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, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd_cb.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 static void kiblnd_peer_alive(kib_peer_t *peer);
  44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
  45 static void kiblnd_check_sends(kib_conn_t *conn);
  46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
  47                                int type, int body_nob);
  48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
  49                             int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
  50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
  51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
  52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
  53
  54 static void
  55 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
  56 {
  57         lnet_msg_t *lntmsg[2];
  58         kib_net_t  *net = ni->ni_data;
  59         int         rc;
  60         int         i;
  61
  62         LASSERT (net != NULL);
  63         LASSERT (!in_interrupt());
  64         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
  65         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
  66         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer response */
  67         LASSERT (tx->tx_pool != NULL);
  68
  69         kiblnd_unmap_tx(ni, tx);
  70
  71         /* tx may have up to 2 lnet msgs to finalise */
  72         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  73         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  74         rc = tx->tx_status;
  75
  76         if (tx->tx_conn != NULL) {
  77                 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
  78
  79                 kiblnd_conn_decref(tx->tx_conn);
  80                 tx->tx_conn = NULL;
  81         }
  82
  83         tx->tx_nwrq = 0;
  84         tx->tx_status = 0;
  85
  86         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  87
  88         /* delay finalize until my descs have been freed */
  89         for (i = 0; i < 2; i++) {
  90                 if (lntmsg[i] == NULL)
  91                         continue;
  92
  93                 lnet_finalize(ni, lntmsg[i], rc);
  94         }
  95 }
  96
  97 void
  98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
  99 {
 100         kib_tx_t *tx;
 101
 102         while (!list_empty(txlist)) {
 103                 tx = list_entry(txlist->next, kib_tx_t, tx_list);
 104
 105                 list_del(&tx->tx_list);
 106                 /* complete now */
 107                 tx->tx_waiting = 0;
 108                 tx->tx_status = status;
 109                 kiblnd_tx_done(ni, tx);
 110         }
 111 }
 112
 113 static kib_tx_t *
 114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
 115 {
 116         kib_net_t               *net = (kib_net_t *)ni->ni_data;
 117         struct list_head        *node;
 118         kib_tx_t                *tx;
 119         kib_tx_poolset_t        *tps;
 120
 121         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
 122         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
 123         if (node == NULL)
 124                 return NULL;
 125         tx = container_of(node, kib_tx_t, tx_list);
 126
 127         LASSERT (tx->tx_nwrq == 0);
 128         LASSERT (!tx->tx_queued);
 129         LASSERT (tx->tx_sending == 0);
 130         LASSERT (!tx->tx_waiting);
 131         LASSERT (tx->tx_status == 0);
 132         LASSERT (tx->tx_conn == NULL);
 133         LASSERT (tx->tx_lntmsg[0] == NULL);
 134         LASSERT (tx->tx_lntmsg[1] == NULL);
 135         LASSERT (tx->tx_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 = conn->ibc_hdev->ibh_mrs;
 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         LASSERT(mr != NULL);
 170
 171         rx->rx_sge.lkey   = mr->lkey;
 172         rx->rx_sge.addr   = rx->rx_msgaddr;
 173         rx->rx_sge.length = IBLND_MSG_SIZE;
 174
 175         rx->rx_wrq.next = NULL;
 176         rx->rx_wrq.sg_list = &rx->rx_sge;
 177         rx->rx_wrq.num_sge = 1;
 178         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 179
 180         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
 181         LASSERT (rx->rx_nob >= 0);              /* not posted */
 182
 183         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 184                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 185                 return 0;
 186         }
 187
 188         rx->rx_nob = -1;                        /* flag posted */
 189
 190         /* NB: need an extra reference after ib_post_recv because we don't
 191          * own this rx (and rx::rx_conn) anymore, LU-5678.
 192          */
 193         kiblnd_conn_addref(conn);
 194         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 195         if (unlikely(rc != 0)) {
 196                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 197                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 198                 rx->rx_nob = 0;
 199         }
 200
 201         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 202                 goto out;
 203
 204         if (unlikely(rc != 0)) {
 205                 kiblnd_close_conn(conn, rc);
 206                 kiblnd_drop_rx(rx);     /* No more posts for this rx */
 207                 goto out;
 208         }
 209
 210         if (credit == IBLND_POSTRX_NO_CREDIT)
 211                 goto out;
 212
 213         spin_lock(&conn->ibc_lock);
 214         if (credit == IBLND_POSTRX_PEER_CREDIT)
 215                 conn->ibc_outstanding_credits++;
 216         else
 217                 conn->ibc_reserved_credits++;
 218         spin_unlock(&conn->ibc_lock);
 219
 220         kiblnd_check_sends(conn);
 221 out:
 222         kiblnd_conn_decref(conn);
 223         return rc;
 224 }
 225
 226 static kib_tx_t *
 227 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
 228 {
 229         struct list_head *tmp;
 230
 231         list_for_each(tmp, &conn->ibc_active_txs) {
 232                 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
 233
 234                 LASSERT(!tx->tx_queued);
 235                 LASSERT(tx->tx_sending != 0 || tx->tx_waiting);
 236
 237                 if (tx->tx_cookie != cookie)
 238                         continue;
 239
 240                 if (tx->tx_waiting &&
 241                     tx->tx_msg->ibm_type == txtype)
 242                         return tx;
 243
 244                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 245                       tx->tx_waiting ? "" : "NOT ",
 246                       tx->tx_msg->ibm_type, txtype);
 247         }
 248         return NULL;
 249 }
 250
 251 static void
 252 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
 253 {
 254         kib_tx_t    *tx;
 255         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 256         int          idle;
 257
 258         spin_lock(&conn->ibc_lock);
 259
 260         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 261         if (tx == NULL) {
 262                 spin_unlock(&conn->ibc_lock);
 263
 264                 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
 265                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 266                 kiblnd_close_conn(conn, -EPROTO);
 267                 return;
 268         }
 269
 270         if (tx->tx_status == 0) {               /* success so far */
 271                 if (status < 0) {               /* failed? */
 272                         tx->tx_status = status;
 273                 } else if (txtype == IBLND_MSG_GET_REQ) {
 274                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 275                 }
 276         }
 277
 278         tx->tx_waiting = 0;
 279
 280         idle = !tx->tx_queued && (tx->tx_sending == 0);
 281         if (idle)
 282                 list_del(&tx->tx_list);
 283
 284         spin_unlock(&conn->ibc_lock);
 285
 286         if (idle)
 287                 kiblnd_tx_done(ni, tx);
 288 }
 289
 290 static void
 291 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
 292 {
 293         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 294         kib_tx_t    *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 295
 296         if (tx == NULL) {
 297                 CERROR("Can't get tx for completion %x for %s\n",
 298                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 299                 return;
 300         }
 301
 302         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 303         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 304         kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
 305
 306         kiblnd_queue_tx(tx, conn);
 307 }
 308
 309 static void
 310 kiblnd_handle_rx (kib_rx_t *rx)
 311 {
 312         kib_msg_t    *msg = rx->rx_msg;
 313         kib_conn_t   *conn = rx->rx_conn;
 314         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 315         int           credits = msg->ibm_credits;
 316         kib_tx_t     *tx;
 317         int           rc = 0;
 318         int           rc2;
 319         int           post_credit;
 320
 321         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 322
 323         CDEBUG (D_NET, "Received %x[%d] from %s\n",
 324                 msg->ibm_type, credits,
 325                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
 326
 327         if (credits != 0) {
 328                 /* Have I received credits that will let me send? */
 329                 spin_lock(&conn->ibc_lock);
 330
 331                 if (conn->ibc_credits + credits >
 332                     conn->ibc_queue_depth) {
 333                         rc2 = conn->ibc_credits;
 334                         spin_unlock(&conn->ibc_lock);
 335
 336                         CERROR("Bad credits from %s: %d + %d > %d\n",
 337                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 338                                rc2, credits,
 339                                conn->ibc_queue_depth);
 340
 341                         kiblnd_close_conn(conn, -EPROTO);
 342                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 343                         return;
 344                 }
 345
 346                 conn->ibc_credits += credits;
 347
 348                 /* This ensures the credit taken by NOOP can be returned */
 349                 if (msg->ibm_type == IBLND_MSG_NOOP &&
 350                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
 351                         conn->ibc_outstanding_credits++;
 352
 353                 spin_unlock(&conn->ibc_lock);
 354                 kiblnd_check_sends(conn);
 355         }
 356
 357         switch (msg->ibm_type) {
 358         default:
 359                 CERROR("Bad IBLND message type %x from %s\n",
 360                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 361                 post_credit = IBLND_POSTRX_NO_CREDIT;
 362                 rc = -EPROTO;
 363                 break;
 364
 365         case IBLND_MSG_NOOP:
 366                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
 367                         post_credit = IBLND_POSTRX_NO_CREDIT;
 368                         break;
 369                 }
 370
 371                 if (credits != 0) /* credit already posted */
 372                         post_credit = IBLND_POSTRX_NO_CREDIT;
 373                 else              /* a keepalive NOOP */
 374                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 375                 break;
 376
 377         case IBLND_MSG_IMMEDIATE:
 378                 post_credit = IBLND_POSTRX_DONT_POST;
 379                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 380                                 msg->ibm_srcnid, rx, 0);
 381                 if (rc < 0)                     /* repost on error */
 382                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 383                 break;
 384
 385         case IBLND_MSG_PUT_REQ:
 386                 post_credit = IBLND_POSTRX_DONT_POST;
 387                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 388                                 msg->ibm_srcnid, rx, 1);
 389                 if (rc < 0)                     /* repost on error */
 390                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 391                 break;
 392
 393         case IBLND_MSG_PUT_NAK:
 394                 CWARN ("PUT_NACK from %s\n",
 395                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 396                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 397                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 398                                          msg->ibm_u.completion.ibcm_status,
 399                                          msg->ibm_u.completion.ibcm_cookie);
 400                 break;
 401
 402         case IBLND_MSG_PUT_ACK:
 403                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 404
 405                 spin_lock(&conn->ibc_lock);
 406                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 407                                         msg->ibm_u.putack.ibpam_src_cookie);
 408                 if (tx != NULL)
 409                         list_del(&tx->tx_list);
 410                 spin_unlock(&conn->ibc_lock);
 411
 412                 if (tx == NULL) {
 413                         CERROR("Unmatched PUT_ACK from %s\n",
 414                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 415                         rc = -EPROTO;
 416                         break;
 417                 }
 418
 419                 LASSERT (tx->tx_waiting);
 420                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
 421                  * (a) I can overwrite tx_msg since my peer has received it!
 422                  * (b) tx_waiting set tells tx_complete() it's not done. */
 423
 424                 tx->tx_nwrq = 0;                /* overwrite PUT_REQ */
 425
 426                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 427                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 428                                        &msg->ibm_u.putack.ibpam_rd,
 429                                        msg->ibm_u.putack.ibpam_dst_cookie);
 430                 if (rc2 < 0)
 431                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 432                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 433
 434                 spin_lock(&conn->ibc_lock);
 435                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
 436                 kiblnd_queue_tx_locked(tx, conn);
 437                 spin_unlock(&conn->ibc_lock);
 438                 break;
 439
 440         case IBLND_MSG_PUT_DONE:
 441                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 442                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 443                                          msg->ibm_u.completion.ibcm_status,
 444                                          msg->ibm_u.completion.ibcm_cookie);
 445                 break;
 446
 447         case IBLND_MSG_GET_REQ:
 448                 post_credit = IBLND_POSTRX_DONT_POST;
 449                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 450                                 msg->ibm_srcnid, rx, 1);
 451                 if (rc < 0)                     /* repost on error */
 452                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 453                 break;
 454
 455         case IBLND_MSG_GET_DONE:
 456                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 457                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 458                                          msg->ibm_u.completion.ibcm_status,
 459                                          msg->ibm_u.completion.ibcm_cookie);
 460                 break;
 461         }
 462
 463         if (rc < 0)                             /* protocol error */
 464                 kiblnd_close_conn(conn, rc);
 465
 466         if (post_credit != IBLND_POSTRX_DONT_POST)
 467                 kiblnd_post_rx(rx, post_credit);
 468 }
 469
 470 static void
 471 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
 472 {
 473         kib_msg_t    *msg = rx->rx_msg;
 474         kib_conn_t   *conn = rx->rx_conn;
 475         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 476         kib_net_t    *net = ni->ni_data;
 477         int           rc;
 478         int           err = -EIO;
 479
 480         LASSERT (net != NULL);
 481         LASSERT (rx->rx_nob < 0);               /* was posted */
 482         rx->rx_nob = 0;                         /* isn't now */
 483
 484         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 485                 goto ignore;
 486
 487         if (status != IB_WC_SUCCESS) {
 488                 CNETERR("Rx from %s failed: %d\n",
 489                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 490                 goto failed;
 491         }
 492
 493         LASSERT (nob >= 0);
 494         rx->rx_nob = nob;
 495
 496         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 497         if (rc != 0) {
 498                 CERROR ("Error %d unpacking rx from %s\n",
 499                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 500                 goto failed;
 501         }
 502
 503         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 504             msg->ibm_dstnid != ni->ni_nid ||
 505             msg->ibm_srcstamp != conn->ibc_incarnation ||
 506             msg->ibm_dststamp != net->ibn_incarnation) {
 507                 CERROR ("Stale rx from %s\n",
 508                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
 509                 err = -ESTALE;
 510                 goto failed;
 511         }
 512
 513         /* set time last known alive */
 514         kiblnd_peer_alive(conn->ibc_peer);
 515
 516         /* racing with connection establishment/teardown! */
 517
 518         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 519                 rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
 520                 unsigned long  flags;
 521
 522                 write_lock_irqsave(g_lock, flags);
 523                 /* must check holding global lock to eliminate race */
 524                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 525                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 526                         write_unlock_irqrestore(g_lock, flags);
 527                         return;
 528                 }
 529                 write_unlock_irqrestore(g_lock, flags);
 530         }
 531         kiblnd_handle_rx(rx);
 532         return;
 533
 534  failed:
 535         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 536         kiblnd_close_conn(conn, err);
 537  ignore:
 538         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
 539 }
 540
 541 static struct page *
 542 kiblnd_kvaddr_to_page (unsigned long vaddr)
 543 {
 544         struct page *page;
 545
 546         if (is_vmalloc_addr((void *)vaddr)) {
 547                 page = vmalloc_to_page ((void *)vaddr);
 548                 LASSERT (page != NULL);
 549                 return page;
 550         }
 551 #ifdef CONFIG_HIGHMEM
 552         if (vaddr >= PKMAP_BASE &&
 553             vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
 554                 /* No highmem pages only used for bulk (kiov) I/O */
 555                 CERROR("find page for address in highmem\n");
 556                 LBUG();
 557         }
 558 #endif
 559         page = virt_to_page (vaddr);
 560         LASSERT (page != NULL);
 561         return page;
 562 }
 563
 564 static int
 565 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
 566 {
 567         kib_hca_dev_t           *hdev;
 568         kib_fmr_poolset_t       *fps;
 569         int                     cpt;
 570         int                     rc;
 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         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 577
 578         fps = net->ibn_fmr_ps[cpt];
 579         rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
 580         if (rc != 0) {
 581                 CERROR("Can't map %u pages: %d\n", nob, rc);
 582                 return rc;
 583         }
 584
 585         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 586          * the rkey */
 587         rd->rd_key = tx->fmr.fmr_key;
 588         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
 589         rd->rd_frags[0].rf_nob   = nob;
 590         rd->rd_nfrags = 1;
 591
 592         return 0;
 593 }
 594
 595 static void
 596 kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
 597 {
 598         kib_net_t  *net = ni->ni_data;
 599
 600         LASSERT(net != NULL);
 601
 602         if (net->ibn_fmr_ps != NULL)
 603                 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
 604
 605         if (tx->tx_nfrags != 0) {
 606                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
 607                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 608                 tx->tx_nfrags = 0;
 609         }
 610 }
 611
 612 static int
 613 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
 614 {
 615         kib_net_t     *net   = ni->ni_data;
 616         kib_hca_dev_t *hdev  = net->ibn_dev->ibd_hdev;
 617         struct ib_mr  *mr    = NULL;
 618         __u32 nob;
 619         int i;
 620
 621         /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
 622          * RDMA sink */
 623         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 624         tx->tx_nfrags = nfrags;
 625
 626         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
 627                                           tx->tx_nfrags, tx->tx_dmadir);
 628
 629         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 630                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 631                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 632                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 633                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 634                 nob += rd->rd_frags[i].rf_nob;
 635         }
 636
 637         mr = kiblnd_find_rd_dma_mr(ni, rd,
 638                                    (tx->tx_conn != NULL) ?
 639                                    tx->tx_conn->ibc_max_frags : -1);
 640         if (mr != NULL) {
 641                 /* found pre-mapping MR */
 642                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 643                 return 0;
 644         }
 645
 646         if (net->ibn_fmr_ps != NULL)
 647                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 648
 649         return -EINVAL;
 650 }
 651
 652
 653 static int
 654 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 655                     unsigned int niov, struct kvec *iov, int offset, int nob)
 656 {
 657         kib_net_t          *net = ni->ni_data;
 658         struct page        *page;
 659         struct scatterlist *sg;
 660         unsigned long       vaddr;
 661         int                 fragnob;
 662         int                 page_offset;
 663
 664         LASSERT (nob > 0);
 665         LASSERT (niov > 0);
 666         LASSERT (net != NULL);
 667
 668         while (offset >= iov->iov_len) {
 669                 offset -= iov->iov_len;
 670                 niov--;
 671                 iov++;
 672                 LASSERT (niov > 0);
 673         }
 674
 675         sg = tx->tx_frags;
 676         do {
 677                 LASSERT (niov > 0);
 678
 679                 vaddr = ((unsigned long)iov->iov_base) + offset;
 680                 page_offset = vaddr & (PAGE_SIZE - 1);
 681                 page = kiblnd_kvaddr_to_page(vaddr);
 682                 if (page == NULL) {
 683                         CERROR ("Can't find page\n");
 684                         return -EFAULT;
 685                 }
 686
 687                 fragnob = min((int)(iov->iov_len - offset), nob);
 688                 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
 689
 690                 sg_set_page(sg, page, fragnob, page_offset);
 691                 sg++;
 692
 693                 if (offset + fragnob < iov->iov_len) {
 694                         offset += fragnob;
 695                 } else {
 696                         offset = 0;
 697                         iov++;
 698                         niov--;
 699                 }
 700                 nob -= fragnob;
 701         } while (nob > 0);
 702
 703         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 704 }
 705
 706 static int
 707 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 708                       int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 709 {
 710         kib_net_t          *net = ni->ni_data;
 711         struct scatterlist *sg;
 712         int                 fragnob;
 713
 714         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 715
 716         LASSERT (nob > 0);
 717         LASSERT (nkiov > 0);
 718         LASSERT (net != NULL);
 719
 720         while (offset >= kiov->kiov_len) {
 721                 offset -= kiov->kiov_len;
 722                 nkiov--;
 723                 kiov++;
 724                 LASSERT (nkiov > 0);
 725         }
 726
 727         sg = tx->tx_frags;
 728         do {
 729                 LASSERT (nkiov > 0);
 730
 731                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 732
 733                 sg_set_page(sg, kiov->kiov_page, fragnob,
 734                             kiov->kiov_offset + offset);
 735                 sg++;
 736
 737                 offset = 0;
 738                 kiov++;
 739                 nkiov--;
 740                 nob -= fragnob;
 741         } while (nob > 0);
 742
 743         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 744 }
 745
 746 static int
 747 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
 748 __must_hold(&conn->ibc_lock)
 749 {
 750         kib_msg_t         *msg = tx->tx_msg;
 751         kib_peer_t        *peer = conn->ibc_peer;
 752         struct lnet_ni    *ni = peer->ibp_ni;
 753         int                ver = conn->ibc_version;
 754         int                rc;
 755         int                done;
 756
 757         LASSERT(tx->tx_queued);
 758         /* We rely on this for QP sizing */
 759         LASSERT(tx->tx_nwrq > 0);
 760         LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
 761
 762         LASSERT(credit == 0 || credit == 1);
 763         LASSERT(conn->ibc_outstanding_credits >= 0);
 764         LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
 765         LASSERT(conn->ibc_credits >= 0);
 766         LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
 767
 768         if (conn->ibc_nsends_posted ==
 769             kiblnd_concurrent_sends(ver, ni)) {
 770                 /* tx completions outstanding... */
 771                 CDEBUG(D_NET, "%s: posted enough\n",
 772                        libcfs_nid2str(peer->ibp_nid));
 773                 return -EAGAIN;
 774         }
 775
 776         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 777                 CDEBUG(D_NET, "%s: no credits\n",
 778                        libcfs_nid2str(peer->ibp_nid));
 779                 return -EAGAIN;
 780         }
 781
 782         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 783             conn->ibc_credits == 1 &&   /* last credit reserved */
 784             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
 785                 CDEBUG(D_NET, "%s: not using last credit\n",
 786                        libcfs_nid2str(peer->ibp_nid));
 787                 return -EAGAIN;
 788         }
 789
 790         /* NB don't drop ibc_lock before bumping tx_sending */
 791         list_del(&tx->tx_list);
 792         tx->tx_queued = 0;
 793
 794         if (msg->ibm_type == IBLND_MSG_NOOP &&
 795             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
 796              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 797               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 798                 /* OK to drop when posted enough NOOPs, since
 799                  * kiblnd_check_sends will queue NOOP again when
 800                  * posted NOOPs complete */
 801                 spin_unlock(&conn->ibc_lock);
 802                 kiblnd_tx_done(peer->ibp_ni, tx);
 803                 spin_lock(&conn->ibc_lock);
 804                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 805                        libcfs_nid2str(peer->ibp_nid),
 806                        conn->ibc_noops_posted);
 807                 return 0;
 808         }
 809
 810         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 811                         peer->ibp_nid, conn->ibc_incarnation);
 812
 813         conn->ibc_credits -= credit;
 814         conn->ibc_outstanding_credits = 0;
 815         conn->ibc_nsends_posted++;
 816         if (msg->ibm_type == IBLND_MSG_NOOP)
 817                 conn->ibc_noops_posted++;
 818
 819         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 820          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 821          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 822          * and then re-queued here.  It's (just) possible that
 823          * tx_sending is non-zero if we've not done the tx_complete()
 824          * from the first send; hence the ++ rather than = below. */
 825         tx->tx_sending++;
 826         list_add(&tx->tx_list, &conn->ibc_active_txs);
 827
 828         /* I'm still holding ibc_lock! */
 829         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
 830                 rc = -ECONNABORTED;
 831         } else if (tx->tx_pool->tpo_pool.po_failed ||
 832                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
 833                 /* close_conn will launch failover */
 834                 rc = -ENETDOWN;
 835         } else {
 836                 struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
 837                 struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
 838                 struct ib_send_wr *wr  = &tx->tx_wrq[0].wr;
 839
 840                 if (frd != NULL) {
 841                         if (!frd->frd_valid) {
 842                                 wr = &frd->frd_inv_wr.wr;
 843                                 wr->next = &frd->frd_fastreg_wr.wr;
 844                         } else {
 845                                 wr = &frd->frd_fastreg_wr.wr;
 846                         }
 847                         frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
 848                 }
 849
 850                 LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
 851                          "bad wr_id "LPX64", opc %d, flags %d, peer: %s\n",
 852                          bad->wr_id, bad->opcode, bad->send_flags,
 853                          libcfs_nid2str(conn->ibc_peer->ibp_nid));
 854
 855                 bad = NULL;
 856                 rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
 857         }
 858
 859         conn->ibc_last_send = jiffies;
 860
 861         if (rc == 0)
 862                 return 0;
 863
 864         /* NB credits are transferred in the actual
 865          * message, which can only be the last work item */
 866         conn->ibc_credits += credit;
 867         conn->ibc_outstanding_credits += msg->ibm_credits;
 868         conn->ibc_nsends_posted--;
 869         if (msg->ibm_type == IBLND_MSG_NOOP)
 870                 conn->ibc_noops_posted--;
 871
 872         tx->tx_status = rc;
 873         tx->tx_waiting = 0;
 874         tx->tx_sending--;
 875
 876         done = (tx->tx_sending == 0);
 877         if (done)
 878                 list_del(&tx->tx_list);
 879
 880         spin_unlock(&conn->ibc_lock);
 881
 882         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 883                 CERROR("Error %d posting transmit to %s\n",
 884                        rc, libcfs_nid2str(peer->ibp_nid));
 885         else
 886                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 887                        rc, libcfs_nid2str(peer->ibp_nid));
 888
 889         kiblnd_close_conn(conn, rc);
 890
 891         if (done)
 892                 kiblnd_tx_done(peer->ibp_ni, tx);
 893
 894         spin_lock(&conn->ibc_lock);
 895
 896         return -EIO;
 897 }
 898
 899 static void
 900 kiblnd_check_sends (kib_conn_t *conn)
 901 {
 902         int        ver = conn->ibc_version;
 903         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 904         kib_tx_t  *tx;
 905
 906         /* Don't send anything until after the connection is established */
 907         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 908                 CDEBUG(D_NET, "%s too soon\n",
 909                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 910                 return;
 911         }
 912
 913         spin_lock(&conn->ibc_lock);
 914
 915         LASSERT(conn->ibc_nsends_posted <=
 916                 kiblnd_concurrent_sends(ver, ni));
 917         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 918                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 919         LASSERT (conn->ibc_reserved_credits >= 0);
 920
 921         while (conn->ibc_reserved_credits > 0 &&
 922                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 923                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 924                                     kib_tx_t, tx_list);
 925                 list_del(&tx->tx_list);
 926                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 927                 conn->ibc_reserved_credits--;
 928         }
 929
 930         if (kiblnd_need_noop(conn)) {
 931                 spin_unlock(&conn->ibc_lock);
 932
 933                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 934                 if (tx != NULL)
 935                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 936
 937                 spin_lock(&conn->ibc_lock);
 938                 if (tx != NULL)
 939                         kiblnd_queue_tx_locked(tx, conn);
 940         }
 941
 942         for (;;) {
 943                 int credit;
 944
 945                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 946                         credit = 0;
 947                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
 948                                             kib_tx_t, tx_list);
 949                 } else if (!list_empty(&conn->ibc_tx_noops)) {
 950                         LASSERT (!IBLND_OOB_CAPABLE(ver));
 951                         credit = 1;
 952                         tx = list_entry(conn->ibc_tx_noops.next,
 953                                         kib_tx_t, tx_list);
 954                 } else if (!list_empty(&conn->ibc_tx_queue)) {
 955                         credit = 1;
 956                         tx = list_entry(conn->ibc_tx_queue.next,
 957                                             kib_tx_t, tx_list);
 958                 } else
 959                         break;
 960
 961                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
 962                         break;
 963         }
 964
 965         spin_unlock(&conn->ibc_lock);
 966 }
 967
 968 static void
 969 kiblnd_tx_complete (kib_tx_t *tx, int status)
 970 {
 971         int           failed = (status != IB_WC_SUCCESS);
 972         kib_conn_t   *conn = tx->tx_conn;
 973         int           idle;
 974
 975         LASSERT (tx->tx_sending > 0);
 976
 977         if (failed) {
 978                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 979                         CNETERR("Tx -> %s cookie "LPX64
 980                                 " sending %d waiting %d: failed %d\n",
 981                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
 982                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
 983                                 status);
 984
 985                 kiblnd_close_conn(conn, -EIO);
 986         } else {
 987                 kiblnd_peer_alive(conn->ibc_peer);
 988         }
 989
 990         spin_lock(&conn->ibc_lock);
 991
 992         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
 993          * gets to free it, which also drops its ref on 'conn'. */
 994
 995         tx->tx_sending--;
 996         conn->ibc_nsends_posted--;
 997         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
 998                 conn->ibc_noops_posted--;
 999
1000         if (failed) {
1001                 tx->tx_waiting = 0;             /* don't wait for peer */
1002                 tx->tx_status = -EIO;
1003         }
1004
1005         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1006                !tx->tx_waiting &&               /* Not waiting for peer */
1007                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1008         if (idle)
1009                 list_del(&tx->tx_list);
1010
1011         kiblnd_conn_addref(conn);               /* 1 ref for me.... */
1012
1013         spin_unlock(&conn->ibc_lock);
1014
1015         if (idle)
1016                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1017
1018         kiblnd_check_sends(conn);
1019
1020         kiblnd_conn_decref(conn);               /* ...until here */
1021 }
1022
1023 static void
1024 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1025 {
1026         kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1027         struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1028         struct ib_rdma_wr *wrq;
1029         int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1030         struct ib_mr *mr = hdev->ibh_mrs;
1031
1032         LASSERT(tx->tx_nwrq >= 0);
1033         LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1034         LASSERT(nob <= IBLND_MSG_SIZE);
1035         LASSERT(mr != NULL);
1036
1037         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1038
1039         sge->lkey   = mr->lkey;
1040         sge->addr   = tx->tx_msgaddr;
1041         sge->length = nob;
1042
1043         wrq = &tx->tx_wrq[tx->tx_nwrq];
1044         memset(wrq, 0, sizeof(*wrq));
1045
1046         wrq->wr.next            = NULL;
1047         wrq->wr.wr_id           = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1048         wrq->wr.sg_list         = sge;
1049         wrq->wr.num_sge         = 1;
1050         wrq->wr.opcode          = IB_WR_SEND;
1051         wrq->wr.send_flags      = IB_SEND_SIGNALED;
1052
1053         tx->tx_nwrq++;
1054 }
1055
1056 static int
1057 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1058                  int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1059 {
1060         kib_msg_t         *ibmsg = tx->tx_msg;
1061         kib_rdma_desc_t   *srcrd = tx->tx_rd;
1062         struct ib_sge     *sge = &tx->tx_sge[0];
1063         struct ib_rdma_wr *wrq;
1064         int                rc  = resid;
1065         int                srcidx;
1066         int                dstidx;
1067         int                wrknob;
1068
1069         LASSERT (!in_interrupt());
1070         LASSERT (tx->tx_nwrq == 0);
1071         LASSERT (type == IBLND_MSG_GET_DONE ||
1072                  type == IBLND_MSG_PUT_DONE);
1073
1074         srcidx = dstidx = 0;
1075
1076         while (resid > 0) {
1077                 if (srcidx >= srcrd->rd_nfrags) {
1078                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1079                         rc = -EPROTO;
1080                         break;
1081                 }
1082
1083                 if (dstidx == dstrd->rd_nfrags) {
1084                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1085                         rc = -EPROTO;
1086                         break;
1087                 }
1088
1089                 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1090                         CERROR("RDMA has too many fragments for peer %s (%d), "
1091                                "src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1092                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1093                                conn->ibc_max_frags,
1094                                srcidx, srcrd->rd_nfrags,
1095                                dstidx, dstrd->rd_nfrags);
1096                         rc = -EMSGSIZE;
1097                         break;
1098                 }
1099
1100                 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1101                                  kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1102
1103                 sge = &tx->tx_sge[tx->tx_nwrq];
1104                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1105                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1106                 sge->length = wrknob;
1107
1108                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1109
1110                 wrq->wr.next            = &(wrq + 1)->wr;
1111                 wrq->wr.wr_id           = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1112                 wrq->wr.sg_list         = sge;
1113                 wrq->wr.num_sge         = 1;
1114                 wrq->wr.opcode          = IB_WR_RDMA_WRITE;
1115                 wrq->wr.send_flags      = 0;
1116
1117 #ifdef HAVE_IB_RDMA_WR
1118                 wrq->remote_addr        = kiblnd_rd_frag_addr(dstrd, dstidx);
1119                 wrq->rkey               = kiblnd_rd_frag_key(dstrd, dstidx);
1120 #else
1121                 wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1122                 wrq->wr.wr.rdma.rkey    = kiblnd_rd_frag_key(dstrd, dstidx);
1123 #endif
1124
1125                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1126                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1127
1128                 resid -= wrknob;
1129
1130                 tx->tx_nwrq++;
1131                 wrq++;
1132                 sge++;
1133         }
1134
1135         if (rc < 0)                             /* no RDMA if completing with failure */
1136                 tx->tx_nwrq = 0;
1137
1138         ibmsg->ibm_u.completion.ibcm_status = rc;
1139         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1140         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1141                            type, sizeof (kib_completion_msg_t));
1142
1143         return rc;
1144 }
1145
1146 static void
1147 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1148 {
1149         struct list_head *q;
1150
1151         LASSERT(tx->tx_nwrq > 0);       /* work items set up */
1152         LASSERT(!tx->tx_queued);        /* not queued for sending already */
1153         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1154
1155         tx->tx_queued = 1;
1156         tx->tx_deadline = jiffies +
1157                           msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1158                                            MSEC_PER_SEC);
1159
1160         if (tx->tx_conn == NULL) {
1161                 kiblnd_conn_addref(conn);
1162                 tx->tx_conn = conn;
1163                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1164         } else {
1165                 /* PUT_DONE first attached to conn as a PUT_REQ */
1166                 LASSERT (tx->tx_conn == conn);
1167                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1168         }
1169
1170         switch (tx->tx_msg->ibm_type) {
1171         default:
1172                 LBUG();
1173
1174         case IBLND_MSG_PUT_REQ:
1175         case IBLND_MSG_GET_REQ:
1176                 q = &conn->ibc_tx_queue_rsrvd;
1177                 break;
1178
1179         case IBLND_MSG_PUT_NAK:
1180         case IBLND_MSG_PUT_ACK:
1181         case IBLND_MSG_PUT_DONE:
1182         case IBLND_MSG_GET_DONE:
1183                 q = &conn->ibc_tx_queue_nocred;
1184                 break;
1185
1186         case IBLND_MSG_NOOP:
1187                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1188                         q = &conn->ibc_tx_queue_nocred;
1189                 else
1190                         q = &conn->ibc_tx_noops;
1191                 break;
1192
1193         case IBLND_MSG_IMMEDIATE:
1194                 q = &conn->ibc_tx_queue;
1195                 break;
1196         }
1197
1198         list_add_tail(&tx->tx_list, q);
1199 }
1200
1201 static void
1202 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1203 {
1204         spin_lock(&conn->ibc_lock);
1205         kiblnd_queue_tx_locked(tx, conn);
1206         spin_unlock(&conn->ibc_lock);
1207
1208         kiblnd_check_sends(conn);
1209 }
1210
1211 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1212                                struct sockaddr_in *srcaddr,
1213                                struct sockaddr_in *dstaddr,
1214                                int timeout_ms)
1215 {
1216         unsigned short port;
1217         int rc;
1218
1219         /* allow the port to be reused */
1220         rc = rdma_set_reuseaddr(cmid, 1);
1221         if (rc != 0) {
1222                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1223                 return rc;
1224         }
1225
1226         /* look for a free privileged port */
1227         for (port = PROT_SOCK-1; port > 0; port--) {
1228                 srcaddr->sin_port = htons(port);
1229                 rc = rdma_resolve_addr(cmid,
1230                                        (struct sockaddr *)srcaddr,
1231                                        (struct sockaddr *)dstaddr,
1232                                        timeout_ms);
1233                 if (rc == 0) {
1234                         CDEBUG(D_NET, "bound to port %hu\n", port);
1235                         return 0;
1236                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1237                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1238                                port, rc);
1239                 } else {
1240                         return rc;
1241                 }
1242         }
1243
1244         CERROR("Failed to bind to a free privileged port\n");
1245         return rc;
1246 }
1247
1248 static void
1249 kiblnd_connect_peer (kib_peer_t *peer)
1250 {
1251         struct rdma_cm_id *cmid;
1252         kib_dev_t         *dev;
1253         kib_net_t         *net = peer->ibp_ni->ni_data;
1254         struct sockaddr_in srcaddr;
1255         struct sockaddr_in dstaddr;
1256         int                rc;
1257
1258         LASSERT (net != NULL);
1259         LASSERT (peer->ibp_connecting > 0);
1260         LASSERT(!peer->ibp_reconnecting);
1261
1262         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1263                                      IB_QPT_RC);
1264
1265         if (IS_ERR(cmid)) {
1266                 CERROR("Can't create CMID for %s: %ld\n",
1267                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1268                 rc = PTR_ERR(cmid);
1269                 goto failed;
1270         }
1271
1272         dev = net->ibn_dev;
1273         memset(&srcaddr, 0, sizeof(srcaddr));
1274         srcaddr.sin_family = AF_INET;
1275         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1276
1277         memset(&dstaddr, 0, sizeof(dstaddr));
1278         dstaddr.sin_family = AF_INET;
1279         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1280         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1281
1282         kiblnd_peer_addref(peer);               /* cmid's ref */
1283
1284         if (*kiblnd_tunables.kib_use_priv_port) {
1285                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1286                                          *kiblnd_tunables.kib_timeout * 1000);
1287         } else {
1288                 rc = rdma_resolve_addr(cmid,
1289                                        (struct sockaddr *)&srcaddr,
1290                                        (struct sockaddr *)&dstaddr,
1291                                        *kiblnd_tunables.kib_timeout * 1000);
1292         }
1293         if (rc != 0) {
1294                 /* Can't initiate address resolution:  */
1295                 CERROR("Can't resolve addr for %s: %d\n",
1296                        libcfs_nid2str(peer->ibp_nid), rc);
1297                 goto failed2;
1298         }
1299
1300         LASSERT (cmid->device != NULL);
1301         CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1302                libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1303                &dev->ibd_ifip, cmid->device->name);
1304
1305         return;
1306
1307  failed2:
1308         kiblnd_peer_connect_failed(peer, 1, rc);
1309         kiblnd_peer_decref(peer);               /* cmid's ref */
1310         rdma_destroy_id(cmid);
1311         return;
1312  failed:
1313         kiblnd_peer_connect_failed(peer, 1, rc);
1314 }
1315
1316 bool
1317 kiblnd_reconnect_peer(kib_peer_t *peer)
1318 {
1319         rwlock_t         *glock = &kiblnd_data.kib_global_lock;
1320         char             *reason = NULL;
1321         struct list_head  txs;
1322         unsigned long     flags;
1323
1324         INIT_LIST_HEAD(&txs);
1325
1326         write_lock_irqsave(glock, flags);
1327         if (peer->ibp_reconnecting == 0) {
1328                 if (peer->ibp_accepting)
1329                         reason = "accepting";
1330                 else if (peer->ibp_connecting)
1331                         reason = "connecting";
1332                 else if (!list_empty(&peer->ibp_conns))
1333                         reason = "connected";
1334                 else /* connected then closed */
1335                         reason = "closed";
1336
1337                 goto no_reconnect;
1338         }
1339
1340         LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1341                 list_empty(&peer->ibp_conns));
1342         peer->ibp_reconnecting = 0;
1343
1344         if (!kiblnd_peer_active(peer)) {
1345                 list_splice_init(&peer->ibp_tx_queue, &txs);
1346                 reason = "unlinked";
1347                 goto no_reconnect;
1348         }
1349
1350         peer->ibp_connecting++;
1351         peer->ibp_reconnected++;
1352         write_unlock_irqrestore(glock, flags);
1353
1354         kiblnd_connect_peer(peer);
1355         return true;
1356
1357  no_reconnect:
1358         write_unlock_irqrestore(glock, flags);
1359
1360         CWARN("Abort reconnection of %s: %s\n",
1361               libcfs_nid2str(peer->ibp_nid), reason);
1362         kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1363         return false;
1364 }
1365
1366 void
1367 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1368 {
1369         kib_peer_t        *peer;
1370         kib_peer_t        *peer2;
1371         kib_conn_t        *conn;
1372         rwlock_t        *g_lock = &kiblnd_data.kib_global_lock;
1373         unsigned long      flags;
1374         int                rc;
1375
1376         /* If I get here, I've committed to send, so I complete the tx with
1377          * failure on any problems */
1378
1379         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1380         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1381
1382         /* First time, just use a read lock since I expect to find my peer
1383          * connected */
1384         read_lock_irqsave(g_lock, flags);
1385
1386         peer = kiblnd_find_peer_locked(nid);
1387         if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1388                 /* Found a peer with an established connection */
1389                 conn = kiblnd_get_conn_locked(peer);
1390                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1391
1392                 read_unlock_irqrestore(g_lock, flags);
1393
1394                 if (tx != NULL)
1395                         kiblnd_queue_tx(tx, conn);
1396                 kiblnd_conn_decref(conn); /* ...to here */
1397                 return;
1398         }
1399
1400         read_unlock(g_lock);
1401         /* Re-try with a write lock */
1402         write_lock(g_lock);
1403
1404         peer = kiblnd_find_peer_locked(nid);
1405         if (peer != NULL) {
1406                 if (list_empty(&peer->ibp_conns)) {
1407                         /* found a peer, but it's still connecting... */
1408                         LASSERT(kiblnd_peer_connecting(peer));
1409                         if (tx != NULL)
1410                                 list_add_tail(&tx->tx_list,
1411                                                   &peer->ibp_tx_queue);
1412                         write_unlock_irqrestore(g_lock, flags);
1413                 } else {
1414                         conn = kiblnd_get_conn_locked(peer);
1415                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1416
1417                         write_unlock_irqrestore(g_lock, flags);
1418
1419                         if (tx != NULL)
1420                                 kiblnd_queue_tx(tx, conn);
1421                         kiblnd_conn_decref(conn); /* ...to here */
1422                 }
1423                 return;
1424         }
1425
1426         write_unlock_irqrestore(g_lock, flags);
1427
1428         /* Allocate a peer ready to add to the peer table and retry */
1429         rc = kiblnd_create_peer(ni, &peer, nid);
1430         if (rc != 0) {
1431                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1432                 if (tx != NULL) {
1433                         tx->tx_status = -EHOSTUNREACH;
1434                         tx->tx_waiting = 0;
1435                         kiblnd_tx_done(ni, tx);
1436                 }
1437                 return;
1438         }
1439
1440         write_lock_irqsave(g_lock, flags);
1441
1442         peer2 = kiblnd_find_peer_locked(nid);
1443         if (peer2 != NULL) {
1444                 if (list_empty(&peer2->ibp_conns)) {
1445                         /* found a peer, but it's still connecting... */
1446                         LASSERT(kiblnd_peer_connecting(peer2));
1447                         if (tx != NULL)
1448                                 list_add_tail(&tx->tx_list,
1449                                                   &peer2->ibp_tx_queue);
1450                         write_unlock_irqrestore(g_lock, flags);
1451                 } else {
1452                         conn = kiblnd_get_conn_locked(peer2);
1453                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1454
1455                         write_unlock_irqrestore(g_lock, flags);
1456
1457                         if (tx != NULL)
1458                                 kiblnd_queue_tx(tx, conn);
1459                         kiblnd_conn_decref(conn); /* ...to here */
1460                 }
1461
1462                 kiblnd_peer_decref(peer);
1463                 return;
1464         }
1465
1466         /* Brand new peer */
1467         LASSERT (peer->ibp_connecting == 0);
1468         peer->ibp_connecting = 1;
1469
1470         /* always called with a ref on ni, which prevents ni being shutdown */
1471         LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1472
1473         if (tx != NULL)
1474                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1475
1476         kiblnd_peer_addref(peer);
1477         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1478
1479         write_unlock_irqrestore(g_lock, flags);
1480
1481         kiblnd_connect_peer(peer);
1482         kiblnd_peer_decref(peer);
1483 }
1484
1485 int
1486 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1487 {
1488         lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
1489         int               type = lntmsg->msg_type;
1490         lnet_process_id_t target = lntmsg->msg_target;
1491         int               target_is_router = lntmsg->msg_target_is_router;
1492         int               routing = lntmsg->msg_routing;
1493         unsigned int      payload_niov = lntmsg->msg_niov;
1494         struct kvec      *payload_iov = lntmsg->msg_iov;
1495         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
1496         unsigned int      payload_offset = lntmsg->msg_offset;
1497         unsigned int      payload_nob = lntmsg->msg_len;
1498         kib_msg_t        *ibmsg;
1499         kib_rdma_desc_t  *rd;
1500         kib_tx_t         *tx;
1501         int               nob;
1502         int               rc;
1503
1504         /* NB 'private' is different depending on what we're sending.... */
1505
1506         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1507                payload_nob, payload_niov, libcfs_id2str(target));
1508
1509         LASSERT (payload_nob == 0 || payload_niov > 0);
1510         LASSERT (payload_niov <= LNET_MAX_IOV);
1511
1512         /* Thread context */
1513         LASSERT (!in_interrupt());
1514         /* payload is either all vaddrs or all pages */
1515         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1516
1517         switch (type) {
1518         default:
1519                 LBUG();
1520                 return (-EIO);
1521
1522         case LNET_MSG_ACK:
1523                 LASSERT (payload_nob == 0);
1524                 break;
1525
1526         case LNET_MSG_GET:
1527                 if (routing || target_is_router)
1528                         break;                  /* send IMMEDIATE */
1529
1530                 /* is the REPLY message too small for RDMA? */
1531                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1532                 if (nob <= IBLND_MSG_SIZE)
1533                         break;                  /* send IMMEDIATE */
1534
1535                 tx = kiblnd_get_idle_tx(ni, target.nid);
1536                 if (tx == NULL) {
1537                         CERROR("Can't allocate txd for GET to %s\n",
1538                                libcfs_nid2str(target.nid));
1539                         return -ENOMEM;
1540                 }
1541
1542                 ibmsg = tx->tx_msg;
1543                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1544                 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1545                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1546                                                  lntmsg->msg_md->md_niov,
1547                                                  lntmsg->msg_md->md_iov.iov,
1548                                                  0, lntmsg->msg_md->md_length);
1549                 else
1550                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1551                                                   lntmsg->msg_md->md_niov,
1552                                                   lntmsg->msg_md->md_iov.kiov,
1553                                                   0, lntmsg->msg_md->md_length);
1554                 if (rc != 0) {
1555                         CERROR("Can't setup GET sink for %s: %d\n",
1556                                libcfs_nid2str(target.nid), rc);
1557                         kiblnd_tx_done(ni, tx);
1558                         return -EIO;
1559                 }
1560
1561                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1562                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1563                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1564
1565                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1566
1567                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1568                 if (tx->tx_lntmsg[1] == NULL) {
1569                         CERROR("Can't create reply for GET -> %s\n",
1570                                libcfs_nid2str(target.nid));
1571                         kiblnd_tx_done(ni, tx);
1572                         return -EIO;
1573                 }
1574
1575                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1576                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1577                 kiblnd_launch_tx(ni, tx, target.nid);
1578                 return 0;
1579
1580         case LNET_MSG_REPLY:
1581         case LNET_MSG_PUT:
1582                 /* Is the payload small enough not to need RDMA? */
1583                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1584                 if (nob <= IBLND_MSG_SIZE)
1585                         break;                  /* send IMMEDIATE */
1586
1587                 tx = kiblnd_get_idle_tx(ni, target.nid);
1588                 if (tx == NULL) {
1589                         CERROR("Can't allocate %s txd for %s\n",
1590                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1591                                libcfs_nid2str(target.nid));
1592                         return -ENOMEM;
1593                 }
1594
1595                 if (payload_kiov == NULL)
1596                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1597                                                  payload_niov, payload_iov,
1598                                                  payload_offset, payload_nob);
1599                 else
1600                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1601                                                   payload_niov, payload_kiov,
1602                                                   payload_offset, payload_nob);
1603                 if (rc != 0) {
1604                         CERROR("Can't setup PUT src for %s: %d\n",
1605                                libcfs_nid2str(target.nid), rc);
1606                         kiblnd_tx_done(ni, tx);
1607                         return -EIO;
1608                 }
1609
1610                 ibmsg = tx->tx_msg;
1611                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1612                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1613                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1614
1615                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1616                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1617                 kiblnd_launch_tx(ni, tx, target.nid);
1618                 return 0;
1619         }
1620
1621         /* send IMMEDIATE */
1622
1623         LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1624                  <= IBLND_MSG_SIZE);
1625
1626         tx = kiblnd_get_idle_tx(ni, target.nid);
1627         if (tx == NULL) {
1628                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1629                         type, libcfs_nid2str(target.nid));
1630                 return -ENOMEM;
1631         }
1632
1633         ibmsg = tx->tx_msg;
1634         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1635
1636         if (payload_kiov != NULL)
1637                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1638                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1639                                     payload_niov, payload_kiov,
1640                                     payload_offset, payload_nob);
1641         else
1642                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1643                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1644                                    payload_niov, payload_iov,
1645                                    payload_offset, payload_nob);
1646
1647         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1648         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1649
1650         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1651         kiblnd_launch_tx(ni, tx, target.nid);
1652         return 0;
1653 }
1654
1655 static void
1656 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1657 {
1658         lnet_process_id_t target = lntmsg->msg_target;
1659         unsigned int      niov = lntmsg->msg_niov;
1660         struct kvec      *iov = lntmsg->msg_iov;
1661         lnet_kiov_t      *kiov = lntmsg->msg_kiov;
1662         unsigned int      offset = lntmsg->msg_offset;
1663         unsigned int      nob = lntmsg->msg_len;
1664         kib_tx_t         *tx;
1665         int               rc;
1666
1667         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1668         if (tx == NULL) {
1669                 CERROR("Can't get tx for REPLY to %s\n",
1670                        libcfs_nid2str(target.nid));
1671                 goto failed_0;
1672         }
1673
1674         if (nob == 0)
1675                 rc = 0;
1676         else if (kiov == NULL)
1677                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1678                                          niov, iov, offset, nob);
1679         else
1680                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1681                                           niov, kiov, offset, nob);
1682
1683         if (rc != 0) {
1684                 CERROR("Can't setup GET src for %s: %d\n",
1685                        libcfs_nid2str(target.nid), rc);
1686                 goto failed_1;
1687         }
1688
1689         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1690                               IBLND_MSG_GET_DONE, nob,
1691                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1692                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1693         if (rc < 0) {
1694                 CERROR("Can't setup rdma for GET from %s: %d\n",
1695                        libcfs_nid2str(target.nid), rc);
1696                 goto failed_1;
1697         }
1698
1699         if (nob == 0) {
1700                 /* No RDMA: local completion may happen now! */
1701                 lnet_finalize(ni, lntmsg, 0);
1702         } else {
1703                 /* RDMA: lnet_finalize(lntmsg) when it
1704                  * completes */
1705                 tx->tx_lntmsg[0] = lntmsg;
1706         }
1707
1708         kiblnd_queue_tx(tx, rx->rx_conn);
1709         return;
1710
1711  failed_1:
1712         kiblnd_tx_done(ni, tx);
1713  failed_0:
1714         lnet_finalize(ni, lntmsg, -EIO);
1715 }
1716
1717 int
1718 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1719             unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1720             unsigned int offset, unsigned int mlen, unsigned int rlen)
1721 {
1722         kib_rx_t    *rx = private;
1723         kib_msg_t   *rxmsg = rx->rx_msg;
1724         kib_conn_t  *conn = rx->rx_conn;
1725         kib_tx_t    *tx;
1726         int          nob;
1727         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1728         int          rc = 0;
1729
1730         LASSERT (mlen <= rlen);
1731         LASSERT (!in_interrupt());
1732         /* Either all pages or all vaddrs */
1733         LASSERT (!(kiov != NULL && iov != NULL));
1734
1735         switch (rxmsg->ibm_type) {
1736         default:
1737                 LBUG();
1738
1739         case IBLND_MSG_IMMEDIATE:
1740                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1741                 if (nob > rx->rx_nob) {
1742                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1743                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1744                                 nob, rx->rx_nob);
1745                         rc = -EPROTO;
1746                         break;
1747                 }
1748
1749                 if (kiov != NULL)
1750                         lnet_copy_flat2kiov(niov, kiov, offset,
1751                                             IBLND_MSG_SIZE, rxmsg,
1752                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1753                                             mlen);
1754                 else
1755                         lnet_copy_flat2iov(niov, iov, offset,
1756                                            IBLND_MSG_SIZE, rxmsg,
1757                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1758                                            mlen);
1759                 lnet_finalize (ni, lntmsg, 0);
1760                 break;
1761
1762         case IBLND_MSG_PUT_REQ: {
1763                 kib_msg_t       *txmsg;
1764                 kib_rdma_desc_t *rd;
1765
1766                 if (mlen == 0) {
1767                         lnet_finalize(ni, lntmsg, 0);
1768                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1769                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1770                         break;
1771                 }
1772
1773                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1774                 if (tx == NULL) {
1775                         CERROR("Can't allocate tx for %s\n",
1776                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1777                         /* Not replying will break the connection */
1778                         rc = -ENOMEM;
1779                         break;
1780                 }
1781
1782                 txmsg = tx->tx_msg;
1783                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1784                 if (kiov == NULL)
1785                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1786                                                  niov, iov, offset, mlen);
1787                 else
1788                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1789                                                   niov, kiov, offset, mlen);
1790                 if (rc != 0) {
1791                         CERROR("Can't setup PUT sink for %s: %d\n",
1792                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1793                         kiblnd_tx_done(ni, tx);
1794                         /* tell peer it's over */
1795                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1796                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1797                         break;
1798                 }
1799
1800                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1801                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1802                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1803
1804                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1805
1806                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1807                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1808                 kiblnd_queue_tx(tx, conn);
1809
1810                 /* reposted buffer reserved for PUT_DONE */
1811                 post_credit = IBLND_POSTRX_NO_CREDIT;
1812                 break;
1813                 }
1814
1815         case IBLND_MSG_GET_REQ:
1816                 if (lntmsg != NULL) {
1817                         /* Optimized GET; RDMA lntmsg's payload */
1818                         kiblnd_reply(ni, rx, lntmsg);
1819                 } else {
1820                         /* GET didn't match anything */
1821                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1822                                                -ENODATA,
1823                                                rxmsg->ibm_u.get.ibgm_cookie);
1824                 }
1825                 break;
1826         }
1827
1828         kiblnd_post_rx(rx, post_credit);
1829         return rc;
1830 }
1831
1832 int
1833 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1834 {
1835         struct task_struct *task = kthread_run(fn, arg, name);
1836
1837         if (IS_ERR(task))
1838                 return PTR_ERR(task);
1839
1840         atomic_inc(&kiblnd_data.kib_nthreads);
1841         return 0;
1842 }
1843
1844 static void
1845 kiblnd_thread_fini (void)
1846 {
1847         atomic_dec (&kiblnd_data.kib_nthreads);
1848 }
1849
1850 static void
1851 kiblnd_peer_alive (kib_peer_t *peer)
1852 {
1853         /* This is racy, but everyone's only writing cfs_time_current() */
1854         peer->ibp_last_alive = cfs_time_current();
1855         smp_mb();
1856 }
1857
1858 static void
1859 kiblnd_peer_notify (kib_peer_t *peer)
1860 {
1861         int           error = 0;
1862         cfs_time_t    last_alive = 0;
1863         unsigned long flags;
1864
1865         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1866
1867         if (kiblnd_peer_idle(peer) && peer->ibp_error != 0) {
1868                 error = peer->ibp_error;
1869                 peer->ibp_error = 0;
1870
1871                 last_alive = peer->ibp_last_alive;
1872         }
1873
1874         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1875
1876         if (error != 0)
1877                 lnet_notify(peer->ibp_ni,
1878                             peer->ibp_nid, 0, last_alive);
1879 }
1880
1881 void
1882 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1883 {
1884         /* This just does the immediate housekeeping.  'error' is zero for a
1885          * normal shutdown which can happen only after the connection has been
1886          * established.  If the connection is established, schedule the
1887          * connection to be finished off by the connd.  Otherwise the connd is
1888          * already dealing with it (either to set it up or tear it down).
1889          * Caller holds kib_global_lock exclusively in irq context */
1890         kib_peer_t       *peer = conn->ibc_peer;
1891         kib_dev_t        *dev;
1892         unsigned long     flags;
1893
1894         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1895
1896         if (error != 0 && conn->ibc_comms_error == 0)
1897                 conn->ibc_comms_error = error;
1898
1899         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1900                 return; /* already being handled  */
1901
1902         if (error == 0 &&
1903             list_empty(&conn->ibc_tx_noops) &&
1904             list_empty(&conn->ibc_tx_queue) &&
1905             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1906             list_empty(&conn->ibc_tx_queue_nocred) &&
1907             list_empty(&conn->ibc_active_txs)) {
1908                 CDEBUG(D_NET, "closing conn to %s\n",
1909                        libcfs_nid2str(peer->ibp_nid));
1910         } else {
1911                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1912                        libcfs_nid2str(peer->ibp_nid), error,
1913                        list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1914                        list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1915                        list_empty(&conn->ibc_tx_queue_rsrvd) ?
1916                                                 "" : "(sending_rsrvd)",
1917                        list_empty(&conn->ibc_tx_queue_nocred) ?
1918                                                  "" : "(sending_nocred)",
1919                        list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1920         }
1921
1922         dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1923         list_del(&conn->ibc_list);
1924         /* connd (see below) takes over ibc_list's ref */
1925
1926         if (list_empty(&peer->ibp_conns) &&    /* no more conns */
1927             kiblnd_peer_active(peer)) {         /* still in peer table */
1928                 kiblnd_unlink_peer_locked(peer);
1929
1930                 /* set/clear error on last conn */
1931                 peer->ibp_error = conn->ibc_comms_error;
1932         }
1933
1934         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1935
1936         if (error != 0 &&
1937             kiblnd_dev_can_failover(dev)) {
1938                 list_add_tail(&dev->ibd_fail_list,
1939                               &kiblnd_data.kib_failed_devs);
1940                 wake_up(&kiblnd_data.kib_failover_waitq);
1941         }
1942
1943         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1944
1945         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1946         wake_up(&kiblnd_data.kib_connd_waitq);
1947
1948         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1949 }
1950
1951 void
1952 kiblnd_close_conn(kib_conn_t *conn, int error)
1953 {
1954         unsigned long flags;
1955
1956         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1957
1958         kiblnd_close_conn_locked(conn, error);
1959
1960         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1961 }
1962
1963 static void
1964 kiblnd_handle_early_rxs(kib_conn_t *conn)
1965 {
1966         unsigned long    flags;
1967         kib_rx_t        *rx;
1968
1969         LASSERT(!in_interrupt());
1970         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1971
1972         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1973         while (!list_empty(&conn->ibc_early_rxs)) {
1974                 rx = list_entry(conn->ibc_early_rxs.next,
1975                                     kib_rx_t, rx_list);
1976                 list_del(&rx->rx_list);
1977                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1978
1979                 kiblnd_handle_rx(rx);
1980
1981                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1982         }
1983         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1984 }
1985
1986 static void
1987 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1988 {
1989         struct list_head         zombies = LIST_HEAD_INIT(zombies);
1990         struct list_head        *tmp;
1991         struct list_head        *nxt;
1992         kib_tx_t                *tx;
1993
1994         spin_lock(&conn->ibc_lock);
1995
1996         list_for_each_safe(tmp, nxt, txs) {
1997                 tx = list_entry(tmp, kib_tx_t, tx_list);
1998
1999                 if (txs == &conn->ibc_active_txs) {
2000                         LASSERT(!tx->tx_queued);
2001                         LASSERT(tx->tx_waiting ||
2002                                 tx->tx_sending != 0);
2003                 } else {
2004                         LASSERT(tx->tx_queued);
2005                 }
2006
2007                 tx->tx_status = -ECONNABORTED;
2008                 tx->tx_waiting = 0;
2009
2010                 if (tx->tx_sending == 0) {
2011                         tx->tx_queued = 0;
2012                         list_del(&tx->tx_list);
2013                         list_add(&tx->tx_list, &zombies);
2014                 }
2015         }
2016
2017         spin_unlock(&conn->ibc_lock);
2018
2019         kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2020 }
2021
2022 static void
2023 kiblnd_finalise_conn (kib_conn_t *conn)
2024 {
2025         LASSERT (!in_interrupt());
2026         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
2027
2028         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2029
2030         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
2031          * for connections that didn't get as far as being connected, because
2032          * rdma_disconnect() does this for free. */
2033         kiblnd_abort_receives(conn);
2034
2035         /* Complete all tx descs not waiting for sends to complete.
2036          * NB we should be safe from RDMA now that the QP has changed state */
2037
2038         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2039         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2040         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2041         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2042         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2043
2044         kiblnd_handle_early_rxs(conn);
2045 }
2046
2047 static void
2048 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2049 {
2050         struct list_head zombies = LIST_HEAD_INIT(zombies);
2051         unsigned long   flags;
2052
2053         LASSERT (error != 0);
2054         LASSERT (!in_interrupt());
2055
2056         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2057
2058         if (active) {
2059                 LASSERT (peer->ibp_connecting > 0);
2060                 peer->ibp_connecting--;
2061         } else {
2062                 LASSERT (peer->ibp_accepting > 0);
2063                 peer->ibp_accepting--;
2064         }
2065
2066         if (kiblnd_peer_connecting(peer)) {
2067                 /* another connection attempt under way... */
2068                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2069                                         flags);
2070                 return;
2071         }
2072
2073         peer->ibp_reconnected = 0;
2074         if (list_empty(&peer->ibp_conns)) {
2075                 /* Take peer's blocked transmits to complete with error */
2076                 list_add(&zombies, &peer->ibp_tx_queue);
2077                 list_del_init(&peer->ibp_tx_queue);
2078
2079                 if (kiblnd_peer_active(peer))
2080                         kiblnd_unlink_peer_locked(peer);
2081
2082                 peer->ibp_error = error;
2083         } else {
2084                 /* Can't have blocked transmits if there are connections */
2085                 LASSERT(list_empty(&peer->ibp_tx_queue));
2086         }
2087
2088         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2089
2090         kiblnd_peer_notify(peer);
2091
2092         if (list_empty(&zombies))
2093                 return;
2094
2095         CNETERR("Deleting messages for %s: connection failed\n",
2096                 libcfs_nid2str(peer->ibp_nid));
2097
2098         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2099 }
2100
2101 static void
2102 kiblnd_connreq_done(kib_conn_t *conn, int status)
2103 {
2104         kib_peer_t       *peer = conn->ibc_peer;
2105         kib_tx_t         *tx;
2106         struct list_head txs;
2107         unsigned long    flags;
2108         int              active;
2109
2110         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2111
2112         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2113                libcfs_nid2str(peer->ibp_nid), active,
2114                conn->ibc_version, status);
2115
2116         LASSERT (!in_interrupt());
2117         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2118                   peer->ibp_connecting > 0) ||
2119                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2120                   peer->ibp_accepting > 0));
2121
2122         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2123         conn->ibc_connvars = NULL;
2124
2125         if (status != 0) {
2126                 /* failed to establish connection */
2127                 kiblnd_peer_connect_failed(peer, active, status);
2128                 kiblnd_finalise_conn(conn);
2129                 return;
2130         }
2131
2132         /* connection established */
2133         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2134
2135         conn->ibc_last_send = jiffies;
2136         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2137         kiblnd_peer_alive(peer);
2138
2139         /* Add conn to peer's list and nuke any dangling conns from a different
2140          * peer instance... */
2141         kiblnd_conn_addref(conn);       /* +1 ref for ibc_list */
2142         list_add(&conn->ibc_list, &peer->ibp_conns);
2143         peer->ibp_reconnected = 0;
2144         if (active)
2145                 peer->ibp_connecting--;
2146         else
2147                 peer->ibp_accepting--;
2148
2149         if (peer->ibp_version == 0) {
2150                 peer->ibp_version     = conn->ibc_version;
2151                 peer->ibp_incarnation = conn->ibc_incarnation;
2152         }
2153
2154         if (peer->ibp_version     != conn->ibc_version ||
2155             peer->ibp_incarnation != conn->ibc_incarnation) {
2156                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2157                                                 conn->ibc_incarnation);
2158                 peer->ibp_version     = conn->ibc_version;
2159                 peer->ibp_incarnation = conn->ibc_incarnation;
2160         }
2161
2162         /* grab pending txs while I have the lock */
2163         list_add(&txs, &peer->ibp_tx_queue);
2164         list_del_init(&peer->ibp_tx_queue);
2165
2166         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2167             conn->ibc_comms_error != 0) {       /* error has happened already */
2168                 lnet_ni_t *ni = peer->ibp_ni;
2169
2170                 /* start to shut down connection */
2171                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2172                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2173
2174                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2175
2176                 return;
2177         }
2178
2179         /* refcount taken by cmid is not reliable after I released the glock
2180          * because this connection is visible to other threads now, another
2181          * thread can find and close this connection right after I released
2182          * the glock, if kiblnd_cm_callback for RDMA_CM_EVENT_DISCONNECTED is
2183          * called, it can release the connection refcount taken by cmid.
2184          * It means the connection could be destroyed before I finish my
2185          * operations on it.
2186          */
2187         kiblnd_conn_addref(conn);
2188         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2189
2190         /* Schedule blocked txs */
2191         spin_lock(&conn->ibc_lock);
2192         while (!list_empty(&txs)) {
2193                 tx = list_entry(txs.next, kib_tx_t, tx_list);
2194                 list_del(&tx->tx_list);
2195
2196                 kiblnd_queue_tx_locked(tx, conn);
2197         }
2198         spin_unlock(&conn->ibc_lock);
2199
2200         kiblnd_check_sends(conn);
2201
2202         /* schedule blocked rxs */
2203         kiblnd_handle_early_rxs(conn);
2204
2205         kiblnd_conn_decref(conn);
2206 }
2207
2208 static void
2209 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2210 {
2211         int          rc;
2212
2213         rc = rdma_reject(cmid, rej, sizeof(*rej));
2214
2215         if (rc != 0)
2216                 CWARN("Error %d sending reject\n", rc);
2217 }
2218
2219 static int
2220 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2221 {
2222         rwlock_t                *g_lock = &kiblnd_data.kib_global_lock;
2223         kib_msg_t             *reqmsg = priv;
2224         kib_msg_t             *ackmsg;
2225         kib_dev_t             *ibdev;
2226         kib_peer_t            *peer;
2227         kib_peer_t            *peer2;
2228         kib_conn_t            *conn;
2229         lnet_ni_t             *ni  = NULL;
2230         kib_net_t             *net = NULL;
2231         lnet_nid_t             nid;
2232         struct rdma_conn_param cp;
2233         kib_rej_t              rej;
2234         int                    version = IBLND_MSG_VERSION;
2235         unsigned long          flags;
2236         int                    rc;
2237         struct sockaddr_in    *peer_addr;
2238         LASSERT (!in_interrupt());
2239
2240         /* cmid inherits 'context' from the corresponding listener id */
2241         ibdev = (kib_dev_t *)cmid->context;
2242         LASSERT (ibdev != NULL);
2243
2244         memset(&rej, 0, sizeof(rej));
2245         rej.ibr_magic                = IBLND_MSG_MAGIC;
2246         rej.ibr_why                  = IBLND_REJECT_FATAL;
2247         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2248
2249         peer_addr = (struct sockaddr_in *)&(cmid->route.addr.dst_addr);
2250         if (*kiblnd_tunables.kib_require_priv_port &&
2251             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2252                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2253                 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2254                        &ip, ntohs(peer_addr->sin_port));
2255                 goto failed;
2256         }
2257
2258         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2259                 CERROR("Short connection request\n");
2260                 goto failed;
2261         }
2262
2263         /* Future protocol version compatibility support!  If the
2264          * o2iblnd-specific protocol changes, or when LNET unifies
2265          * protocols over all LNDs, the initial connection will
2266          * negotiate a protocol version.  I trap this here to avoid
2267          * console errors; the reject tells the peer which protocol I
2268          * speak. */
2269         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2270             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2271                 goto failed;
2272         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2273             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2274             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2275                 goto failed;
2276         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2277             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2278             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2279                 goto failed;
2280
2281         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2282         if (rc != 0) {
2283                 CERROR("Can't parse connection request: %d\n", rc);
2284                 goto failed;
2285         }
2286
2287         nid = reqmsg->ibm_srcnid;
2288         ni  = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2289
2290         if (ni != NULL) {
2291                 net = (kib_net_t *)ni->ni_data;
2292                 rej.ibr_incarnation = net->ibn_incarnation;
2293         }
2294
2295         if (ni == NULL ||                         /* no matching net */
2296             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2297             net->ibn_dev != ibdev) {              /* wrong device */
2298                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
2299                        "bad dst nid %s\n", libcfs_nid2str(nid),
2300                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2301                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2302                         &ibdev->ibd_ifip,
2303                        libcfs_nid2str(reqmsg->ibm_dstnid));
2304
2305                 goto failed;
2306         }
2307
2308        /* check time stamp as soon as possible */
2309         if (reqmsg->ibm_dststamp != 0 &&
2310             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2311                 CWARN("Stale connection request\n");
2312                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2313                 goto failed;
2314         }
2315
2316         /* I can accept peer's version */
2317         version = reqmsg->ibm_version;
2318
2319         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2320                 CERROR("Unexpected connreq msg type: %x from %s\n",
2321                        reqmsg->ibm_type, libcfs_nid2str(nid));
2322                 goto failed;
2323         }
2324
2325         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2326             kiblnd_msg_queue_size(version, ni)) {
2327                 CERROR("Can't accept conn from %s, queue depth too large: "
2328                        " %d (<=%d wanted)\n",
2329                        libcfs_nid2str(nid),
2330                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2331                        kiblnd_msg_queue_size(version, ni));
2332
2333                 if (version == IBLND_MSG_VERSION)
2334                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2335
2336                 goto failed;
2337         }
2338
2339         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2340             kiblnd_rdma_frags(version, ni)) {
2341                 CWARN("Can't accept conn from %s (version %x): "
2342                       "max_frags %d too large (%d wanted)\n",
2343                       libcfs_nid2str(nid), version,
2344                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2345                       kiblnd_rdma_frags(version, ni));
2346
2347                 if (version >= IBLND_MSG_VERSION)
2348                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2349
2350                 goto failed;
2351         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2352                    kiblnd_rdma_frags(version, ni) &&
2353                    net->ibn_fmr_ps == NULL) {
2354                 CWARN("Can't accept conn from %s (version %x): "
2355                       "max_frags %d incompatible without FMR pool "
2356                       "(%d wanted)\n",
2357                       libcfs_nid2str(nid), version,
2358                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2359                       kiblnd_rdma_frags(version, ni));
2360
2361                 if (version == IBLND_MSG_VERSION)
2362                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2363
2364                 goto failed;
2365         }
2366
2367         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2368                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2369                        libcfs_nid2str(nid),
2370                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2371                        IBLND_MSG_SIZE);
2372                 goto failed;
2373         }
2374
2375         /* assume 'nid' is a new peer; create  */
2376         rc = kiblnd_create_peer(ni, &peer, nid);
2377         if (rc != 0) {
2378                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2379                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2380                 goto failed;
2381         }
2382
2383         /* We have validated the peer's parameters so use those */
2384         peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2385         peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2386
2387         write_lock_irqsave(g_lock, flags);
2388
2389         peer2 = kiblnd_find_peer_locked(nid);
2390         if (peer2 != NULL) {
2391                 if (peer2->ibp_version == 0) {
2392                         peer2->ibp_version     = version;
2393                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2394                 }
2395
2396                 /* not the guy I've talked with */
2397                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2398                     peer2->ibp_version     != version) {
2399                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2400
2401                         if (kiblnd_peer_active(peer2)) {
2402                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2403                                 peer2->ibp_version = version;
2404                         }
2405                         write_unlock_irqrestore(g_lock, flags);
2406
2407                         CWARN("Conn stale %s version %x/%x incarnation "LPU64"/"LPU64"\n",
2408                               libcfs_nid2str(nid), peer2->ibp_version, version,
2409                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2410
2411                         kiblnd_peer_decref(peer);
2412                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2413                         goto failed;
2414                 }
2415
2416                 /* tie-break connection race in favour of the higher NID */
2417                 if (peer2->ibp_connecting != 0 &&
2418                     nid < ni->ni_nid) {
2419                         write_unlock_irqrestore(g_lock, flags);
2420
2421                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2422
2423                         kiblnd_peer_decref(peer);
2424                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2425                         goto failed;
2426                 }
2427                 /*
2428                  * passive connection is allowed even this peer is waiting for
2429                  * reconnection.
2430                  */
2431                 peer2->ibp_reconnecting = 0;
2432                 peer2->ibp_accepting++;
2433                 kiblnd_peer_addref(peer2);
2434
2435                 /* Race with kiblnd_launch_tx (active connect) to create peer
2436                  * so copy validated parameters since we now know what the
2437                  * peer's limits are */
2438                 peer2->ibp_max_frags = peer->ibp_max_frags;
2439                 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2440
2441                 write_unlock_irqrestore(g_lock, flags);
2442                 kiblnd_peer_decref(peer);
2443                 peer = peer2;
2444         } else {
2445                 /* Brand new peer */
2446                 LASSERT (peer->ibp_accepting == 0);
2447                 LASSERT (peer->ibp_version == 0 &&
2448                          peer->ibp_incarnation == 0);
2449
2450                 peer->ibp_accepting   = 1;
2451                 peer->ibp_version     = version;
2452                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2453
2454                 /* I have a ref on ni that prevents it being shutdown */
2455                 LASSERT (net->ibn_shutdown == 0);
2456
2457                 kiblnd_peer_addref(peer);
2458                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2459
2460                 write_unlock_irqrestore(g_lock, flags);
2461         }
2462
2463         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2464         if (conn == NULL) {
2465                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2466                 kiblnd_peer_decref(peer);
2467                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2468                 goto failed;
2469         }
2470
2471         /* conn now "owns" cmid, so I return success from here on to ensure the
2472          * CM callback doesn't destroy cmid. */
2473         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2474         conn->ibc_credits          = conn->ibc_queue_depth;
2475         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2476         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2477                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2478
2479         ackmsg = &conn->ibc_connvars->cv_msg;
2480         memset(ackmsg, 0, sizeof(*ackmsg));
2481
2482         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2483                         sizeof(ackmsg->ibm_u.connparams));
2484         ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2485         ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2486         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2487
2488         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2489
2490         memset(&cp, 0, sizeof(cp));
2491         cp.private_data        = ackmsg;
2492         cp.private_data_len    = ackmsg->ibm_nob;
2493         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2494         cp.initiator_depth     = 0;
2495         cp.flow_control        = 1;
2496         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2497         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2498
2499         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2500
2501         rc = rdma_accept(cmid, &cp);
2502         if (rc != 0) {
2503                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2504                 rej.ibr_version = version;
2505                 rej.ibr_why     = IBLND_REJECT_FATAL;
2506
2507                 kiblnd_reject(cmid, &rej);
2508                 kiblnd_connreq_done(conn, rc);
2509                 kiblnd_conn_decref(conn);
2510         }
2511
2512         lnet_ni_decref(ni);
2513         return 0;
2514
2515  failed:
2516         if (ni != NULL)
2517                 lnet_ni_decref(ni);
2518
2519         rej.ibr_version = version;
2520         rej.ibr_cp.ibcp_queue_depth = kiblnd_msg_queue_size(version, ni);
2521         rej.ibr_cp.ibcp_max_frags   = kiblnd_rdma_frags(version, ni);
2522         kiblnd_reject(cmid, &rej);
2523
2524         return -ECONNREFUSED;
2525 }
2526
2527 static void
2528 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2529                        __u64 incarnation, int why, kib_connparams_t *cp)
2530 {
2531         rwlock_t        *glock = &kiblnd_data.kib_global_lock;
2532         kib_peer_t      *peer = conn->ibc_peer;
2533         char            *reason;
2534         int              msg_size = IBLND_MSG_SIZE;
2535         int              frag_num = -1;
2536         int              queue_dep = -1;
2537         bool             reconnect;
2538         unsigned long    flags;
2539
2540         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2541         LASSERT(peer->ibp_connecting > 0);      /* 'conn' at least */
2542         LASSERT(!peer->ibp_reconnecting);
2543
2544         if (cp) {
2545                 msg_size        = cp->ibcp_max_msg_size;
2546                 frag_num        = cp->ibcp_max_frags;
2547                 queue_dep       = cp->ibcp_queue_depth;
2548         }
2549
2550         write_lock_irqsave(glock, flags);
2551         /* retry connection if it's still needed and no other connection
2552          * attempts (active or passive) are in progress
2553          * NB: reconnect is still needed even when ibp_tx_queue is
2554          * empty if ibp_version != version because reconnect may be
2555          * initiated by kiblnd_query() */
2556         reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2557                      peer->ibp_version != version) &&
2558                     peer->ibp_connecting == 1 &&
2559                     peer->ibp_accepting == 0;
2560         if (!reconnect) {
2561                 reason = "no need";
2562                 goto out;
2563         }
2564
2565         switch (why) {
2566         default:
2567                 reason = "Unknown";
2568                 break;
2569
2570         case IBLND_REJECT_RDMA_FRAGS: {
2571                 struct lnet_ioctl_config_lnd_tunables *tunables;
2572
2573                 if (!cp) {
2574                         reason = "can't negotiate max frags";
2575                         goto out;
2576                 }
2577                 tunables = peer->ibp_ni->ni_lnd_tunables;
2578                 if (!tunables->lt_tun_u.lt_o2ib.lnd_map_on_demand) {
2579                         reason = "map_on_demand must be enabled";
2580                         goto out;
2581                 }
2582                 if (conn->ibc_max_frags <= frag_num) {
2583                         reason = "unsupported max frags";
2584                         goto out;
2585                 }
2586
2587                 peer->ibp_max_frags = frag_num;
2588                 reason = "rdma fragments";
2589                 break;
2590         }
2591         case IBLND_REJECT_MSG_QUEUE_SIZE:
2592                 if (!cp) {
2593                         reason = "can't negotiate queue depth";
2594                         goto out;
2595                 }
2596                 if (conn->ibc_queue_depth <= queue_dep) {
2597                         reason = "unsupported queue depth";
2598                         goto out;
2599                 }
2600
2601                 peer->ibp_queue_depth = queue_dep;
2602                 reason = "queue depth";
2603                 break;
2604
2605         case IBLND_REJECT_CONN_STALE:
2606                 reason = "stale";
2607                 break;
2608
2609         case IBLND_REJECT_CONN_RACE:
2610                 reason = "conn race";
2611                 break;
2612
2613         case IBLND_REJECT_CONN_UNCOMPAT:
2614                 reason = "version negotiation";
2615                 break;
2616         }
2617
2618         conn->ibc_reconnect = 1;
2619         peer->ibp_reconnecting = 1;
2620         peer->ibp_version = version;
2621         if (incarnation != 0)
2622                 peer->ibp_incarnation = incarnation;
2623  out:
2624         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2625
2626         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2627                 libcfs_nid2str(peer->ibp_nid),
2628                 reconnect ? "reconnect" : "don't reconnect",
2629                 reason, IBLND_MSG_VERSION, version, msg_size,
2630                 conn->ibc_queue_depth, queue_dep,
2631                 conn->ibc_max_frags, frag_num);
2632         /*
2633          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2634          * while destroying the zombie
2635          */
2636 }
2637
2638 static void
2639 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2640 {
2641         kib_peer_t    *peer = conn->ibc_peer;
2642
2643         LASSERT (!in_interrupt());
2644         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2645
2646         switch (reason) {
2647         case IB_CM_REJ_STALE_CONN:
2648                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2649                                        IBLND_REJECT_CONN_STALE, NULL);
2650                 break;
2651
2652         case IB_CM_REJ_INVALID_SERVICE_ID:
2653                 CNETERR("%s rejected: no listener at %d\n",
2654                         libcfs_nid2str(peer->ibp_nid),
2655                         *kiblnd_tunables.kib_service);
2656                 break;
2657
2658         case IB_CM_REJ_CONSUMER_DEFINED:
2659                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2660                         kib_rej_t        *rej         = priv;
2661                         kib_connparams_t *cp          = NULL;
2662                         int               flip        = 0;
2663                         __u64             incarnation = -1;
2664
2665                         /* NB. default incarnation is -1 because:
2666                          * a) V1 will ignore dst incarnation in connreq.
2667                          * b) V2 will provide incarnation while rejecting me,
2668                          *    -1 will be overwrote.
2669                          *
2670                          * if I try to connect to a V1 peer with V2 protocol,
2671                          * it rejected me then upgrade to V2, I have no idea
2672                          * about the upgrading and try to reconnect with V1,
2673                          * in this case upgraded V2 can find out I'm trying to
2674                          * talk to the old guy and reject me(incarnation is -1).
2675                          */
2676
2677                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2678                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2679                                 __swab32s(&rej->ibr_magic);
2680                                 __swab16s(&rej->ibr_version);
2681                                 flip = 1;
2682                         }
2683
2684                         if (priv_nob >= sizeof(kib_rej_t) &&
2685                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2686                                 /* priv_nob is always 148 in current version
2687                                  * of OFED, so we still need to check version.
2688                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2689                                 cp = &rej->ibr_cp;
2690
2691                                 if (flip) {
2692                                         __swab64s(&rej->ibr_incarnation);
2693                                         __swab16s(&cp->ibcp_queue_depth);
2694                                         __swab16s(&cp->ibcp_max_frags);
2695                                         __swab32s(&cp->ibcp_max_msg_size);
2696                                 }
2697
2698                                 incarnation = rej->ibr_incarnation;
2699                         }
2700
2701                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2702                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2703                                 CERROR("%s rejected: consumer defined fatal error\n",
2704                                        libcfs_nid2str(peer->ibp_nid));
2705                                 break;
2706                         }
2707
2708                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2709                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2710                                 CERROR("%s rejected: o2iblnd version %x error\n",
2711                                        libcfs_nid2str(peer->ibp_nid),
2712                                        rej->ibr_version);
2713                                 break;
2714                         }
2715
2716                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2717                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2718                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2719                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2720
2721                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2722                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2723                         }
2724
2725                         switch (rej->ibr_why) {
2726                         case IBLND_REJECT_CONN_RACE:
2727                         case IBLND_REJECT_CONN_STALE:
2728                         case IBLND_REJECT_CONN_UNCOMPAT:
2729                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2730                         case IBLND_REJECT_RDMA_FRAGS:
2731                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2732                                                 incarnation, rej->ibr_why, cp);
2733                                 break;
2734
2735                         case IBLND_REJECT_NO_RESOURCES:
2736                                 CERROR("%s rejected: o2iblnd no resources\n",
2737                                        libcfs_nid2str(peer->ibp_nid));
2738                                 break;
2739
2740                         case IBLND_REJECT_FATAL:
2741                                 CERROR("%s rejected: o2iblnd fatal error\n",
2742                                        libcfs_nid2str(peer->ibp_nid));
2743                                 break;
2744
2745                         default:
2746                                 CERROR("%s rejected: o2iblnd reason %d\n",
2747                                        libcfs_nid2str(peer->ibp_nid),
2748                                        rej->ibr_why);
2749                                 break;
2750                         }
2751                         break;
2752                 }
2753                 /* fall through */
2754         default:
2755                 CNETERR("%s rejected: reason %d, size %d\n",
2756                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2757                 break;
2758         }
2759
2760         kiblnd_connreq_done(conn, -ECONNREFUSED);
2761 }
2762
2763 static void
2764 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2765 {
2766         kib_peer_t    *peer = conn->ibc_peer;
2767         lnet_ni_t     *ni   = peer->ibp_ni;
2768         kib_net_t     *net  = ni->ni_data;
2769         kib_msg_t     *msg  = priv;
2770         int            ver  = conn->ibc_version;
2771         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2772         unsigned long  flags;
2773
2774         LASSERT (net != NULL);
2775
2776         if (rc != 0) {
2777                 CERROR("Can't unpack connack from %s: %d\n",
2778                        libcfs_nid2str(peer->ibp_nid), rc);
2779                 goto failed;
2780         }
2781
2782         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2783                 CERROR("Unexpected message %d from %s\n",
2784                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2785                 rc = -EPROTO;
2786                 goto failed;
2787         }
2788
2789         if (ver != msg->ibm_version) {
2790                 CERROR("%s replied version %x is different with "
2791                        "requested version %x\n",
2792                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2793                 rc = -EPROTO;
2794                 goto failed;
2795         }
2796
2797         if (msg->ibm_u.connparams.ibcp_queue_depth >
2798             conn->ibc_queue_depth) {
2799                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2800                        libcfs_nid2str(peer->ibp_nid),
2801                        msg->ibm_u.connparams.ibcp_queue_depth,
2802                        conn->ibc_queue_depth);
2803                 rc = -EPROTO;
2804                 goto failed;
2805         }
2806
2807         if (msg->ibm_u.connparams.ibcp_max_frags >
2808             conn->ibc_max_frags) {
2809                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2810                        libcfs_nid2str(peer->ibp_nid),
2811                        msg->ibm_u.connparams.ibcp_max_frags,
2812                        conn->ibc_max_frags);
2813                 rc = -EPROTO;
2814                 goto failed;
2815         }
2816
2817         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2818                 CERROR("%s max message size %d too big (%d max)\n",
2819                        libcfs_nid2str(peer->ibp_nid),
2820                        msg->ibm_u.connparams.ibcp_max_msg_size,
2821                        IBLND_MSG_SIZE);
2822                 rc = -EPROTO;
2823                 goto failed;
2824         }
2825
2826         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2827         if (msg->ibm_dstnid == ni->ni_nid &&
2828             msg->ibm_dststamp == net->ibn_incarnation)
2829                 rc = 0;
2830         else
2831                 rc = -ESTALE;
2832         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2833
2834         if (rc != 0) {
2835                 CERROR("Bad connection reply from %s, rc = %d, "
2836                        "version: %x max_frags: %d\n",
2837                        libcfs_nid2str(peer->ibp_nid), rc,
2838                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2839                 goto failed;
2840         }
2841
2842         conn->ibc_incarnation      = msg->ibm_srcstamp;
2843         conn->ibc_credits          = msg->ibm_u.connparams.ibcp_queue_depth;
2844         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2845         conn->ibc_queue_depth      = msg->ibm_u.connparams.ibcp_queue_depth;
2846         conn->ibc_max_frags        = msg->ibm_u.connparams.ibcp_max_frags;
2847         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2848                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2849
2850         kiblnd_connreq_done(conn, 0);
2851         return;
2852
2853  failed:
2854         /* NB My QP has already established itself, so I handle anything going
2855          * wrong here by setting ibc_comms_error.
2856          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2857          * immediately tears it down. */
2858
2859         LASSERT (rc != 0);
2860         conn->ibc_comms_error = rc;
2861         kiblnd_connreq_done(conn, 0);
2862 }
2863
2864 static int
2865 kiblnd_active_connect (struct rdma_cm_id *cmid)
2866 {
2867         kib_peer_t              *peer = (kib_peer_t *)cmid->context;
2868         kib_conn_t              *conn;
2869         kib_msg_t               *msg;
2870         struct rdma_conn_param   cp;
2871         int                      version;
2872         __u64                    incarnation;
2873         unsigned long            flags;
2874         int                      rc;
2875
2876         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2877
2878         incarnation = peer->ibp_incarnation;
2879         version     = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
2880                                                  peer->ibp_version;
2881
2882         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2883
2884         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2885                                   version);
2886         if (conn == NULL) {
2887                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2888                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2889                 return -ENOMEM;
2890         }
2891
2892         /* conn "owns" cmid now, so I return success from here on to ensure the
2893          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2894          * on peer */
2895
2896         msg = &conn->ibc_connvars->cv_msg;
2897
2898         memset(msg, 0, sizeof(*msg));
2899         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2900         msg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
2901         msg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
2902         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2903
2904         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2905                         0, peer->ibp_nid, incarnation);
2906
2907         memset(&cp, 0, sizeof(cp));
2908         cp.private_data        = msg;
2909         cp.private_data_len    = msg->ibm_nob;
2910         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2911         cp.initiator_depth     = 0;
2912         cp.flow_control        = 1;
2913         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2914         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2915
2916         LASSERT(cmid->context == (void *)conn);
2917         LASSERT(conn->ibc_cmid == cmid);
2918
2919         rc = rdma_connect(cmid, &cp);
2920         if (rc != 0) {
2921                 CERROR("Can't connect to %s: %d\n",
2922                        libcfs_nid2str(peer->ibp_nid), rc);
2923                 kiblnd_connreq_done(conn, rc);
2924                 kiblnd_conn_decref(conn);
2925         }
2926
2927         return 0;
2928 }
2929
2930 int
2931 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2932 {
2933         kib_peer_t  *peer;
2934         kib_conn_t  *conn;
2935         int          rc;
2936
2937         switch (event->event) {
2938         default:
2939                 CERROR("Unexpected event: %d, status: %d\n",
2940                        event->event, event->status);
2941                 LBUG();
2942
2943         case RDMA_CM_EVENT_CONNECT_REQUEST:
2944                 /* destroy cmid on failure */
2945                 rc = kiblnd_passive_connect(cmid,
2946                                             (void *)KIBLND_CONN_PARAM(event),
2947                                             KIBLND_CONN_PARAM_LEN(event));
2948                 CDEBUG(D_NET, "connreq: %d\n", rc);
2949                 return rc;
2950
2951         case RDMA_CM_EVENT_ADDR_ERROR:
2952                 peer = (kib_peer_t *)cmid->context;
2953                 CNETERR("%s: ADDR ERROR %d\n",
2954                        libcfs_nid2str(peer->ibp_nid), event->status);
2955                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2956                 kiblnd_peer_decref(peer);
2957                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
2958
2959         case RDMA_CM_EVENT_ADDR_RESOLVED:
2960                 peer = (kib_peer_t *)cmid->context;
2961
2962                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2963                        libcfs_nid2str(peer->ibp_nid), event->status);
2964
2965                 if (event->status != 0) {
2966                         CNETERR("Can't resolve address for %s: %d\n",
2967                                 libcfs_nid2str(peer->ibp_nid), event->status);
2968                         rc = event->status;
2969                 } else {
2970                         rc = rdma_resolve_route(
2971                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2972                         if (rc == 0)
2973                                 return 0;
2974                         /* Can't initiate route resolution */
2975                         CERROR("Can't resolve route for %s: %d\n",
2976                                libcfs_nid2str(peer->ibp_nid), rc);
2977                 }
2978                 kiblnd_peer_connect_failed(peer, 1, rc);
2979                 kiblnd_peer_decref(peer);
2980                 return rc;                      /* rc != 0 destroys cmid */
2981
2982         case RDMA_CM_EVENT_ROUTE_ERROR:
2983                 peer = (kib_peer_t *)cmid->context;
2984                 CNETERR("%s: ROUTE ERROR %d\n",
2985                         libcfs_nid2str(peer->ibp_nid), event->status);
2986                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2987                 kiblnd_peer_decref(peer);
2988                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
2989
2990         case RDMA_CM_EVENT_ROUTE_RESOLVED:
2991                 peer = (kib_peer_t *)cmid->context;
2992                 CDEBUG(D_NET,"%s Route resolved: %d\n",
2993                        libcfs_nid2str(peer->ibp_nid), event->status);
2994
2995                 if (event->status == 0)
2996                         return kiblnd_active_connect(cmid);
2997
2998                 CNETERR("Can't resolve route for %s: %d\n",
2999                        libcfs_nid2str(peer->ibp_nid), event->status);
3000                 kiblnd_peer_connect_failed(peer, 1, event->status);
3001                 kiblnd_peer_decref(peer);
3002                 return event->status;           /* rc != 0 destroys cmid */
3003
3004         case RDMA_CM_EVENT_UNREACHABLE:
3005                 conn = (kib_conn_t *)cmid->context;
3006                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3007                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3008                 CNETERR("%s: UNREACHABLE %d\n",
3009                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3010                 kiblnd_connreq_done(conn, -ENETDOWN);
3011                 kiblnd_conn_decref(conn);
3012                 return 0;
3013
3014         case RDMA_CM_EVENT_CONNECT_ERROR:
3015                 conn = (kib_conn_t *)cmid->context;
3016                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3017                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3018                 CNETERR("%s: CONNECT ERROR %d\n",
3019                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3020                 kiblnd_connreq_done(conn, -ENOTCONN);
3021                 kiblnd_conn_decref(conn);
3022                 return 0;
3023
3024         case RDMA_CM_EVENT_REJECTED:
3025                 conn = (kib_conn_t *)cmid->context;
3026                 switch (conn->ibc_state) {
3027                 default:
3028                         LBUG();
3029
3030                 case IBLND_CONN_PASSIVE_WAIT:
3031                         CERROR ("%s: REJECTED %d\n",
3032                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
3033                                 event->status);
3034                         kiblnd_connreq_done(conn, -ECONNRESET);
3035                         break;
3036
3037                 case IBLND_CONN_ACTIVE_CONNECT:
3038                         kiblnd_rejected(conn, event->status,
3039                                         (void *)KIBLND_CONN_PARAM(event),
3040                                         KIBLND_CONN_PARAM_LEN(event));
3041                         break;
3042                 }
3043                 kiblnd_conn_decref(conn);
3044                 return 0;
3045
3046         case RDMA_CM_EVENT_ESTABLISHED:
3047                 conn = (kib_conn_t *)cmid->context;
3048                 switch (conn->ibc_state) {
3049                 default:
3050                         LBUG();
3051
3052                 case IBLND_CONN_PASSIVE_WAIT:
3053                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3054                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3055                         kiblnd_connreq_done(conn, 0);
3056                         break;
3057
3058                 case IBLND_CONN_ACTIVE_CONNECT:
3059                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3060                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3061                         kiblnd_check_connreply(conn,
3062                                                (void *)KIBLND_CONN_PARAM(event),
3063                                                KIBLND_CONN_PARAM_LEN(event));
3064                         break;
3065                 }
3066                 /* net keeps its ref on conn! */
3067                 return 0;
3068
3069         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3070                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3071                 return 0;
3072
3073         case RDMA_CM_EVENT_DISCONNECTED:
3074                 conn = (kib_conn_t *)cmid->context;
3075                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3076                         CERROR("%s DISCONNECTED\n",
3077                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3078                         kiblnd_connreq_done(conn, -ECONNRESET);
3079                 } else {
3080                         kiblnd_close_conn(conn, 0);
3081                 }
3082                 kiblnd_conn_decref(conn);
3083                 cmid->context = NULL;
3084                 return 0;
3085
3086         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3087                 LCONSOLE_ERROR_MSG(0x131,
3088                                    "Received notification of device removal\n"
3089                                    "Please shutdown LNET to allow this to proceed\n");
3090                 /* Can't remove network from underneath LNET for now, so I have
3091                  * to ignore this */
3092                 return 0;
3093
3094         case RDMA_CM_EVENT_ADDR_CHANGE:
3095                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3096                 return 0;
3097         }
3098 }
3099
3100 static int
3101 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3102 {
3103         kib_tx_t         *tx;
3104         struct list_head *ttmp;
3105
3106         list_for_each(ttmp, txs) {
3107                 tx = list_entry(ttmp, kib_tx_t, tx_list);
3108
3109                 if (txs != &conn->ibc_active_txs) {
3110                         LASSERT(tx->tx_queued);
3111                 } else {
3112                         LASSERT(!tx->tx_queued);
3113                         LASSERT(tx->tx_waiting || tx->tx_sending != 0);
3114                 }
3115
3116                 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3117                         CERROR("Timed out tx: %s, %lu seconds\n",
3118                                kiblnd_queue2str(conn, txs),
3119                                cfs_duration_sec(jiffies - tx->tx_deadline));
3120                         return 1;
3121                 }
3122         }
3123
3124         return 0;
3125 }
3126
3127 static int
3128 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3129 {
3130         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3131                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3132                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3133                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3134                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3135 }
3136
3137 static void
3138 kiblnd_check_conns (int idx)
3139 {
3140         struct list_head  closes = LIST_HEAD_INIT(closes);
3141         struct list_head  checksends = LIST_HEAD_INIT(checksends);
3142         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3143         struct list_head *ptmp;
3144         kib_peer_t       *peer;
3145         kib_conn_t       *conn;
3146         struct list_head *ctmp;
3147         unsigned long     flags;
3148
3149         /* NB. We expect to have a look at all the peers and not find any
3150          * RDMAs to time out, so we just use a shared lock while we
3151          * take a look... */
3152         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3153
3154         list_for_each(ptmp, peers) {
3155                 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3156
3157                 list_for_each(ctmp, &peer->ibp_conns) {
3158                         int timedout;
3159                         int sendnoop;
3160
3161                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
3162
3163                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3164
3165                         spin_lock(&conn->ibc_lock);
3166
3167                         sendnoop = kiblnd_need_noop(conn);
3168                         timedout = kiblnd_conn_timed_out_locked(conn);
3169                         if (!sendnoop && !timedout) {
3170                                 spin_unlock(&conn->ibc_lock);
3171                                 continue;
3172                         }
3173
3174                         if (timedout) {
3175                                 CERROR("Timed out RDMA with %s (%lu): "
3176                                        "c: %u, oc: %u, rc: %u\n",
3177                                        libcfs_nid2str(peer->ibp_nid),
3178                                        cfs_duration_sec(cfs_time_current() -
3179                                                         peer->ibp_last_alive),
3180                                        conn->ibc_credits,
3181                                        conn->ibc_outstanding_credits,
3182                                        conn->ibc_reserved_credits);
3183                                 list_add(&conn->ibc_connd_list, &closes);
3184                         } else {
3185                                 list_add(&conn->ibc_connd_list, &checksends);
3186                         }
3187                         /* +ref for 'closes' or 'checksends' */
3188                         kiblnd_conn_addref(conn);
3189
3190                         spin_unlock(&conn->ibc_lock);
3191                 }
3192         }
3193
3194         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3195
3196         /* Handle timeout by closing the whole
3197          * connection. We can only be sure RDMA activity
3198          * has ceased once the QP has been modified. */
3199         while (!list_empty(&closes)) {
3200                 conn = list_entry(closes.next,
3201                                   kib_conn_t, ibc_connd_list);
3202                 list_del(&conn->ibc_connd_list);
3203                 kiblnd_close_conn(conn, -ETIMEDOUT);
3204                 kiblnd_conn_decref(conn);
3205         }
3206
3207         /* In case we have enough credits to return via a
3208          * NOOP, but there were no non-blocking tx descs
3209          * free to do it last time... */
3210         while (!list_empty(&checksends)) {
3211                 conn = list_entry(checksends.next,
3212                                   kib_conn_t, ibc_connd_list);
3213                 list_del(&conn->ibc_connd_list);
3214                 kiblnd_check_sends(conn);
3215                 kiblnd_conn_decref(conn);
3216         }
3217 }
3218
3219 static void
3220 kiblnd_disconnect_conn (kib_conn_t *conn)
3221 {
3222         LASSERT (!in_interrupt());
3223         LASSERT (current == kiblnd_data.kib_connd);
3224         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
3225
3226         rdma_disconnect(conn->ibc_cmid);
3227         kiblnd_finalise_conn(conn);
3228
3229         kiblnd_peer_notify(conn->ibc_peer);
3230 }
3231
3232 /*
3233  * High-water for reconnection to the same peer, reconnection attempt should
3234  * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3235  */
3236 #define KIB_RECONN_HIGH_RACE    10
3237 /*
3238  * Allow connd to take a break and handle other things after consecutive
3239  * reconnection attemps.
3240  */
3241 #define KIB_RECONN_BREAK        100
3242
3243 int
3244 kiblnd_connd (void *arg)
3245 {
3246         spinlock_t        *lock= &kiblnd_data.kib_connd_lock;
3247         wait_queue_t       wait;
3248         unsigned long      flags;
3249         kib_conn_t        *conn;
3250         int                timeout;
3251         int                i;
3252         int                dropped_lock;
3253         int                peer_index = 0;
3254         unsigned long      deadline = jiffies;
3255
3256         cfs_block_allsigs();
3257
3258         init_waitqueue_entry(&wait, current);
3259         kiblnd_data.kib_connd = current;
3260
3261         spin_lock_irqsave(lock, flags);
3262
3263         while (!kiblnd_data.kib_shutdown) {
3264                 int reconn = 0;
3265
3266                 dropped_lock = 0;
3267
3268                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3269                         kib_peer_t *peer = NULL;
3270
3271                         conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3272                                           kib_conn_t, ibc_list);
3273                         list_del(&conn->ibc_list);
3274                         if (conn->ibc_reconnect) {
3275                                 peer = conn->ibc_peer;
3276                                 kiblnd_peer_addref(peer);
3277                         }
3278
3279                         spin_unlock_irqrestore(lock, flags);
3280                         dropped_lock = 1;
3281
3282                         kiblnd_destroy_conn(conn, !peer);
3283
3284                         spin_lock_irqsave(lock, flags);
3285                         if (!peer)
3286                                 continue;
3287
3288                         conn->ibc_peer = peer;
3289                         if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3290                                 list_add_tail(&conn->ibc_list,
3291                                               &kiblnd_data.kib_reconn_list);
3292                         else
3293                                 list_add_tail(&conn->ibc_list,
3294                                               &kiblnd_data.kib_reconn_wait);
3295                 }
3296
3297                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3298                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3299                                               kib_conn_t, ibc_list);
3300                         list_del(&conn->ibc_list);
3301
3302                         spin_unlock_irqrestore(lock, flags);
3303                         dropped_lock = 1;
3304
3305                         kiblnd_disconnect_conn(conn);
3306                         kiblnd_conn_decref(conn);
3307
3308                         spin_lock_irqsave(lock, flags);
3309                 }
3310
3311                 while (reconn < KIB_RECONN_BREAK) {
3312                         if (kiblnd_data.kib_reconn_sec != get_seconds()) {
3313                                 kiblnd_data.kib_reconn_sec = get_seconds();
3314                                 list_splice_init(&kiblnd_data.kib_reconn_wait,
3315                                                  &kiblnd_data.kib_reconn_list);
3316                         }
3317
3318                         if (list_empty(&kiblnd_data.kib_reconn_list))
3319                                 break;
3320
3321                         conn = list_entry(kiblnd_data.kib_reconn_list.next,
3322                                           kib_conn_t, ibc_list);
3323                         list_del(&conn->ibc_list);
3324
3325                         spin_unlock_irqrestore(lock, flags);
3326                         dropped_lock = 1;
3327
3328                         reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3329                         kiblnd_peer_decref(conn->ibc_peer);
3330                         LIBCFS_FREE(conn, sizeof(*conn));
3331
3332                         spin_lock_irqsave(lock, flags);
3333                 }
3334
3335                 /* careful with the jiffy wrap... */
3336                 timeout = (int)(deadline - jiffies);
3337                 if (timeout <= 0) {
3338                         const int n = 4;
3339                         const int p = 1;
3340                         int       chunk = kiblnd_data.kib_peer_hash_size;
3341
3342                         spin_unlock_irqrestore(lock, flags);
3343                         dropped_lock = 1;
3344
3345                         /* Time to check for RDMA timeouts on a few more
3346                          * peers: I do checks every 'p' seconds on a
3347                          * proportion of the peer table and I need to check
3348                          * every connection 'n' times within a timeout
3349                          * interval, to ensure I detect a timeout on any
3350                          * connection within (n+1)/n times the timeout
3351                          * interval. */
3352
3353                         if (*kiblnd_tunables.kib_timeout > n * p)
3354                                 chunk = (chunk * n * p) /
3355                                         *kiblnd_tunables.kib_timeout;
3356                         if (chunk == 0)
3357                                 chunk = 1;
3358
3359                         for (i = 0; i < chunk; i++) {
3360                                 kiblnd_check_conns(peer_index);
3361                                 peer_index = (peer_index + 1) %
3362                                              kiblnd_data.kib_peer_hash_size;
3363                         }
3364
3365                         deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3366                         spin_lock_irqsave(lock, flags);
3367                 }
3368
3369                 if (dropped_lock)
3370                         continue;
3371
3372                 /* Nothing to do for 'timeout'  */
3373                 set_current_state(TASK_INTERRUPTIBLE);
3374                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3375                 spin_unlock_irqrestore(lock, flags);
3376
3377                 schedule_timeout(timeout);
3378
3379                 set_current_state(TASK_RUNNING);
3380                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3381                 spin_lock_irqsave(lock, flags);
3382         }
3383
3384         spin_unlock_irqrestore(lock, flags);
3385
3386         kiblnd_thread_fini();
3387         return 0;
3388 }
3389
3390 void
3391 kiblnd_qp_event(struct ib_event *event, void *arg)
3392 {
3393         kib_conn_t *conn = arg;
3394
3395         switch (event->event) {
3396         case IB_EVENT_COMM_EST:
3397                 CDEBUG(D_NET, "%s established\n",
3398                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3399                 return;
3400
3401         default:
3402                 CERROR("%s: Async QP event type %d\n",
3403                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3404                 return;
3405         }
3406 }
3407
3408 static void
3409 kiblnd_complete (struct ib_wc *wc)
3410 {
3411         switch (kiblnd_wreqid2type(wc->wr_id)) {
3412         default:
3413                 LBUG();
3414
3415         case IBLND_WID_MR:
3416                 if (wc->status != IB_WC_SUCCESS &&
3417                     wc->status != IB_WC_WR_FLUSH_ERR)
3418                         CNETERR("FastReg failed: %d\n", wc->status);
3419                 return;
3420
3421         case IBLND_WID_RDMA:
3422                 /* We only get RDMA completion notification if it fails.  All
3423                  * subsequent work items, including the final SEND will fail
3424                  * too.  However we can't print out any more info about the
3425                  * failing RDMA because 'tx' might be back on the idle list or
3426                  * even reused already if we didn't manage to post all our work
3427                  * items */
3428                 CNETERR("RDMA (tx: %p) failed: %d\n",
3429                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3430                 return;
3431
3432         case IBLND_WID_TX:
3433                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3434                 return;
3435
3436         case IBLND_WID_RX:
3437                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3438                                    wc->byte_len);
3439                 return;
3440         }
3441 }
3442
3443 void
3444 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3445 {
3446         /* NB I'm not allowed to schedule this conn once its refcount has
3447          * reached 0.  Since fundamentally I'm racing with scheduler threads
3448          * consuming my CQ I could be called after all completions have
3449          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3450          * and this CQ is about to be destroyed so I NOOP. */
3451         kib_conn_t              *conn = (kib_conn_t *)arg;
3452         struct kib_sched_info   *sched = conn->ibc_sched;
3453         unsigned long           flags;
3454
3455         LASSERT(cq == conn->ibc_cq);
3456
3457         spin_lock_irqsave(&sched->ibs_lock, flags);
3458
3459         conn->ibc_ready = 1;
3460
3461         if (!conn->ibc_scheduled &&
3462             (conn->ibc_nrx > 0 ||
3463              conn->ibc_nsends_posted > 0)) {
3464                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3465                 conn->ibc_scheduled = 1;
3466                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3467
3468                 if (waitqueue_active(&sched->ibs_waitq))
3469                         wake_up(&sched->ibs_waitq);
3470         }
3471
3472         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3473 }
3474
3475 void
3476 kiblnd_cq_event(struct ib_event *event, void *arg)
3477 {
3478         kib_conn_t *conn = arg;
3479
3480         CERROR("%s: async CQ event type %d\n",
3481                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3482 }
3483
3484 int
3485 kiblnd_scheduler(void *arg)
3486 {
3487         long                    id = (long)arg;
3488         struct kib_sched_info   *sched;
3489         kib_conn_t              *conn;
3490         wait_queue_t            wait;
3491         unsigned long           flags;
3492         struct ib_wc            wc;
3493         int                     did_something;
3494         int                     busy_loops = 0;
3495         int                     rc;
3496
3497         cfs_block_allsigs();
3498
3499         init_waitqueue_entry(&wait, current);
3500
3501         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3502
3503         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3504         if (rc != 0) {
3505                 CWARN("Failed to bind on CPT %d, please verify whether "
3506                       "all CPUs are healthy and reload modules if necessary, "
3507                       "otherwise your system might under risk of low "
3508                       "performance\n", sched->ibs_cpt);
3509         }
3510
3511         spin_lock_irqsave(&sched->ibs_lock, flags);
3512
3513         while (!kiblnd_data.kib_shutdown) {
3514                 if (busy_loops++ >= IBLND_RESCHED) {
3515                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3516
3517                         cond_resched();
3518                         busy_loops = 0;
3519
3520                         spin_lock_irqsave(&sched->ibs_lock, flags);
3521                 }
3522
3523                 did_something = 0;
3524
3525                 if (!list_empty(&sched->ibs_conns)) {
3526                         conn = list_entry(sched->ibs_conns.next,
3527                                               kib_conn_t, ibc_sched_list);
3528                         /* take over kib_sched_conns' ref on conn... */
3529                         LASSERT(conn->ibc_scheduled);
3530                         list_del(&conn->ibc_sched_list);
3531                         conn->ibc_ready = 0;
3532
3533                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3534
3535                         wc.wr_id = IBLND_WID_INVAL;
3536
3537                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3538                         if (rc == 0) {
3539                                 rc = ib_req_notify_cq(conn->ibc_cq,
3540                                                       IB_CQ_NEXT_COMP);
3541                                 if (rc < 0) {
3542                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3543                                               "closing connection\n",
3544                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3545                                         kiblnd_close_conn(conn, -EIO);
3546                                         kiblnd_conn_decref(conn);
3547                                         spin_lock_irqsave(&sched->ibs_lock,
3548                                                               flags);
3549                                         continue;
3550                                 }
3551
3552                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3553                         }
3554
3555                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3556                                 LCONSOLE_ERROR(
3557                                         "ib_poll_cq (rc: %d) returned invalid "
3558                                         "wr_id, opcode %d, status: %d, "
3559                                         "vendor_err: %d, conn: %s status: %d\n"
3560                                         "please upgrade firmware and OFED or "
3561                                         "contact vendor.\n", rc,
3562                                         wc.opcode, wc.status, wc.vendor_err,
3563                                         libcfs_nid2str(conn->ibc_peer->ibp_nid),
3564                                         conn->ibc_state);
3565                                 rc = -EINVAL;
3566                         }
3567
3568                         if (rc < 0) {
3569                                 CWARN("%s: ib_poll_cq failed: %d, "
3570                                       "closing connection\n",
3571                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3572                                       rc);
3573                                 kiblnd_close_conn(conn, -EIO);
3574                                 kiblnd_conn_decref(conn);
3575                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3576                                 continue;
3577                         }
3578
3579                         spin_lock_irqsave(&sched->ibs_lock, flags);
3580
3581                         if (rc != 0 || conn->ibc_ready) {
3582                                 /* There may be another completion waiting; get
3583                                  * another scheduler to check while I handle
3584                                  * this one... */
3585                                 /* +1 ref for sched_conns */
3586                                 kiblnd_conn_addref(conn);
3587                                 list_add_tail(&conn->ibc_sched_list,
3588                                                   &sched->ibs_conns);
3589                                 if (waitqueue_active(&sched->ibs_waitq))
3590                                         wake_up(&sched->ibs_waitq);
3591                         } else {
3592                                 conn->ibc_scheduled = 0;
3593                         }
3594
3595                         if (rc != 0) {
3596                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3597                                 kiblnd_complete(&wc);
3598
3599                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3600                         }
3601
3602                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3603                         did_something = 1;
3604                 }
3605
3606                 if (did_something)
3607                         continue;
3608
3609                 set_current_state(TASK_INTERRUPTIBLE);
3610                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3611                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3612
3613                 schedule();
3614                 busy_loops = 0;
3615
3616                 remove_wait_queue(&sched->ibs_waitq, &wait);
3617                 set_current_state(TASK_RUNNING);
3618                 spin_lock_irqsave(&sched->ibs_lock, flags);
3619         }
3620
3621         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3622
3623         kiblnd_thread_fini();
3624         return 0;
3625 }
3626
3627 int
3628 kiblnd_failover_thread(void *arg)
3629 {
3630         rwlock_t                *glock = &kiblnd_data.kib_global_lock;
3631         kib_dev_t               *dev;
3632         wait_queue_t            wait;
3633         unsigned long           flags;
3634         int                     rc;
3635
3636         LASSERT(*kiblnd_tunables.kib_dev_failover != 0);
3637
3638         cfs_block_allsigs();
3639
3640         init_waitqueue_entry(&wait, current);
3641         write_lock_irqsave(glock, flags);
3642
3643         while (!kiblnd_data.kib_shutdown) {
3644                 int     do_failover = 0;
3645                 int     long_sleep;
3646
3647                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3648                                     ibd_fail_list) {
3649                         if (cfs_time_before(cfs_time_current(),
3650                                             dev->ibd_next_failover))
3651                                 continue;
3652                         do_failover = 1;
3653                         break;
3654                 }
3655
3656                 if (do_failover) {
3657                         list_del_init(&dev->ibd_fail_list);
3658                         dev->ibd_failover = 1;
3659                         write_unlock_irqrestore(glock, flags);
3660
3661                         rc = kiblnd_dev_failover(dev);
3662
3663                         write_lock_irqsave(glock, flags);
3664
3665                         LASSERT (dev->ibd_failover);
3666                         dev->ibd_failover = 0;
3667                         if (rc >= 0) { /* Device is OK or failover succeed */
3668                                 dev->ibd_next_failover = cfs_time_shift(3);
3669                                 continue;
3670                         }
3671
3672                         /* failed to failover, retry later */
3673                         dev->ibd_next_failover =
3674                                 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3675                         if (kiblnd_dev_can_failover(dev)) {
3676                                 list_add_tail(&dev->ibd_fail_list,
3677                                               &kiblnd_data.kib_failed_devs);
3678                         }
3679
3680                         continue;
3681                 }
3682
3683                 /* long sleep if no more pending failover */
3684                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3685
3686                 set_current_state(TASK_INTERRUPTIBLE);
3687                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3688                 write_unlock_irqrestore(glock, flags);
3689
3690                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3691                                                    cfs_time_seconds(1));
3692                 set_current_state(TASK_RUNNING);
3693                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3694                 write_lock_irqsave(glock, flags);
3695
3696                 if (!long_sleep || rc != 0)
3697                         continue;
3698
3699                 /* have a long sleep, routine check all active devices,
3700                  * we need checking like this because if there is not active
3701                  * connection on the dev and no SEND from local, we may listen
3702                  * on wrong HCA for ever while there is a bonding failover */
3703                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3704                         if (kiblnd_dev_can_failover(dev)) {
3705                                 list_add_tail(&dev->ibd_fail_list,
3706                                               &kiblnd_data.kib_failed_devs);
3707                         }
3708                 }
3709         }
3710
3711         write_unlock_irqrestore(glock, flags);
3712
3713         kiblnd_thread_fini();
3714         return 0;
3715 }