long    flags;
2223         int              active;
2224
2225         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2226
2227         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2228                libcfs_nid2str(peer_ni->ibp_nid), active,
2229                conn->ibc_version, status);
2230
2231         LASSERT (!in_interrupt());
2232         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2233                   peer_ni->ibp_connecting > 0) ||
2234                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2235                   peer_ni->ibp_accepting > 0));
2236
2237         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2238         conn->ibc_connvars = NULL;
2239
2240         if (status != 0) {
2241                 /* failed to establish connection */
2242                 kiblnd_peer_connect_failed(peer_ni, active, status);
2243                 kiblnd_finalise_conn(conn);
2244                 return;
2245         }
2246
2247         /* connection established */
2248         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2249
2250         conn->ibc_last_send = ktime_get();
2251         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2252         kiblnd_peer_alive(peer_ni);
2253
2254         /* Add conn to peer_ni's list and nuke any dangling conns from a different
2255          * peer_ni instance... */
2256         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2257         list_add(&conn->ibc_list, &peer_ni->ibp_conns);
2258         peer_ni->ibp_reconnected = 0;
2259         if (active)
2260                 peer_ni->ibp_connecting--;
2261         else
2262                 peer_ni->ibp_accepting--;
2263
2264         if (peer_ni->ibp_version == 0) {
2265                 peer_ni->ibp_version     = conn->ibc_version;
2266                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2267         }
2268
2269         if (peer_ni->ibp_version     != conn->ibc_version ||
2270             peer_ni->ibp_incarnation != conn->ibc_incarnation) {
2271                 kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
2272                                                 conn->ibc_incarnation);
2273                 peer_ni->ibp_version     = conn->ibc_version;
2274                 peer_ni->ibp_incarnation = conn->ibc_incarnation;
2275         }
2276
2277         /* grab pending txs while I have the lock */
2278         list_splice_init(&peer_ni->ibp_tx_queue, &txs);
2279
2280         if (!kiblnd_peer_active(peer_ni) ||        /* peer_ni has been deleted */
2281             conn->ibc_comms_error != 0) {       /* error has happened already */
2282
2283                 /* start to shut down connection */
2284                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2285                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2286
2287                 kiblnd_txlist_done(&txs, -ECONNABORTED,
2288                                    LNET_MSG_STATUS_LOCAL_ERROR);
2289
2290                 return;
2291         }
2292
2293         /* +1 ref for myself, this connection is visible to other threads
2294          * now, refcount of peer:ibp_conns can be released by connection
2295          * close from either a different thread, or the calling of
2296          * kiblnd_check_sends_locked() below. See bz21911 for details.
2297          */
2298         kiblnd_conn_addref(conn);
2299         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2300
2301         /* Schedule blocked txs
2302          * Note: if we are running with conns_per_peer > 1, these blocked
2303          * txs will all get scheduled to the first connection which gets
2304          * scheduled.  We won't be using round robin on this first batch.
2305          */
2306         spin_lock(&conn->ibc_lock);
2307         while (!list_empty(&txs)) {
2308                 tx = list_entry(txs.next, struct kib_tx, tx_list);
2309                 list_del(&tx->tx_list);
2310
2311                 kiblnd_queue_tx_locked(tx, conn);
2312         }
2313         kiblnd_check_sends_locked(conn);
2314         spin_unlock(&conn->ibc_lock);
2315
2316         /* schedule blocked rxs */
2317         kiblnd_handle_early_rxs(conn);
2318         kiblnd_conn_decref(conn);
2319 }
2320
2321 static void
2322 kiblnd_reject(struct rdma_cm_id *cmid, struct kib_rej *rej)
2323 {
2324         int          rc;
2325
2326 #ifdef HAVE_RDMA_REJECT_4ARGS
2327         rc = rdma_reject(cmid, rej, sizeof(*rej), IB_CM_REJ_CONSUMER_DEFINED);
2328 #else
2329         rc = rdma_reject(cmid, rej, sizeof(*rej));
2330 #endif
2331
2332         if (rc != 0)
2333                 CWARN("Error %d sending reject\n", rc);
2334 }
2335
2336 static int
2337 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2338 {
2339         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2340         struct kib_msg *reqmsg = priv;
2341         struct kib_msg *ackmsg;
2342         struct kib_dev *ibdev;
2343         struct kib_peer_ni *peer_ni;
2344         struct kib_peer_ni *peer2;
2345         struct kib_conn *conn;
2346         struct lnet_ni *ni = NULL;
2347         struct kib_net *net = NULL;
2348         lnet_nid_t             nid;
2349         struct rdma_conn_param cp;
2350         struct kib_rej rej;
2351         int                    version = IBLND_MSG_VERSION;
2352         unsigned long          flags;
2353         int                    rc;
2354         struct sockaddr_in    *peer_addr;
2355         LASSERT (!in_interrupt());
2356
2357         /* cmid inherits 'context' from the corresponding listener id */
2358         ibdev = cmid->context;
2359         LASSERT(ibdev);
2360
2361         memset(&rej, 0, sizeof(rej));
2362         rej.ibr_magic                = IBLND_MSG_MAGIC;
2363         rej.ibr_why                  = IBLND_REJECT_FATAL;
2364         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2365
2366         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2367         if (*kiblnd_tunables.kib_require_priv_port &&
2368             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2369                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2370                 CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
2371                        &ip, ntohs(peer_addr->sin_port));
2372                 goto failed;
2373         }
2374
2375         if (priv_nob < offsetof(struct kib_msg, ibm_type)) {
2376                 CERROR("Short connection request\n");
2377                 goto failed;
2378         }
2379
2380         /* Future protocol version compatibility support!  If the
2381          * o2iblnd-specific protocol changes, or when LNET unifies
2382          * protocols over all LNDs, the initial connection will
2383          * negotiate a protocol version.  I trap this here to avoid
2384          * console errors; the reject tells the peer_ni which protocol I
2385          * speak. */
2386         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2387             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2388                 goto failed;
2389         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2390             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2391             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2392                 goto failed;
2393         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2394             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2395             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2396                 goto failed;
2397
2398         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2399         if (rc != 0) {
2400                 CERROR("Can't parse connection request: %d\n", rc);
2401                 goto failed;
2402         }
2403
2404         nid = reqmsg->ibm_srcnid;
2405         ni  = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
2406
2407         if (ni != NULL) {
2408                 net = (struct kib_net *)ni->ni_data;
2409                 rej.ibr_incarnation = net->ibn_incarnation;
2410         }
2411
2412         if (ni == NULL ||                         /* no matching net */
2413             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2414             net->ibn_dev != ibdev) {              /* wrong device */
2415                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2416                        "bad dst nid %s\n", libcfs_nid2str(nid),
2417                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2418                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2419                         &ibdev->ibd_ifip,
2420                        libcfs_nid2str(reqmsg->ibm_dstnid));
2421
2422                 goto failed;
2423         }
2424
2425        /* check time stamp as soon as possible */
2426         if (reqmsg->ibm_dststamp != 0 &&
2427             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2428                 CWARN("Stale connection request\n");
2429                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2430                 goto failed;
2431         }
2432
2433         /* I can accept peer_ni's version */
2434         version = reqmsg->ibm_version;
2435
2436         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2437                 CERROR("Unexpected connreq msg type: %x from %s\n",
2438                        reqmsg->ibm_type, libcfs_nid2str(nid));
2439                 goto failed;
2440         }
2441
2442         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2443             kiblnd_msg_queue_size(version, ni)) {
2444                 CERROR("Can't accept conn from %s, queue depth too large: "
2445                        " %d (<=%d wanted)\n",
2446                        libcfs_nid2str(nid),
2447                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2448                        kiblnd_msg_queue_size(version, ni));
2449
2450                 if (version == IBLND_MSG_VERSION)
2451                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2452
2453                 goto failed;
2454         }
2455
2456         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2457             IBLND_MAX_RDMA_FRAGS) {
2458                 CWARN("Can't accept conn from %s (version %x): "
2459                       "max_frags %d too large (%d wanted)\n",
2460                       libcfs_nid2str(nid), version,
2461                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2462                       IBLND_MAX_RDMA_FRAGS);
2463
2464                 if (version >= IBLND_MSG_VERSION)
2465                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2466
2467                 goto failed;
2468         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2469                    IBLND_MAX_RDMA_FRAGS &&
2470                    net->ibn_fmr_ps == NULL) {
2471                 CWARN("Can't accept conn from %s (version %x): "
2472                       "max_frags %d incompatible without FMR pool "
2473                       "(%d wanted)\n",
2474                       libcfs_nid2str(nid), version,
2475                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2476                       IBLND_MAX_RDMA_FRAGS);
2477
2478                 if (version == IBLND_MSG_VERSION)
2479                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2480
2481                 goto failed;
2482         }
2483
2484         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2485                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2486                        libcfs_nid2str(nid),
2487                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2488                        IBLND_MSG_SIZE);
2489                 goto failed;
2490         }
2491
2492         /* assume 'nid' is a new peer_ni; create  */
2493         rc = kiblnd_create_peer(ni, &peer_ni, nid);
2494         if (rc != 0) {
2495                 CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
2496                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2497                 goto failed;
2498         }
2499
2500         /* We have validated the peer's parameters so use those */
2501         peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2502         peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2503
2504         write_lock_irqsave(g_lock, flags);
2505
2506         peer2 = kiblnd_find_peer_locked(ni, nid);
2507         if (peer2 != NULL) {
2508                 if (peer2->ibp_version == 0) {
2509                         peer2->ibp_version     = version;
2510                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2511                 }
2512
2513                 /* not the guy I've talked with */
2514                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2515                     peer2->ibp_version     != version) {
2516                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2517
2518                         if (kiblnd_peer_active(peer2)) {
2519                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2520                                 peer2->ibp_version = version;
2521                         }
2522                         write_unlock_irqrestore(g_lock, flags);
2523
2524                         CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2525                               libcfs_nid2str(nid), peer2->ibp_version, version,
2526                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2527
2528                         kiblnd_peer_decref(peer_ni);
2529                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2530                         goto failed;
2531                 }
2532
2533                 /* Tie-break connection race in favour of the higher NID.
2534                  * If we keep running into a race condition multiple times,
2535                  * we have to assume that the connection attempt with the
2536                  * higher NID is stuck in a connecting state and will never
2537                  * recover.  As such, we pass through this if-block and let
2538                  * the lower NID connection win so we can move forward.
2539                  */
2540                 if (peer2->ibp_connecting != 0 &&
2541                     nid < ni->ni_nid && peer2->ibp_races <
2542                     MAX_CONN_RACES_BEFORE_ABORT) {
2543                         peer2->ibp_races++;
2544                         write_unlock_irqrestore(g_lock, flags);
2545
2546                         CDEBUG(D_NET, "Conn race %s\n",
2547                                libcfs_nid2str(peer2->ibp_nid));
2548
2549                         kiblnd_peer_decref(peer_ni);
2550                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2551                         goto failed;
2552                 }
2553                 if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
2554                         CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
2555                                 libcfs_nid2str(peer2->ibp_nid),
2556                                 MAX_CONN_RACES_BEFORE_ABORT);
2557                 /*
2558                  * passive connection is allowed even this peer_ni is waiting for
2559                  * reconnection.
2560                  */
2561                 peer2->ibp_reconnecting = 0;
2562                 peer2->ibp_races = 0;
2563                 peer2->ibp_accepting++;
2564                 kiblnd_peer_addref(peer2);
2565
2566                 /* Race with kiblnd_launch_tx (active connect) to create peer_ni
2567                  * so copy validated parameters since we now know what the
2568                  * peer_ni's limits are */
2569                 peer2->ibp_max_frags = peer_ni->ibp_max_frags;
2570                 peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
2571
2572                 write_unlock_irqrestore(g_lock, flags);
2573                 kiblnd_peer_decref(peer_ni);
2574                 peer_ni = peer2;
2575         } else {
2576                 /* Brand new peer_ni */
2577                 LASSERT (peer_ni->ibp_accepting == 0);
2578                 LASSERT (peer_ni->ibp_version == 0 &&
2579                          peer_ni->ibp_incarnation == 0);
2580
2581                 peer_ni->ibp_accepting   = 1;
2582                 peer_ni->ibp_version     = version;
2583                 peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
2584
2585                 /* I have a ref on ni that prevents it being shutdown */
2586                 LASSERT (net->ibn_shutdown == 0);
2587
2588                 kiblnd_peer_addref(peer_ni);
2589                 list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
2590
2591                 write_unlock_irqrestore(g_lock, flags);
2592         }
2593
2594         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2595         if (conn == NULL) {
2596                 kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
2597                 kiblnd_peer_decref(peer_ni);
2598                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2599                 goto failed;
2600         }
2601
2602         /* conn now "owns" cmid, so I return success from here on to ensure the
2603          * CM callback doesn't destroy cmid. */
2604         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2605         conn->ibc_credits          = conn->ibc_queue_depth;
2606         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2607         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2608                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2609
2610         ackmsg = &conn->ibc_connvars->cv_msg;
2611         memset(ackmsg, 0, sizeof(*ackmsg));
2612
2613         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2614                         sizeof(ackmsg->ibm_u.connparams));
2615         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2616         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2617         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2618
2619         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2620
2621         memset(&cp, 0, sizeof(cp));
2622         cp.private_data        = ackmsg;
2623         cp.private_data_len    = ackmsg->ibm_nob;
2624         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2625         cp.initiator_depth     = 0;
2626         cp.flow_control        = 1;
2627         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2628         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2629
2630         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2631
2632         rc = rdma_accept(cmid, &cp);
2633         if (rc != 0) {
2634                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2635                 rej.ibr_version = version;
2636                 rej.ibr_why     = IBLND_REJECT_FATAL;
2637
2638                 kiblnd_reject(cmid, &rej);
2639                 kiblnd_connreq_done(conn, rc);
2640                 kiblnd_conn_decref(conn);
2641         }
2642
2643         lnet_ni_decref(ni);
2644         return 0;
2645
2646  failed:
2647         if (ni != NULL) {
2648                 rej.ibr_cp.ibcp_queue_depth =
2649                         kiblnd_msg_queue_size(version, ni);
2650                 rej.ibr_cp.ibcp_max_frags   = IBLND_MAX_RDMA_FRAGS;
2651                 lnet_ni_decref(ni);
2652         }
2653
2654         rej.ibr_version = version;
2655         kiblnd_reject(cmid, &rej);
2656
2657         return -ECONNREFUSED;
2658 }
2659
2660 static void
2661 kiblnd_check_reconnect(struct kib_conn *conn, int version,
2662                        u64 incarnation, int why, struct kib_connparams *cp)
2663 {
2664         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
2665         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2666         char            *reason;
2667         int              msg_size = IBLND_MSG_SIZE;
2668         int              frag_num = -1;
2669         int              queue_dep = -1;
2670         bool             reconnect;
2671         unsigned long    flags;
2672
2673         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2674         LASSERT(peer_ni->ibp_connecting > 0);   /* 'conn' at least */
2675
2676         if (cp) {
2677                 msg_size        = cp->ibcp_max_msg_size;
2678                 frag_num        = cp->ibcp_max_frags;
2679                 queue_dep       = cp->ibcp_queue_depth;
2680         }
2681
2682         write_lock_irqsave(glock, flags);
2683         /* retry connection if it's still needed and no other connection
2684          * attempts (active or passive) are in progress
2685          * NB: reconnect is still needed even when ibp_tx_queue is
2686          * empty if ibp_version != version because reconnect may be
2687          * initiated.
2688          */
2689         reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
2690                      peer_ni->ibp_version != version) &&
2691                     peer_ni->ibp_connecting &&
2692                     peer_ni->ibp_accepting == 0;
2693         if (!reconnect) {
2694                 reason = "no need";
2695                 goto out;
2696         }
2697
2698         switch (why) {
2699         default:
2700                 reason = "Unknown";
2701                 break;
2702
2703         case IBLND_REJECT_RDMA_FRAGS: {
2704                 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
2705
2706                 if (!cp) {
2707                         reason = "can't negotiate max frags";
2708                         goto out;
2709                 }
2710                 tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
2711 #ifdef HAVE_IB_GET_DMA_MR
2712                 /*
2713                  * This check only makes sense if the kernel supports global
2714                  * memory registration. Otherwise, map_on_demand will never == 0
2715                  */
2716                 if (!tunables->lnd_map_on_demand) {
2717                         reason = "map_on_demand must be enabled";
2718                         goto out;
2719                 }
2720 #endif
2721                 if (conn->ibc_max_frags <= frag_num) {
2722                         reason = "unsupported max frags";
2723                         goto out;
2724                 }
2725
2726                 peer_ni->ibp_max_frags = frag_num;
2727                 reason = "rdma fragments";
2728                 break;
2729         }
2730         case IBLND_REJECT_MSG_QUEUE_SIZE:
2731                 if (!cp) {
2732                         reason = "can't negotiate queue depth";
2733                         goto out;
2734                 }
2735                 if (conn->ibc_queue_depth <= queue_dep) {
2736                         reason = "unsupported queue depth";
2737                         goto out;
2738                 }
2739
2740                 peer_ni->ibp_queue_depth = queue_dep;
2741                 reason = "queue depth";
2742                 break;
2743
2744         case IBLND_REJECT_CONN_STALE:
2745                 reason = "stale";
2746                 break;
2747
2748         case IBLND_REJECT_CONN_RACE:
2749                 reason = "conn race";
2750                 break;
2751
2752         case IBLND_REJECT_CONN_UNCOMPAT:
2753                 reason = "version negotiation";
2754                 break;
2755
2756         case IBLND_REJECT_INVALID_SRV_ID:
2757                 reason = "invalid service id";
2758                 break;
2759         }
2760
2761         conn->ibc_reconnect = 1;
2762         peer_ni->ibp_reconnecting++;
2763         peer_ni->ibp_version = version;
2764         if (incarnation != 0)
2765                 peer_ni->ibp_incarnation = incarnation;
2766  out:
2767         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2768
2769         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2770                 libcfs_nid2str(peer_ni->ibp_nid),
2771                 reconnect ? "reconnect" : "don't reconnect",
2772                 reason, IBLND_MSG_VERSION, version, msg_size,
2773                 conn->ibc_queue_depth, queue_dep,
2774                 conn->ibc_max_frags, frag_num);
2775         /*
2776          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
2777          * while destroying the zombie
2778          */
2779 }
2780
2781 static void
2782 kiblnd_rejected(struct kib_conn *conn, int reason, void *priv, int priv_nob)
2783 {
2784         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2785
2786         LASSERT (!in_interrupt());
2787         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2788
2789         switch (reason) {
2790         case IB_CM_REJ_STALE_CONN:
2791                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2792                                        IBLND_REJECT_CONN_STALE, NULL);
2793                 break;
2794
2795         case IB_CM_REJ_INVALID_SERVICE_ID:
2796                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2797                                        IBLND_REJECT_INVALID_SRV_ID, NULL);
2798                 CNETERR("%s rejected: no listener at %d\n",
2799                         libcfs_nid2str(peer_ni->ibp_nid),
2800                         *kiblnd_tunables.kib_service);
2801                 break;
2802
2803         case IB_CM_REJ_CONSUMER_DEFINED:
2804                 if (priv_nob >= offsetof(struct kib_rej, ibr_padding)) {
2805                         struct kib_rej *rej = priv;
2806                         struct kib_connparams *cp = NULL;
2807                         int               flip        = 0;
2808                         __u64             incarnation = -1;
2809
2810                         /* NB. default incarnation is -1 because:
2811                          * a) V1 will ignore dst incarnation in connreq.
2812                          * b) V2 will provide incarnation while rejecting me,
2813                          *    -1 will be overwrote.
2814                          *
2815                          * if I try to connect to a V1 peer_ni with V2 protocol,
2816                          * it rejected me then upgrade to V2, I have no idea
2817                          * about the upgrading and try to reconnect with V1,
2818                          * in this case upgraded V2 can find out I'm trying to
2819                          * talk to the old guy and reject me(incarnation is -1).
2820                          */
2821
2822                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2823                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2824                                 __swab32s(&rej->ibr_magic);
2825                                 __swab16s(&rej->ibr_version);
2826                                 flip = 1;
2827                         }
2828
2829                         if (priv_nob >= sizeof(struct kib_rej) &&
2830                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2831                                 /* priv_nob is always 148 in current version
2832                                  * of OFED, so we still need to check version.
2833                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2834                                 cp = &rej->ibr_cp;
2835
2836                                 if (flip) {
2837                                         __swab64s(&rej->ibr_incarnation);
2838                                         __swab16s(&cp->ibcp_queue_depth);
2839                                         __swab16s(&cp->ibcp_max_frags);
2840                                         __swab32s(&cp->ibcp_max_msg_size);
2841                                 }
2842
2843                                 incarnation = rej->ibr_incarnation;
2844                         }
2845
2846                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2847                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2848                                 CERROR("%s rejected: consumer defined fatal error\n",
2849                                        libcfs_nid2str(peer_ni->ibp_nid));
2850                                 break;
2851                         }
2852
2853                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2854                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2855                                 CERROR("%s rejected: o2iblnd version %x error\n",
2856                                        libcfs_nid2str(peer_ni->ibp_nid),
2857                                        rej->ibr_version);
2858                                 break;
2859                         }
2860
2861                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2862                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2863                                 CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
2864                                        libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
2865
2866                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2867                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2868                         }
2869
2870                         switch (rej->ibr_why) {
2871                         case IBLND_REJECT_CONN_RACE:
2872                         case IBLND_REJECT_CONN_STALE:
2873                         case IBLND_REJECT_CONN_UNCOMPAT:
2874                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2875                         case IBLND_REJECT_RDMA_FRAGS:
2876                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2877                                                 incarnation, rej->ibr_why, cp);
2878                                 break;
2879
2880                         case IBLND_REJECT_NO_RESOURCES:
2881                                 CERROR("%s rejected: o2iblnd no resources\n",
2882                                        libcfs_nid2str(peer_ni->ibp_nid));
2883                                 break;
2884
2885                         case IBLND_REJECT_FATAL:
2886                                 CERROR("%s rejected: o2iblnd fatal error\n",
2887                                        libcfs_nid2str(peer_ni->ibp_nid));
2888                                 break;
2889
2890                         default:
2891                                 CERROR("%s rejected: o2iblnd reason %d\n",
2892                                        libcfs_nid2str(peer_ni->ibp_nid),
2893                                        rej->ibr_why);
2894                                 break;
2895                         }
2896                         break;
2897                 }
2898                 /* fall through */
2899         default:
2900                 CNETERR("%s rejected: reason %d, size %d\n",
2901                         libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
2902                 break;
2903         }
2904
2905         kiblnd_connreq_done(conn, -ECONNREFUSED);
2906 }
2907
2908 static void
2909 kiblnd_check_connreply(struct kib_conn *conn, void *priv, int priv_nob)
2910 {
2911         struct kib_peer_ni *peer_ni = conn->ibc_peer;
2912         struct lnet_ni *ni = peer_ni->ibp_ni;
2913         struct kib_net *net = ni->ni_data;
2914         struct kib_msg *msg = priv;
2915         int            ver  = conn->ibc_version;
2916         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2917         unsigned long  flags;
2918
2919         LASSERT (net != NULL);
2920
2921         if (rc != 0) {
2922                 CERROR("Can't unpack connack from %s: %d\n",
2923                        libcfs_nid2str(peer_ni->ibp_nid), rc);
2924                 goto failed;
2925         }
2926
2927         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2928                 CERROR("Unexpected message %d from %s\n",
2929                        msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
2930                 rc = -EPROTO;
2931                 goto failed;
2932         }
2933
2934         if (ver != msg->ibm_version) {
2935                 CERROR("%s replied version %x is different with "
2936                        "requested version %x\n",
2937                        libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
2938                 rc = -EPROTO;
2939                 goto failed;
2940         }
2941
2942         if (msg->ibm_u.connparams.ibcp_queue_depth >
2943             conn->ibc_queue_depth) {
2944                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2945                        libcfs_nid2str(peer_ni->ibp_nid),
2946                        msg->ibm_u.connparams.ibcp_queue_depth,
2947                        conn->ibc_queue_depth);
2948                 rc = -EPROTO;
2949                 goto failed;
2950         }
2951
2952         if (msg->ibm_u.connparams.ibcp_max_frags >
2953             conn->ibc_max_frags) {
2954                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2955                        libcfs_nid2str(peer_ni->ibp_nid),
2956                        msg->ibm_u.connparams.ibcp_max_frags,
2957                        conn->ibc_max_frags);
2958                 rc = -EPROTO;
2959                 goto failed;
2960         }
2961
2962         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2963                 CERROR("%s max message size %d too big (%d max)\n",
2964                        libcfs_nid2str(peer_ni->ibp_nid),
2965                        msg->ibm_u.connparams.ibcp_max_msg_size,
2966                        IBLND_MSG_SIZE);
2967                 rc = -EPROTO;
2968                 goto failed;
2969         }
2970
2971         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2972         if (msg->ibm_dstnid == ni->ni_nid &&
2973             msg->ibm_dststamp == net->ibn_incarnation)
2974                 rc = 0;
2975         else
2976                 rc = -ESTALE;
2977         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2978
2979         if (rc != 0) {
2980                 CERROR("Bad connection reply from %s, rc = %d, "
2981                        "version: %x max_frags: %d\n",
2982                        libcfs_nid2str(peer_ni->ibp_nid), rc,
2983                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2984                 goto failed;
2985         }
2986
2987         conn->ibc_incarnation      = msg->ibm_srcstamp;
2988         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
2989         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2990         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
2991         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
2992         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2993                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2994
2995         kiblnd_connreq_done(conn, 0);
2996         return;
2997
2998  failed:
2999         /* NB My QP has already established itself, so I handle anything going
3000          * wrong here by setting ibc_comms_error.
3001          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
3002          * immediately tears it down. */
3003
3004         LASSERT (rc != 0);
3005         conn->ibc_comms_error = rc;
3006         kiblnd_connreq_done(conn, 0);
3007 }
3008
3009 static int
3010 kiblnd_active_connect(struct rdma_cm_id *cmid)
3011 {
3012         struct kib_peer_ni *peer_ni = cmid->context;
3013         struct kib_conn *conn;
3014         struct kib_msg *msg;
3015         struct rdma_conn_param cp;
3016         int                      version;
3017         __u64                    incarnation;
3018         unsigned long            flags;
3019         int                      rc;
3020
3021         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3022
3023         incarnation = peer_ni->ibp_incarnation;
3024         version     = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
3025                                                  peer_ni->ibp_version;
3026
3027         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3028
3029         conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
3030                                   version);
3031         if (conn == NULL) {
3032                 kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
3033                 kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
3034                 return -ENOMEM;
3035         }
3036
3037         /* conn "owns" cmid now, so I return success from here on to ensure the
3038          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
3039          * on peer_ni */
3040
3041         msg = &conn->ibc_connvars->cv_msg;
3042
3043         memset(msg, 0, sizeof(*msg));
3044         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
3045         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
3046         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
3047         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
3048
3049         kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
3050                         0, peer_ni->ibp_nid, incarnation);
3051
3052         memset(&cp, 0, sizeof(cp));
3053         cp.private_data        = msg;
3054         cp.private_data_len    = msg->ibm_nob;
3055         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
3056         cp.initiator_depth     = 0;
3057         cp.flow_control        = 1;
3058         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
3059         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
3060
3061         LASSERT(cmid->context == (void *)conn);
3062         LASSERT(conn->ibc_cmid == cmid);
3063
3064         rc = rdma_connect(cmid, &cp);
3065         if (rc != 0) {
3066                 CERROR("Can't connect to %s: %d\n",
3067                        libcfs_nid2str(peer_ni->ibp_nid), rc);
3068                 kiblnd_connreq_done(conn, rc);
3069                 kiblnd_conn_decref(conn);
3070         }
3071
3072         return 0;
3073 }
3074
3075 int
3076 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
3077 {
3078         struct kib_peer_ni *peer_ni;
3079         struct kib_conn *conn;
3080         int rc;
3081
3082         switch (event->event) {
3083         default:
3084                 CERROR("Unexpected event: %d, status: %d\n",
3085                        event->event, event->status);
3086                 LBUG();
3087
3088         case RDMA_CM_EVENT_CONNECT_REQUEST:
3089                 /* destroy cmid on failure */
3090                 rc = kiblnd_passive_connect(cmid,
3091                                             (void *)KIBLND_CONN_PARAM(event),
3092                                             KIBLND_CONN_PARAM_LEN(event));
3093                 CDEBUG(D_NET, "connreq: %d\n", rc);
3094                 return rc;
3095
3096         case RDMA_CM_EVENT_ADDR_ERROR:
3097                 peer_ni = cmid->context;
3098                 CNETERR("%s: ADDR ERROR %d\n",
3099                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3100                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3101                 kiblnd_peer_decref(peer_ni);
3102                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
3103
3104         case RDMA_CM_EVENT_ADDR_RESOLVED:
3105                 peer_ni = cmid->context;
3106
3107                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
3108                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3109
3110                 if (event->status != 0) {
3111                         CNETERR("Can't resolve address for %s: %d\n",
3112                                 libcfs_nid2str(peer_ni->ibp_nid), event->status);
3113                         rc = event->status;
3114                 } else {
3115                         rc = rdma_resolve_route(
3116                                 cmid, kiblnd_timeout() * 1000);
3117                         if (rc == 0) {
3118                                 struct kib_net *net = peer_ni->ibp_ni->ni_data;
3119                                 struct kib_dev *dev = net->ibn_dev;
3120
3121                                 CDEBUG(D_NET, "%s: connection bound to "\
3122                                        "%s:%pI4h:%s\n",
3123                                        libcfs_nid2str(peer_ni->ibp_nid),
3124                                        dev->ibd_ifname,
3125                                        &dev->ibd_ifip, cmid->device->name);
3126
3127                                 return 0;
3128                         }
3129
3130                         /* Can't initiate route resolution */
3131                         CERROR("Can't resolve route for %s: %d\n",
3132                                libcfs_nid2str(peer_ni->ibp_nid), rc);
3133                 }
3134                 kiblnd_peer_connect_failed(peer_ni, 1, rc);
3135                 kiblnd_peer_decref(peer_ni);
3136                 return rc;                      /* rc != 0 destroys cmid */
3137
3138         case RDMA_CM_EVENT_ROUTE_ERROR:
3139                 peer_ni = cmid->context;
3140                 CNETERR("%s: ROUTE ERROR %d\n",
3141                         libcfs_nid2str(peer_ni->ibp_nid), event->status);
3142                 kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
3143                 kiblnd_peer_decref(peer_ni);
3144                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
3145
3146         case RDMA_CM_EVENT_ROUTE_RESOLVED:
3147                 peer_ni = cmid->context;
3148                 CDEBUG(D_NET,"%s Route resolved: %d\n",
3149                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3150
3151                 if (event->status == 0)
3152                         return kiblnd_active_connect(cmid);
3153
3154                 CNETERR("Can't resolve route for %s: %d\n",
3155                        libcfs_nid2str(peer_ni->ibp_nid), event->status);
3156                 kiblnd_peer_connect_failed(peer_ni, 1, event->status);
3157                 kiblnd_peer_decref(peer_ni);
3158                 return event->status;           /* rc != 0 destroys cmid */
3159
3160         case RDMA_CM_EVENT_UNREACHABLE:
3161                 conn = cmid->context;
3162                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3163                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3164                 CNETERR("%s: UNREACHABLE %d\n",
3165                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3166                 kiblnd_connreq_done(conn, -ENETDOWN);
3167                 kiblnd_conn_decref(conn);
3168                 return 0;
3169
3170         case RDMA_CM_EVENT_CONNECT_ERROR:
3171                 conn = cmid->context;
3172                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3173                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3174                 CNETERR("%s: CONNECT ERROR %d\n",
3175                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3176                 kiblnd_connreq_done(conn, -ENOTCONN);
3177                 kiblnd_conn_decref(conn);
3178                 return 0;
3179
3180         case RDMA_CM_EVENT_REJECTED:
3181                 conn = cmid->context;
3182                 switch (conn->ibc_state) {
3183                 default:
3184                         LBUG();
3185
3186                 case IBLND_CONN_PASSIVE_WAIT:
3187                         CERROR ("%s: REJECTED %d\n",
3188                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3189                                 event->status);
3190                         kiblnd_connreq_done(conn, -ECONNRESET);
3191                         break;
3192
3193                 case IBLND_CONN_ACTIVE_CONNECT:
3194                         kiblnd_rejected(conn, event->status,
3195                                         (void *)KIBLND_CONN_PARAM(event),
3196                                         KIBLND_CONN_PARAM_LEN(event));
3197                         break;
3198                 }
3199                 kiblnd_conn_decref(conn);
3200                 return 0;
3201
3202         case RDMA_CM_EVENT_ESTABLISHED:
3203                 conn = cmid->context;
3204                 switch (conn->ibc_state) {
3205                 default:
3206                         LBUG();
3207
3208                 case IBLND_CONN_PASSIVE_WAIT:
3209                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3210                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3211                         kiblnd_connreq_done(conn, 0);
3212                         break;
3213
3214                 case IBLND_CONN_ACTIVE_CONNECT:
3215                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3216                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3217                         kiblnd_check_connreply(conn,
3218                                                (void *)KIBLND_CONN_PARAM(event),
3219                                                KIBLND_CONN_PARAM_LEN(event));
3220                         break;
3221                 }
3222                 /* net keeps its ref on conn! */
3223                 return 0;
3224
3225         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3226                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3227                 return 0;
3228
3229         case RDMA_CM_EVENT_DISCONNECTED:
3230                 conn = cmid->context;
3231                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3232                         CERROR("%s DISCONNECTED\n",
3233                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3234                         kiblnd_connreq_done(conn, -ECONNRESET);
3235                 } else {
3236                         kiblnd_close_conn(conn, 0);
3237                 }
3238                 kiblnd_conn_decref(conn);
3239                 cmid->context = NULL;
3240                 return 0;
3241
3242         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3243                 LCONSOLE_ERROR_MSG(0x131,
3244                                    "Received notification of device removal\n"
3245                                    "Please shutdown LNET to allow this to proceed\n");
3246                 /* Can't remove network from underneath LNET for now, so I have
3247                  * to ignore this */
3248                 return 0;
3249
3250         case RDMA_CM_EVENT_ADDR_CHANGE:
3251                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3252                 return 0;
3253         }
3254 }
3255
3256 static int
3257 kiblnd_check_txs_locked(struct kib_conn *conn, struct list_head *txs)
3258 {
3259         struct kib_tx *tx;
3260         struct list_head *ttmp;
3261
3262         list_for_each(ttmp, txs) {
3263                 tx = list_entry(ttmp, struct kib_tx, tx_list);
3264
3265                 if (txs != &conn->ibc_active_txs) {
3266                         LASSERT(tx->tx_queued);
3267                 } else {
3268                         LASSERT(!tx->tx_queued);
3269                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3270                 }
3271
3272                 if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3273                         CERROR("Timed out tx: %s, %lld seconds\n",
3274                                kiblnd_queue2str(conn, txs),
3275                                ktime_ms_delta(ktime_get(),
3276                                               tx->tx_deadline) / MSEC_PER_SEC);
3277                         return 1;
3278                 }
3279         }
3280
3281         return 0;
3282 }
3283
3284 static int
3285 kiblnd_conn_timed_out_locked(struct kib_conn *conn)
3286 {
3287         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3288                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3289                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3290                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3291                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3292 }
3293
3294 static void
3295 kiblnd_check_conns (int idx)
3296 {
3297         LIST_HEAD(closes);
3298         LIST_HEAD(checksends);
3299         LIST_HEAD(timedout_txs);
3300         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3301         struct list_head *ptmp;
3302         struct kib_peer_ni *peer_ni;
3303         struct kib_conn *conn;
3304         struct kib_tx *tx, *tx_tmp;
3305         struct list_head *ctmp;
3306         unsigned long     flags;
3307
3308         /* NB. We expect to have a look at all the peers and not find any
3309          * RDMAs to time out, so we just use a shared lock while we
3310          * take a look... */
3311         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3312
3313         list_for_each(ptmp, peers) {
3314                 peer_ni = list_entry(ptmp, struct kib_peer_ni, ibp_list);
3315
3316                 /* Check tx_deadline */
3317                 list_for_each_entry_safe(tx, tx_tmp, &peer_ni->ibp_tx_queue, tx_list) {
3318                         if (ktime_compare(ktime_get(), tx->tx_deadline) >= 0) {
3319                                 CWARN("Timed out tx for %s: %lld seconds\n",
3320                                       libcfs_nid2str(peer_ni->ibp_nid),
3321                                       ktime_ms_delta(ktime_get(),
3322                                                      tx->tx_deadline) / MSEC_PER_SEC);
3323                                 list_move(&tx->tx_list, &timedout_txs);
3324                         }
3325                 }
3326
3327                 list_for_each(ctmp, &peer_ni->ibp_conns) {
3328                         int timedout;
3329                         int sendnoop;
3330
3331                         conn = list_entry(ctmp, struct kib_conn, ibc_list);
3332
3333                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3334
3335                         spin_lock(&conn->ibc_lock);
3336
3337                         sendnoop = kiblnd_need_noop(conn);
3338                         timedout = kiblnd_conn_timed_out_locked(conn);
3339                         if (!sendnoop && !timedout) {
3340                                 spin_unlock(&conn->ibc_lock);
3341                                 continue;
3342                         }
3343
3344                         if (timedout) {
3345                                 CERROR("Timed out RDMA with %s (%lld): "
3346                                        "c: %u, oc: %u, rc: %u\n",
3347                                        libcfs_nid2str(peer_ni->ibp_nid),
3348                                        ktime_get_seconds() - peer_ni->ibp_last_alive,
3349                                        conn->ibc_credits,
3350                                        conn->ibc_outstanding_credits,
3351                                        conn->ibc_reserved_credits);
3352                                 list_add(&conn->ibc_connd_list, &closes);
3353                         } else {
3354                                 list_add(&conn->ibc_connd_list, &checksends);
3355                         }
3356                         /* +ref for 'closes' or 'checksends' */
3357                         kiblnd_conn_addref(conn);
3358
3359                         spin_unlock(&conn->ibc_lock);
3360                 }
3361         }
3362
3363         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3364
3365         if (!list_empty(&timedout_txs))
3366                 kiblnd_txlist_done(&timedout_txs, -ETIMEDOUT,
3367                                    LNET_MSG_STATUS_LOCAL_TIMEOUT);
3368
3369         /* Handle timeout by closing the whole
3370          * connection. We can only be sure RDMA activity
3371          * has ceased once the QP has been modified. */
3372         while (!list_empty(&closes)) {
3373                 conn = list_entry(closes.next,
3374                                   struct kib_conn, ibc_connd_list);
3375                 list_del(&conn->ibc_connd_list);
3376                 kiblnd_close_conn(conn, -ETIMEDOUT);
3377                 kiblnd_conn_decref(conn);
3378         }
3379
3380         /* In case we have enough credits to return via a
3381          * NOOP, but there were no non-blocking tx descs
3382          * free to do it last time... */
3383         while (!list_empty(&checksends)) {
3384                 conn = list_entry(checksends.next,
3385                                   struct kib_conn, ibc_connd_list);
3386                 list_del(&conn->ibc_connd_list);
3387
3388                 spin_lock(&conn->ibc_lock);
3389                 kiblnd_check_sends_locked(conn);
3390                 spin_unlock(&conn->ibc_lock);
3391
3392                 kiblnd_conn_decref(conn);
3393         }
3394 }
3395
3396 static void
3397 kiblnd_disconnect_conn(struct kib_conn *conn)
3398 {
3399         LASSERT (!in_interrupt());
3400         LASSERT (current == kiblnd_data.kib_connd);
3401         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3402
3403         rdma_disconnect(conn->ibc_cmid);
3404         kiblnd_finalise_conn(conn);
3405
3406         kiblnd_peer_notify(conn->ibc_peer);
3407 }
3408
3409 /*
3410  * High-water for reconnection to the same peer_ni, reconnection attempt should
3411  * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3412  */
3413 #define KIB_RECONN_HIGH_RACE    10
3414 /*
3415  * Allow connd to take a break and handle other things after consecutive
3416  * reconnection attemps.
3417  */
3418 #define KIB_RECONN_BREAK        100
3419
3420 int
3421 kiblnd_connd (void *arg)
3422 {
3423         spinlock_t        *lock= &kiblnd_data.kib_connd_lock;
3424         wait_queue_entry_t wait;
3425         unsigned long      flags;
3426         struct kib_conn *conn;
3427         int                timeout;
3428         int                i;
3429         int                dropped_lock;
3430         int                peer_index = 0;
3431         unsigned long      deadline = jiffies;
3432
3433         init_waitqueue_entry(&wait, current);
3434         kiblnd_data.kib_connd = current;
3435
3436         spin_lock_irqsave(lock, flags);
3437
3438         while (!kiblnd_data.kib_shutdown) {
3439                 int reconn = 0;
3440
3441                 dropped_lock = 0;
3442
3443                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3444                         struct kib_peer_ni *peer_ni = NULL;
3445
3446                         conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3447                                           struct kib_conn, ibc_list);
3448                         list_del(&conn->ibc_list);
3449                         if (conn->ibc_reconnect) {
3450                                 peer_ni = conn->ibc_peer;
3451                                 kiblnd_peer_addref(peer_ni);
3452                         }
3453
3454                         spin_unlock_irqrestore(lock, flags);
3455                         dropped_lock = 1;
3456
3457                         kiblnd_destroy_conn(conn);
3458
3459                         spin_lock_irqsave(lock, flags);
3460                         if (!peer_ni) {
3461                                 LIBCFS_FREE(conn, sizeof(*conn));
3462                                 continue;
3463                         }
3464
3465                         conn->ibc_peer = peer_ni;
3466                         if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3467                                 list_add_tail(&conn->ibc_list,
3468                                               &kiblnd_data.kib_reconn_list);
3469                         else
3470                                 list_add_tail(&conn->ibc_list,
3471                                               &kiblnd_data.kib_reconn_wait);
3472                 }
3473
3474                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3475                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3476                                           struct kib_conn, ibc_list);
3477                         list_del(&conn->ibc_list);
3478
3479                         spin_unlock_irqrestore(lock, flags);
3480                         dropped_lock = 1;
3481
3482                         kiblnd_disconnect_conn(conn);
3483                         kiblnd_conn_decref(conn);
3484
3485                         spin_lock_irqsave(lock, flags);
3486                 }
3487
3488                 while (reconn < KIB_RECONN_BREAK) {
3489                         if (kiblnd_data.kib_reconn_sec !=
3490                             ktime_get_real_seconds()) {
3491                                 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3492                                 list_splice_init(&kiblnd_data.kib_reconn_wait,
3493                                                  &kiblnd_data.kib_reconn_list);
3494                         }
3495
3496                         if (list_empty(&kiblnd_data.kib_reconn_list))
3497                                 break;
3498
3499                         conn = list_entry(kiblnd_data.kib_reconn_list.next,
3500                                           struct kib_conn, ibc_list);
3501                         list_del(&conn->ibc_list);
3502
3503                         spin_unlock_irqrestore(lock, flags);
3504                         dropped_lock = 1;
3505
3506                         reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3507                         kiblnd_peer_decref(conn->ibc_peer);
3508                         LIBCFS_FREE(conn, sizeof(*conn));
3509
3510                         spin_lock_irqsave(lock, flags);
3511                 }
3512
3513                 /* careful with the jiffy wrap... */
3514                 timeout = (int)(deadline - jiffies);
3515                 if (timeout <= 0) {
3516                         const int n = 4;
3517                         const int p = 1;
3518                         int       chunk = kiblnd_data.kib_peer_hash_size;
3519                         unsigned int lnd_timeout;
3520
3521                         spin_unlock_irqrestore(lock, flags);
3522                         dropped_lock = 1;
3523
3524                         /* Time to check for RDMA timeouts on a few more
3525                          * peers: I do checks every 'p' seconds on a
3526                          * proportion of the peer_ni table and I need to check
3527                          * every connection 'n' times within a timeout
3528                          * interval, to ensure I detect a timeout on any
3529                          * connection within (n+1)/n times the timeout
3530                          * interval. */
3531
3532                         lnd_timeout = kiblnd_timeout();
3533                         if (lnd_timeout > n * p)
3534                                 chunk = (chunk * n * p) / lnd_timeout;
3535                         if (chunk == 0)
3536                                 chunk = 1;
3537
3538                         for (i = 0; i < chunk; i++) {
3539                                 kiblnd_check_conns(peer_index);
3540                                 peer_index = (peer_index + 1) %
3541                                              kiblnd_data.kib_peer_hash_size;
3542                         }
3543
3544                         deadline += cfs_time_seconds(p);
3545                         spin_lock_irqsave(lock, flags);
3546                 }
3547
3548                 if (dropped_lock)
3549                         continue;
3550
3551                 /* Nothing to do for 'timeout'  */
3552                 set_current_state(TASK_INTERRUPTIBLE);
3553                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3554                 spin_unlock_irqrestore(lock, flags);
3555
3556                 schedule_timeout(timeout);
3557
3558                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3559                 spin_lock_irqsave(lock, flags);
3560         }
3561
3562         spin_unlock_irqrestore(lock, flags);
3563
3564         kiblnd_thread_fini();
3565         return 0;
3566 }
3567
3568 void
3569 kiblnd_qp_event(struct ib_event *event, void *arg)
3570 {
3571         struct kib_conn *conn = arg;
3572
3573         switch (event->event) {
3574         case IB_EVENT_COMM_EST:
3575                 CDEBUG(D_NET, "%s established\n",
3576                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3577                 /* We received a packet but connection isn't established
3578                  * probably handshake packet was lost, so free to
3579                  * force make connection established */
3580                 rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
3581                 return;
3582
3583         case IB_EVENT_PORT_ERR:
3584         case IB_EVENT_DEVICE_FATAL:
3585                 CERROR("Fatal device error for NI %s\n",
3586                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3587                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 1);
3588                 return;
3589
3590         case IB_EVENT_PORT_ACTIVE:
3591                 CERROR("Port reactivated for NI %s\n",
3592                        libcfs_nid2str(conn->ibc_peer->ibp_ni->ni_nid));
3593                 atomic_set(&conn->ibc_peer->ibp_ni->ni_fatal_error_on, 0);
3594                 return;
3595
3596         default:
3597                 CERROR("%s: Async QP event type %d\n",
3598                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3599                 return;
3600         }
3601 }
3602
3603 static void
3604 kiblnd_complete (struct ib_wc *wc)
3605 {
3606         switch (kiblnd_wreqid2type(wc->wr_id)) {
3607         default:
3608                 LBUG();
3609
3610         case IBLND_WID_MR:
3611                 if (wc->status != IB_WC_SUCCESS &&
3612                     wc->status != IB_WC_WR_FLUSH_ERR)
3613                         CNETERR("FastReg failed: %d\n", wc->status);
3614                 return;
3615
3616         case IBLND_WID_RDMA:
3617                 /* We only get RDMA completion notification if it fails.  All
3618                  * subsequent work items, including the final SEND will fail
3619                  * too.  However we can't print out any more info about the
3620                  * failing RDMA because 'tx' might be back on the idle list or
3621                  * even reused already if we didn't manage to post all our work
3622                  * items */
3623                 CNETERR("RDMA (tx: %p) failed: %d\n",
3624                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3625                 return;
3626
3627         case IBLND_WID_TX:
3628                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3629                 return;
3630
3631         case IBLND_WID_RX:
3632                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3633                                    wc->byte_len);
3634                 return;
3635         }
3636 }
3637
3638 void
3639 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3640 {
3641         /* NB I'm not allowed to schedule this conn once its refcount has
3642          * reached 0.  Since fundamentally I'm racing with scheduler threads
3643          * consuming my CQ I could be called after all completions have
3644          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3645          * and this CQ is about to be destroyed so I NOOP. */
3646         struct kib_conn *conn = arg;
3647         struct kib_sched_info *sched = conn->ibc_sched;
3648         unsigned long flags;
3649
3650         LASSERT(cq == conn->ibc_cq);
3651
3652         spin_lock_irqsave(&sched->ibs_lock, flags);
3653
3654         conn->ibc_ready = 1;
3655
3656         if (!conn->ibc_scheduled &&
3657             (conn->ibc_nrx > 0 ||
3658              conn->ibc_nsends_posted > 0)) {
3659                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3660                 conn->ibc_scheduled = 1;
3661                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3662
3663                 if (waitqueue_active(&sched->ibs_waitq))
3664                         wake_up(&sched->ibs_waitq);
3665         }
3666
3667         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3668 }
3669
3670 void
3671 kiblnd_cq_event(struct ib_event *event, void *arg)
3672 {
3673         struct kib_conn *conn = arg;
3674
3675         CERROR("%s: async CQ event type %d\n",
3676                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3677 }
3678
3679 int
3680 kiblnd_scheduler(void *arg)
3681 {
3682         long                    id = (long)arg;
3683         struct kib_sched_info   *sched;
3684         struct kib_conn *conn;
3685         wait_queue_entry_t      wait;
3686         unsigned long           flags;
3687         struct ib_wc            wc;
3688         int                     did_something;
3689         int                     rc;
3690
3691         init_waitqueue_entry(&wait, current);
3692
3693         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3694
3695         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3696         if (rc != 0) {
3697                 CWARN("Unable to bind on CPU partition %d, please verify "
3698                       "whether all CPUs are healthy and reload modules if "
3699                       "necessary, otherwise your system might under risk of "
3700                       "low performance\n", sched->ibs_cpt);
3701         }
3702
3703         spin_lock_irqsave(&sched->ibs_lock, flags);
3704
3705         while (!kiblnd_data.kib_shutdown) {
3706                 if (need_resched()) {
3707                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3708
3709                         cond_resched();
3710
3711                         spin_lock_irqsave(&sched->ibs_lock, flags);
3712                 }
3713
3714                 did_something = 0;
3715
3716                 if (!list_empty(&sched->ibs_conns)) {
3717                         conn = list_entry(sched->ibs_conns.next,
3718                                           struct kib_conn, ibc_sched_list);
3719                         /* take over kib_sched_conns' ref on conn... */
3720                         LASSERT(conn->ibc_scheduled);
3721                         list_del(&conn->ibc_sched_list);
3722                         conn->ibc_ready = 0;
3723
3724                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3725
3726                         wc.wr_id = IBLND_WID_INVAL;
3727
3728                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3729                         if (rc == 0) {
3730                                 rc = ib_req_notify_cq(conn->ibc_cq,
3731                                                       IB_CQ_NEXT_COMP);
3732                                 if (rc < 0) {
3733                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3734                                               "closing connection\n",
3735                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3736                                         kiblnd_close_conn(conn, -EIO);
3737                                         kiblnd_conn_decref(conn);
3738                                         spin_lock_irqsave(&sched->ibs_lock,
3739                                                               flags);
3740                                         continue;
3741                                 }
3742
3743                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3744                         }
3745
3746                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3747                                 LCONSOLE_ERROR(
3748                                         "ib_poll_cq (rc: %d) returned invalid "
3749                                         "wr_id, opcode %d, status: %d, "
3750                                         "vendor_err: %d, conn: %s status: %d\n"
3751                                         "please upgrade firmware and OFED or "
3752                                         "contact vendor.\n", rc,
3753                                         wc.opcode, wc.status, wc.vendor_err,
3754                                         libcfs_nid2str(conn->ibc_peer->ibp_nid),
3755                                         conn->ibc_state);
3756                                 rc = -EINVAL;
3757                         }
3758
3759                         if (rc < 0) {
3760                                 CWARN("%s: ib_poll_cq failed: %d, "
3761                                       "closing connection\n",
3762                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3763                                       rc);
3764                                 kiblnd_close_conn(conn, -EIO);
3765                                 kiblnd_conn_decref(conn);
3766                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3767                                 continue;
3768                         }
3769
3770                         spin_lock_irqsave(&sched->ibs_lock, flags);
3771
3772                         if (rc != 0 || conn->ibc_ready) {
3773                                 /* There may be another completion waiting; get
3774                                  * another scheduler to check while I handle
3775                                  * this one... */
3776                                 /* +1 ref for sched_conns */
3777                                 kiblnd_conn_addref(conn);
3778                                 list_add_tail(&conn->ibc_sched_list,
3779                                                   &sched->ibs_conns);
3780                                 if (waitqueue_active(&sched->ibs_waitq))
3781                                         wake_up(&sched->ibs_waitq);
3782                         } else {
3783                                 conn->ibc_scheduled = 0;
3784                         }
3785
3786                         if (rc != 0) {
3787                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3788                                 kiblnd_complete(&wc);
3789
3790                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3791                         }
3792
3793                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3794                         did_something = 1;
3795                 }
3796
3797                 if (did_something)
3798                         continue;
3799
3800                 set_current_state(TASK_INTERRUPTIBLE);
3801                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3802                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3803
3804                 schedule();
3805
3806                 remove_wait_queue(&sched->ibs_waitq, &wait);
3807                 set_current_state(TASK_RUNNING);
3808                 spin_lock_irqsave(&sched->ibs_lock, flags);
3809         }
3810
3811         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3812
3813         kiblnd_thread_fini();
3814         return 0;
3815 }
3816
3817 int
3818 kiblnd_failover_thread(void *arg)
3819 {
3820         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
3821         struct kib_dev *dev;
3822         struct net *ns = arg;
3823         wait_queue_entry_t wait;
3824         unsigned long    flags;
3825         int              rc;
3826
3827         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3828
3829         init_waitqueue_entry(&wait, current);
3830         write_lock_irqsave(glock, flags);
3831
3832         while (!kiblnd_data.kib_shutdown) {
3833                 int     do_failover = 0;
3834                 int     long_sleep;
3835
3836                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3837                                     ibd_fail_list) {
3838                         if (ktime_get_seconds() < dev->ibd_next_failover)
3839                                 continue;
3840                         do_failover = 1;
3841                         break;
3842                 }
3843
3844                 if (do_failover) {
3845                         list_del_init(&dev->ibd_fail_list);
3846                         dev->ibd_failover = 1;
3847                         write_unlock_irqrestore(glock, flags);
3848
3849                         rc = kiblnd_dev_failover(dev, ns);
3850
3851                         write_lock_irqsave(glock, flags);
3852
3853                         LASSERT (dev->ibd_failover);
3854                         dev->ibd_failover = 0;
3855                         if (rc >= 0) { /* Device is OK or failover succeed */
3856                                 dev->ibd_next_failover = ktime_get_seconds() + 3;
3857                                 continue;
3858                         }
3859
3860                         /* failed to failover, retry later */
3861                         dev->ibd_next_failover = ktime_get_seconds() +
3862                                                  min(dev->ibd_failed_failover, 10);
3863                         if (kiblnd_dev_can_failover(dev)) {
3864                                 list_add_tail(&dev->ibd_fail_list,
3865                                               &kiblnd_data.kib_failed_devs);
3866                         }
3867
3868                         continue;
3869                 }
3870
3871                 /* long sleep if no more pending failover */
3872                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3873
3874                 set_current_state(TASK_INTERRUPTIBLE);
3875                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3876                 write_unlock_irqrestore(glock, flags);
3877
3878                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3879                                                    cfs_time_seconds(1));
3880                 set_current_state(TASK_RUNNING);
3881                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3882                 write_lock_irqsave(glock, flags);
3883
3884                 if (!long_sleep || rc != 0)
3885                         continue;
3886
3887                 /* have a long sleep, routine check all active devices,
3888                  * we need checking like this because if there is not active
3889                  * connection on the dev and no SEND from local, we may listen
3890                  * on wrong HCA for ever while there is a bonding failover */
3891                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3892                         if (kiblnd_dev_can_failover(dev)) {
3893                                 list_add_tail(&dev->ibd_fail_list,
3894                                               &kiblnd_data.kib_failed_devs);
3895                         }
3896                 }
3897         }
3898
3899         write_unlock_irqrestore(glock, flags);
3900
3901         kiblnd_thread_fini();
3902         return 0;
3903 }