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