lnet/klnds/o2iblnd/o2iblnd_cb.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd_cb.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 static void kiblnd_peer_alive(kib_peer_t *peer);
  44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
  45 static void kiblnd_check_sends(kib_conn_t *conn);
  46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
  47                                int type, int body_nob);
  48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
  49                             int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
  50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
  51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
  52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
  53
  54 static void
  55 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
  56 {
  57         lnet_msg_t *lntmsg[2];
  58         kib_net_t  *net = ni->ni_data;
  59         int         rc;
  60         int         i;
  61
  62         LASSERT (net != NULL);
  63         LASSERT (!in_interrupt());
  64         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
  65         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
  66         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer response */
  67         LASSERT (tx->tx_pool != NULL);
  68
  69         kiblnd_unmap_tx(ni, tx);
  70
  71         /* tx may have up to 2 lnet msgs to finalise */
  72         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  73         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  74         rc = tx->tx_status;
  75
  76         if (tx->tx_conn != NULL) {
  77                 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
  78
  79                 kiblnd_conn_decref(tx->tx_conn);
  80                 tx->tx_conn = NULL;
  81         }
  82
  83         tx->tx_nwrq = 0;
  84         tx->tx_status = 0;
  85
  86         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  87
  88         /* delay finalize until my descs have been freed */
  89         for (i = 0; i < 2; i++) {
  90                 if (lntmsg[i] == NULL)
  91                         continue;
  92
  93                 lnet_finalize(ni, lntmsg[i], rc);
  94         }
  95 }
  96
  97 void
  98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
  99 {
 100         kib_tx_t *tx;
 101
 102         while (!list_empty(txlist)) {
 103                 tx = list_entry(txlist->next, kib_tx_t, tx_list);
 104
 105                 list_del(&tx->tx_list);
 106                 /* complete now */
 107                 tx->tx_waiting = 0;
 108                 tx->tx_status = status;
 109                 kiblnd_tx_done(ni, tx);
 110         }
 111 }
 112
 113 static kib_tx_t *
 114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
 115 {
 116         kib_net_t               *net = (kib_net_t *)ni->ni_data;
 117         struct list_head        *node;
 118         kib_tx_t                *tx;
 119         kib_tx_poolset_t        *tps;
 120
 121         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
 122         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
 123         if (node == NULL)
 124                 return NULL;
 125         tx = container_of(node, kib_tx_t, tx_list);
 126
 127         LASSERT (tx->tx_nwrq == 0);
 128         LASSERT (!tx->tx_queued);
 129         LASSERT (tx->tx_sending == 0);
 130         LASSERT (!tx->tx_waiting);
 131         LASSERT (tx->tx_status == 0);
 132         LASSERT (tx->tx_conn == NULL);
 133         LASSERT (tx->tx_lntmsg[0] == NULL);
 134         LASSERT (tx->tx_lntmsg[1] == NULL);
 135         LASSERT (tx->tx_u.pmr == NULL);
 136         LASSERT (tx->tx_nfrags == 0);
 137
 138         return tx;
 139 }
 140
 141 static void
 142 kiblnd_drop_rx(kib_rx_t *rx)
 143 {
 144         kib_conn_t              *conn   = rx->rx_conn;
 145         struct kib_sched_info   *sched  = conn->ibc_sched;
 146         unsigned long           flags;
 147
 148         spin_lock_irqsave(&sched->ibs_lock, flags);
 149         LASSERT(conn->ibc_nrx > 0);
 150         conn->ibc_nrx--;
 151         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 152
 153         kiblnd_conn_decref(conn);
 154 }
 155
 156 int
 157 kiblnd_post_rx (kib_rx_t *rx, int credit)
 158 {
 159         kib_conn_t         *conn = rx->rx_conn;
 160         kib_net_t          *net = conn->ibc_peer->ibp_ni->ni_data;
 161         struct ib_recv_wr  *bad_wrq = NULL;
 162         struct ib_mr       *mr;
 163         int                 rc;
 164
 165         LASSERT (net != NULL);
 166         LASSERT (!in_interrupt());
 167         LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
 168                  credit == IBLND_POSTRX_PEER_CREDIT ||
 169                  credit == IBLND_POSTRX_RSRVD_CREDIT);
 170
 171         mr = kiblnd_find_dma_mr(conn->ibc_hdev, rx->rx_msgaddr, IBLND_MSG_SIZE);
 172         LASSERT (mr != NULL);
 173
 174         rx->rx_sge.lkey   = mr->lkey;
 175         rx->rx_sge.addr   = rx->rx_msgaddr;
 176         rx->rx_sge.length = IBLND_MSG_SIZE;
 177
 178         rx->rx_wrq.next = NULL;
 179         rx->rx_wrq.sg_list = &rx->rx_sge;
 180         rx->rx_wrq.num_sge = 1;
 181         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 182
 183         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
 184         LASSERT (rx->rx_nob >= 0);              /* not posted */
 185
 186         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 187                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 188                 return 0;
 189         }
 190
 191         rx->rx_nob = -1;                        /* flag posted */
 192
 193         /* NB: need an extra reference after ib_post_recv because we don't
 194          * own this rx (and rx::rx_conn) anymore, LU-5678.
 195          */
 196         kiblnd_conn_addref(conn);
 197         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 198         if (unlikely(rc != 0)) {
 199                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 200                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 201                 rx->rx_nob = 0;
 202         }
 203
 204         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 205                 goto out;
 206
 207         if (unlikely(rc != 0)) {
 208                 kiblnd_close_conn(conn, rc);
 209                 kiblnd_drop_rx(rx);     /* No more posts for this rx */
 210                 goto out;
 211         }
 212
 213         if (credit == IBLND_POSTRX_NO_CREDIT)
 214                 goto out;
 215
 216         spin_lock(&conn->ibc_lock);
 217         if (credit == IBLND_POSTRX_PEER_CREDIT)
 218                 conn->ibc_outstanding_credits++;
 219         else
 220                 conn->ibc_reserved_credits++;
 221         spin_unlock(&conn->ibc_lock);
 222
 223         kiblnd_check_sends(conn);
 224 out:
 225         kiblnd_conn_decref(conn);
 226         return rc;
 227 }
 228
 229 static kib_tx_t *
 230 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
 231 {
 232         struct list_head *tmp;
 233
 234         list_for_each(tmp, &conn->ibc_active_txs) {
 235                 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
 236
 237                 LASSERT(!tx->tx_queued);
 238                 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
 239
 240                 if (tx->tx_cookie != cookie)
 241                         continue;
 242
 243                 if (tx->tx_waiting &&
 244                     tx->tx_msg->ibm_type == txtype)
 245                         return tx;
 246
 247                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 248                       tx->tx_waiting ? "" : "NOT ",
 249                       tx->tx_msg->ibm_type, txtype);
 250         }
 251         return NULL;
 252 }
 253
 254 static void
 255 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
 256 {
 257         kib_tx_t    *tx;
 258         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 259         int          idle;
 260
 261         spin_lock(&conn->ibc_lock);
 262
 263         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 264         if (tx == NULL) {
 265                 spin_unlock(&conn->ibc_lock);
 266
 267                 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
 268                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 269                 kiblnd_close_conn(conn, -EPROTO);
 270                 return;
 271         }
 272
 273         if (tx->tx_status == 0) {               /* success so far */
 274                 if (status < 0) {               /* failed? */
 275                         tx->tx_status = status;
 276                 } else if (txtype == IBLND_MSG_GET_REQ) {
 277                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 278                 }
 279         }
 280
 281         tx->tx_waiting = 0;
 282
 283         idle = !tx->tx_queued && (tx->tx_sending == 0);
 284         if (idle)
 285                 list_del(&tx->tx_list);
 286
 287         spin_unlock(&conn->ibc_lock);
 288
 289         if (idle)
 290                 kiblnd_tx_done(ni, tx);
 291 }
 292
 293 static void
 294 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
 295 {
 296         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 297         kib_tx_t    *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 298
 299         if (tx == NULL) {
 300                 CERROR("Can't get tx for completion %x for %s\n",
 301                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 302                 return;
 303         }
 304
 305         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 306         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 307         kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
 308
 309         kiblnd_queue_tx(tx, conn);
 310 }
 311
 312 static void
 313 kiblnd_handle_rx (kib_rx_t *rx)
 314 {
 315         kib_msg_t    *msg = rx->rx_msg;
 316         kib_conn_t   *conn = rx->rx_conn;
 317         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 318         int           credits = msg->ibm_credits;
 319         kib_tx_t     *tx;
 320         int           rc = 0;
 321         int           rc2;
 322         int           post_credit;
 323
 324         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 325
 326         CDEBUG (D_NET, "Received %x[%d] from %s\n",
 327                 msg->ibm_type, credits,
 328                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
 329
 330         if (credits != 0) {
 331                 /* Have I received credits that will let me send? */
 332                 spin_lock(&conn->ibc_lock);
 333
 334                 if (conn->ibc_credits + credits >
 335                     IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
 336                         rc2 = conn->ibc_credits;
 337                         spin_unlock(&conn->ibc_lock);
 338
 339                         CERROR("Bad credits from %s: %d + %d > %d\n",
 340                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 341                                rc2, credits,
 342                                IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
 343
 344                         kiblnd_close_conn(conn, -EPROTO);
 345                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 346                         return;
 347                 }
 348
 349                 conn->ibc_credits += credits;
 350
 351                 /* This ensures the credit taken by NOOP can be returned */
 352                 if (msg->ibm_type == IBLND_MSG_NOOP &&
 353                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
 354                         conn->ibc_outstanding_credits++;
 355
 356                 spin_unlock(&conn->ibc_lock);
 357                 kiblnd_check_sends(conn);
 358         }
 359
 360         switch (msg->ibm_type) {
 361         default:
 362                 CERROR("Bad IBLND message type %x from %s\n",
 363                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 364                 post_credit = IBLND_POSTRX_NO_CREDIT;
 365                 rc = -EPROTO;
 366                 break;
 367
 368         case IBLND_MSG_NOOP:
 369                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
 370                         post_credit = IBLND_POSTRX_NO_CREDIT;
 371                         break;
 372                 }
 373
 374                 if (credits != 0) /* credit already posted */
 375                         post_credit = IBLND_POSTRX_NO_CREDIT;
 376                 else              /* a keepalive NOOP */
 377                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 378                 break;
 379
 380         case IBLND_MSG_IMMEDIATE:
 381                 post_credit = IBLND_POSTRX_DONT_POST;
 382                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 383                                 msg->ibm_srcnid, rx, 0);
 384                 if (rc < 0)                     /* repost on error */
 385                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 386                 break;
 387
 388         case IBLND_MSG_PUT_REQ:
 389                 post_credit = IBLND_POSTRX_DONT_POST;
 390                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 391                                 msg->ibm_srcnid, rx, 1);
 392                 if (rc < 0)                     /* repost on error */
 393                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 394                 break;
 395
 396         case IBLND_MSG_PUT_NAK:
 397                 CWARN ("PUT_NACK from %s\n",
 398                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 399                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 400                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 401                                          msg->ibm_u.completion.ibcm_status,
 402                                          msg->ibm_u.completion.ibcm_cookie);
 403                 break;
 404
 405         case IBLND_MSG_PUT_ACK:
 406                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 407
 408                 spin_lock(&conn->ibc_lock);
 409                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 410                                         msg->ibm_u.putack.ibpam_src_cookie);
 411                 if (tx != NULL)
 412                         list_del(&tx->tx_list);
 413                 spin_unlock(&conn->ibc_lock);
 414
 415                 if (tx == NULL) {
 416                         CERROR("Unmatched PUT_ACK from %s\n",
 417                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 418                         rc = -EPROTO;
 419                         break;
 420                 }
 421
 422                 LASSERT (tx->tx_waiting);
 423                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
 424                  * (a) I can overwrite tx_msg since my peer has received it!
 425                  * (b) tx_waiting set tells tx_complete() it's not done. */
 426
 427                 tx->tx_nwrq = 0;                /* overwrite PUT_REQ */
 428
 429                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 430                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 431                                        &msg->ibm_u.putack.ibpam_rd,
 432                                        msg->ibm_u.putack.ibpam_dst_cookie);
 433                 if (rc2 < 0)
 434                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 435                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 436
 437                 spin_lock(&conn->ibc_lock);
 438                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
 439                 kiblnd_queue_tx_locked(tx, conn);
 440                 spin_unlock(&conn->ibc_lock);
 441                 break;
 442
 443         case IBLND_MSG_PUT_DONE:
 444                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 445                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 446                                          msg->ibm_u.completion.ibcm_status,
 447                                          msg->ibm_u.completion.ibcm_cookie);
 448                 break;
 449
 450         case IBLND_MSG_GET_REQ:
 451                 post_credit = IBLND_POSTRX_DONT_POST;
 452                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 453                                 msg->ibm_srcnid, rx, 1);
 454                 if (rc < 0)                     /* repost on error */
 455                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 456                 break;
 457
 458         case IBLND_MSG_GET_DONE:
 459                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 460                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 461                                          msg->ibm_u.completion.ibcm_status,
 462                                          msg->ibm_u.completion.ibcm_cookie);
 463                 break;
 464         }
 465
 466         if (rc < 0)                             /* protocol error */
 467                 kiblnd_close_conn(conn, rc);
 468
 469         if (post_credit != IBLND_POSTRX_DONT_POST)
 470                 kiblnd_post_rx(rx, post_credit);
 471 }
 472
 473 static void
 474 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
 475 {
 476         kib_msg_t    *msg = rx->rx_msg;
 477         kib_conn_t   *conn = rx->rx_conn;
 478         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 479         kib_net_t    *net = ni->ni_data;
 480         int           rc;
 481         int           err = -EIO;
 482
 483         LASSERT (net != NULL);
 484         LASSERT (rx->rx_nob < 0);               /* was posted */
 485         rx->rx_nob = 0;                         /* isn't now */
 486
 487         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 488                 goto ignore;
 489
 490         if (status != IB_WC_SUCCESS) {
 491                 CNETERR("Rx from %s failed: %d\n",
 492                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 493                 goto failed;
 494         }
 495
 496         LASSERT (nob >= 0);
 497         rx->rx_nob = nob;
 498
 499         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 500         if (rc != 0) {
 501                 CERROR ("Error %d unpacking rx from %s\n",
 502                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 503                 goto failed;
 504         }
 505
 506         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 507             msg->ibm_dstnid != ni->ni_nid ||
 508             msg->ibm_srcstamp != conn->ibc_incarnation ||
 509             msg->ibm_dststamp != net->ibn_incarnation) {
 510                 CERROR ("Stale rx from %s\n",
 511                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
 512                 err = -ESTALE;
 513                 goto failed;
 514         }
 515
 516         /* set time last known alive */
 517         kiblnd_peer_alive(conn->ibc_peer);
 518
 519         /* racing with connection establishment/teardown! */
 520
 521         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 522                 rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
 523                 unsigned long  flags;
 524
 525                 write_lock_irqsave(g_lock, flags);
 526                 /* must check holding global lock to eliminate race */
 527                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 528                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 529                         write_unlock_irqrestore(g_lock, flags);
 530                         return;
 531                 }
 532                 write_unlock_irqrestore(g_lock, flags);
 533         }
 534         kiblnd_handle_rx(rx);
 535         return;
 536
 537  failed:
 538         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 539         kiblnd_close_conn(conn, err);
 540  ignore:
 541         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
 542 }
 543
 544 static struct page *
 545 kiblnd_kvaddr_to_page (unsigned long vaddr)
 546 {
 547         struct page *page;
 548
 549         if (is_vmalloc_addr((void *)vaddr)) {
 550                 page = vmalloc_to_page ((void *)vaddr);
 551                 LASSERT (page != NULL);
 552                 return page;
 553         }
 554 #ifdef CONFIG_HIGHMEM
 555         if (vaddr >= PKMAP_BASE &&
 556             vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
 557                 /* No highmem pages only used for bulk (kiov) I/O */
 558                 CERROR("find page for address in highmem\n");
 559                 LBUG();
 560         }
 561 #endif
 562         page = virt_to_page (vaddr);
 563         LASSERT (page != NULL);
 564         return page;
 565 }
 566
 567 static int
 568 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 569 {
 570         kib_hca_dev_t           *hdev;
 571         __u64                   *pages = tx->tx_pages;
 572         kib_fmr_poolset_t       *fps;
 573         int                     npages;
 574         int                     size;
 575         int                     cpt;
 576         int                     rc;
 577         int                     i;
 578
 579         LASSERT(tx->tx_pool != NULL);
 580         LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
 581
 582         hdev  = tx->tx_pool->tpo_hdev;
 583
 584         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
 585                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
 586                                size += hdev->ibh_page_size) {
 587                         pages[npages ++] = (rd->rd_frags[i].rf_addr &
 588                                             hdev->ibh_page_mask) + size;
 589                 }
 590         }
 591
 592         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 593
 594         fps = net->ibn_fmr_ps[cpt];
 595         rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->tx_u.fmr);
 596         if (rc != 0) {
 597                 CERROR ("Can't map %d pages: %d\n", npages, rc);
 598                 return rc;
 599         }
 600
 601         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 602          * the rkey */
 603         rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
 604                                          tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
 605         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
 606         rd->rd_frags[0].rf_nob   = nob;
 607         rd->rd_nfrags = 1;
 608
 609         return 0;
 610 }
 611
 612 static int
 613 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 614 {
 615         kib_hca_dev_t           *hdev;
 616         kib_pmr_poolset_t       *pps;
 617         __u64                   iova;
 618         int                     cpt;
 619         int                     rc;
 620
 621         LASSERT(tx->tx_pool != NULL);
 622         LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
 623
 624         hdev = tx->tx_pool->tpo_hdev;
 625
 626         iova = rd->rd_frags[0].rf_addr & ~hdev->ibh_page_mask;
 627
 628         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 629
 630         pps = net->ibn_pmr_ps[cpt];
 631         rc = kiblnd_pmr_pool_map(pps, hdev, rd, &iova, &tx->tx_u.pmr);
 632         if (rc != 0) {
 633                 CERROR("Failed to create MR by phybuf: %d\n", rc);
 634                 return rc;
 635         }
 636
 637         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 638          * the rkey */
 639         rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
 640                                          tx->tx_u.pmr->pmr_mr->lkey;
 641         rd->rd_nfrags = 1;
 642         rd->rd_frags[0].rf_addr = iova;
 643         rd->rd_frags[0].rf_nob  = nob;
 644
 645         return 0;
 646 }
 647
 648 static void
 649 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
 650 {
 651         kib_net_t  *net = ni->ni_data;
 652
 653         LASSERT(net != NULL);
 654
 655         if (net->ibn_fmr_ps != NULL && tx->tx_u.fmr.fmr_pfmr != NULL) {
 656                 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
 657                 tx->tx_u.fmr.fmr_pfmr = NULL;
 658
 659         } else if (net->ibn_pmr_ps != NULL && tx->tx_u.pmr != NULL) {
 660                 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
 661                 tx->tx_u.pmr = NULL;
 662         }
 663
 664         if (tx->tx_nfrags != 0) {
 665                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
 666                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 667                 tx->tx_nfrags = 0;
 668         }
 669 }
 670
 671 static int
 672 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
 673 {
 674         kib_hca_dev_t *hdev  = tx->tx_pool->tpo_hdev;
 675         kib_net_t     *net   = ni->ni_data;
 676         struct ib_mr  *mr    = NULL;
 677         __u32          nob;
 678         int            i;
 679
 680         /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
 681          * RDMA sink */
 682         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 683         tx->tx_nfrags = nfrags;
 684
 685         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
 686                                           tx->tx_nfrags, tx->tx_dmadir);
 687
 688         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 689                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 690                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 691                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 692                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 693                 nob += rd->rd_frags[i].rf_nob;
 694         }
 695
 696         /* looking for pre-mapping MR */
 697         mr = kiblnd_find_rd_dma_mr(hdev, rd);
 698         if (mr != NULL) {
 699                 /* found pre-mapping MR */
 700                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 701                 return 0;
 702         }
 703
 704         if (net->ibn_fmr_ps != NULL)
 705                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 706         else if (net->ibn_pmr_ps != NULL)
 707                 return kiblnd_pmr_map_tx(net, tx, rd, nob);
 708
 709         return -EINVAL;
 710 }
 711
 712
 713 static int
 714 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 715                     unsigned int niov, struct iovec *iov, int offset, int nob)
 716 {
 717         kib_net_t          *net = ni->ni_data;
 718         struct page        *page;
 719         struct scatterlist *sg;
 720         unsigned long       vaddr;
 721         int                 fragnob;
 722         int                 page_offset;
 723
 724         LASSERT (nob > 0);
 725         LASSERT (niov > 0);
 726         LASSERT (net != NULL);
 727
 728         while (offset >= iov->iov_len) {
 729                 offset -= iov->iov_len;
 730                 niov--;
 731                 iov++;
 732                 LASSERT (niov > 0);
 733         }
 734
 735         sg = tx->tx_frags;
 736         do {
 737                 LASSERT (niov > 0);
 738
 739                 vaddr = ((unsigned long)iov->iov_base) + offset;
 740                 page_offset = vaddr & (PAGE_SIZE - 1);
 741                 page = kiblnd_kvaddr_to_page(vaddr);
 742                 if (page == NULL) {
 743                         CERROR ("Can't find page\n");
 744                         return -EFAULT;
 745                 }
 746
 747                 fragnob = min((int)(iov->iov_len - offset), nob);
 748                 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
 749
 750                 sg_set_page(sg, page, fragnob, page_offset);
 751                 sg++;
 752
 753                 if (offset + fragnob < iov->iov_len) {
 754                         offset += fragnob;
 755                 } else {
 756                         offset = 0;
 757                         iov++;
 758                         niov--;
 759                 }
 760                 nob -= fragnob;
 761         } while (nob > 0);
 762
 763         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 764 }
 765
 766 static int
 767 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 768                       int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 769 {
 770         kib_net_t          *net = ni->ni_data;
 771         struct scatterlist *sg;
 772         int                 fragnob;
 773
 774         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 775
 776         LASSERT (nob > 0);
 777         LASSERT (nkiov > 0);
 778         LASSERT (net != NULL);
 779
 780         while (offset >= kiov->kiov_len) {
 781                 offset -= kiov->kiov_len;
 782                 nkiov--;
 783                 kiov++;
 784                 LASSERT (nkiov > 0);
 785         }
 786
 787         sg = tx->tx_frags;
 788         do {
 789                 LASSERT (nkiov > 0);
 790
 791                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 792
 793                 sg_set_page(sg, kiov->kiov_page, fragnob,
 794                             kiov->kiov_offset + offset);
 795                 sg++;
 796
 797                 offset = 0;
 798                 kiov++;
 799                 nkiov--;
 800                 nob -= fragnob;
 801         } while (nob > 0);
 802
 803         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 804 }
 805
 806 static int
 807 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
 808 __must_hold(&conn->ibc_lock)
 809 {
 810         kib_msg_t         *msg = tx->tx_msg;
 811         kib_peer_t        *peer = conn->ibc_peer;
 812         int                ver = conn->ibc_version;
 813         int                rc;
 814         int                done;
 815         struct ib_send_wr *bad_wrq;
 816
 817         LASSERT (tx->tx_queued);
 818         /* We rely on this for QP sizing */
 819         LASSERT (tx->tx_nwrq > 0);
 820         LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
 821
 822         LASSERT (credit == 0 || credit == 1);
 823         LASSERT (conn->ibc_outstanding_credits >= 0);
 824         LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 825         LASSERT (conn->ibc_credits >= 0);
 826         LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 827
 828         if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
 829                 /* tx completions outstanding... */
 830                 CDEBUG(D_NET, "%s: posted enough\n",
 831                        libcfs_nid2str(peer->ibp_nid));
 832                 return -EAGAIN;
 833         }
 834
 835         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 836                 CDEBUG(D_NET, "%s: no credits\n",
 837                        libcfs_nid2str(peer->ibp_nid));
 838                 return -EAGAIN;
 839         }
 840
 841         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 842             conn->ibc_credits == 1 &&   /* last credit reserved */
 843             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
 844                 CDEBUG(D_NET, "%s: not using last credit\n",
 845                        libcfs_nid2str(peer->ibp_nid));
 846                 return -EAGAIN;
 847         }
 848
 849         /* NB don't drop ibc_lock before bumping tx_sending */
 850         list_del(&tx->tx_list);
 851         tx->tx_queued = 0;
 852
 853         if (msg->ibm_type == IBLND_MSG_NOOP &&
 854             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
 855              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 856               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 857                 /* OK to drop when posted enough NOOPs, since
 858                  * kiblnd_check_sends will queue NOOP again when
 859                  * posted NOOPs complete */
 860                 spin_unlock(&conn->ibc_lock);
 861                 kiblnd_tx_done(peer->ibp_ni, tx);
 862                 spin_lock(&conn->ibc_lock);
 863                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 864                        libcfs_nid2str(peer->ibp_nid),
 865                        conn->ibc_noops_posted);
 866                 return 0;
 867         }
 868
 869         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 870                         peer->ibp_nid, conn->ibc_incarnation);
 871
 872         conn->ibc_credits -= credit;
 873         conn->ibc_outstanding_credits = 0;
 874         conn->ibc_nsends_posted++;
 875         if (msg->ibm_type == IBLND_MSG_NOOP)
 876                 conn->ibc_noops_posted++;
 877
 878         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 879          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 880          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 881          * and then re-queued here.  It's (just) possible that
 882          * tx_sending is non-zero if we've not done the tx_complete()
 883          * from the first send; hence the ++ rather than = below. */
 884         tx->tx_sending++;
 885         list_add(&tx->tx_list, &conn->ibc_active_txs);
 886
 887         /* I'm still holding ibc_lock! */
 888         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
 889                 rc = -ECONNABORTED;
 890         } else if (tx->tx_pool->tpo_pool.po_failed ||
 891                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
 892                 /* close_conn will launch failover */
 893                 rc = -ENETDOWN;
 894         } else {
 895                 rc = ib_post_send(conn->ibc_cmid->qp,
 896                                   tx->tx_wrq, &bad_wrq);
 897         }
 898
 899         conn->ibc_last_send = jiffies;
 900
 901         if (rc == 0)
 902                 return 0;
 903
 904         /* NB credits are transferred in the actual
 905          * message, which can only be the last work item */
 906         conn->ibc_credits += credit;
 907         conn->ibc_outstanding_credits += msg->ibm_credits;
 908         conn->ibc_nsends_posted--;
 909         if (msg->ibm_type == IBLND_MSG_NOOP)
 910                 conn->ibc_noops_posted--;
 911
 912         tx->tx_status = rc;
 913         tx->tx_waiting = 0;
 914         tx->tx_sending--;
 915
 916         done = (tx->tx_sending == 0);
 917         if (done)
 918                 list_del(&tx->tx_list);
 919
 920         spin_unlock(&conn->ibc_lock);
 921
 922         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 923                 CERROR("Error %d posting transmit to %s\n",
 924                        rc, libcfs_nid2str(peer->ibp_nid));
 925         else
 926                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 927                        rc, libcfs_nid2str(peer->ibp_nid));
 928
 929         kiblnd_close_conn(conn, rc);
 930
 931         if (done)
 932                 kiblnd_tx_done(peer->ibp_ni, tx);
 933
 934         spin_lock(&conn->ibc_lock);
 935
 936         return -EIO;
 937 }
 938
 939 static void
 940 kiblnd_check_sends (kib_conn_t *conn)
 941 {
 942         int        ver = conn->ibc_version;
 943         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 944         kib_tx_t  *tx;
 945
 946         /* Don't send anything until after the connection is established */
 947         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 948                 CDEBUG(D_NET, "%s too soon\n",
 949                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 950                 return;
 951         }
 952
 953         spin_lock(&conn->ibc_lock);
 954
 955         LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
 956         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 957                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 958         LASSERT (conn->ibc_reserved_credits >= 0);
 959
 960         while (conn->ibc_reserved_credits > 0 &&
 961                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 962                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 963                                     kib_tx_t, tx_list);
 964                 list_del(&tx->tx_list);
 965                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 966                 conn->ibc_reserved_credits--;
 967         }
 968
 969         if (kiblnd_need_noop(conn)) {
 970                 spin_unlock(&conn->ibc_lock);
 971
 972                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 973                 if (tx != NULL)
 974                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 975
 976                 spin_lock(&conn->ibc_lock);
 977                 if (tx != NULL)
 978                         kiblnd_queue_tx_locked(tx, conn);
 979         }
 980
 981         kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
 982
 983         for (;;) {
 984                 int credit;
 985
 986                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 987                         credit = 0;
 988                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
 989                                             kib_tx_t, tx_list);
 990                 } else if (!list_empty(&conn->ibc_tx_noops)) {
 991                         LASSERT (!IBLND_OOB_CAPABLE(ver));
 992                         credit = 1;
 993                         tx = list_entry(conn->ibc_tx_noops.next,
 994                                         kib_tx_t, tx_list);
 995                 } else if (!list_empty(&conn->ibc_tx_queue)) {
 996                         credit = 1;
 997                         tx = list_entry(conn->ibc_tx_queue.next,
 998                                             kib_tx_t, tx_list);
 999                 } else
1000                         break;
1001
1002                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
1003                         break;
1004         }
1005
1006         spin_unlock(&conn->ibc_lock);
1007
1008         kiblnd_conn_decref(conn); /* ...until here */
1009 }
1010
1011 static void
1012 kiblnd_tx_complete (kib_tx_t *tx, int status)
1013 {
1014         int           failed = (status != IB_WC_SUCCESS);
1015         kib_conn_t   *conn = tx->tx_conn;
1016         int           idle;
1017
1018         LASSERT (tx->tx_sending > 0);
1019
1020         if (failed) {
1021                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
1022                         CNETERR("Tx -> %s cookie "LPX64
1023                                 " sending %d waiting %d: failed %d\n",
1024                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
1025                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
1026                                 status);
1027
1028                 kiblnd_close_conn(conn, -EIO);
1029         } else {
1030                 kiblnd_peer_alive(conn->ibc_peer);
1031         }
1032
1033         spin_lock(&conn->ibc_lock);
1034
1035         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
1036          * gets to free it, which also drops its ref on 'conn'. */
1037
1038         tx->tx_sending--;
1039         conn->ibc_nsends_posted--;
1040         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1041                 conn->ibc_noops_posted--;
1042
1043         if (failed) {
1044                 tx->tx_waiting = 0;             /* don't wait for peer */
1045                 tx->tx_status = -EIO;
1046         }
1047
1048         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1049                !tx->tx_waiting &&               /* Not waiting for peer */
1050                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1051         if (idle)
1052                 list_del(&tx->tx_list);
1053
1054         kiblnd_conn_addref(conn);               /* 1 ref for me.... */
1055
1056         spin_unlock(&conn->ibc_lock);
1057
1058         if (idle)
1059                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1060
1061         kiblnd_check_sends(conn);
1062
1063         kiblnd_conn_decref(conn);               /* ...until here */
1064 }
1065
1066 static void
1067 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1068 {
1069         kib_hca_dev_t     *hdev = tx->tx_pool->tpo_hdev;
1070         struct ib_sge     *sge = &tx->tx_sge[tx->tx_nwrq];
1071         struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1072         int                nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1073         struct ib_mr      *mr;
1074
1075         LASSERT (tx->tx_nwrq >= 0);
1076         LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1077         LASSERT (nob <= IBLND_MSG_SIZE);
1078
1079         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1080
1081         mr = kiblnd_find_dma_mr(hdev, tx->tx_msgaddr, nob);
1082         LASSERT (mr != NULL);
1083
1084         sge->lkey   = mr->lkey;
1085         sge->addr   = tx->tx_msgaddr;
1086         sge->length = nob;
1087
1088         memset(wrq, 0, sizeof(*wrq));
1089
1090         wrq->next       = NULL;
1091         wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1092         wrq->sg_list    = sge;
1093         wrq->num_sge    = 1;
1094         wrq->opcode     = IB_WR_SEND;
1095         wrq->send_flags = IB_SEND_SIGNALED;
1096
1097         tx->tx_nwrq++;
1098 }
1099
1100 static int
1101 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1102                  int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1103 {
1104         kib_msg_t         *ibmsg = tx->tx_msg;
1105         kib_rdma_desc_t   *srcrd = tx->tx_rd;
1106         struct ib_sge     *sge = &tx->tx_sge[0];
1107         struct ib_send_wr *wrq = &tx->tx_wrq[0];
1108         int                rc  = resid;
1109         int                srcidx;
1110         int                dstidx;
1111         int                wrknob;
1112
1113         LASSERT (!in_interrupt());
1114         LASSERT (tx->tx_nwrq == 0);
1115         LASSERT (type == IBLND_MSG_GET_DONE ||
1116                  type == IBLND_MSG_PUT_DONE);
1117
1118         srcidx = dstidx = 0;
1119
1120         while (resid > 0) {
1121                 if (srcidx >= srcrd->rd_nfrags) {
1122                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1123                         rc = -EPROTO;
1124                         break;
1125                 }
1126
1127                 if (dstidx == dstrd->rd_nfrags) {
1128                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1129                         rc = -EPROTO;
1130                         break;
1131                 }
1132
1133                 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1134                         CERROR("RDMA too fragmented for %s (%d): "
1135                                "%d/%d src %d/%d dst frags\n",
1136                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1137                                IBLND_RDMA_FRAGS(conn->ibc_version),
1138                                srcidx, srcrd->rd_nfrags,
1139                                dstidx, dstrd->rd_nfrags);
1140                         rc = -EMSGSIZE;
1141                         break;
1142                 }
1143
1144                 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1145                                  kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1146
1147                 sge = &tx->tx_sge[tx->tx_nwrq];
1148                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1149                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1150                 sge->length = wrknob;
1151
1152                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1153
1154                 wrq->next       = wrq + 1;
1155                 wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1156                 wrq->sg_list    = sge;
1157                 wrq->num_sge    = 1;
1158                 wrq->opcode     = IB_WR_RDMA_WRITE;
1159                 wrq->send_flags = 0;
1160
1161                 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1162                 wrq->wr.rdma.rkey        = kiblnd_rd_frag_key(dstrd, dstidx);
1163
1164                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1165                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1166
1167                 resid -= wrknob;
1168
1169                 tx->tx_nwrq++;
1170                 wrq++;
1171                 sge++;
1172         }
1173
1174         if (rc < 0)                             /* no RDMA if completing with failure */
1175                 tx->tx_nwrq = 0;
1176
1177         ibmsg->ibm_u.completion.ibcm_status = rc;
1178         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1179         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1180                            type, sizeof (kib_completion_msg_t));
1181
1182         return rc;
1183 }
1184
1185 static void
1186 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1187 {
1188         struct list_head *q;
1189
1190         LASSERT(tx->tx_nwrq > 0);       /* work items set up */
1191         LASSERT(!tx->tx_queued);        /* not queued for sending already */
1192         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1193
1194         tx->tx_queued = 1;
1195         tx->tx_deadline = jiffies +
1196                           msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1197                                            MSEC_PER_SEC);
1198
1199         if (tx->tx_conn == NULL) {
1200                 kiblnd_conn_addref(conn);
1201                 tx->tx_conn = conn;
1202                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1203         } else {
1204                 /* PUT_DONE first attached to conn as a PUT_REQ */
1205                 LASSERT (tx->tx_conn == conn);
1206                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1207         }
1208
1209         switch (tx->tx_msg->ibm_type) {
1210         default:
1211                 LBUG();
1212
1213         case IBLND_MSG_PUT_REQ:
1214         case IBLND_MSG_GET_REQ:
1215                 q = &conn->ibc_tx_queue_rsrvd;
1216                 break;
1217
1218         case IBLND_MSG_PUT_NAK:
1219         case IBLND_MSG_PUT_ACK:
1220         case IBLND_MSG_PUT_DONE:
1221         case IBLND_MSG_GET_DONE:
1222                 q = &conn->ibc_tx_queue_nocred;
1223                 break;
1224
1225         case IBLND_MSG_NOOP:
1226                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1227                         q = &conn->ibc_tx_queue_nocred;
1228                 else
1229                         q = &conn->ibc_tx_noops;
1230                 break;
1231
1232         case IBLND_MSG_IMMEDIATE:
1233                 q = &conn->ibc_tx_queue;
1234                 break;
1235         }
1236
1237         list_add_tail(&tx->tx_list, q);
1238 }
1239
1240 static void
1241 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1242 {
1243         spin_lock(&conn->ibc_lock);
1244         kiblnd_queue_tx_locked(tx, conn);
1245         spin_unlock(&conn->ibc_lock);
1246
1247         kiblnd_check_sends(conn);
1248 }
1249
1250 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1251                                struct sockaddr_in *srcaddr,
1252                                struct sockaddr_in *dstaddr,
1253                                int timeout_ms)
1254 {
1255         unsigned short port;
1256         int rc;
1257
1258         /* allow the port to be reused */
1259         rc = rdma_set_reuseaddr(cmid, 1);
1260         if (rc != 0) {
1261                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1262                 return rc;
1263         }
1264
1265         /* look for a free privileged port */
1266         for (port = PROT_SOCK-1; port > 0; port--) {
1267                 srcaddr->sin_port = htons(port);
1268                 rc = rdma_resolve_addr(cmid,
1269                                        (struct sockaddr *)srcaddr,
1270                                        (struct sockaddr *)dstaddr,
1271                                        timeout_ms);
1272                 if (rc == 0) {
1273                         CDEBUG(D_NET, "bound to port %hu\n", port);
1274                         return 0;
1275                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1276                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1277                                port, rc);
1278                 } else {
1279                         return rc;
1280                 }
1281         }
1282
1283         CERROR("Failed to bind to a free privileged port\n");
1284         return rc;
1285 }
1286
1287 void
1288 kiblnd_connect_peer (kib_peer_t *peer)
1289 {
1290         struct rdma_cm_id *cmid;
1291         kib_dev_t         *dev;
1292         kib_net_t         *net = peer->ibp_ni->ni_data;
1293         struct sockaddr_in srcaddr;
1294         struct sockaddr_in dstaddr;
1295         int                rc;
1296
1297         LASSERT (net != NULL);
1298         LASSERT (peer->ibp_connecting > 0);
1299
1300         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1301                                      IB_QPT_RC);
1302
1303         if (IS_ERR(cmid)) {
1304                 CERROR("Can't create CMID for %s: %ld\n",
1305                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1306                 rc = PTR_ERR(cmid);
1307                 goto failed;
1308         }
1309
1310         dev = net->ibn_dev;
1311         memset(&srcaddr, 0, sizeof(srcaddr));
1312         srcaddr.sin_family = AF_INET;
1313         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1314
1315         memset(&dstaddr, 0, sizeof(dstaddr));
1316         dstaddr.sin_family = AF_INET;
1317         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1318         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1319
1320         kiblnd_peer_addref(peer);               /* cmid's ref */
1321
1322         if (*kiblnd_tunables.kib_use_priv_port) {
1323                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1324                                          *kiblnd_tunables.kib_timeout * 1000);
1325         } else {
1326                 rc = rdma_resolve_addr(cmid,
1327                                        (struct sockaddr *)&srcaddr,
1328                                        (struct sockaddr *)&dstaddr,
1329                                        *kiblnd_tunables.kib_timeout * 1000);
1330         }
1331         if (rc != 0) {
1332                 /* Can't initiate address resolution:  */
1333                 CERROR("Can't resolve addr for %s: %d\n",
1334                        libcfs_nid2str(peer->ibp_nid), rc);
1335                 goto failed2;
1336         }
1337
1338         LASSERT (cmid->device != NULL);
1339         CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1340                libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1341                &dev->ibd_ifip, cmid->device->name);
1342
1343         return;
1344
1345  failed2:
1346         kiblnd_peer_decref(peer);               /* cmid's ref */
1347         rdma_destroy_id(cmid);
1348  failed:
1349         kiblnd_peer_connect_failed(peer, 1, rc);
1350 }
1351
1352 void
1353 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1354 {
1355         kib_peer_t        *peer;
1356         kib_peer_t        *peer2;
1357         kib_conn_t        *conn;
1358         rwlock_t        *g_lock = &kiblnd_data.kib_global_lock;
1359         unsigned long      flags;
1360         int                rc;
1361
1362         /* If I get here, I've committed to send, so I complete the tx with
1363          * failure on any problems */
1364
1365         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1366         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1367
1368         /* First time, just use a read lock since I expect to find my peer
1369          * connected */
1370         read_lock_irqsave(g_lock, flags);
1371
1372         peer = kiblnd_find_peer_locked(nid);
1373         if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1374                 /* Found a peer with an established connection */
1375                 conn = kiblnd_get_conn_locked(peer);
1376                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1377
1378                 read_unlock_irqrestore(g_lock, flags);
1379
1380                 if (tx != NULL)
1381                         kiblnd_queue_tx(tx, conn);
1382                 kiblnd_conn_decref(conn); /* ...to here */
1383                 return;
1384         }
1385
1386         read_unlock(g_lock);
1387         /* Re-try with a write lock */
1388         write_lock(g_lock);
1389
1390         peer = kiblnd_find_peer_locked(nid);
1391         if (peer != NULL) {
1392                 if (list_empty(&peer->ibp_conns)) {
1393                         /* found a peer, but it's still connecting... */
1394                         LASSERT (peer->ibp_connecting != 0 ||
1395                                  peer->ibp_accepting != 0);
1396                         if (tx != NULL)
1397                                 list_add_tail(&tx->tx_list,
1398                                                   &peer->ibp_tx_queue);
1399                         write_unlock_irqrestore(g_lock, flags);
1400                 } else {
1401                         conn = kiblnd_get_conn_locked(peer);
1402                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1403
1404                         write_unlock_irqrestore(g_lock, flags);
1405
1406                         if (tx != NULL)
1407                                 kiblnd_queue_tx(tx, conn);
1408                         kiblnd_conn_decref(conn); /* ...to here */
1409                 }
1410                 return;
1411         }
1412
1413         write_unlock_irqrestore(g_lock, flags);
1414
1415         /* Allocate a peer ready to add to the peer table and retry */
1416         rc = kiblnd_create_peer(ni, &peer, nid);
1417         if (rc != 0) {
1418                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1419                 if (tx != NULL) {
1420                         tx->tx_status = -EHOSTUNREACH;
1421                         tx->tx_waiting = 0;
1422                         kiblnd_tx_done(ni, tx);
1423                 }
1424                 return;
1425         }
1426
1427         write_lock_irqsave(g_lock, flags);
1428
1429         peer2 = kiblnd_find_peer_locked(nid);
1430         if (peer2 != NULL) {
1431                 if (list_empty(&peer2->ibp_conns)) {
1432                         /* found a peer, but it's still connecting... */
1433                         LASSERT (peer2->ibp_connecting != 0 ||
1434                                  peer2->ibp_accepting != 0);
1435                         if (tx != NULL)
1436                                 list_add_tail(&tx->tx_list,
1437                                                   &peer2->ibp_tx_queue);
1438                         write_unlock_irqrestore(g_lock, flags);
1439                 } else {
1440                         conn = kiblnd_get_conn_locked(peer2);
1441                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1442
1443                         write_unlock_irqrestore(g_lock, flags);
1444
1445                         if (tx != NULL)
1446                                 kiblnd_queue_tx(tx, conn);
1447                         kiblnd_conn_decref(conn); /* ...to here */
1448                 }
1449
1450                 kiblnd_peer_decref(peer);
1451                 return;
1452         }
1453
1454         /* Brand new peer */
1455         LASSERT (peer->ibp_connecting == 0);
1456         peer->ibp_connecting = 1;
1457
1458         /* always called with a ref on ni, which prevents ni being shutdown */
1459         LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1460
1461         if (tx != NULL)
1462                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1463
1464         kiblnd_peer_addref(peer);
1465         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1466
1467         write_unlock_irqrestore(g_lock, flags);
1468
1469         kiblnd_connect_peer(peer);
1470         kiblnd_peer_decref(peer);
1471 }
1472
1473 int
1474 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1475 {
1476         lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
1477         int               type = lntmsg->msg_type;
1478         lnet_process_id_t target = lntmsg->msg_target;
1479         int               target_is_router = lntmsg->msg_target_is_router;
1480         int               routing = lntmsg->msg_routing;
1481         unsigned int      payload_niov = lntmsg->msg_niov;
1482         struct iovec     *payload_iov = lntmsg->msg_iov;
1483         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
1484         unsigned int      payload_offset = lntmsg->msg_offset;
1485         unsigned int      payload_nob = lntmsg->msg_len;
1486         kib_msg_t        *ibmsg;
1487         kib_rdma_desc_t  *rd;
1488         kib_tx_t         *tx;
1489         int               nob;
1490         int               rc;
1491
1492         /* NB 'private' is different depending on what we're sending.... */
1493
1494         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1495                payload_nob, payload_niov, libcfs_id2str(target));
1496
1497         LASSERT (payload_nob == 0 || payload_niov > 0);
1498         LASSERT (payload_niov <= LNET_MAX_IOV);
1499
1500         /* Thread context */
1501         LASSERT (!in_interrupt());
1502         /* payload is either all vaddrs or all pages */
1503         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1504
1505         switch (type) {
1506         default:
1507                 LBUG();
1508                 return (-EIO);
1509
1510         case LNET_MSG_ACK:
1511                 LASSERT (payload_nob == 0);
1512                 break;
1513
1514         case LNET_MSG_GET:
1515                 if (routing || target_is_router)
1516                         break;                  /* send IMMEDIATE */
1517
1518                 /* is the REPLY message too small for RDMA? */
1519                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1520                 if (nob <= IBLND_MSG_SIZE)
1521                         break;                  /* send IMMEDIATE */
1522
1523                 tx = kiblnd_get_idle_tx(ni, target.nid);
1524                 if (tx == NULL) {
1525                         CERROR("Can't allocate txd for GET to %s\n",
1526                                libcfs_nid2str(target.nid));
1527                         return -ENOMEM;
1528                 }
1529
1530                 ibmsg = tx->tx_msg;
1531                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1532                 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1533                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1534                                                  lntmsg->msg_md->md_niov,
1535                                                  lntmsg->msg_md->md_iov.iov,
1536                                                  0, lntmsg->msg_md->md_length);
1537                 else
1538                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1539                                                   lntmsg->msg_md->md_niov,
1540                                                   lntmsg->msg_md->md_iov.kiov,
1541                                                   0, lntmsg->msg_md->md_length);
1542                 if (rc != 0) {
1543                         CERROR("Can't setup GET sink for %s: %d\n",
1544                                libcfs_nid2str(target.nid), rc);
1545                         kiblnd_tx_done(ni, tx);
1546                         return -EIO;
1547                 }
1548
1549                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1550                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1551                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1552
1553                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1554
1555                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1556                 if (tx->tx_lntmsg[1] == NULL) {
1557                         CERROR("Can't create reply for GET -> %s\n",
1558                                libcfs_nid2str(target.nid));
1559                         kiblnd_tx_done(ni, tx);
1560                         return -EIO;
1561                 }
1562
1563                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1564                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1565                 kiblnd_launch_tx(ni, tx, target.nid);
1566                 return 0;
1567
1568         case LNET_MSG_REPLY:
1569         case LNET_MSG_PUT:
1570                 /* Is the payload small enough not to need RDMA? */
1571                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1572                 if (nob <= IBLND_MSG_SIZE)
1573                         break;                  /* send IMMEDIATE */
1574
1575                 tx = kiblnd_get_idle_tx(ni, target.nid);
1576                 if (tx == NULL) {
1577                         CERROR("Can't allocate %s txd for %s\n",
1578                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1579                                libcfs_nid2str(target.nid));
1580                         return -ENOMEM;
1581                 }
1582
1583                 if (payload_kiov == NULL)
1584                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1585                                                  payload_niov, payload_iov,
1586                                                  payload_offset, payload_nob);
1587                 else
1588                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1589                                                   payload_niov, payload_kiov,
1590                                                   payload_offset, payload_nob);
1591                 if (rc != 0) {
1592                         CERROR("Can't setup PUT src for %s: %d\n",
1593                                libcfs_nid2str(target.nid), rc);
1594                         kiblnd_tx_done(ni, tx);
1595                         return -EIO;
1596                 }
1597
1598                 ibmsg = tx->tx_msg;
1599                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1600                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1601                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1602
1603                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1604                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1605                 kiblnd_launch_tx(ni, tx, target.nid);
1606                 return 0;
1607         }
1608
1609         /* send IMMEDIATE */
1610
1611         LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1612                  <= IBLND_MSG_SIZE);
1613
1614         tx = kiblnd_get_idle_tx(ni, target.nid);
1615         if (tx == NULL) {
1616                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1617                         type, libcfs_nid2str(target.nid));
1618                 return -ENOMEM;
1619         }
1620
1621         ibmsg = tx->tx_msg;
1622         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1623
1624         if (payload_kiov != NULL)
1625                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1626                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1627                                     payload_niov, payload_kiov,
1628                                     payload_offset, payload_nob);
1629         else
1630                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1631                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1632                                    payload_niov, payload_iov,
1633                                    payload_offset, payload_nob);
1634
1635         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1636         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1637
1638         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1639         kiblnd_launch_tx(ni, tx, target.nid);
1640         return 0;
1641 }
1642
1643 static void
1644 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1645 {
1646         lnet_process_id_t target = lntmsg->msg_target;
1647         unsigned int      niov = lntmsg->msg_niov;
1648         struct iovec     *iov = lntmsg->msg_iov;
1649         lnet_kiov_t      *kiov = lntmsg->msg_kiov;
1650         unsigned int      offset = lntmsg->msg_offset;
1651         unsigned int      nob = lntmsg->msg_len;
1652         kib_tx_t         *tx;
1653         int               rc;
1654
1655         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1656         if (tx == NULL) {
1657                 CERROR("Can't get tx for REPLY to %s\n",
1658                        libcfs_nid2str(target.nid));
1659                 goto failed_0;
1660         }
1661
1662         if (nob == 0)
1663                 rc = 0;
1664         else if (kiov == NULL)
1665                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1666                                          niov, iov, offset, nob);
1667         else
1668                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1669                                           niov, kiov, offset, nob);
1670
1671         if (rc != 0) {
1672                 CERROR("Can't setup GET src for %s: %d\n",
1673                        libcfs_nid2str(target.nid), rc);
1674                 goto failed_1;
1675         }
1676
1677         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1678                               IBLND_MSG_GET_DONE, nob,
1679                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1680                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1681         if (rc < 0) {
1682                 CERROR("Can't setup rdma for GET from %s: %d\n",
1683                        libcfs_nid2str(target.nid), rc);
1684                 goto failed_1;
1685         }
1686
1687         if (nob == 0) {
1688                 /* No RDMA: local completion may happen now! */
1689                 lnet_finalize(ni, lntmsg, 0);
1690         } else {
1691                 /* RDMA: lnet_finalize(lntmsg) when it
1692                  * completes */
1693                 tx->tx_lntmsg[0] = lntmsg;
1694         }
1695
1696         kiblnd_queue_tx(tx, rx->rx_conn);
1697         return;
1698
1699  failed_1:
1700         kiblnd_tx_done(ni, tx);
1701  failed_0:
1702         lnet_finalize(ni, lntmsg, -EIO);
1703 }
1704
1705 int
1706 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1707              unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1708              unsigned int offset, unsigned int mlen, unsigned int rlen)
1709 {
1710         kib_rx_t    *rx = private;
1711         kib_msg_t   *rxmsg = rx->rx_msg;
1712         kib_conn_t  *conn = rx->rx_conn;
1713         kib_tx_t    *tx;
1714         int          nob;
1715         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1716         int          rc = 0;
1717
1718         LASSERT (mlen <= rlen);
1719         LASSERT (!in_interrupt());
1720         /* Either all pages or all vaddrs */
1721         LASSERT (!(kiov != NULL && iov != NULL));
1722
1723         switch (rxmsg->ibm_type) {
1724         default:
1725                 LBUG();
1726
1727         case IBLND_MSG_IMMEDIATE:
1728                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1729                 if (nob > rx->rx_nob) {
1730                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1731                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1732                                 nob, rx->rx_nob);
1733                         rc = -EPROTO;
1734                         break;
1735                 }
1736
1737                 if (kiov != NULL)
1738                         lnet_copy_flat2kiov(niov, kiov, offset,
1739                                             IBLND_MSG_SIZE, rxmsg,
1740                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1741                                             mlen);
1742                 else
1743                         lnet_copy_flat2iov(niov, iov, offset,
1744                                            IBLND_MSG_SIZE, rxmsg,
1745                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1746                                            mlen);
1747                 lnet_finalize (ni, lntmsg, 0);
1748                 break;
1749
1750         case IBLND_MSG_PUT_REQ: {
1751                 kib_msg_t       *txmsg;
1752                 kib_rdma_desc_t *rd;
1753
1754                 if (mlen == 0) {
1755                         lnet_finalize(ni, lntmsg, 0);
1756                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1757                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1758                         break;
1759                 }
1760
1761                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1762                 if (tx == NULL) {
1763                         CERROR("Can't allocate tx for %s\n",
1764                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1765                         /* Not replying will break the connection */
1766                         rc = -ENOMEM;
1767                         break;
1768                 }
1769
1770                 txmsg = tx->tx_msg;
1771                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1772                 if (kiov == NULL)
1773                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1774                                                  niov, iov, offset, mlen);
1775                 else
1776                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1777                                                   niov, kiov, offset, mlen);
1778                 if (rc != 0) {
1779                         CERROR("Can't setup PUT sink for %s: %d\n",
1780                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1781                         kiblnd_tx_done(ni, tx);
1782                         /* tell peer it's over */
1783                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1784                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1785                         break;
1786                 }
1787
1788                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1789                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1790                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1791
1792                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1793
1794                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1795                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1796                 kiblnd_queue_tx(tx, conn);
1797
1798                 /* reposted buffer reserved for PUT_DONE */
1799                 post_credit = IBLND_POSTRX_NO_CREDIT;
1800                 break;
1801                 }
1802
1803         case IBLND_MSG_GET_REQ:
1804                 if (lntmsg != NULL) {
1805                         /* Optimized GET; RDMA lntmsg's payload */
1806                         kiblnd_reply(ni, rx, lntmsg);
1807                 } else {
1808                         /* GET didn't match anything */
1809                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1810                                                -ENODATA,
1811                                                rxmsg->ibm_u.get.ibgm_cookie);
1812                 }
1813                 break;
1814         }
1815
1816         kiblnd_post_rx(rx, post_credit);
1817         return rc;
1818 }
1819
1820 int
1821 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1822 {
1823         struct task_struct *task = kthread_run(fn, arg, name);
1824
1825         if (IS_ERR(task))
1826                 return PTR_ERR(task);
1827
1828         atomic_inc(&kiblnd_data.kib_nthreads);
1829         return 0;
1830 }
1831
1832 static void
1833 kiblnd_thread_fini (void)
1834 {
1835         atomic_dec (&kiblnd_data.kib_nthreads);
1836 }
1837
1838 static void
1839 kiblnd_peer_alive (kib_peer_t *peer)
1840 {
1841         /* This is racy, but everyone's only writing cfs_time_current() */
1842         peer->ibp_last_alive = cfs_time_current();
1843         smp_mb();
1844 }
1845
1846 static void
1847 kiblnd_peer_notify (kib_peer_t *peer)
1848 {
1849         int           error = 0;
1850         cfs_time_t    last_alive = 0;
1851         unsigned long flags;
1852
1853         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1854
1855         if (list_empty(&peer->ibp_conns) &&
1856             peer->ibp_accepting == 0 &&
1857             peer->ibp_connecting == 0 &&
1858             peer->ibp_error != 0) {
1859                 error = peer->ibp_error;
1860                 peer->ibp_error = 0;
1861
1862                 last_alive = peer->ibp_last_alive;
1863         }
1864
1865         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1866
1867         if (error != 0)
1868                 lnet_notify(peer->ibp_ni,
1869                             peer->ibp_nid, 0, last_alive);
1870 }
1871
1872 void
1873 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1874 {
1875         /* This just does the immediate housekeeping.  'error' is zero for a
1876          * normal shutdown which can happen only after the connection has been
1877          * established.  If the connection is established, schedule the
1878          * connection to be finished off by the connd.  Otherwise the connd is
1879          * already dealing with it (either to set it up or tear it down).
1880          * Caller holds kib_global_lock exclusively in irq context */
1881         kib_peer_t       *peer = conn->ibc_peer;
1882         kib_dev_t        *dev;
1883         unsigned long     flags;
1884
1885         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1886
1887         if (error != 0 && conn->ibc_comms_error == 0)
1888                 conn->ibc_comms_error = error;
1889
1890         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1891                 return; /* already being handled  */
1892
1893         if (error == 0 &&
1894             list_empty(&conn->ibc_tx_noops) &&
1895             list_empty(&conn->ibc_tx_queue) &&
1896             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1897             list_empty(&conn->ibc_tx_queue_nocred) &&
1898             list_empty(&conn->ibc_active_txs)) {
1899                 CDEBUG(D_NET, "closing conn to %s\n",
1900                        libcfs_nid2str(peer->ibp_nid));
1901         } else {
1902                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1903                        libcfs_nid2str(peer->ibp_nid), error,
1904                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1905                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1906                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
1907                                                 "" : "(sending_rsrvd)",
1908                        list_empty(&conn->ibc_tx_queue_nocred) ?
1909                                                  "" : "(sending_nocred)",
1910                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1911         }
1912
1913         dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1914         list_del(&conn->ibc_list);
1915         /* connd (see below) takes over ibc_list's ref */
1916
1917         if (list_empty(&peer->ibp_conns) &&    /* no more conns */
1918             kiblnd_peer_active(peer)) {         /* still in peer table */
1919                 kiblnd_unlink_peer_locked(peer);
1920
1921                 /* set/clear error on last conn */
1922                 peer->ibp_error = conn->ibc_comms_error;
1923         }
1924
1925         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1926
1927         if (error != 0 &&
1928             kiblnd_dev_can_failover(dev)) {
1929                 list_add_tail(&dev->ibd_fail_list,
1930                               &kiblnd_data.kib_failed_devs);
1931                 wake_up(&kiblnd_data.kib_failover_waitq);
1932         }
1933
1934         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1935
1936         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1937         wake_up(&kiblnd_data.kib_connd_waitq);
1938
1939         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1940 }
1941
1942 void
1943 kiblnd_close_conn(kib_conn_t *conn, int error)
1944 {
1945         unsigned long flags;
1946
1947         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1948
1949         kiblnd_close_conn_locked(conn, error);
1950
1951         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1952 }
1953
1954 static void
1955 kiblnd_handle_early_rxs(kib_conn_t *conn)
1956 {
1957         unsigned long    flags;
1958         kib_rx_t        *rx;
1959
1960         LASSERT(!in_interrupt());
1961         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1962
1963         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1964         while (!list_empty(&conn->ibc_early_rxs)) {
1965                 rx = list_entry(conn->ibc_early_rxs.next,
1966                                     kib_rx_t, rx_list);
1967                 list_del(&rx->rx_list);
1968                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1969
1970                 kiblnd_handle_rx(rx);
1971
1972                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1973         }
1974         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1975 }
1976
1977 static void
1978 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1979 {
1980         struct list_head         zombies = LIST_HEAD_INIT(zombies);
1981         struct list_head        *tmp;
1982         struct list_head        *nxt;
1983         kib_tx_t                *tx;
1984
1985         spin_lock(&conn->ibc_lock);
1986
1987         list_for_each_safe(tmp, nxt, txs) {
1988                 tx = list_entry(tmp, kib_tx_t, tx_list);
1989
1990                 if (txs == &conn->ibc_active_txs) {
1991                         LASSERT(!tx->tx_queued);
1992                         LASSERT(tx->tx_waiting ||
1993                                 tx->tx_sending != 0);
1994                 } else {
1995                         LASSERT(tx->tx_queued);
1996                 }
1997
1998                 tx->tx_status = -ECONNABORTED;
1999                 tx->tx_waiting = 0;
2000
2001                 if (tx->tx_sending == 0) {
2002                         tx->tx_queued = 0;
2003                         list_del(&tx->tx_list);
2004                         list_add(&tx->tx_list, &zombies);
2005                 }
2006         }
2007
2008         spin_unlock(&conn->ibc_lock);
2009
2010         kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2011 }
2012
2013 static void
2014 kiblnd_finalise_conn (kib_conn_t *conn)
2015 {
2016         LASSERT (!in_interrupt());
2017         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2018
2019         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2020
2021         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2022          * for connections that didn't get as far as being connected, because
2023          * rdma_disconnect() does this for free. */
2024         kiblnd_abort_receives(conn);
2025
2026         /* Complete all tx descs not waiting for sends to complete.
2027          * NB we should be safe from RDMA now that the QP has changed state */
2028
2029         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2030         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2031         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2032         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2033         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2034
2035         kiblnd_handle_early_rxs(conn);
2036 }
2037
2038 static void
2039 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2040 {
2041         struct list_head zombies = LIST_HEAD_INIT(zombies);
2042         unsigned long   flags;
2043
2044         LASSERT (error != 0);
2045         LASSERT (!in_interrupt());
2046
2047         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2048
2049         if (active) {
2050                 LASSERT (peer->ibp_connecting > 0);
2051                 peer->ibp_connecting--;
2052         } else {
2053                 LASSERT (peer->ibp_accepting > 0);
2054                 peer->ibp_accepting--;
2055         }
2056
2057         if (peer->ibp_connecting != 0 ||
2058             peer->ibp_accepting != 0) {
2059                 /* another connection attempt under way... */
2060                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2061                                         flags);
2062                 return;
2063         }
2064
2065         if (list_empty(&peer->ibp_conns)) {
2066                 /* Take peer's blocked transmits to complete with error */
2067                 list_add(&zombies, &peer->ibp_tx_queue);
2068                 list_del_init(&peer->ibp_tx_queue);
2069
2070                 if (kiblnd_peer_active(peer))
2071                         kiblnd_unlink_peer_locked(peer);
2072
2073                 peer->ibp_error = error;
2074         } else {
2075                 /* Can't have blocked transmits if there are connections */
2076                 LASSERT(list_empty(&peer->ibp_tx_queue));
2077         }
2078
2079         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2080
2081         kiblnd_peer_notify(peer);
2082
2083         if (list_empty(&zombies))
2084                 return;
2085
2086         CNETERR("Deleting messages for %s: connection failed\n",
2087                 libcfs_nid2str(peer->ibp_nid));
2088
2089         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2090 }
2091
2092 static void
2093 kiblnd_connreq_done(kib_conn_t *conn, int status)
2094 {
2095         kib_peer_t       *peer = conn->ibc_peer;
2096         kib_tx_t         *tx;
2097         struct list_head txs;
2098         unsigned long    flags;
2099         int              active;
2100
2101         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2102
2103         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2104                libcfs_nid2str(peer->ibp_nid), active,
2105                conn->ibc_version, status);
2106
2107         LASSERT (!in_interrupt());
2108         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2109                   peer->ibp_connecting > 0) ||
2110                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2111                   peer->ibp_accepting > 0));
2112
2113         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2114         conn->ibc_connvars = NULL;
2115
2116         if (status != 0) {
2117                 /* failed to establish connection */
2118                 kiblnd_peer_connect_failed(peer, active, status);
2119                 kiblnd_finalise_conn(conn);
2120                 return;
2121         }
2122
2123         /* connection established */
2124         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2125
2126         conn->ibc_last_send = jiffies;
2127         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2128         kiblnd_peer_alive(peer);
2129
2130         /* Add conn to peer's list and nuke any dangling conns from a different
2131          * peer instance... */
2132         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2133         list_add(&conn->ibc_list, &peer->ibp_conns);
2134         if (active)
2135                 peer->ibp_connecting--;
2136         else
2137                 peer->ibp_accepting--;
2138
2139         if (peer->ibp_version == 0) {
2140                 peer->ibp_version     = conn->ibc_version;
2141                 peer->ibp_incarnation = conn->ibc_incarnation;
2142         }
2143
2144         if (peer->ibp_version     != conn->ibc_version ||
2145             peer->ibp_incarnation != conn->ibc_incarnation) {
2146                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2147                                                 conn->ibc_incarnation);
2148                 peer->ibp_version     = conn->ibc_version;
2149                 peer->ibp_incarnation = conn->ibc_incarnation;
2150         }
2151
2152         /* grab pending txs while I have the lock */
2153         list_add(&txs, &peer->ibp_tx_queue);
2154         list_del_init(&peer->ibp_tx_queue);
2155
2156         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2157             conn->ibc_comms_error != 0) {       /* error has happened already */
2158                 lnet_ni_t *ni = peer->ibp_ni;
2159
2160                 /* start to shut down connection */
2161                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2162                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2163
2164                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2165
2166                 return;
2167         }
2168
2169         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2170
2171         /* Schedule blocked txs */
2172         spin_lock(&conn->ibc_lock);
2173         while (!list_empty(&txs)) {
2174                 tx = list_entry(txs.next, kib_tx_t, tx_list);
2175                 list_del(&tx->tx_list);
2176
2177                 kiblnd_queue_tx_locked(tx, conn);
2178         }
2179         spin_unlock(&conn->ibc_lock);
2180
2181         kiblnd_check_sends(conn);
2182
2183         /* schedule blocked rxs */
2184         kiblnd_handle_early_rxs(conn);
2185 }
2186
2187 static void
2188 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2189 {
2190         int          rc;
2191
2192         rc = rdma_reject(cmid, rej, sizeof(*rej));
2193
2194         if (rc != 0)
2195                 CWARN("Error %d sending reject\n", rc);
2196 }
2197
2198 static int
2199 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2200 {
2201         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2202         kib_msg_t             *reqmsg = priv;
2203         kib_msg_t             *ackmsg;
2204         kib_dev_t             *ibdev;
2205         kib_peer_t            *peer;
2206         kib_peer_t            *peer2;
2207         kib_conn_t            *conn;
2208         lnet_ni_t             *ni  = NULL;
2209         kib_net_t             *net = NULL;
2210         lnet_nid_t             nid;
2211         struct rdma_conn_param cp;
2212         kib_rej_t              rej;
2213         int                    version = IBLND_MSG_VERSION;
2214         unsigned long          flags;
2215         int                    rc;
2216         struct sockaddr_in    *peer_addr;
2217         LASSERT (!in_interrupt());
2218
2219         /* cmid inherits 'context' from the corresponding listener id */
2220         ibdev = (kib_dev_t *)cmid->context;
2221         LASSERT (ibdev != NULL);
2222
2223         memset(&rej, 0, sizeof(rej));
2224         rej.ibr_magic                = IBLND_MSG_MAGIC;
2225         rej.ibr_why                  = IBLND_REJECT_FATAL;
2226         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2227
2228         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2229         if (*kiblnd_tunables.kib_require_priv_port &&
2230             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2231                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2232                 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2233                        &ip, ntohs(peer_addr->sin_port));
2234                 goto failed;
2235         }
2236
2237         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2238                 CERROR("Short connection request\n");
2239                 goto failed;
2240         }
2241
2242         /* Future protocol version compatibility support!  If the
2243          * o2iblnd-specific protocol changes, or when LNET unifies
2244          * protocols over all LNDs, the initial connection will
2245          * negotiate a protocol version.  I trap this here to avoid
2246          * console errors; the reject tells the peer which protocol I
2247          * speak. */
2248         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2249             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2250                 goto failed;
2251         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2252             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2253             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2254                 goto failed;
2255         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2256             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2257             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2258                 goto failed;
2259
2260         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2261         if (rc != 0) {
2262                 CERROR("Can't parse connection request: %d\n", rc);
2263                 goto failed;
2264         }
2265
2266         nid = reqmsg->ibm_srcnid;
2267         ni  = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2268
2269         if (ni != NULL) {
2270                 net = (kib_net_t *)ni->ni_data;
2271                 rej.ibr_incarnation = net->ibn_incarnation;
2272         }
2273
2274         if (ni == NULL ||                         /* no matching net */
2275             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2276             net->ibn_dev != ibdev) {              /* wrong device */
2277                 CERROR("Can't accept %s on %s (%s:%d:%pI4h): "
2278                        "bad dst nid %s\n", libcfs_nid2str(nid),
2279                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2280                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2281                         &ibdev->ibd_ifip,
2282                        libcfs_nid2str(reqmsg->ibm_dstnid));
2283
2284                 goto failed;
2285         }
2286
2287        /* check time stamp as soon as possible */
2288         if (reqmsg->ibm_dststamp != 0 &&
2289             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2290                 CWARN("Stale connection request\n");
2291                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2292                 goto failed;
2293         }
2294
2295         /* I can accept peer's version */
2296         version = reqmsg->ibm_version;
2297
2298         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2299                 CERROR("Unexpected connreq msg type: %x from %s\n",
2300                        reqmsg->ibm_type, libcfs_nid2str(nid));
2301                 goto failed;
2302         }
2303
2304         if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2305             IBLND_MSG_QUEUE_SIZE(version)) {
2306                 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2307                        libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2308                        IBLND_MSG_QUEUE_SIZE(version));
2309
2310                 if (version == IBLND_MSG_VERSION)
2311                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2312
2313                 goto failed;
2314         }
2315
2316         if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2317             IBLND_RDMA_FRAGS(version)) {
2318                 CERROR("Can't accept %s(version %x): "
2319                        "incompatible max_frags %d (%d wanted)\n",
2320                        libcfs_nid2str(nid), version,
2321                        reqmsg->ibm_u.connparams.ibcp_max_frags,
2322                        IBLND_RDMA_FRAGS(version));
2323
2324                 if (version == IBLND_MSG_VERSION)
2325                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2326
2327                 goto failed;
2328
2329         }
2330
2331         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2332                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2333                        libcfs_nid2str(nid),
2334                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2335                        IBLND_MSG_SIZE);
2336                 goto failed;
2337         }
2338
2339         /* assume 'nid' is a new peer; create  */
2340         rc = kiblnd_create_peer(ni, &peer, nid);
2341         if (rc != 0) {
2342                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2343                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2344                 goto failed;
2345         }
2346
2347         write_lock_irqsave(g_lock, flags);
2348
2349         peer2 = kiblnd_find_peer_locked(nid);
2350         if (peer2 != NULL) {
2351                 if (peer2->ibp_version == 0) {
2352                         peer2->ibp_version     = version;
2353                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2354                 }
2355
2356                 /* not the guy I've talked with */
2357                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2358                     peer2->ibp_version     != version) {
2359                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2360                         write_unlock_irqrestore(g_lock, flags);
2361
2362                         CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2363                               libcfs_nid2str(nid), peer2->ibp_version, version);
2364
2365                         kiblnd_peer_decref(peer);
2366                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2367                         goto failed;
2368                 }
2369
2370                 /* tie-break connection race in favour of the higher NID */
2371                 if (peer2->ibp_connecting != 0 &&
2372                     nid < ni->ni_nid) {
2373                         write_unlock_irqrestore(g_lock, flags);
2374
2375                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2376
2377                         kiblnd_peer_decref(peer);
2378                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2379                         goto failed;
2380                 }
2381
2382                 peer2->ibp_accepting++;
2383                 kiblnd_peer_addref(peer2);
2384
2385                 write_unlock_irqrestore(g_lock, flags);
2386                 kiblnd_peer_decref(peer);
2387                 peer = peer2;
2388         } else {
2389                 /* Brand new peer */
2390                 LASSERT (peer->ibp_accepting == 0);
2391                 LASSERT (peer->ibp_version == 0 &&
2392                          peer->ibp_incarnation == 0);
2393
2394                 peer->ibp_accepting   = 1;
2395                 peer->ibp_version     = version;
2396                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2397
2398                 /* I have a ref on ni that prevents it being shutdown */
2399                 LASSERT (net->ibn_shutdown == 0);
2400
2401                 kiblnd_peer_addref(peer);
2402                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2403
2404                 write_unlock_irqrestore(g_lock, flags);
2405         }
2406
2407         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2408         if (conn == NULL) {
2409                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2410                 kiblnd_peer_decref(peer);
2411                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2412                 goto failed;
2413         }
2414
2415         /* conn now "owns" cmid, so I return success from here on to ensure the
2416          * CM callback doesn't destroy cmid. */
2417
2418         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2419         conn->ibc_credits          = IBLND_MSG_QUEUE_SIZE(version);
2420         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2421         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2422                  <= IBLND_RX_MSGS(version));
2423
2424         ackmsg = &conn->ibc_connvars->cv_msg;
2425         memset(ackmsg, 0, sizeof(*ackmsg));
2426
2427         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2428                         sizeof(ackmsg->ibm_u.connparams));
2429         ackmsg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2430         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2431         ackmsg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2432
2433         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2434
2435         memset(&cp, 0, sizeof(cp));
2436         cp.private_data        = ackmsg;
2437         cp.private_data_len    = ackmsg->ibm_nob;
2438         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2439         cp.initiator_depth     = 0;
2440         cp.flow_control        = 1;
2441         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2442         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2443
2444         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2445
2446         rc = rdma_accept(cmid, &cp);
2447         if (rc != 0) {
2448                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2449                 rej.ibr_version = version;
2450                 rej.ibr_why     = IBLND_REJECT_FATAL;
2451
2452                 kiblnd_reject(cmid, &rej);
2453                 kiblnd_connreq_done(conn, rc);
2454                 kiblnd_conn_decref(conn);
2455         }
2456
2457         lnet_ni_decref(ni);
2458         return 0;
2459
2460  failed:
2461         if (ni != NULL)
2462                 lnet_ni_decref(ni);
2463
2464         rej.ibr_version = version;
2465         rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2466         rej.ibr_cp.ibcp_max_frags   = IBLND_RDMA_FRAGS(version);
2467         kiblnd_reject(cmid, &rej);
2468
2469         return -ECONNREFUSED;
2470 }
2471
2472 static void
2473 kiblnd_reconnect (kib_conn_t *conn, int version,
2474                   __u64 incarnation, int why, kib_connparams_t *cp)
2475 {
2476         kib_peer_t      *peer = conn->ibc_peer;
2477         char            *reason;
2478         int              retry_now = 0;
2479         unsigned long    flags;
2480
2481         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2482         LASSERT (peer->ibp_connecting > 0);     /* 'conn' at least */
2483
2484         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2485
2486         /* retry connection if it's still needed and no other connection
2487          * attempts (active or passive) are in progress
2488          * NB: reconnect is still needed even when ibp_tx_queue is
2489          * empty if ibp_version != version because reconnect may be
2490          * initiated by kiblnd_query() */
2491         if ((!list_empty(&peer->ibp_tx_queue) ||
2492              peer->ibp_version != version) &&
2493             peer->ibp_connecting == 1 &&
2494             peer->ibp_accepting == 0) {
2495                 if (why == IBLND_REJECT_CONN_RACE) {
2496                         /* don't reconnect immediately, intensive reconnecting
2497                          * may consume a lot of memory. kiblnd_destroy_conn
2498                          * will reconnect after releasing all resources of
2499                          * this connection */
2500                         conn->ibc_conn_race = 1;
2501                 } else {
2502                         retry_now = 1;
2503                 }
2504                 peer->ibp_connecting++;
2505
2506                 peer->ibp_version     = version;
2507                 peer->ibp_incarnation = incarnation;
2508         }
2509
2510         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2511
2512         if (!retry_now)
2513                 return;
2514
2515         switch (why) {
2516         default:
2517                 reason = "Unknown";
2518                 break;
2519
2520         case IBLND_REJECT_CONN_STALE:
2521                 reason = "stale";
2522                 break;
2523
2524         case IBLND_REJECT_CONN_UNCOMPAT:
2525                 reason = "version negotiation";
2526                 break;
2527         }
2528
2529         CNETERR("%s: retrying (%s), %x, %x, "
2530                 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2531                 libcfs_nid2str(peer->ibp_nid),
2532                 reason, IBLND_MSG_VERSION, version,
2533                 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2534                 cp != NULL? cp->ibcp_max_frags   : IBLND_RDMA_FRAGS(version),
2535                 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2536
2537         kiblnd_connect_peer(peer);
2538 }
2539
2540 static void
2541 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2542 {
2543         kib_peer_t    *peer = conn->ibc_peer;
2544
2545         LASSERT (!in_interrupt());
2546         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2547
2548         switch (reason) {
2549         case IB_CM_REJ_STALE_CONN:
2550                 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2551                                  IBLND_REJECT_CONN_STALE, NULL);
2552                 break;
2553
2554         case IB_CM_REJ_INVALID_SERVICE_ID:
2555                 CNETERR("%s rejected: no listener at %d\n",
2556                         libcfs_nid2str(peer->ibp_nid),
2557                         *kiblnd_tunables.kib_service);
2558                 break;
2559
2560         case IB_CM_REJ_CONSUMER_DEFINED:
2561                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2562                         kib_rej_t        *rej         = priv;
2563                         kib_connparams_t *cp          = NULL;
2564                         int               flip        = 0;
2565                         __u64             incarnation = -1;
2566
2567                         /* NB. default incarnation is -1 because:
2568                          * a) V1 will ignore dst incarnation in connreq.
2569                          * b) V2 will provide incarnation while rejecting me,
2570                          *    -1 will be overwrote.
2571                          *
2572                          * if I try to connect to a V1 peer with V2 protocol,
2573                          * it rejected me then upgrade to V2, I have no idea
2574                          * about the upgrading and try to reconnect with V1,
2575                          * in this case upgraded V2 can find out I'm trying to
2576                          * talk to the old guy and reject me(incarnation is -1).
2577                          */
2578
2579                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2580                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2581                                 __swab32s(&rej->ibr_magic);
2582                                 __swab16s(&rej->ibr_version);
2583                                 flip = 1;
2584                         }
2585
2586                         if (priv_nob >= sizeof(kib_rej_t) &&
2587                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2588                                 /* priv_nob is always 148 in current version
2589                                  * of OFED, so we still need to check version.
2590                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2591                                 cp = &rej->ibr_cp;
2592
2593                                 if (flip) {
2594                                         __swab64s(&rej->ibr_incarnation);
2595                                         __swab16s(&cp->ibcp_queue_depth);
2596                                         __swab16s(&cp->ibcp_max_frags);
2597                                         __swab32s(&cp->ibcp_max_msg_size);
2598                                 }
2599
2600                                 incarnation = rej->ibr_incarnation;
2601                         }
2602
2603                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2604                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2605                                 CERROR("%s rejected: consumer defined fatal error\n",
2606                                        libcfs_nid2str(peer->ibp_nid));
2607                                 break;
2608                         }
2609
2610                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2611                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2612                                 CERROR("%s rejected: o2iblnd version %x error\n",
2613                                        libcfs_nid2str(peer->ibp_nid),
2614                                        rej->ibr_version);
2615                                 break;
2616                         }
2617
2618                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2619                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2620                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2621                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2622
2623                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2624                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2625                         }
2626
2627                         switch (rej->ibr_why) {
2628                         case IBLND_REJECT_CONN_RACE:
2629                         case IBLND_REJECT_CONN_STALE:
2630                         case IBLND_REJECT_CONN_UNCOMPAT:
2631                                 kiblnd_reconnect(conn, rej->ibr_version,
2632                                                  incarnation, rej->ibr_why, cp);
2633                                 break;
2634
2635                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2636                                 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2637                                        libcfs_nid2str(peer->ibp_nid),
2638                                        cp != NULL ? cp->ibcp_queue_depth :
2639                                        IBLND_MSG_QUEUE_SIZE(rej->ibr_version),
2640                                        IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2641                                 break;
2642
2643                         case IBLND_REJECT_RDMA_FRAGS:
2644                                 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2645                                        libcfs_nid2str(peer->ibp_nid),
2646                                        cp != NULL ? cp->ibcp_max_frags :
2647                                        IBLND_RDMA_FRAGS(rej->ibr_version),
2648                                        IBLND_RDMA_FRAGS(conn->ibc_version));
2649                                 break;
2650
2651                         case IBLND_REJECT_NO_RESOURCES:
2652                                 CERROR("%s rejected: o2iblnd no resources\n",
2653                                        libcfs_nid2str(peer->ibp_nid));
2654                                 break;
2655
2656                         case IBLND_REJECT_FATAL:
2657                                 CERROR("%s rejected: o2iblnd fatal error\n",
2658                                        libcfs_nid2str(peer->ibp_nid));
2659                                 break;
2660
2661                         default:
2662                                 CERROR("%s rejected: o2iblnd reason %d\n",
2663                                        libcfs_nid2str(peer->ibp_nid),
2664                                        rej->ibr_why);
2665                                 break;
2666                         }
2667                         break;
2668                 }
2669                 /* fall through */
2670         default:
2671                 CNETERR("%s rejected: reason %d, size %d\n",
2672                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2673                 break;
2674         }
2675
2676         kiblnd_connreq_done(conn, -ECONNREFUSED);
2677 }
2678
2679 static void
2680 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2681 {
2682         kib_peer_t    *peer = conn->ibc_peer;
2683         lnet_ni_t     *ni   = peer->ibp_ni;
2684         kib_net_t     *net  = ni->ni_data;
2685         kib_msg_t     *msg  = priv;
2686         int            ver  = conn->ibc_version;
2687         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2688         unsigned long  flags;
2689
2690         LASSERT (net != NULL);
2691
2692         if (rc != 0) {
2693                 CERROR("Can't unpack connack from %s: %d\n",
2694                        libcfs_nid2str(peer->ibp_nid), rc);
2695                 goto failed;
2696         }
2697
2698         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2699                 CERROR("Unexpected message %d from %s\n",
2700                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2701                 rc = -EPROTO;
2702                 goto failed;
2703         }
2704
2705         if (ver != msg->ibm_version) {
2706                 CERROR("%s replied version %x is different with "
2707                        "requested version %x\n",
2708                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2709                 rc = -EPROTO;
2710                 goto failed;
2711         }
2712
2713         if (msg->ibm_u.connparams.ibcp_queue_depth !=
2714             IBLND_MSG_QUEUE_SIZE(ver)) {
2715                 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2716                        libcfs_nid2str(peer->ibp_nid),
2717                        msg->ibm_u.connparams.ibcp_queue_depth,
2718                        IBLND_MSG_QUEUE_SIZE(ver));
2719                 rc = -EPROTO;
2720                 goto failed;
2721         }
2722
2723         if (msg->ibm_u.connparams.ibcp_max_frags !=
2724             IBLND_RDMA_FRAGS(ver)) {
2725                 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2726                        libcfs_nid2str(peer->ibp_nid),
2727                        msg->ibm_u.connparams.ibcp_max_frags,
2728                        IBLND_RDMA_FRAGS(ver));
2729                 rc = -EPROTO;
2730                 goto failed;
2731         }
2732
2733         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2734                 CERROR("%s max message size %d too big (%d max)\n",
2735                        libcfs_nid2str(peer->ibp_nid),
2736                        msg->ibm_u.connparams.ibcp_max_msg_size,
2737                        IBLND_MSG_SIZE);
2738                 rc = -EPROTO;
2739                 goto failed;
2740         }
2741
2742         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2743         if (msg->ibm_dstnid == ni->ni_nid &&
2744             msg->ibm_dststamp == net->ibn_incarnation)
2745                 rc = 0;
2746         else
2747                 rc = -ESTALE;
2748         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2749
2750         if (rc != 0) {
2751                 CERROR("Bad connection reply from %s, rc = %d, "
2752                        "version: %x max_frags: %d\n",
2753                        libcfs_nid2str(peer->ibp_nid), rc,
2754                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2755                 goto failed;
2756         }
2757
2758         conn->ibc_incarnation      = msg->ibm_srcstamp;
2759         conn->ibc_credits          =
2760         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2761         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2762                  <= IBLND_RX_MSGS(ver));
2763
2764         kiblnd_connreq_done(conn, 0);
2765         return;
2766
2767  failed:
2768         /* NB My QP has already established itself, so I handle anything going
2769          * wrong here by setting ibc_comms_error.
2770          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2771          * immediately tears it down. */
2772
2773         LASSERT (rc != 0);
2774         conn->ibc_comms_error = rc;
2775         kiblnd_connreq_done(conn, 0);
2776 }
2777
2778 static int
2779 kiblnd_active_connect (struct rdma_cm_id *cmid)
2780 {
2781         kib_peer_t              *peer = (kib_peer_t *)cmid->context;
2782         kib_conn_t              *conn;
2783         kib_msg_t               *msg;
2784         struct rdma_conn_param   cp;
2785         int                      version;
2786         __u64                    incarnation;
2787         unsigned long            flags;
2788         int                      rc;
2789
2790         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2791
2792         incarnation = peer->ibp_incarnation;
2793         version     = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2794                                                  peer->ibp_version;
2795
2796         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2797
2798         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2799         if (conn == NULL) {
2800                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2801                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2802                 return -ENOMEM;
2803         }
2804
2805         /* conn "owns" cmid now, so I return success from here on to ensure the
2806          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2807          * on peer */
2808
2809         msg = &conn->ibc_connvars->cv_msg;
2810
2811         memset(msg, 0, sizeof(*msg));
2812         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2813         msg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2814         msg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2815         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2816
2817         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2818                         0, peer->ibp_nid, incarnation);
2819
2820         memset(&cp, 0, sizeof(cp));
2821         cp.private_data        = msg;
2822         cp.private_data_len    = msg->ibm_nob;
2823         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2824         cp.initiator_depth     = 0;
2825         cp.flow_control        = 1;
2826         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2827         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2828
2829         LASSERT(cmid->context == (void *)conn);
2830         LASSERT(conn->ibc_cmid == cmid);
2831
2832         rc = rdma_connect(cmid, &cp);
2833         if (rc != 0) {
2834                 CERROR("Can't connect to %s: %d\n",
2835                        libcfs_nid2str(peer->ibp_nid), rc);
2836                 kiblnd_connreq_done(conn, rc);
2837                 kiblnd_conn_decref(conn);
2838         }
2839
2840         return 0;
2841 }
2842
2843 int
2844 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2845 {
2846         kib_peer_t  *peer;
2847         kib_conn_t  *conn;
2848         int          rc;
2849
2850         switch (event->event) {
2851         default:
2852                 CERROR("Unexpected event: %d, status: %d\n",
2853                        event->event, event->status);
2854                 LBUG();
2855
2856         case RDMA_CM_EVENT_CONNECT_REQUEST:
2857                 /* destroy cmid on failure */
2858                 rc = kiblnd_passive_connect(cmid,
2859                                             (void *)KIBLND_CONN_PARAM(event),
2860                                             KIBLND_CONN_PARAM_LEN(event));
2861                 CDEBUG(D_NET, "connreq: %d\n", rc);
2862                 return rc;
2863
2864         case RDMA_CM_EVENT_ADDR_ERROR:
2865                 peer = (kib_peer_t *)cmid->context;
2866                 CNETERR("%s: ADDR ERROR %d\n",
2867                        libcfs_nid2str(peer->ibp_nid), event->status);
2868                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2869                 kiblnd_peer_decref(peer);
2870                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
2871
2872         case RDMA_CM_EVENT_ADDR_RESOLVED:
2873                 peer = (kib_peer_t *)cmid->context;
2874
2875                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2876                        libcfs_nid2str(peer->ibp_nid), event->status);
2877
2878                 if (event->status != 0) {
2879                         CNETERR("Can't resolve address for %s: %d\n",
2880                                 libcfs_nid2str(peer->ibp_nid), event->status);
2881                         rc = event->status;
2882                 } else {
2883                         rc = rdma_resolve_route(
2884                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2885                         if (rc == 0)
2886                                 return 0;
2887                         /* Can't initiate route resolution */
2888                         CERROR("Can't resolve route for %s: %d\n",
2889                                libcfs_nid2str(peer->ibp_nid), rc);
2890                 }
2891                 kiblnd_peer_connect_failed(peer, 1, rc);
2892                 kiblnd_peer_decref(peer);
2893                 return rc;                      /* rc != 0 destroys cmid */
2894
2895         case RDMA_CM_EVENT_ROUTE_ERROR:
2896                 peer = (kib_peer_t *)cmid->context;
2897                 CNETERR("%s: ROUTE ERROR %d\n",
2898                         libcfs_nid2str(peer->ibp_nid), event->status);
2899                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2900                 kiblnd_peer_decref(peer);
2901                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
2902
2903         case RDMA_CM_EVENT_ROUTE_RESOLVED:
2904                 peer = (kib_peer_t *)cmid->context;
2905                 CDEBUG(D_NET,"%s Route resolved: %d\n",
2906                        libcfs_nid2str(peer->ibp_nid), event->status);
2907
2908                 if (event->status == 0)
2909                         return kiblnd_active_connect(cmid);
2910
2911                 CNETERR("Can't resolve route for %s: %d\n",
2912                        libcfs_nid2str(peer->ibp_nid), event->status);
2913                 kiblnd_peer_connect_failed(peer, 1, event->status);
2914                 kiblnd_peer_decref(peer);
2915                 return event->status;           /* rc != 0 destroys cmid */
2916
2917         case RDMA_CM_EVENT_UNREACHABLE:
2918                 conn = (kib_conn_t *)cmid->context;
2919                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2920                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2921                 CNETERR("%s: UNREACHABLE %d\n",
2922                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2923                 kiblnd_connreq_done(conn, -ENETDOWN);
2924                 kiblnd_conn_decref(conn);
2925                 return 0;
2926
2927         case RDMA_CM_EVENT_CONNECT_ERROR:
2928                 conn = (kib_conn_t *)cmid->context;
2929                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2930                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2931                 CNETERR("%s: CONNECT ERROR %d\n",
2932                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2933                 kiblnd_connreq_done(conn, -ENOTCONN);
2934                 kiblnd_conn_decref(conn);
2935                 return 0;
2936
2937         case RDMA_CM_EVENT_REJECTED:
2938                 conn = (kib_conn_t *)cmid->context;
2939                 switch (conn->ibc_state) {
2940                 default:
2941                         LBUG();
2942
2943                 case IBLND_CONN_PASSIVE_WAIT:
2944                         CERROR ("%s: REJECTED %d\n",
2945                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2946                                 event->status);
2947                         kiblnd_connreq_done(conn, -ECONNRESET);
2948                         break;
2949
2950                 case IBLND_CONN_ACTIVE_CONNECT:
2951                         kiblnd_rejected(conn, event->status,
2952                                         (void *)KIBLND_CONN_PARAM(event),
2953                                         KIBLND_CONN_PARAM_LEN(event));
2954                         break;
2955                 }
2956                 kiblnd_conn_decref(conn);
2957                 return 0;
2958
2959         case RDMA_CM_EVENT_ESTABLISHED:
2960                 conn = (kib_conn_t *)cmid->context;
2961                 switch (conn->ibc_state) {
2962                 default:
2963                         LBUG();
2964
2965                 case IBLND_CONN_PASSIVE_WAIT:
2966                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2967                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2968                         kiblnd_connreq_done(conn, 0);
2969                         break;
2970
2971                 case IBLND_CONN_ACTIVE_CONNECT:
2972                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2973                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2974                         kiblnd_check_connreply(conn,
2975                                                (void *)KIBLND_CONN_PARAM(event),
2976                                                KIBLND_CONN_PARAM_LEN(event));
2977                         break;
2978                 }
2979                 /* net keeps its ref on conn! */
2980                 return 0;
2981
2982         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2983                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2984                 return 0;
2985
2986         case RDMA_CM_EVENT_DISCONNECTED:
2987                 conn = (kib_conn_t *)cmid->context;
2988                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2989                         CERROR("%s DISCONNECTED\n",
2990                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2991                         kiblnd_connreq_done(conn, -ECONNRESET);
2992                 } else {
2993                         kiblnd_close_conn(conn, 0);
2994                 }
2995                 kiblnd_conn_decref(conn);
2996                 cmid->context = NULL;
2997                 return 0;
2998
2999         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3000                 LCONSOLE_ERROR_MSG(0x131,
3001                                    "Received notification of device removal\n"
3002                                    "Please shutdown LNET to allow this to proceed\n");
3003                 /* Can't remove network from underneath LNET for now, so I have
3004                  * to ignore this */
3005                 return 0;
3006
3007         case RDMA_CM_EVENT_ADDR_CHANGE:
3008                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3009                 return 0;
3010         }
3011 }
3012
3013 static int
3014 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3015 {
3016         kib_tx_t         *tx;
3017         struct list_head *ttmp;
3018
3019         list_for_each(ttmp, txs) {
3020                 tx = list_entry(ttmp, kib_tx_t, tx_list);
3021
3022                 if (txs != &conn->ibc_active_txs) {
3023                         LASSERT(tx->tx_queued);
3024                 } else {
3025                         LASSERT(!tx->tx_queued);
3026                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3027                 }
3028
3029                 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3030                         CERROR("Timed out tx: %s, %lu seconds\n",
3031                                kiblnd_queue2str(conn, txs),
3032                                cfs_duration_sec(jiffies - tx->tx_deadline));
3033                         return 1;
3034                 }
3035         }
3036
3037         return 0;
3038 }
3039
3040 static int
3041 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3042 {
3043         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3044                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3045                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3046                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3047                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3048 }
3049
3050 static void
3051 kiblnd_check_conns (int idx)
3052 {
3053         struct list_head  closes = LIST_HEAD_INIT(closes);
3054         struct list_head  checksends = LIST_HEAD_INIT(checksends);
3055         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3056         struct list_head *ptmp;
3057         kib_peer_t       *peer;
3058         kib_conn_t       *conn;
3059         struct list_head *ctmp;
3060         unsigned long     flags;
3061
3062         /* NB. We expect to have a look at all the peers and not find any
3063          * RDMAs to time out, so we just use a shared lock while we
3064          * take a look... */
3065         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3066
3067         list_for_each(ptmp, peers) {
3068                 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3069
3070                 list_for_each(ctmp, &peer->ibp_conns) {
3071                         int timedout;
3072                         int sendnoop;
3073
3074                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
3075
3076                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3077
3078                         spin_lock(&conn->ibc_lock);
3079
3080                         sendnoop = kiblnd_need_noop(conn);
3081                         timedout = kiblnd_conn_timed_out_locked(conn);
3082                         if (!sendnoop && !timedout) {
3083                                 spin_unlock(&conn->ibc_lock);
3084                                 continue;
3085                         }
3086
3087                         if (timedout) {
3088                                 CERROR("Timed out RDMA with %s (%lu): "
3089                                        "c: %u, oc: %u, rc: %u\n",
3090                                        libcfs_nid2str(peer->ibp_nid),
3091                                        cfs_duration_sec(cfs_time_current() -
3092                                                         peer->ibp_last_alive),
3093                                        conn->ibc_credits,
3094                                        conn->ibc_outstanding_credits,
3095                                        conn->ibc_reserved_credits);
3096                                 list_add(&conn->ibc_connd_list, &closes);
3097                         } else {
3098                                 list_add(&conn->ibc_connd_list, &checksends);
3099                         }
3100                         /* +ref for 'closes' or 'checksends' */
3101                         kiblnd_conn_addref(conn);
3102
3103                         spin_unlock(&conn->ibc_lock);
3104                 }
3105         }
3106
3107         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3108
3109         /* Handle timeout by closing the whole
3110          * connection. We can only be sure RDMA activity
3111          * has ceased once the QP has been modified. */
3112         while (!list_empty(&closes)) {
3113                 conn = list_entry(closes.next,
3114                                   kib_conn_t, ibc_connd_list);
3115                 list_del(&conn->ibc_connd_list);
3116                 kiblnd_close_conn(conn, -ETIMEDOUT);
3117                 kiblnd_conn_decref(conn);
3118         }
3119
3120         /* In case we have enough credits to return via a
3121          * NOOP, but there were no non-blocking tx descs
3122          * free to do it last time... */
3123         while (!list_empty(&checksends)) {
3124                 conn = list_entry(checksends.next,
3125                                   kib_conn_t, ibc_connd_list);
3126                 list_del(&conn->ibc_connd_list);
3127                 kiblnd_check_sends(conn);
3128                 kiblnd_conn_decref(conn);
3129         }
3130 }
3131
3132 static void
3133 kiblnd_disconnect_conn (kib_conn_t *conn)
3134 {
3135         LASSERT (!in_interrupt());
3136         LASSERT (current == kiblnd_data.kib_connd);
3137         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3138
3139         rdma_disconnect(conn->ibc_cmid);
3140         kiblnd_finalise_conn(conn);
3141
3142         kiblnd_peer_notify(conn->ibc_peer);
3143 }
3144
3145 int
3146 kiblnd_connd (void *arg)
3147 {
3148         wait_queue_t     wait;
3149         unsigned long      flags;
3150         kib_conn_t        *conn;
3151         int                timeout;
3152         int                i;
3153         int                dropped_lock;
3154         int                peer_index = 0;
3155         unsigned long      deadline = jiffies;
3156
3157         cfs_block_allsigs();
3158
3159         init_waitqueue_entry(&wait, current);
3160         kiblnd_data.kib_connd = current;
3161
3162         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3163
3164         while (!kiblnd_data.kib_shutdown) {
3165
3166                 dropped_lock = 0;
3167
3168                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3169                         conn = list_entry(kiblnd_data. \
3170                                               kib_connd_zombies.next,
3171                                               kib_conn_t, ibc_list);
3172                         list_del(&conn->ibc_list);
3173
3174                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3175                                                flags);
3176                         dropped_lock = 1;
3177
3178                         kiblnd_destroy_conn(conn);
3179
3180                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3181                 }
3182
3183                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3184                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3185                                               kib_conn_t, ibc_list);
3186                         list_del(&conn->ibc_list);
3187
3188                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
3189                                                flags);
3190                         dropped_lock = 1;
3191
3192                         kiblnd_disconnect_conn(conn);
3193                         kiblnd_conn_decref(conn);
3194
3195                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3196                 }
3197
3198                 /* careful with the jiffy wrap... */
3199                 timeout = (int)(deadline - jiffies);
3200                 if (timeout <= 0) {
3201                         const int n = 4;
3202                         const int p = 1;
3203                         int       chunk = kiblnd_data.kib_peer_hash_size;
3204
3205                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3206                         dropped_lock = 1;
3207
3208                         /* Time to check for RDMA timeouts on a few more
3209                          * peers: I do checks every 'p' seconds on a
3210                          * proportion of the peer table and I need to check
3211                          * every connection 'n' times within a timeout
3212                          * interval, to ensure I detect a timeout on any
3213                          * connection within (n+1)/n times the timeout
3214                          * interval. */
3215
3216                         if (*kiblnd_tunables.kib_timeout > n * p)
3217                                 chunk = (chunk * n * p) /
3218                                         *kiblnd_tunables.kib_timeout;
3219                         if (chunk == 0)
3220                                 chunk = 1;
3221
3222                         for (i = 0; i < chunk; i++) {
3223                                 kiblnd_check_conns(peer_index);
3224                                 peer_index = (peer_index + 1) %
3225                                              kiblnd_data.kib_peer_hash_size;
3226                         }
3227
3228                         deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3229                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3230                 }
3231
3232                 if (dropped_lock)
3233                         continue;
3234
3235                 /* Nothing to do for 'timeout'  */
3236                 set_current_state(TASK_INTERRUPTIBLE);
3237                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3238                 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3239
3240                 schedule_timeout(timeout);
3241
3242                 set_current_state(TASK_RUNNING);
3243                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3244                 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3245         }
3246
3247         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3248
3249         kiblnd_thread_fini();
3250         return 0;
3251 }
3252
3253 void
3254 kiblnd_qp_event(struct ib_event *event, void *arg)
3255 {
3256         kib_conn_t *conn = arg;
3257
3258         switch (event->event) {
3259         case IB_EVENT_COMM_EST:
3260                 CDEBUG(D_NET, "%s established\n",
3261                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3262                 return;
3263
3264         default:
3265                 CERROR("%s: Async QP event type %d\n",
3266                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3267                 return;
3268         }
3269 }
3270
3271 static void
3272 kiblnd_complete (struct ib_wc *wc)
3273 {
3274         switch (kiblnd_wreqid2type(wc->wr_id)) {
3275         default:
3276                 LBUG();
3277
3278         case IBLND_WID_RDMA:
3279                 /* We only get RDMA completion notification if it fails.  All
3280                  * subsequent work items, including the final SEND will fail
3281                  * too.  However we can't print out any more info about the
3282                  * failing RDMA because 'tx' might be back on the idle list or
3283                  * even reused already if we didn't manage to post all our work
3284                  * items */
3285                 CNETERR("RDMA (tx: %p) failed: %d\n",
3286                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3287                 return;
3288
3289         case IBLND_WID_TX:
3290                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3291                 return;
3292
3293         case IBLND_WID_RX:
3294                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3295                                    wc->byte_len);
3296                 return;
3297         }
3298 }
3299
3300 void
3301 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3302 {
3303         /* NB I'm not allowed to schedule this conn once its refcount has
3304          * reached 0.  Since fundamentally I'm racing with scheduler threads
3305          * consuming my CQ I could be called after all completions have
3306          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3307          * and this CQ is about to be destroyed so I NOOP. */
3308         kib_conn_t              *conn = (kib_conn_t *)arg;
3309         struct kib_sched_info   *sched = conn->ibc_sched;
3310         unsigned long           flags;
3311
3312         LASSERT(cq == conn->ibc_cq);
3313
3314         spin_lock_irqsave(&sched->ibs_lock, flags);
3315
3316         conn->ibc_ready = 1;
3317
3318         if (!conn->ibc_scheduled &&
3319             (conn->ibc_nrx > 0 ||
3320              conn->ibc_nsends_posted > 0)) {
3321                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3322                 conn->ibc_scheduled = 1;
3323                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3324
3325                 if (waitqueue_active(&sched->ibs_waitq))
3326                         wake_up(&sched->ibs_waitq);
3327         }
3328
3329         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3330 }
3331
3332 void
3333 kiblnd_cq_event(struct ib_event *event, void *arg)
3334 {
3335         kib_conn_t *conn = arg;
3336
3337         CERROR("%s: async CQ event type %d\n",
3338                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3339 }
3340
3341 int
3342 kiblnd_scheduler(void *arg)
3343 {
3344         long                    id = (long)arg;
3345         struct kib_sched_info   *sched;
3346         kib_conn_t              *conn;
3347         wait_queue_t            wait;
3348         unsigned long           flags;
3349         struct ib_wc            wc;
3350         int                     did_something;
3351         int                     busy_loops = 0;
3352         int                     rc;
3353
3354         cfs_block_allsigs();
3355
3356         init_waitqueue_entry(&wait, current);
3357
3358         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3359
3360         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3361         if (rc != 0) {
3362                 CWARN("Failed to bind on CPT %d, please verify whether "
3363                       "all CPUs are healthy and reload modules if necessary, "
3364                       "otherwise your system might under risk of low "
3365                       "performance\n", sched->ibs_cpt);
3366         }
3367
3368         spin_lock_irqsave(&sched->ibs_lock, flags);
3369
3370         while (!kiblnd_data.kib_shutdown) {
3371                 if (busy_loops++ >= IBLND_RESCHED) {
3372                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3373
3374                         cond_resched();
3375                         busy_loops = 0;
3376
3377                         spin_lock_irqsave(&sched->ibs_lock, flags);
3378                 }
3379
3380                 did_something = 0;
3381
3382                 if (!list_empty(&sched->ibs_conns)) {
3383                         conn = list_entry(sched->ibs_conns.next,
3384                                               kib_conn_t, ibc_sched_list);
3385                         /* take over kib_sched_conns' ref on conn... */
3386                         LASSERT(conn->ibc_scheduled);
3387                         list_del(&conn->ibc_sched_list);
3388                         conn->ibc_ready = 0;
3389
3390                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3391
3392                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3393                         if (rc == 0) {
3394                                 rc = ib_req_notify_cq(conn->ibc_cq,
3395                                                       IB_CQ_NEXT_COMP);
3396                                 if (rc < 0) {
3397                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3398                                               "closing connection\n",
3399                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3400                                         kiblnd_close_conn(conn, -EIO);
3401                                         kiblnd_conn_decref(conn);
3402                                         spin_lock_irqsave(&sched->ibs_lock,
3403                                                               flags);
3404                                         continue;
3405                                 }
3406
3407                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3408                         }
3409
3410                         if (rc < 0) {
3411                                 CWARN("%s: ib_poll_cq failed: %d, "
3412                                       "closing connection\n",
3413                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3414                                       rc);
3415                                 kiblnd_close_conn(conn, -EIO);
3416                                 kiblnd_conn_decref(conn);
3417                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3418                                 continue;
3419                         }
3420
3421                         spin_lock_irqsave(&sched->ibs_lock, flags);
3422
3423                         if (rc != 0 || conn->ibc_ready) {
3424                                 /* There may be another completion waiting; get
3425                                  * another scheduler to check while I handle
3426                                  * this one... */
3427                                 /* +1 ref for sched_conns */
3428                                 kiblnd_conn_addref(conn);
3429                                 list_add_tail(&conn->ibc_sched_list,
3430                                                   &sched->ibs_conns);
3431                                 if (waitqueue_active(&sched->ibs_waitq))
3432                                         wake_up(&sched->ibs_waitq);
3433                         } else {
3434                                 conn->ibc_scheduled = 0;
3435                         }
3436
3437                         if (rc != 0) {
3438                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3439                                 kiblnd_complete(&wc);
3440
3441                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3442                         }
3443
3444                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3445                         did_something = 1;
3446                 }
3447
3448                 if (did_something)
3449                         continue;
3450
3451                 set_current_state(TASK_INTERRUPTIBLE);
3452                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3453                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3454
3455                 schedule();
3456                 busy_loops = 0;
3457
3458                 remove_wait_queue(&sched->ibs_waitq, &wait);
3459                 set_current_state(TASK_RUNNING);
3460                 spin_lock_irqsave(&sched->ibs_lock, flags);
3461         }
3462
3463         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3464
3465         kiblnd_thread_fini();
3466         return 0;
3467 }
3468
3469 int
3470 kiblnd_failover_thread(void *arg)
3471 {
3472         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
3473         kib_dev_t               *dev;
3474         wait_queue_t            wait;
3475         unsigned long           flags;
3476         int                     rc;
3477
3478         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3479
3480         cfs_block_allsigs();
3481
3482         init_waitqueue_entry(&wait, current);
3483         write_lock_irqsave(glock, flags);
3484
3485         while (!kiblnd_data.kib_shutdown) {
3486                 int     do_failover = 0;
3487                 int     long_sleep;
3488
3489                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3490                                     ibd_fail_list) {
3491                         if (cfs_time_before(cfs_time_current(),
3492                                             dev->ibd_next_failover))
3493                                 continue;
3494                         do_failover = 1;
3495                         break;
3496                 }
3497
3498                 if (do_failover) {
3499                         list_del_init(&dev->ibd_fail_list);
3500                         dev->ibd_failover = 1;
3501                         write_unlock_irqrestore(glock, flags);
3502
3503                         rc = kiblnd_dev_failover(dev);
3504
3505                         write_lock_irqsave(glock, flags);
3506
3507                         LASSERT (dev->ibd_failover);
3508                         dev->ibd_failover = 0;
3509                         if (rc >= 0) { /* Device is OK or failover succeed */
3510                                 dev->ibd_next_failover = cfs_time_shift(3);
3511                                 continue;
3512                         }
3513
3514                         /* failed to failover, retry later */
3515                         dev->ibd_next_failover =
3516                                 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3517                         if (kiblnd_dev_can_failover(dev)) {
3518                                 list_add_tail(&dev->ibd_fail_list,
3519                                               &kiblnd_data.kib_failed_devs);
3520                         }
3521
3522                         continue;
3523                 }
3524
3525                 /* long sleep if no more pending failover */
3526                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3527
3528                 set_current_state(TASK_INTERRUPTIBLE);
3529                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3530                 write_unlock_irqrestore(glock, flags);
3531
3532                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3533                                                    cfs_time_seconds(1));
3534                 set_current_state(TASK_RUNNING);
3535                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3536                 write_lock_irqsave(glock, flags);
3537
3538                 if (!long_sleep || rc != 0)
3539                         continue;
3540
3541                 /* have a long sleep, routine check all active devices,
3542                  * we need checking like this because if there is not active
3543                  * connection on the dev and no SEND from local, we may listen
3544                  * on wrong HCA for ever while there is a bonding failover */
3545                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3546                         if (kiblnd_dev_can_failover(dev)) {
3547                                 list_add_tail(&dev->ibd_fail_list,
3548                                               &kiblnd_data.kib_failed_devs);
3549                         }
3550                 }
3551         }
3552
3553         write_unlock_irqrestore(glock, flags);
3554
3555         kiblnd_thread_fini();
3556         return 0;
3557 }