lnet/klnds/o2iblnd/o2iblnd_cb.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 void
  44 kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
  45 {
  46         lnet_msg_t *lntmsg[2];
  47         kib_net_t  *net = ni->ni_data;
  48         int         rc;
  49         int         i;
  50
  51         LASSERT (net != NULL);
  52         LASSERT (!in_interrupt());
  53         LASSERT (!tx->tx_queued);               /* mustn't be queued for sending */
  54         LASSERT (tx->tx_sending == 0);          /* mustn't be awaiting sent callback */
  55         LASSERT (!tx->tx_waiting);              /* mustn't be awaiting peer response */
  56         LASSERT (tx->tx_pool != NULL);
  57
  58         kiblnd_unmap_tx(ni, tx);
  59
  60         /* tx may have up to 2 lnet msgs to finalise */
  61         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  62         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  63         rc = tx->tx_status;
  64
  65         if (tx->tx_conn != NULL) {
  66                 LASSERT (ni == tx->tx_conn->ibc_peer->ibp_ni);
  67
  68                 kiblnd_conn_decref(tx->tx_conn);
  69                 tx->tx_conn = NULL;
  70         }
  71
  72         tx->tx_nwrq = 0;
  73         tx->tx_status = 0;
  74
  75         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  76
  77         /* delay finalize until my descs have been freed */
  78         for (i = 0; i < 2; i++) {
  79                 if (lntmsg[i] == NULL)
  80                         continue;
  81
  82                 lnet_finalize(ni, lntmsg[i], rc);
  83         }
  84 }
  85
  86 void
  87 kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status)
  88 {
  89         kib_tx_t *tx;
  90
  91         while (!list_empty (txlist)) {
  92                 tx = list_entry (txlist->next, kib_tx_t, tx_list);
  93
  94                 list_del(&tx->tx_list);
  95                 /* complete now */
  96                 tx->tx_waiting = 0;
  97                 tx->tx_status = status;
  98                 kiblnd_tx_done(ni, tx);
  99         }
 100 }
 101
 102 kib_tx_t *
 103 kiblnd_get_idle_tx (lnet_ni_t *ni)
 104 {
 105         kib_net_t        *net = (kib_net_t *)ni->ni_data;
 106         struct list_head *node;
 107         kib_tx_t         *tx;
 108
 109         node = kiblnd_pool_alloc_node(&net->ibn_tx_ps.tps_poolset);
 110         if (node == NULL)
 111                 return NULL;
 112         tx = container_of(node, kib_tx_t, tx_list);
 113
 114         LASSERT (tx->tx_nwrq == 0);
 115         LASSERT (!tx->tx_queued);
 116         LASSERT (tx->tx_sending == 0);
 117         LASSERT (!tx->tx_waiting);
 118         LASSERT (tx->tx_status == 0);
 119         LASSERT (tx->tx_conn == NULL);
 120         LASSERT (tx->tx_lntmsg[0] == NULL);
 121         LASSERT (tx->tx_lntmsg[1] == NULL);
 122         LASSERT (tx->tx_u.pmr == NULL);
 123         LASSERT (tx->tx_nfrags == 0);
 124
 125         return tx;
 126 }
 127
 128 void
 129 kiblnd_drop_rx (kib_rx_t *rx)
 130 {
 131         kib_conn_t         *conn = rx->rx_conn;
 132         unsigned long       flags;
 133
 134         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
 135         LASSERT (conn->ibc_nrx > 0);
 136         conn->ibc_nrx--;
 137         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
 138
 139         kiblnd_conn_decref(conn);
 140 }
 141
 142 int
 143 kiblnd_post_rx (kib_rx_t *rx, int credit)
 144 {
 145         kib_conn_t         *conn = rx->rx_conn;
 146         kib_net_t          *net = conn->ibc_peer->ibp_ni->ni_data;
 147         struct ib_recv_wr  *bad_wrq = NULL;
 148         struct ib_mr       *mr;
 149         int                 rc;
 150
 151         LASSERT (net != NULL);
 152         LASSERT (!in_interrupt());
 153         LASSERT (credit == IBLND_POSTRX_NO_CREDIT ||
 154                  credit == IBLND_POSTRX_PEER_CREDIT ||
 155                  credit == IBLND_POSTRX_RSRVD_CREDIT);
 156
 157         mr = kiblnd_find_dma_mr(net, rx->rx_msgaddr, IBLND_MSG_SIZE);
 158         LASSERT (mr != NULL);
 159
 160         rx->rx_sge.lkey   = mr->lkey;
 161         rx->rx_sge.addr   = rx->rx_msgaddr;
 162         rx->rx_sge.length = IBLND_MSG_SIZE;
 163
 164         rx->rx_wrq.next = NULL;
 165         rx->rx_wrq.sg_list = &rx->rx_sge;
 166         rx->rx_wrq.num_sge = 1;
 167         rx->rx_wrq.wr_id = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 168
 169         LASSERT (conn->ibc_state >= IBLND_CONN_INIT);
 170         LASSERT (rx->rx_nob >= 0);              /* not posted */
 171
 172         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 173                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 174                 return 0;
 175         }
 176
 177         rx->rx_nob = -1;                        /* flag posted */
 178
 179         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 180         if (rc != 0) {
 181                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 182                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 183                 rx->rx_nob = 0;
 184         }
 185
 186         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 187                 return rc;
 188
 189         if (rc != 0) {
 190                 kiblnd_close_conn(conn, rc);
 191                 kiblnd_drop_rx(rx);             /* No more posts for this rx */
 192                 return rc;
 193         }
 194
 195         if (credit == IBLND_POSTRX_NO_CREDIT)
 196                 return 0;
 197
 198         spin_lock(&conn->ibc_lock);
 199         if (credit == IBLND_POSTRX_PEER_CREDIT)
 200                 conn->ibc_outstanding_credits++;
 201         else
 202                 conn->ibc_reserved_credits++;
 203         spin_unlock(&conn->ibc_lock);
 204
 205         kiblnd_check_sends(conn);
 206         return 0;
 207 }
 208
 209 kib_tx_t *
 210 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
 211 {
 212         struct list_head   *tmp;
 213
 214         list_for_each(tmp, &conn->ibc_active_txs) {
 215                 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
 216
 217                 LASSERT (!tx->tx_queued);
 218                 LASSERT (tx->tx_sending != 0 || tx->tx_waiting);
 219
 220                 if (tx->tx_cookie != cookie)
 221                         continue;
 222
 223                 if (tx->tx_waiting &&
 224                     tx->tx_msg->ibm_type == txtype)
 225                         return tx;
 226
 227                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 228                       tx->tx_waiting ? "" : "NOT ",
 229                       tx->tx_msg->ibm_type, txtype);
 230         }
 231         return NULL;
 232 }
 233
 234 void
 235 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
 236 {
 237         kib_tx_t    *tx;
 238         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 239         int          idle;
 240
 241         spin_lock(&conn->ibc_lock);
 242
 243         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 244         if (tx == NULL) {
 245                 spin_unlock(&conn->ibc_lock);
 246
 247                 CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
 248                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 249                 kiblnd_close_conn(conn, -EPROTO);
 250                 return;
 251         }
 252
 253         if (tx->tx_status == 0) {               /* success so far */
 254                 if (status < 0) {               /* failed? */
 255                         tx->tx_status = status;
 256                 } else if (txtype == IBLND_MSG_GET_REQ) {
 257                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 258                 }
 259         }
 260
 261         tx->tx_waiting = 0;
 262
 263         idle = !tx->tx_queued && (tx->tx_sending == 0);
 264         if (idle)
 265                 list_del(&tx->tx_list);
 266
 267         spin_unlock(&conn->ibc_lock);
 268
 269         if (idle)
 270                 kiblnd_tx_done(ni, tx);
 271 }
 272
 273 void
 274 kiblnd_send_completion (kib_conn_t *conn, int type, int status, __u64 cookie)
 275 {
 276         lnet_ni_t   *ni = conn->ibc_peer->ibp_ni;
 277         kib_tx_t    *tx = kiblnd_get_idle_tx(ni);
 278
 279         if (tx == NULL) {
 280                 CERROR("Can't get tx for completion %x for %s\n",
 281                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 282                 return;
 283         }
 284
 285         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 286         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 287         kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
 288
 289         kiblnd_queue_tx(tx, conn);
 290 }
 291
 292 void
 293 kiblnd_handle_rx (kib_rx_t *rx)
 294 {
 295         kib_msg_t    *msg = rx->rx_msg;
 296         kib_conn_t   *conn = rx->rx_conn;
 297         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 298         int           credits = msg->ibm_credits;
 299         kib_tx_t     *tx;
 300         int           rc = 0;
 301         int           rc2;
 302         int           post_credit;
 303
 304         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 305
 306         CDEBUG (D_NET, "Received %x[%d] from %s\n",
 307                 msg->ibm_type, credits, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 308
 309         if (credits != 0) {
 310                 /* Have I received credits that will let me send? */
 311                 spin_lock(&conn->ibc_lock);
 312
 313                 if (conn->ibc_credits + credits >
 314                     IBLND_MSG_QUEUE_SIZE(conn->ibc_version)) {
 315                         rc2 = conn->ibc_credits;
 316                         spin_unlock(&conn->ibc_lock);
 317
 318                         CERROR("Bad credits from %s: %d + %d > %d\n",
 319                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 320                                rc2, credits, IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
 321
 322                         kiblnd_close_conn(conn, -EPROTO);
 323                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 324                         return;
 325                 }
 326
 327                 conn->ibc_credits += credits;
 328
 329                 spin_unlock(&conn->ibc_lock);
 330                 kiblnd_check_sends(conn);
 331         }
 332
 333         switch (msg->ibm_type) {
 334         default:
 335                 CERROR("Bad IBLND message type %x from %s\n",
 336                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 337                 post_credit = IBLND_POSTRX_NO_CREDIT;
 338                 rc = -EPROTO;
 339                 break;
 340
 341         case IBLND_MSG_NOOP:
 342                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
 343                         post_credit = IBLND_POSTRX_NO_CREDIT;
 344                 else
 345                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 346                 break;
 347
 348         case IBLND_MSG_IMMEDIATE:
 349                 post_credit = IBLND_POSTRX_DONT_POST;
 350                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 351                                 msg->ibm_srcnid, rx, 0);
 352                 if (rc < 0)                     /* repost on error */
 353                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 354                 break;
 355
 356         case IBLND_MSG_PUT_REQ:
 357                 post_credit = IBLND_POSTRX_DONT_POST;
 358                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 359                                 msg->ibm_srcnid, rx, 1);
 360                 if (rc < 0)                     /* repost on error */
 361                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 362                 break;
 363
 364         case IBLND_MSG_PUT_NAK:
 365                 CWARN ("PUT_NACK from %s\n", libcfs_nid2str(conn->ibc_peer->ibp_nid));
 366                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 367                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 368                                          msg->ibm_u.completion.ibcm_status,
 369                                          msg->ibm_u.completion.ibcm_cookie);
 370                 break;
 371
 372         case IBLND_MSG_PUT_ACK:
 373                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 374
 375                 spin_lock(&conn->ibc_lock);
 376                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 377                                                    msg->ibm_u.putack.ibpam_src_cookie);
 378                 if (tx != NULL)
 379                         list_del(&tx->tx_list);
 380                 spin_unlock(&conn->ibc_lock);
 381
 382                 if (tx == NULL) {
 383                         CERROR("Unmatched PUT_ACK from %s\n",
 384                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 385                         rc = -EPROTO;
 386                         break;
 387                 }
 388
 389                 LASSERT (tx->tx_waiting);
 390                 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
 391                  * (a) I can overwrite tx_msg since my peer has received it!
 392                  * (b) tx_waiting set tells tx_complete() it's not done. */
 393
 394                 tx->tx_nwrq = 0;                /* overwrite PUT_REQ */
 395
 396                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 397                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 398                                        &msg->ibm_u.putack.ibpam_rd,
 399                                        msg->ibm_u.putack.ibpam_dst_cookie);
 400                 if (rc2 < 0)
 401                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 402                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 403
 404                 spin_lock(&conn->ibc_lock);
 405                 tx->tx_waiting = 0;             /* clear waiting and queue atomically */
 406                 kiblnd_queue_tx_locked(tx, conn);
 407                 spin_unlock(&conn->ibc_lock);
 408                 break;
 409
 410         case IBLND_MSG_PUT_DONE:
 411                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 412                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 413                                          msg->ibm_u.completion.ibcm_status,
 414                                          msg->ibm_u.completion.ibcm_cookie);
 415                 break;
 416
 417         case IBLND_MSG_GET_REQ:
 418                 post_credit = IBLND_POSTRX_DONT_POST;
 419                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 420                                 msg->ibm_srcnid, rx, 1);
 421                 if (rc < 0)                     /* repost on error */
 422                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 423                 break;
 424
 425         case IBLND_MSG_GET_DONE:
 426                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 427                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 428                                          msg->ibm_u.completion.ibcm_status,
 429                                          msg->ibm_u.completion.ibcm_cookie);
 430                 break;
 431         }
 432
 433         if (rc < 0)                             /* protocol error */
 434                 kiblnd_close_conn(conn, rc);
 435
 436         if (post_credit != IBLND_POSTRX_DONT_POST)
 437                 kiblnd_post_rx(rx, post_credit);
 438 }
 439
 440 void
 441 kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
 442 {
 443         kib_msg_t    *msg = rx->rx_msg;
 444         kib_conn_t   *conn = rx->rx_conn;
 445         lnet_ni_t    *ni = conn->ibc_peer->ibp_ni;
 446         kib_net_t    *net = ni->ni_data;
 447         int           rc;
 448         int           err = -EIO;
 449
 450         LASSERT (net != NULL);
 451         LASSERT (rx->rx_nob < 0);               /* was posted */
 452         rx->rx_nob = 0;                         /* isn't now */
 453
 454         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 455                 goto ignore;
 456
 457         if (status != IB_WC_SUCCESS) {
 458                 CNETERR("Rx from %s failed: %d\n",
 459                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 460                 goto failed;
 461         }
 462
 463         LASSERT (nob >= 0);
 464         rx->rx_nob = nob;
 465
 466         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 467         if (rc != 0) {
 468                 CERROR ("Error %d unpacking rx from %s\n",
 469                         rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 470                 goto failed;
 471         }
 472
 473         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 474             msg->ibm_dstnid != ni->ni_nid ||
 475             msg->ibm_srcstamp != conn->ibc_incarnation ||
 476             msg->ibm_dststamp != net->ibn_incarnation) {
 477                 CERROR ("Stale rx from %s\n",
 478                         libcfs_nid2str(conn->ibc_peer->ibp_nid));
 479                 err = -ESTALE;
 480                 goto failed;
 481         }
 482
 483         /* set time last known alive */
 484         kiblnd_peer_alive(conn->ibc_peer);
 485
 486         /* racing with connection establishment/teardown! */
 487
 488         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 489                 rwlock_t      *g_lock = &kiblnd_data.kib_global_lock;
 490                 unsigned long  flags;
 491
 492                 write_lock_irqsave(g_lock, flags);
 493                 /* must check holding global lock to eliminate race */
 494                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 495                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 496                         write_unlock_irqrestore(g_lock, flags);
 497                         return;
 498                 }
 499                 write_unlock_irqrestore(g_lock, flags);
 500         }
 501         kiblnd_handle_rx(rx);
 502         return;
 503
 504  failed:
 505         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 506         kiblnd_close_conn(conn, err);
 507  ignore:
 508         kiblnd_drop_rx(rx);                     /* Don't re-post rx. */
 509 }
 510
 511 struct page *
 512 kiblnd_kvaddr_to_page (unsigned long vaddr)
 513 {
 514         struct page *page;
 515
 516         if (vaddr >= VMALLOC_START &&
 517             vaddr < VMALLOC_END) {
 518                 page = vmalloc_to_page ((void *)vaddr);
 519                 LASSERT (page != NULL);
 520                 return page;
 521         }
 522 #ifdef CONFIG_HIGHMEM
 523         if (vaddr >= PKMAP_BASE &&
 524             vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
 525                 /* No highmem pages only used for bulk (kiov) I/O */
 526                 CERROR("find page for address in highmem\n");
 527                 LBUG();
 528         }
 529 #endif
 530         page = virt_to_page (vaddr);
 531         LASSERT (page != NULL);
 532         return page;
 533 }
 534
 535 static int
 536 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 537 {
 538         kib_dev_t          *ibdev  = net->ibn_dev;
 539         __u64              *pages  = tx->tx_pages;
 540         int                 npages;
 541         int                 size;
 542         int                 rc;
 543         int                 i;
 544
 545         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
 546                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
 547                                size += ibdev->ibd_page_size) {
 548                         pages[npages ++] = (rd->rd_frags[i].rf_addr &
 549                                             ibdev->ibd_page_mask) + size;
 550                 }
 551         }
 552
 553         rc = kiblnd_fmr_pool_map(&net->ibn_fmr_ps, pages, npages, 0, &tx->tx_u.fmr);
 554         if (rc != 0) {
 555                 CERROR ("Can't map %d pages: %d\n", npages, rc);
 556                 return rc;
 557         }
 558
 559         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 560          * the rkey */
 561         rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.fmr.fmr_pfmr->fmr->rkey :
 562                                          tx->tx_u.fmr.fmr_pfmr->fmr->lkey;
 563         rd->rd_frags[0].rf_addr &= ~ibdev->ibd_page_mask;
 564         rd->rd_frags[0].rf_nob   = nob;
 565         rd->rd_nfrags = 1;
 566
 567         return 0;
 568 }
 569
 570 static int
 571 kiblnd_pmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 572 {
 573         __u64   iova;
 574         int     rc;
 575
 576         iova = rd->rd_frags[0].rf_addr & ~net->ibn_dev->ibd_page_mask;
 577
 578         rc = kiblnd_pmr_pool_map(&net->ibn_pmr_ps, rd, &iova, &tx->tx_u.pmr);
 579         if (rc != 0) {
 580                 CERROR("Failed to create MR by phybuf: %d\n", rc);
 581                 return rc;
 582         }
 583
 584         /* If rd is not tx_rd, it's going to get sent to a peer, who will need
 585          * the rkey */
 586         rd->rd_key = (rd != tx->tx_rd) ? tx->tx_u.pmr->pmr_mr->rkey :
 587                                          tx->tx_u.pmr->pmr_mr->lkey;
 588         rd->rd_nfrags = 1;
 589         rd->rd_frags[0].rf_addr = iova;
 590         rd->rd_frags[0].rf_nob  = nob;
 591
 592         return 0;
 593 }
 594
 595 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_with_fmr && tx->tx_u.fmr.fmr_pfmr != NULL) {
 603                 kiblnd_fmr_pool_unmap(&tx->tx_u.fmr, tx->tx_status);
 604                 tx->tx_u.fmr.fmr_pfmr = NULL;
 605         } else if (net->ibn_with_pmr && tx->tx_u.pmr != NULL) {
 606                 kiblnd_pmr_pool_unmap(tx->tx_u.pmr);
 607                 tx->tx_u.pmr = NULL;
 608         }
 609
 610         if (tx->tx_nfrags != 0) {
 611                 kiblnd_dma_unmap_sg(net->ibn_dev->ibd_cmid->device,
 612                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 613                 tx->tx_nfrags = 0;
 614         }
 615 }
 616
 617 int
 618 kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
 619               kib_rdma_desc_t *rd, int nfrags)
 620 {
 621         kib_net_t          *net   = ni->ni_data;
 622         struct ib_mr       *mr    = NULL;
 623         __u32               nob;
 624         int                 i;
 625
 626         /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
 627          * RDMA sink */
 628         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 629         tx->tx_nfrags = nfrags;
 630
 631         rd->rd_nfrags =
 632                 kiblnd_dma_map_sg(net->ibn_dev->ibd_cmid->device,
 633                                   tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 634
 635         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 636                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 637                         net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
 638                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 639                         net->ibn_dev->ibd_cmid->device, &tx->tx_frags[i]);
 640                 nob += rd->rd_frags[i].rf_nob;
 641         }
 642
 643         /* looking for pre-mapping MR */
 644         mr = kiblnd_find_rd_dma_mr(net, rd);
 645         if (mr != NULL) {
 646                 /* found pre-mapping MR */
 647                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 648                 return 0;
 649         }
 650
 651         if (net->ibn_with_fmr)
 652                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 653         else if (net->ibn_with_pmr)
 654                 return kiblnd_pmr_map_tx(net, tx, rd, nob);
 655
 656         return -EINVAL;
 657 }
 658
 659
 660 int
 661 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 662                     unsigned int niov, struct iovec *iov, int offset, int nob)
 663 {
 664         kib_net_t          *net = ni->ni_data;
 665         struct page        *page;
 666         struct scatterlist *sg;
 667         unsigned long       vaddr;
 668         int                 fragnob;
 669         int                 page_offset;
 670
 671         LASSERT (nob > 0);
 672         LASSERT (niov > 0);
 673         LASSERT (net != NULL);
 674
 675         while (offset >= iov->iov_len) {
 676                 offset -= iov->iov_len;
 677                 niov--;
 678                 iov++;
 679                 LASSERT (niov > 0);
 680         }
 681
 682         sg = tx->tx_frags;
 683         do {
 684                 LASSERT (niov > 0);
 685
 686                 vaddr = ((unsigned long)iov->iov_base) + offset;
 687                 page_offset = vaddr & (PAGE_SIZE - 1);
 688                 page = kiblnd_kvaddr_to_page(vaddr);
 689                 if (page == NULL) {
 690                         CERROR ("Can't find page\n");
 691                         return -EFAULT;
 692                 }
 693
 694                 fragnob = min((int)(iov->iov_len - offset), nob);
 695                 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
 696
 697                 sg_set_page(sg, page, fragnob, page_offset);
 698                 sg++;
 699
 700                 if (offset + fragnob < iov->iov_len) {
 701                         offset += fragnob;
 702                 } else {
 703                         offset = 0;
 704                         iov++;
 705                         niov--;
 706                 }
 707                 nob -= fragnob;
 708         } while (nob > 0);
 709
 710         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 711 }
 712
 713 static inline int
 714 get_kiov_length (int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 715 {
 716         int fragnob;
 717         int count = 0;
 718
 719         do {
 720                 LASSERT (nkiov > 0);
 721
 722                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 723
 724                 count++;
 725
 726                 offset = 0;
 727                 kiov++;
 728                 nkiov--;
 729                 nob -= fragnob;
 730         } while (nob > 0);
 731
 732         return count;
 733 }
 734
 735 int
 736 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 737                       int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 738 {
 739         kib_net_t          *net = ni->ni_data;
 740         struct scatterlist *sg;
 741         int                 fragnob;
 742
 743         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 744
 745         LASSERT (nob > 0);
 746         LASSERT (nkiov > 0);
 747         LASSERT (net != NULL);
 748
 749         while (offset >= kiov->kiov_len) {
 750                 offset -= kiov->kiov_len;
 751                 nkiov--;
 752                 kiov++;
 753                 LASSERT (nkiov > 0);
 754         }
 755
 756         sg = tx->tx_frags;
 757         tx->tx_nfrags = get_kiov_length(nkiov, kiov, offset, nob);
 758         sg_init_table(sg, tx->tx_nfrags);
 759         do {
 760                 LASSERT (nkiov > 0);
 761
 762                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 763
 764                 sg_set_page(sg, kiov->kiov_page, fragnob,
 765                             kiov->kiov_offset + offset);
 766                 sg++;
 767
 768                 offset = 0;
 769                 kiov++;
 770                 nkiov--;
 771                 nob -= fragnob;
 772         } while (nob > 0);
 773
 774         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 775 }
 776
 777 int
 778 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
 779 {
 780         kib_msg_t         *msg = tx->tx_msg;
 781         kib_peer_t        *peer = conn->ibc_peer;
 782         int                ver = conn->ibc_version;
 783         int                rc;
 784         int                done;
 785         struct ib_send_wr *bad_wrq;
 786
 787         LASSERT (tx->tx_queued);
 788         /* We rely on this for QP sizing */
 789         LASSERT (tx->tx_nwrq > 0);
 790         LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
 791
 792         LASSERT (credit == 0 || credit == 1);
 793         LASSERT (conn->ibc_outstanding_credits >= 0);
 794         LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 795         LASSERT (conn->ibc_credits >= 0);
 796         LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 797
 798         if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
 799                 /* tx completions outstanding... */
 800                 CDEBUG(D_NET, "%s: posted enough\n",
 801                        libcfs_nid2str(peer->ibp_nid));
 802                 return -EAGAIN;
 803         }
 804
 805         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 806                 CDEBUG(D_NET, "%s: no credits\n",
 807                        libcfs_nid2str(peer->ibp_nid));
 808                 return -EAGAIN;
 809         }
 810
 811         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 812             conn->ibc_credits == 1 &&   /* last credit reserved for */
 813             conn->ibc_outstanding_credits == 0) { /* giving back credits */
 814                 CDEBUG(D_NET, "%s: not using last credit\n",
 815                        libcfs_nid2str(peer->ibp_nid));
 816                 return -EAGAIN;
 817         }
 818
 819         /* NB don't drop ibc_lock before bumping tx_sending */
 820         list_del(&tx->tx_list);
 821         tx->tx_queued = 0;
 822
 823         if (msg->ibm_type == IBLND_MSG_NOOP &&
 824             (!kiblnd_send_noop(conn) ||     /* redundant NOOP */
 825              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 826               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 827                 /* OK to drop when posted enough NOOPs, since
 828                  * kiblnd_check_sends will queue NOOP again when
 829                  * posted NOOPs complete */
 830                 spin_unlock(&conn->ibc_lock);
 831                 kiblnd_tx_done(peer->ibp_ni, tx);
 832                 spin_lock(&conn->ibc_lock);
 833                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 834                        libcfs_nid2str(peer->ibp_nid),
 835                        conn->ibc_noops_posted);
 836                 return 0;
 837         }
 838
 839         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 840                         peer->ibp_nid, conn->ibc_incarnation);
 841
 842         conn->ibc_credits -= credit;
 843         conn->ibc_outstanding_credits = 0;
 844         conn->ibc_nsends_posted++;
 845         if (msg->ibm_type == IBLND_MSG_NOOP)
 846                 conn->ibc_noops_posted++;
 847
 848         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 849          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 850          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 851          * and then re-queued here.  It's (just) possible that
 852          * tx_sending is non-zero if we've not done the tx_complete()
 853          * from the first send; hence the ++ rather than = below. */
 854         tx->tx_sending++;
 855         list_add(&tx->tx_list, &conn->ibc_active_txs);
 856
 857         /* I'm still holding ibc_lock! */
 858         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
 859                 rc = -ECONNABORTED;
 860         else
 861                 rc = ib_post_send(conn->ibc_cmid->qp,
 862                                   tx->tx_wrq, &bad_wrq);
 863         conn->ibc_last_send = jiffies;
 864
 865         if (rc == 0)
 866                 return 0;
 867
 868         /* NB credits are transferred in the actual
 869          * message, which can only be the last work item */
 870         conn->ibc_credits += credit;
 871         conn->ibc_outstanding_credits += msg->ibm_credits;
 872         conn->ibc_nsends_posted--;
 873         if (msg->ibm_type == IBLND_MSG_NOOP)
 874                 conn->ibc_noops_posted--;
 875
 876         tx->tx_status = rc;
 877         tx->tx_waiting = 0;
 878         tx->tx_sending--;
 879
 880         done = (tx->tx_sending == 0);
 881         if (done)
 882                 list_del(&tx->tx_list);
 883
 884         spin_unlock(&conn->ibc_lock);
 885
 886         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 887                 CERROR("Error %d posting transmit to %s\n",
 888                        rc, libcfs_nid2str(peer->ibp_nid));
 889         else
 890                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 891                        rc, libcfs_nid2str(peer->ibp_nid));
 892
 893         kiblnd_close_conn(conn, rc);
 894
 895         if (done)
 896                 kiblnd_tx_done(peer->ibp_ni, tx);
 897
 898         spin_lock(&conn->ibc_lock);
 899
 900         return -EIO;
 901 }
 902
 903 void
 904 kiblnd_check_sends (kib_conn_t *conn)
 905 {
 906         int        ver = conn->ibc_version;
 907         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 908         kib_tx_t  *tx;
 909
 910         /* Don't send anything until after the connection is established */
 911         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 912                 CDEBUG(D_NET, "%s too soon\n",
 913                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 914                 return;
 915         }
 916
 917         spin_lock(&conn->ibc_lock);
 918
 919         LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
 920         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 921                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 922         LASSERT (conn->ibc_reserved_credits >= 0);
 923
 924         while (conn->ibc_reserved_credits > 0 &&
 925                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 926                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 927                                 kib_tx_t, tx_list);
 928                 list_del(&tx->tx_list);
 929                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 930                 conn->ibc_reserved_credits--;
 931         }
 932
 933         if (kiblnd_send_noop(conn)) {
 934                 spin_unlock(&conn->ibc_lock);
 935
 936                 tx = kiblnd_get_idle_tx(ni);
 937                 if (tx != NULL)
 938                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 939
 940                 spin_lock(&conn->ibc_lock);
 941                 if (tx != NULL)
 942                         kiblnd_queue_tx_locked(tx, conn);
 943         }
 944
 945         kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
 946
 947         for (;;) {
 948                 int credit;
 949
 950                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 951                         credit = 0;
 952                         tx = list_entry(conn->ibc_tx_queue_nocred.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         kiblnd_conn_decref(conn); /* ...until here */
 968 }
 969
 970 void
 971 kiblnd_tx_complete (kib_tx_t *tx, int status)
 972 {
 973         int           failed = (status != IB_WC_SUCCESS);
 974         kib_conn_t   *conn = tx->tx_conn;
 975         int           idle;
 976
 977         LASSERT (tx->tx_sending > 0);
 978
 979         if (failed) {
 980                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 981                         CNETERR("Tx -> %s cookie "LPX64
 982                                 " sending %d waiting %d: failed %d\n",
 983                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
 984                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
 985                                 status);
 986
 987                 kiblnd_close_conn(conn, -EIO);
 988         } else {
 989                 kiblnd_peer_alive(conn->ibc_peer);
 990         }
 991
 992         spin_lock(&conn->ibc_lock);
 993
 994         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
 995          * gets to free it, which also drops its ref on 'conn'. */
 996
 997         tx->tx_sending--;
 998         conn->ibc_nsends_posted--;
 999         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1000                 conn->ibc_noops_posted--;
1001
1002         if (failed) {
1003                 tx->tx_waiting = 0;             /* don't wait for peer */
1004                 tx->tx_status = -EIO;
1005         }
1006
1007         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
1008                !tx->tx_waiting &&               /* Not waiting for peer */
1009                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
1010         if (idle)
1011                 list_del(&tx->tx_list);
1012
1013         kiblnd_conn_addref(conn);               /* 1 ref for me.... */
1014
1015         spin_unlock(&conn->ibc_lock);
1016
1017         if (idle)
1018                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1019
1020         kiblnd_check_sends(conn);
1021
1022         kiblnd_conn_decref(conn);               /* ...until here */
1023 }
1024
1025 void
1026 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1027 {
1028         kib_net_t         *net = ni->ni_data;
1029         struct ib_sge     *sge = &tx->tx_sge[tx->tx_nwrq];
1030         struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1031         int                nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1032         struct ib_mr      *mr;
1033
1034         LASSERT (net != NULL);
1035         LASSERT (tx->tx_nwrq >= 0);
1036         LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1037         LASSERT (nob <= IBLND_MSG_SIZE);
1038
1039         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1040
1041         mr = kiblnd_find_dma_mr(net, tx->tx_msgaddr, nob);
1042         LASSERT (mr != NULL);
1043
1044         sge->lkey   = mr->lkey;
1045         sge->addr   = tx->tx_msgaddr;
1046         sge->length = nob;
1047
1048         memset(wrq, 0, sizeof(*wrq));
1049
1050         wrq->next       = NULL;
1051         wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1052         wrq->sg_list    = sge;
1053         wrq->num_sge    = 1;
1054         wrq->opcode     = IB_WR_SEND;
1055         wrq->send_flags = IB_SEND_SIGNALED;
1056
1057         tx->tx_nwrq++;
1058 }
1059
1060 int
1061 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1062                   int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1063 {
1064         kib_msg_t         *ibmsg = tx->tx_msg;
1065         kib_rdma_desc_t   *srcrd = tx->tx_rd;
1066         struct ib_sge     *sge = &tx->tx_sge[0];
1067         struct ib_send_wr *wrq = &tx->tx_wrq[0];
1068         int                rc  = resid;
1069         int                srcidx;
1070         int                dstidx;
1071         int                wrknob;
1072
1073         LASSERT (!in_interrupt());
1074         LASSERT (tx->tx_nwrq == 0);
1075         LASSERT (type == IBLND_MSG_GET_DONE ||
1076                  type == IBLND_MSG_PUT_DONE);
1077
1078         srcidx = dstidx = 0;
1079
1080         while (resid > 0) {
1081                 if (srcidx >= srcrd->rd_nfrags) {
1082                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1083                         rc = -EPROTO;
1084                         break;
1085                 }
1086
1087                 if (dstidx == dstrd->rd_nfrags) {
1088                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1089                         rc = -EPROTO;
1090                         break;
1091                 }
1092
1093                 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1094                         CERROR("RDMA too fragmented for %s (%d): "
1095                                "%d/%d src %d/%d dst frags\n",
1096                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1097                                IBLND_RDMA_FRAGS(conn->ibc_version),
1098                                srcidx, srcrd->rd_nfrags,
1099                                dstidx, dstrd->rd_nfrags);
1100                         rc = -EMSGSIZE;
1101                         break;
1102                 }
1103
1104                 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1105                                  kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1106
1107                 sge = &tx->tx_sge[tx->tx_nwrq];
1108                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1109                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1110                 sge->length = wrknob;
1111
1112                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1113
1114                 wrq->next       = wrq + 1;
1115                 wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1116                 wrq->sg_list    = sge;
1117                 wrq->num_sge    = 1;
1118                 wrq->opcode     = IB_WR_RDMA_WRITE;
1119                 wrq->send_flags = 0;
1120
1121                 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1122                 wrq->wr.rdma.rkey        = kiblnd_rd_frag_key(dstrd, dstidx);
1123
1124                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1125                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1126
1127                 resid -= wrknob;
1128
1129                 tx->tx_nwrq++;
1130                 wrq++;
1131                 sge++;
1132         }
1133
1134         if (rc < 0)                             /* no RDMA if completing with failure */
1135                 tx->tx_nwrq = 0;
1136
1137         ibmsg->ibm_u.completion.ibcm_status = rc;
1138         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1139         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1140                            type, sizeof (kib_completion_msg_t));
1141
1142         return rc;
1143 }
1144
1145 void
1146 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1147 {
1148         struct list_head   *q;
1149
1150         LASSERT (tx->tx_nwrq > 0);              /* work items set up */
1151         LASSERT (!tx->tx_queued);               /* not queued for sending already */
1152         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1153
1154         tx->tx_queued = 1;
1155         tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1156
1157         if (tx->tx_conn == NULL) {
1158                 kiblnd_conn_addref(conn);
1159                 tx->tx_conn = conn;
1160                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1161         } else {
1162                 /* PUT_DONE first attached to conn as a PUT_REQ */
1163                 LASSERT (tx->tx_conn == conn);
1164                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1165         }
1166
1167         switch (tx->tx_msg->ibm_type) {
1168         default:
1169                 LBUG();
1170
1171         case IBLND_MSG_PUT_REQ:
1172         case IBLND_MSG_GET_REQ:
1173                 q = &conn->ibc_tx_queue_rsrvd;
1174                 break;
1175
1176         case IBLND_MSG_PUT_NAK:
1177         case IBLND_MSG_PUT_ACK:
1178         case IBLND_MSG_PUT_DONE:
1179         case IBLND_MSG_GET_DONE:
1180                 q = &conn->ibc_tx_queue_nocred;
1181                 break;
1182
1183         case IBLND_MSG_NOOP:
1184                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1185                         q = &conn->ibc_tx_queue_nocred;
1186                 else
1187                         q = &conn->ibc_tx_queue;
1188                 break;
1189
1190         case IBLND_MSG_IMMEDIATE:
1191                 q = &conn->ibc_tx_queue;
1192                 break;
1193         }
1194
1195         list_add_tail(&tx->tx_list, q);
1196 }
1197
1198 void
1199 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1200 {
1201         spin_lock(&conn->ibc_lock);
1202         kiblnd_queue_tx_locked(tx, conn);
1203         spin_unlock(&conn->ibc_lock);
1204
1205         kiblnd_check_sends(conn);
1206 }
1207
1208 void
1209 kiblnd_connect_peer (kib_peer_t *peer)
1210 {
1211         struct rdma_cm_id *cmid;
1212         kib_dev_t         *dev;
1213         kib_net_t         *net = peer->ibp_ni->ni_data;
1214         struct sockaddr_in srcaddr;
1215         struct sockaddr_in dstaddr;
1216         int                rc;
1217
1218         LASSERT (net != NULL);
1219         LASSERT (peer->ibp_connecting > 0);
1220
1221         cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP);
1222         if (IS_ERR(cmid)) {
1223                 CERROR("Can't create CMID for %s: %ld\n",
1224                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1225                 rc = PTR_ERR(cmid);
1226                 goto failed;
1227         }
1228
1229         dev = net->ibn_dev;
1230         memset(&srcaddr, 0, sizeof(srcaddr));
1231         srcaddr.sin_family = AF_INET;
1232         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1233
1234         memset(&dstaddr, 0, sizeof(dstaddr));
1235         dstaddr.sin_family = AF_INET;
1236         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1237         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1238
1239         kiblnd_peer_addref(peer);               /* cmid's ref */
1240
1241         rc = rdma_resolve_addr(cmid,
1242                                (struct sockaddr *)&srcaddr,
1243                                (struct sockaddr *)&dstaddr,
1244                                *kiblnd_tunables.kib_timeout * 1000);
1245         if (rc == 0) {
1246                 LASSERT (cmid->device != NULL);
1247                 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1248                        libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1249                        HIPQUAD(dev->ibd_ifip), cmid->device->name);
1250                 return;
1251         }
1252
1253         /* Can't initiate address resolution:  */
1254         CERROR("Can't resolve addr for %s: %d\n",
1255                libcfs_nid2str(peer->ibp_nid), rc);
1256
1257         kiblnd_peer_decref(peer);               /* cmid's ref */
1258         rdma_destroy_id(cmid);
1259  failed:
1260         kiblnd_peer_connect_failed(peer, 1, rc);
1261 }
1262
1263 void
1264 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1265 {
1266         kib_peer_t        *peer;
1267         kib_peer_t        *peer2;
1268         kib_conn_t        *conn;
1269         rwlock_t          *g_lock = &kiblnd_data.kib_global_lock;
1270         unsigned long      flags;
1271         int                rc;
1272
1273         /* If I get here, I've committed to send, so I complete the tx with
1274          * failure on any problems */
1275
1276         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1277         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1278
1279         /* First time, just use a read lock since I expect to find my peer
1280          * connected */
1281         read_lock_irqsave(g_lock, flags);
1282
1283         peer = kiblnd_find_peer_locked(nid);
1284         if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1285                 /* Found a peer with an established connection */
1286                 conn = kiblnd_get_conn_locked(peer);
1287                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1288
1289                 read_unlock_irqrestore(g_lock, flags);
1290
1291                 if (tx != NULL)
1292                         kiblnd_queue_tx(tx, conn);
1293                 kiblnd_conn_decref(conn); /* ...to here */
1294                 return;
1295         }
1296
1297         read_unlock(g_lock);
1298         /* Re-try with a write lock */
1299         write_lock(g_lock);
1300
1301         peer = kiblnd_find_peer_locked(nid);
1302         if (peer != NULL) {
1303                 if (list_empty(&peer->ibp_conns)) {
1304                         /* found a peer, but it's still connecting... */
1305                         LASSERT (peer->ibp_connecting != 0 ||
1306                                  peer->ibp_accepting != 0);
1307                         if (tx != NULL)
1308                                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1309                         write_unlock_irqrestore(g_lock, flags);
1310                 } else {
1311                         conn = kiblnd_get_conn_locked(peer);
1312                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1313
1314                         write_unlock_irqrestore(g_lock, flags);
1315
1316                         if (tx != NULL)
1317                                 kiblnd_queue_tx(tx, conn);
1318                         kiblnd_conn_decref(conn); /* ...to here */
1319                 }
1320                 return;
1321         }
1322
1323         write_unlock_irqrestore(g_lock, flags);
1324
1325         /* Allocate a peer ready to add to the peer table and retry */
1326         rc = kiblnd_create_peer(ni, &peer, nid);
1327         if (rc != 0) {
1328                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1329                 if (tx != NULL) {
1330                         tx->tx_status = -EHOSTUNREACH;
1331                         tx->tx_waiting = 0;
1332                         kiblnd_tx_done(ni, tx);
1333                 }
1334                 return;
1335         }
1336
1337         write_lock_irqsave(g_lock, flags);
1338
1339         peer2 = kiblnd_find_peer_locked(nid);
1340         if (peer2 != NULL) {
1341                 if (list_empty(&peer2->ibp_conns)) {
1342                         /* found a peer, but it's still connecting... */
1343                         LASSERT (peer2->ibp_connecting != 0 ||
1344                                  peer2->ibp_accepting != 0);
1345                         if (tx != NULL)
1346                                 list_add_tail(&tx->tx_list, &peer2->ibp_tx_queue);
1347                         write_unlock_irqrestore(g_lock, flags);
1348                 } else {
1349                         conn = kiblnd_get_conn_locked(peer2);
1350                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1351
1352                         write_unlock_irqrestore(g_lock, flags);
1353
1354                         if (tx != NULL)
1355                                 kiblnd_queue_tx(tx, conn);
1356                         kiblnd_conn_decref(conn); /* ...to here */
1357                 }
1358
1359                 kiblnd_peer_decref(peer);
1360                 return;
1361         }
1362
1363         /* Brand new peer */
1364         LASSERT (peer->ibp_connecting == 0);
1365         peer->ibp_connecting = 1;
1366
1367         /* always called with a ref on ni, which prevents ni being shutdown */
1368         LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1369
1370         if (tx != NULL)
1371                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1372
1373         kiblnd_peer_addref(peer);
1374         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1375
1376         write_unlock_irqrestore(g_lock, flags);
1377
1378         kiblnd_connect_peer(peer);
1379         kiblnd_peer_decref(peer);
1380 }
1381
1382 int
1383 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1384 {
1385         lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
1386         int               type = lntmsg->msg_type;
1387         lnet_process_id_t target = lntmsg->msg_target;
1388         int               target_is_router = lntmsg->msg_target_is_router;
1389         int               routing = lntmsg->msg_routing;
1390         unsigned int      payload_niov = lntmsg->msg_niov;
1391         struct iovec     *payload_iov = lntmsg->msg_iov;
1392         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
1393         unsigned int      payload_offset = lntmsg->msg_offset;
1394         unsigned int      payload_nob = lntmsg->msg_len;
1395         kib_msg_t        *ibmsg;
1396         kib_tx_t         *tx;
1397         int               nob;
1398         int               rc;
1399
1400         /* NB 'private' is different depending on what we're sending.... */
1401
1402         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1403                payload_nob, payload_niov, libcfs_id2str(target));
1404
1405         LASSERT (payload_nob == 0 || payload_niov > 0);
1406         LASSERT (payload_niov <= LNET_MAX_IOV);
1407
1408         /* Thread context */
1409         LASSERT (!in_interrupt());
1410         /* payload is either all vaddrs or all pages */
1411         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1412
1413         switch (type) {
1414         default:
1415                 LBUG();
1416                 return (-EIO);
1417
1418         case LNET_MSG_ACK:
1419                 LASSERT (payload_nob == 0);
1420                 break;
1421
1422         case LNET_MSG_GET:
1423                 if (routing || target_is_router)
1424                         break;                  /* send IMMEDIATE */
1425
1426                 /* is the REPLY message too small for RDMA? */
1427                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1428                 if (nob <= IBLND_MSG_SIZE)
1429                         break;                  /* send IMMEDIATE */
1430
1431                 tx = kiblnd_get_idle_tx(ni);
1432                 if (tx == NULL) {
1433                         CERROR("Can't allocate txd for GET to %s: \n",
1434                                libcfs_nid2str(target.nid));
1435                         return -ENOMEM;
1436                 }
1437
1438                 ibmsg = tx->tx_msg;
1439
1440                 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1441                         rc = kiblnd_setup_rd_iov(ni, tx,
1442                                                  &ibmsg->ibm_u.get.ibgm_rd,
1443                                                  lntmsg->msg_md->md_niov,
1444                                                  lntmsg->msg_md->md_iov.iov,
1445                                                  0, lntmsg->msg_md->md_length);
1446                 else
1447                         rc = kiblnd_setup_rd_kiov(ni, tx,
1448                                                   &ibmsg->ibm_u.get.ibgm_rd,
1449                                                   lntmsg->msg_md->md_niov,
1450                                                   lntmsg->msg_md->md_iov.kiov,
1451                                                   0, lntmsg->msg_md->md_length);
1452                 if (rc != 0) {
1453                         CERROR("Can't setup GET sink for %s: %d\n",
1454                                libcfs_nid2str(target.nid), rc);
1455                         kiblnd_tx_done(ni, tx);
1456                         return -EIO;
1457                 }
1458
1459                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1460                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1461                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1462
1463                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1464
1465                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1466                 if (tx->tx_lntmsg[1] == NULL) {
1467                         CERROR("Can't create reply for GET -> %s\n",
1468                                libcfs_nid2str(target.nid));
1469                         kiblnd_tx_done(ni, tx);
1470                         return -EIO;
1471                 }
1472
1473                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1474                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1475                 kiblnd_launch_tx(ni, tx, target.nid);
1476                 return 0;
1477
1478         case LNET_MSG_REPLY:
1479         case LNET_MSG_PUT:
1480                 /* Is the payload small enough not to need RDMA? */
1481                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1482                 if (nob <= IBLND_MSG_SIZE)
1483                         break;                  /* send IMMEDIATE */
1484
1485                 tx = kiblnd_get_idle_tx(ni);
1486                 if (tx == NULL) {
1487                         CERROR("Can't allocate %s txd for %s\n",
1488                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1489                                libcfs_nid2str(target.nid));
1490                         return -ENOMEM;
1491                 }
1492
1493                 if (payload_kiov == NULL)
1494                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1495                                                  payload_niov, payload_iov,
1496                                                  payload_offset, payload_nob);
1497                 else
1498                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1499                                                   payload_niov, payload_kiov,
1500                                                   payload_offset, payload_nob);
1501                 if (rc != 0) {
1502                         CERROR("Can't setup PUT src for %s: %d\n",
1503                                libcfs_nid2str(target.nid), rc);
1504                         kiblnd_tx_done(ni, tx);
1505                         return -EIO;
1506                 }
1507
1508                 ibmsg = tx->tx_msg;
1509                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1510                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1511                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1512
1513                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1514                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1515                 kiblnd_launch_tx(ni, tx, target.nid);
1516                 return 0;
1517         }
1518
1519         /* send IMMEDIATE */
1520
1521         LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1522                  <= IBLND_MSG_SIZE);
1523
1524         tx = kiblnd_get_idle_tx(ni);
1525         if (tx == NULL) {
1526                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1527                         type, libcfs_nid2str(target.nid));
1528                 return -ENOMEM;
1529         }
1530
1531         ibmsg = tx->tx_msg;
1532         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1533
1534         if (payload_kiov != NULL)
1535                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1536                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1537                                     payload_niov, payload_kiov,
1538                                     payload_offset, payload_nob);
1539         else
1540                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1541                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1542                                    payload_niov, payload_iov,
1543                                    payload_offset, payload_nob);
1544
1545         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1546         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1547
1548         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1549         kiblnd_launch_tx(ni, tx, target.nid);
1550         return 0;
1551 }
1552
1553 void
1554 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1555 {
1556         lnet_process_id_t target = lntmsg->msg_target;
1557         unsigned int      niov = lntmsg->msg_niov;
1558         struct iovec     *iov = lntmsg->msg_iov;
1559         lnet_kiov_t      *kiov = lntmsg->msg_kiov;
1560         unsigned int      offset = lntmsg->msg_offset;
1561         unsigned int      nob = lntmsg->msg_len;
1562         kib_tx_t         *tx;
1563         int               rc;
1564
1565         tx = kiblnd_get_idle_tx(ni);
1566         if (tx == NULL) {
1567                 CERROR("Can't get tx for REPLY to %s\n",
1568                        libcfs_nid2str(target.nid));
1569                 goto failed_0;
1570         }
1571
1572         if (nob == 0)
1573                 rc = 0;
1574         else if (kiov == NULL)
1575                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1576                                          niov, iov, offset, nob);
1577         else
1578                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1579                                           niov, kiov, offset, nob);
1580
1581         if (rc != 0) {
1582                 CERROR("Can't setup GET src for %s: %d\n",
1583                        libcfs_nid2str(target.nid), rc);
1584                 goto failed_1;
1585         }
1586
1587         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1588                               IBLND_MSG_GET_DONE, nob,
1589                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1590                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1591         if (rc < 0) {
1592                 CERROR("Can't setup rdma for GET from %s: %d\n",
1593                        libcfs_nid2str(target.nid), rc);
1594                 goto failed_1;
1595         }
1596
1597         if (nob == 0) {
1598                 /* No RDMA: local completion may happen now! */
1599                 lnet_finalize(ni, lntmsg, 0);
1600         } else {
1601                 /* RDMA: lnet_finalize(lntmsg) when it
1602                  * completes */
1603                 tx->tx_lntmsg[0] = lntmsg;
1604         }
1605
1606         kiblnd_queue_tx(tx, rx->rx_conn);
1607         return;
1608
1609  failed_1:
1610         kiblnd_tx_done(ni, tx);
1611  failed_0:
1612         lnet_finalize(ni, lntmsg, -EIO);
1613 }
1614
1615 int
1616 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1617              unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1618              unsigned int offset, unsigned int mlen, unsigned int rlen)
1619 {
1620         kib_rx_t    *rx = private;
1621         kib_msg_t   *rxmsg = rx->rx_msg;
1622         kib_conn_t  *conn = rx->rx_conn;
1623         kib_tx_t    *tx;
1624         kib_msg_t   *txmsg;
1625         int          nob;
1626         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1627         int          rc = 0;
1628
1629         LASSERT (mlen <= rlen);
1630         LASSERT (!in_interrupt());
1631         /* Either all pages or all vaddrs */
1632         LASSERT (!(kiov != NULL && iov != NULL));
1633
1634         switch (rxmsg->ibm_type) {
1635         default:
1636                 LBUG();
1637
1638         case IBLND_MSG_IMMEDIATE:
1639                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1640                 if (nob > rx->rx_nob) {
1641                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1642                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1643                                 nob, rx->rx_nob);
1644                         rc = -EPROTO;
1645                         break;
1646                 }
1647
1648                 if (kiov != NULL)
1649                         lnet_copy_flat2kiov(niov, kiov, offset,
1650                                             IBLND_MSG_SIZE, rxmsg,
1651                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1652                                             mlen);
1653                 else
1654                         lnet_copy_flat2iov(niov, iov, offset,
1655                                            IBLND_MSG_SIZE, rxmsg,
1656                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1657                                            mlen);
1658                 lnet_finalize (ni, lntmsg, 0);
1659                 break;
1660
1661         case IBLND_MSG_PUT_REQ:
1662                 if (mlen == 0) {
1663                         lnet_finalize(ni, lntmsg, 0);
1664                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1665                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1666                         break;
1667                 }
1668
1669                 tx = kiblnd_get_idle_tx(ni);
1670                 if (tx == NULL) {
1671                         CERROR("Can't allocate tx for %s\n",
1672                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1673                         /* Not replying will break the connection */
1674                         rc = -ENOMEM;
1675                         break;
1676                 }
1677
1678                 txmsg = tx->tx_msg;
1679                 if (kiov == NULL)
1680                         rc = kiblnd_setup_rd_iov(ni, tx,
1681                                                  &txmsg->ibm_u.putack.ibpam_rd,
1682                                                  niov, iov, offset, mlen);
1683                 else
1684                         rc = kiblnd_setup_rd_kiov(ni, tx,
1685                                                   &txmsg->ibm_u.putack.ibpam_rd,
1686                                                   niov, kiov, offset, mlen);
1687                 if (rc != 0) {
1688                         CERROR("Can't setup PUT sink for %s: %d\n",
1689                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1690                         kiblnd_tx_done(ni, tx);
1691                         /* tell peer it's over */
1692                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1693                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1694                         break;
1695                 }
1696
1697                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1698                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1699                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1700
1701                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1702
1703                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1704                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1705                 kiblnd_queue_tx(tx, conn);
1706
1707                 /* reposted buffer reserved for PUT_DONE */
1708                 post_credit = IBLND_POSTRX_NO_CREDIT;
1709                 break;
1710
1711         case IBLND_MSG_GET_REQ:
1712                 if (lntmsg != NULL) {
1713                         /* Optimized GET; RDMA lntmsg's payload */
1714                         kiblnd_reply(ni, rx, lntmsg);
1715                 } else {
1716                         /* GET didn't match anything */
1717                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1718                                                -ENODATA,
1719                                                rxmsg->ibm_u.get.ibgm_cookie);
1720                 }
1721                 break;
1722         }
1723
1724         kiblnd_post_rx(rx, post_credit);
1725         return rc;
1726 }
1727
1728 int
1729 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1730 {
1731         long    pid = kernel_thread (fn, arg, 0);
1732
1733         if (pid < 0)
1734                 return ((int)pid);
1735
1736         atomic_inc (&kiblnd_data.kib_nthreads);
1737         return (0);
1738 }
1739
1740 void
1741 kiblnd_thread_fini (void)
1742 {
1743         atomic_dec (&kiblnd_data.kib_nthreads);
1744 }
1745
1746 void
1747 kiblnd_peer_alive (kib_peer_t *peer)
1748 {
1749         /* This is racy, but everyone's only writing cfs_time_current() */
1750         peer->ibp_last_alive = cfs_time_current();
1751         mb();
1752 }
1753
1754 void
1755 kiblnd_peer_notify (kib_peer_t *peer)
1756 {
1757         int           error = 0;
1758         cfs_time_t    last_alive = 0;
1759         unsigned long flags;
1760
1761         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1762
1763         if (list_empty(&peer->ibp_conns) &&
1764             peer->ibp_accepting == 0 &&
1765             peer->ibp_connecting == 0 &&
1766             peer->ibp_error != 0) {
1767                 error = peer->ibp_error;
1768                 peer->ibp_error = 0;
1769
1770                 last_alive = peer->ibp_last_alive;
1771         }
1772
1773         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1774
1775         if (error != 0)
1776                 lnet_notify(peer->ibp_ni,
1777                             peer->ibp_nid, 0, last_alive);
1778 }
1779
1780 void
1781 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1782 {
1783         /* This just does the immediate housekeeping.  'error' is zero for a
1784          * normal shutdown which can happen only after the connection has been
1785          * established.  If the connection is established, schedule the
1786          * connection to be finished off by the connd.  Otherwise the connd is
1787          * already dealing with it (either to set it up or tear it down).
1788          * Caller holds kib_global_lock exclusively in irq context */
1789         unsigned long     flags;
1790         kib_peer_t       *peer = conn->ibc_peer;
1791
1792         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1793
1794         if (error != 0 && conn->ibc_comms_error == 0)
1795                 conn->ibc_comms_error = error;
1796
1797         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1798                 return; /* already being handled  */
1799
1800         if (error == 0 &&
1801             list_empty(&conn->ibc_tx_queue) &&
1802             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1803             list_empty(&conn->ibc_tx_queue_nocred) &&
1804             list_empty(&conn->ibc_active_txs)) {
1805                 CDEBUG(D_NET, "closing conn to %s\n",
1806                        libcfs_nid2str(peer->ibp_nid));
1807         } else {
1808                 CNETERR("Closing conn to %s: error %d%s%s%s%s\n",
1809                         libcfs_nid2str(peer->ibp_nid), error,
1810                         list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1811                         list_empty(&conn->ibc_tx_queue_rsrvd) ?
1812                                 "" : "(sending_rsrvd)",
1813                         list_empty(&conn->ibc_tx_queue_nocred) ?
1814                                 "" : "(sending_nocred)",
1815                         list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1816         }
1817
1818         list_del(&conn->ibc_list);
1819         /* connd (see below) takes over ibc_list's ref */
1820
1821         if (list_empty (&peer->ibp_conns) &&    /* no more conns */
1822             kiblnd_peer_active(peer)) {         /* still in peer table */
1823                 kiblnd_unlink_peer_locked(peer);
1824
1825                 /* set/clear error on last conn */
1826                 peer->ibp_error = conn->ibc_comms_error;
1827         }
1828
1829         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1830
1831         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1832
1833         list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1834         wake_up (&kiblnd_data.kib_connd_waitq);
1835
1836         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1837 }
1838
1839 void
1840 kiblnd_close_conn (kib_conn_t *conn, int error)
1841 {
1842         unsigned long flags;
1843
1844         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1845
1846         kiblnd_close_conn_locked(conn, error);
1847
1848         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1849 }
1850
1851 void
1852 kiblnd_handle_early_rxs(kib_conn_t *conn)
1853 {
1854         unsigned long    flags;
1855         kib_rx_t        *rx;
1856
1857         LASSERT (!in_interrupt());
1858         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1859
1860         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1861         while (!list_empty(&conn->ibc_early_rxs)) {
1862                 rx = list_entry(conn->ibc_early_rxs.next,
1863                                 kib_rx_t, rx_list);
1864                 list_del(&rx->rx_list);
1865                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1866
1867                 kiblnd_handle_rx(rx);
1868
1869                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1870         }
1871         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1872 }
1873
1874 void
1875 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1876 {
1877         LIST_HEAD           (zombies);
1878         struct list_head    *tmp;
1879         struct list_head    *nxt;
1880         kib_tx_t            *tx;
1881
1882         spin_lock(&conn->ibc_lock);
1883
1884         list_for_each_safe (tmp, nxt, txs) {
1885                 tx = list_entry (tmp, kib_tx_t, tx_list);
1886
1887                 if (txs == &conn->ibc_active_txs) {
1888                         LASSERT (!tx->tx_queued);
1889                         LASSERT (tx->tx_waiting ||
1890                                  tx->tx_sending != 0);
1891                 } else {
1892                         LASSERT (tx->tx_queued);
1893                 }
1894
1895                 tx->tx_status = -ECONNABORTED;
1896                 tx->tx_waiting = 0;
1897
1898                 if (tx->tx_sending == 0) {
1899                         tx->tx_queued = 0;
1900                         list_del (&tx->tx_list);
1901                         list_add (&tx->tx_list, &zombies);
1902                 }
1903         }
1904
1905         spin_unlock(&conn->ibc_lock);
1906
1907         kiblnd_txlist_done(conn->ibc_peer->ibp_ni,
1908                            &zombies, -ECONNABORTED);
1909 }
1910
1911 void
1912 kiblnd_finalise_conn (kib_conn_t *conn)
1913 {
1914         LASSERT (!in_interrupt());
1915         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1916
1917         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1918
1919         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
1920          * for connections that didn't get as far as being connected, because
1921          * rdma_disconnect() does this for free. */
1922         kiblnd_abort_receives(conn);
1923
1924         /* Complete all tx descs not waiting for sends to complete.
1925          * NB we should be safe from RDMA now that the QP has changed state */
1926
1927         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1928         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1929         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1930         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1931
1932         kiblnd_handle_early_rxs(conn);
1933 }
1934
1935 void
1936 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
1937 {
1938         LIST_HEAD        (zombies);
1939         unsigned long     flags;
1940
1941         LASSERT (error != 0);
1942         LASSERT (!in_interrupt());
1943
1944         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1945
1946         if (active) {
1947                 LASSERT (peer->ibp_connecting > 0);
1948                 peer->ibp_connecting--;
1949         } else {
1950                 LASSERT (peer->ibp_accepting > 0);
1951                 peer->ibp_accepting--;
1952         }
1953
1954         if (peer->ibp_connecting != 0 ||
1955             peer->ibp_accepting != 0) {
1956                 /* another connection attempt under way... */
1957                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1958                 return;
1959         }
1960
1961         if (list_empty(&peer->ibp_conns)) {
1962                 /* Take peer's blocked transmits to complete with error */
1963                 list_add(&zombies, &peer->ibp_tx_queue);
1964                 list_del_init(&peer->ibp_tx_queue);
1965
1966                 if (kiblnd_peer_active(peer))
1967                         kiblnd_unlink_peer_locked(peer);
1968
1969                 peer->ibp_error = error;
1970         } else {
1971                 /* Can't have blocked transmits if there are connections */
1972                 LASSERT (list_empty(&peer->ibp_tx_queue));
1973         }
1974
1975         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1976
1977         kiblnd_peer_notify(peer);
1978
1979         if (list_empty (&zombies))
1980                 return;
1981
1982         CNETERR("Deleting messages for %s: connection failed\n",
1983                 libcfs_nid2str(peer->ibp_nid));
1984
1985         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
1986 }
1987
1988 void
1989 kiblnd_connreq_done(kib_conn_t *conn, int status)
1990 {
1991         kib_peer_t        *peer = conn->ibc_peer;
1992         kib_tx_t          *tx;
1993         struct list_head   txs;
1994         unsigned long      flags;
1995         int                active;
1996
1997         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
1998
1999         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
2000                libcfs_nid2str(peer->ibp_nid), active,
2001                conn->ibc_version, status);
2002
2003         LASSERT (!in_interrupt());
2004         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2005                   peer->ibp_connecting > 0) ||
2006                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2007                   peer->ibp_accepting > 0));
2008
2009         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2010         conn->ibc_connvars = NULL;
2011
2012         if (status != 0) {
2013                 /* failed to establish connection */
2014                 kiblnd_peer_connect_failed(peer, active, status);
2015                 kiblnd_finalise_conn(conn);
2016                 return;
2017         }
2018
2019         /* connection established */
2020         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2021
2022         conn->ibc_last_send = jiffies;
2023         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2024         kiblnd_peer_alive(peer);
2025
2026         /* Add conn to peer's list and nuke any dangling conns from a different
2027          * peer instance... */
2028         kiblnd_conn_addref(conn);               /* +1 ref for ibc_list */
2029         list_add(&conn->ibc_list, &peer->ibp_conns);
2030         if (active)
2031                 peer->ibp_connecting--;
2032         else
2033                 peer->ibp_accepting--;
2034
2035         if (peer->ibp_version == 0) {
2036                 peer->ibp_version     = conn->ibc_version;
2037                 peer->ibp_incarnation = conn->ibc_incarnation;
2038         }
2039
2040         if (peer->ibp_version     != conn->ibc_version ||
2041             peer->ibp_incarnation != conn->ibc_incarnation) {
2042                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2043                                                 conn->ibc_incarnation);
2044                 peer->ibp_version     = conn->ibc_version;
2045                 peer->ibp_incarnation = conn->ibc_incarnation;
2046         }
2047
2048         /* grab pending txs while I have the lock */
2049         list_add(&txs, &peer->ibp_tx_queue);
2050         list_del_init(&peer->ibp_tx_queue);
2051
2052         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2053             conn->ibc_comms_error != 0) {       /* error has happened already */
2054                 lnet_ni_t *ni = peer->ibp_ni;
2055
2056                 /* start to shut down connection */
2057                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2058                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2059
2060                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2061
2062                 return;
2063         }
2064
2065         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2066
2067         /* Schedule blocked txs */
2068         spin_lock (&conn->ibc_lock);
2069         while (!list_empty (&txs)) {
2070                 tx = list_entry (txs.next, kib_tx_t, tx_list);
2071                 list_del(&tx->tx_list);
2072
2073                 kiblnd_queue_tx_locked(tx, conn);
2074         }
2075         spin_unlock (&conn->ibc_lock);
2076
2077         kiblnd_check_sends(conn);
2078
2079         /* schedule blocked rxs */
2080         kiblnd_handle_early_rxs(conn);
2081 }
2082
2083 void
2084 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2085 {
2086         int          rc;
2087
2088         rc = rdma_reject(cmid, rej, sizeof(*rej));
2089
2090         if (rc != 0)
2091                 CWARN("Error %d sending reject\n", rc);
2092 }
2093
2094 int
2095 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2096 {
2097         rwlock_t              *g_lock = &kiblnd_data.kib_global_lock;
2098         kib_msg_t             *reqmsg = priv;
2099         kib_msg_t             *ackmsg;
2100         kib_dev_t             *ibdev;
2101         kib_peer_t            *peer;
2102         kib_peer_t            *peer2;
2103         kib_conn_t            *conn;
2104         lnet_ni_t             *ni  = NULL;
2105         kib_net_t             *net = NULL;
2106         lnet_nid_t             nid;
2107         struct rdma_conn_param cp;
2108         kib_rej_t              rej;
2109         int                    version = IBLND_MSG_VERSION;
2110         unsigned long          flags;
2111         int                    rc;
2112
2113         LASSERT (!in_interrupt());
2114
2115         /* cmid inherits 'context' from the corresponding listener id */
2116         ibdev = (kib_dev_t *)cmid->context;
2117         LASSERT (ibdev != NULL);
2118
2119         memset(&rej, 0, sizeof(rej));
2120         rej.ibr_magic                = IBLND_MSG_MAGIC;
2121         rej.ibr_why                  = IBLND_REJECT_FATAL;
2122         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2123
2124         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2125                 CERROR("Short connection request\n");
2126                 goto failed;
2127         }
2128
2129         /* Future protocol version compatibility support!  If the
2130          * o2iblnd-specific protocol changes, or when LNET unifies
2131          * protocols over all LNDs, the initial connection will
2132          * negotiate a protocol version.  I trap this here to avoid
2133          * console errors; the reject tells the peer which protocol I
2134          * speak. */
2135         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2136             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2137                 goto failed;
2138         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2139             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2140             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2141                 goto failed;
2142         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2143             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2144             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2145                 goto failed;
2146
2147         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2148         if (rc != 0) {
2149                 CERROR("Can't parse connection request: %d\n", rc);
2150                 goto failed;
2151         }
2152
2153         nid = reqmsg->ibm_srcnid;
2154         ni  = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2155
2156         if (ni != NULL) {
2157                 net = (kib_net_t *)ni->ni_data;
2158                 rej.ibr_incarnation = net->ibn_incarnation;
2159         }
2160
2161         if (ni == NULL ||                         /* no matching net */
2162             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2163             net->ibn_dev != ibdev) {              /* wrong device */
2164                 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2165                        "bad dst nid %s\n", libcfs_nid2str(nid),
2166                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2167                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2168                        HIPQUAD(ibdev->ibd_ifip),
2169                        libcfs_nid2str(reqmsg->ibm_dstnid));
2170
2171                 goto failed;
2172         }
2173
2174        /* check time stamp as soon as possible */
2175         if (reqmsg->ibm_dststamp != 0 &&
2176             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2177                 CWARN("Stale connection request\n");
2178                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2179                 goto failed;
2180         }
2181
2182         /* I can accept peer's version */
2183         version = reqmsg->ibm_version;
2184
2185         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2186                 CERROR("Unexpected connreq msg type: %x from %s\n",
2187                        reqmsg->ibm_type, libcfs_nid2str(nid));
2188                 goto failed;
2189         }
2190
2191         if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2192             IBLND_MSG_QUEUE_SIZE(version)) {
2193                 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2194                        libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2195                        IBLND_MSG_QUEUE_SIZE(version));
2196
2197                 if (version == IBLND_MSG_VERSION)
2198                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2199
2200                 goto failed;
2201         }
2202
2203         if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2204             IBLND_RDMA_FRAGS(version)) {
2205                 CERROR("Can't accept %s(version %x): "
2206                        "incompatible max_frags %d (%d wanted)\n",
2207                        libcfs_nid2str(nid), version,
2208                        reqmsg->ibm_u.connparams.ibcp_max_frags,
2209                        IBLND_RDMA_FRAGS(version));
2210
2211                 if (version == IBLND_MSG_VERSION)
2212                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2213
2214                 goto failed;
2215
2216         }
2217
2218         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2219                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2220                        libcfs_nid2str(nid),
2221                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2222                        IBLND_MSG_SIZE);
2223                 goto failed;
2224         }
2225
2226         /* assume 'nid' is a new peer; create  */
2227         rc = kiblnd_create_peer(ni, &peer, nid);
2228         if (rc != 0) {
2229                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2230                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2231                 goto failed;
2232         }
2233
2234         write_lock_irqsave(g_lock, flags);
2235
2236         peer2 = kiblnd_find_peer_locked(nid);
2237         if (peer2 != NULL) {
2238                 if (peer2->ibp_version == 0) {
2239                         peer2->ibp_version     = version;
2240                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2241                 }
2242
2243                 /* not the guy I've talked with */
2244                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2245                     peer2->ibp_version     != version) {
2246                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2247                         write_unlock_irqrestore(g_lock, flags);
2248
2249                         CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2250                               libcfs_nid2str(nid), peer2->ibp_version, version);
2251
2252                         kiblnd_peer_decref(peer);
2253                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2254                         goto failed;
2255                 }
2256
2257                 /* tie-break connection race in favour of the higher NID */
2258                 if (peer2->ibp_connecting != 0 &&
2259                     nid < ni->ni_nid) {
2260                         write_unlock_irqrestore(g_lock, flags);
2261
2262                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2263
2264                         kiblnd_peer_decref(peer);
2265                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2266                         goto failed;
2267                 }
2268
2269                 peer2->ibp_accepting++;
2270                 kiblnd_peer_addref(peer2);
2271
2272                 write_unlock_irqrestore(g_lock, flags);
2273                 kiblnd_peer_decref(peer);
2274                 peer = peer2;
2275         } else {
2276                 /* Brand new peer */
2277                 LASSERT (peer->ibp_accepting == 0);
2278                 LASSERT (peer->ibp_version == 0 &&
2279                          peer->ibp_incarnation == 0);
2280
2281                 peer->ibp_accepting   = 1;
2282                 peer->ibp_version     = version;
2283                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2284
2285                 /* I have a ref on ni that prevents it being shutdown */
2286                 LASSERT (net->ibn_shutdown == 0);
2287
2288                 kiblnd_peer_addref(peer);
2289                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2290
2291                 write_unlock_irqrestore(g_lock, flags);
2292         }
2293
2294         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2295         if (conn == NULL) {
2296                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2297                 kiblnd_peer_decref(peer);
2298                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2299                 goto failed;
2300         }
2301
2302         /* conn now "owns" cmid, so I return success from here on to ensure the
2303          * CM callback doesn't destroy cmid. */
2304
2305         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2306         conn->ibc_credits          = IBLND_MSG_QUEUE_SIZE(version);
2307         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2308         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2309                  <= IBLND_RX_MSGS(version));
2310
2311         ackmsg = &conn->ibc_connvars->cv_msg;
2312         memset(ackmsg, 0, sizeof(*ackmsg));
2313
2314         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2315                         sizeof(ackmsg->ibm_u.connparams));
2316         ackmsg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2317         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2318         ackmsg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2319
2320         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2321
2322         memset(&cp, 0, sizeof(cp));
2323         cp.private_data        = ackmsg;
2324         cp.private_data_len    = ackmsg->ibm_nob;
2325         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2326         cp.initiator_depth     = 0;
2327         cp.flow_control        = 1;
2328         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2329         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2330
2331         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2332
2333         rc = rdma_accept(cmid, &cp);
2334         if (rc != 0) {
2335                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2336                 rej.ibr_version = version;
2337                 rej.ibr_why     = IBLND_REJECT_FATAL;
2338
2339                 kiblnd_reject(cmid, &rej);
2340                 kiblnd_connreq_done(conn, rc);
2341                 kiblnd_conn_decref(conn);
2342         }
2343
2344         lnet_ni_decref(ni);
2345         return 0;
2346
2347  failed:
2348         if (ni != NULL)
2349                 lnet_ni_decref(ni);
2350
2351         rej.ibr_version = version;
2352         rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2353         rej.ibr_cp.ibcp_max_frags   = IBLND_RDMA_FRAGS(version);
2354         kiblnd_reject(cmid, &rej);
2355
2356         return -ECONNREFUSED;
2357 }
2358
2359 void
2360 kiblnd_reconnect (kib_conn_t *conn, int version,
2361                   __u64 incarnation, int why, kib_connparams_t *cp)
2362 {
2363         kib_peer_t    *peer = conn->ibc_peer;
2364         char          *reason;
2365         int            retry = 0;
2366         unsigned long  flags;
2367
2368         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2369         LASSERT (peer->ibp_connecting > 0);     /* 'conn' at least */
2370
2371         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2372
2373         /* retry connection if it's still needed and no other connection
2374          * attempts (active or passive) are in progress */
2375         if (!list_empty(&peer->ibp_tx_queue) &&
2376             peer->ibp_connecting == 1 &&
2377             peer->ibp_accepting == 0) {
2378                 retry = 1;
2379                 peer->ibp_connecting++;
2380
2381                 peer->ibp_version     = version;
2382                 peer->ibp_incarnation = incarnation;
2383         }
2384
2385         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2386
2387         if (!retry)
2388                 return;
2389
2390         switch (why) {
2391         default:
2392                 reason = "Unknown";
2393                 break;
2394
2395         case IBLND_REJECT_CONN_STALE:
2396                 reason = "stale";
2397                 break;
2398
2399         case IBLND_REJECT_CONN_RACE:
2400                 reason = "conn race";
2401                 break;
2402
2403         case IBLND_REJECT_CONN_UNCOMPAT:
2404                 reason = "version negotiation";
2405                 break;
2406         }
2407
2408         CNETERR("%s: retrying (%s), %x, %x, "
2409                 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2410                 libcfs_nid2str(peer->ibp_nid),
2411                 reason, IBLND_MSG_VERSION, version,
2412                 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2413                 cp != NULL? cp->ibcp_max_frags   : IBLND_RDMA_FRAGS(version),
2414                 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2415
2416         kiblnd_connect_peer(peer);
2417 }
2418
2419 void
2420 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2421 {
2422         kib_peer_t    *peer = conn->ibc_peer;
2423
2424         LASSERT (!in_interrupt());
2425         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2426
2427         switch (reason) {
2428         case IB_CM_REJ_STALE_CONN:
2429                 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2430                                  IBLND_REJECT_CONN_STALE, NULL);
2431                 break;
2432
2433         case IB_CM_REJ_INVALID_SERVICE_ID:
2434                 CNETERR("%s rejected: no listener at %d\n",
2435                         libcfs_nid2str(peer->ibp_nid),
2436                         *kiblnd_tunables.kib_service);
2437                 break;
2438
2439         case IB_CM_REJ_CONSUMER_DEFINED:
2440                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2441                         kib_rej_t        *rej         = priv;
2442                         kib_connparams_t *cp          = NULL;
2443                         int               flip        = 0;
2444                         __u64             incarnation = -1;
2445
2446                         /* NB. default incarnation is -1 because:
2447                          * a) V1 will ignore dst incarnation in connreq.
2448                          * b) V2 will provide incarnation while rejecting me,
2449                          *    -1 will be overwrote.
2450                          *
2451                          * if I try to connect to a V1 peer with V2 protocol,
2452                          * it rejected me then upgrade to V2, I have no idea
2453                          * about the upgrading and try to reconnect with V1,
2454                          * in this case upgraded V2 can find out I'm trying to
2455                          * talk to the old guy and reject me(incarnation is -1).
2456                          */
2457
2458                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2459                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2460                                 __swab32s(&rej->ibr_magic);
2461                                 __swab16s(&rej->ibr_version);
2462                                 flip = 1;
2463                         }
2464
2465                         if (priv_nob >= sizeof(kib_rej_t) &&
2466                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2467                                 /* priv_nob is always 148 in current version
2468                                  * of OFED, so we still need to check version.
2469                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2470                                 cp = &rej->ibr_cp;
2471
2472                                 if (flip) {
2473                                         __swab64s(&rej->ibr_incarnation);
2474                                         __swab16s(&cp->ibcp_queue_depth);
2475                                         __swab16s(&cp->ibcp_max_frags);
2476                                         __swab32s(&cp->ibcp_max_msg_size);
2477                                 }
2478
2479                                 incarnation = rej->ibr_incarnation;
2480                         }
2481
2482                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2483                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2484                                 CERROR("%s rejected: consumer defined fatal error\n",
2485                                        libcfs_nid2str(peer->ibp_nid));
2486                                 break;
2487                         }
2488
2489                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2490                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2491                                 CERROR("%s rejected: o2iblnd version %x error\n",
2492                                        libcfs_nid2str(peer->ibp_nid),
2493                                        rej->ibr_version);
2494                                 break;
2495                         }
2496
2497                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2498                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2499                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2500                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2501
2502                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2503                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2504                         }
2505
2506                         switch (rej->ibr_why) {
2507                         case IBLND_REJECT_CONN_RACE:
2508                         case IBLND_REJECT_CONN_STALE:
2509                         case IBLND_REJECT_CONN_UNCOMPAT:
2510                                 kiblnd_reconnect(conn, rej->ibr_version,
2511                                                  incarnation, rej->ibr_why, cp);
2512                                 break;
2513
2514                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2515                                 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2516                                        libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2517                                        IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2518                                 break;
2519
2520                         case IBLND_REJECT_RDMA_FRAGS:
2521                                 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2522                                        libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2523                                        IBLND_RDMA_FRAGS(conn->ibc_version));
2524                                 break;
2525
2526                         case IBLND_REJECT_NO_RESOURCES:
2527                                 CERROR("%s rejected: o2iblnd no resources\n",
2528                                        libcfs_nid2str(peer->ibp_nid));
2529                                 break;
2530
2531                         case IBLND_REJECT_FATAL:
2532                                 CERROR("%s rejected: o2iblnd fatal error\n",
2533                                        libcfs_nid2str(peer->ibp_nid));
2534                                 break;
2535
2536                         default:
2537                                 CERROR("%s rejected: o2iblnd reason %d\n",
2538                                        libcfs_nid2str(peer->ibp_nid),
2539                                        rej->ibr_why);
2540                                 break;
2541                         }
2542                         break;
2543                 }
2544                 /* fall through */
2545         default:
2546                 CNETERR("%s rejected: reason %d, size %d\n",
2547                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2548                 break;
2549         }
2550
2551         kiblnd_connreq_done(conn, -ECONNREFUSED);
2552 }
2553
2554 void
2555 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2556 {
2557         kib_peer_t    *peer = conn->ibc_peer;
2558         lnet_ni_t     *ni   = peer->ibp_ni;
2559         kib_net_t     *net  = ni->ni_data;
2560         kib_msg_t     *msg  = priv;
2561         int            ver  = conn->ibc_version;
2562         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2563         unsigned long  flags;
2564
2565         LASSERT (net != NULL);
2566
2567         if (rc != 0) {
2568                 CERROR("Can't unpack connack from %s: %d\n",
2569                        libcfs_nid2str(peer->ibp_nid), rc);
2570                 goto failed;
2571         }
2572
2573         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2574                 CERROR("Unexpected message %d from %s\n",
2575                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2576                 rc = -EPROTO;
2577                 goto failed;
2578         }
2579
2580         if (ver != msg->ibm_version) {
2581                 CERROR("%s replied version %x is different with "
2582                        "requested version %x\n",
2583                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2584                 rc = -EPROTO;
2585                 goto failed;
2586         }
2587
2588         if (msg->ibm_u.connparams.ibcp_queue_depth !=
2589             IBLND_MSG_QUEUE_SIZE(ver)) {
2590                 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2591                        libcfs_nid2str(peer->ibp_nid),
2592                        msg->ibm_u.connparams.ibcp_queue_depth,
2593                        IBLND_MSG_QUEUE_SIZE(ver));
2594                 rc = -EPROTO;
2595                 goto failed;
2596         }
2597
2598         if (msg->ibm_u.connparams.ibcp_max_frags !=
2599             IBLND_RDMA_FRAGS(ver)) {
2600                 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2601                        libcfs_nid2str(peer->ibp_nid),
2602                        msg->ibm_u.connparams.ibcp_max_frags,
2603                        IBLND_RDMA_FRAGS(ver));
2604                 rc = -EPROTO;
2605                 goto failed;
2606         }
2607
2608         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2609                 CERROR("%s max message size %d too big (%d max)\n",
2610                        libcfs_nid2str(peer->ibp_nid),
2611                        msg->ibm_u.connparams.ibcp_max_msg_size,
2612                        IBLND_MSG_SIZE);
2613                 rc = -EPROTO;
2614                 goto failed;
2615         }
2616
2617         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2618         if (msg->ibm_dstnid == ni->ni_nid &&
2619             msg->ibm_dststamp == net->ibn_incarnation)
2620                 rc = 0;
2621         else
2622                 rc = -ESTALE;
2623         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2624
2625         if (rc != 0) {
2626                 CERROR("Bad connection reply from %s, rc = %d, "
2627                        "version: %x max_frags: %d\n",
2628                        libcfs_nid2str(peer->ibp_nid), rc,
2629                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2630                 goto failed;
2631         }
2632
2633         conn->ibc_incarnation      = msg->ibm_srcstamp;
2634         conn->ibc_credits          =
2635         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2636         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2637                  <= IBLND_RX_MSGS(ver));
2638
2639         kiblnd_connreq_done(conn, 0);
2640         return;
2641
2642  failed:
2643         /* NB My QP has already established itself, so I handle anything going
2644          * wrong here by setting ibc_comms_error.
2645          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2646          * immediately tears it down. */
2647
2648         LASSERT (rc != 0);
2649         conn->ibc_comms_error = rc;
2650         kiblnd_connreq_done(conn, 0);
2651 }
2652
2653 int
2654 kiblnd_active_connect (struct rdma_cm_id *cmid)
2655 {
2656         kib_peer_t              *peer = (kib_peer_t *)cmid->context;
2657         kib_conn_t              *conn;
2658         kib_msg_t               *msg;
2659         struct rdma_conn_param   cp;
2660         int                      version;
2661         __u64                    incarnation;
2662         unsigned long            flags;
2663         int                      rc;
2664
2665         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2666
2667         incarnation = peer->ibp_incarnation;
2668         version     = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : peer->ibp_version;
2669
2670         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2671
2672         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2673         if (conn == NULL) {
2674                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2675                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2676                 return -ENOMEM;
2677         }
2678
2679         /* conn "owns" cmid now, so I return success from here on to ensure the
2680          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2681          * on peer */
2682
2683         msg = &conn->ibc_connvars->cv_msg;
2684
2685         memset(msg, 0, sizeof(*msg));
2686         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2687         msg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2688         msg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2689         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2690
2691         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2692                         0, peer->ibp_nid, incarnation);
2693
2694         memset(&cp, 0, sizeof(cp));
2695         cp.private_data        = msg;
2696         cp.private_data_len    = msg->ibm_nob;
2697         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2698         cp.initiator_depth     = 0;
2699         cp.flow_control        = 1;
2700         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2701         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2702
2703         LASSERT(cmid->context == (void *)conn);
2704         LASSERT(conn->ibc_cmid == cmid);
2705
2706         rc = rdma_connect(cmid, &cp);
2707         if (rc != 0) {
2708                 CERROR("Can't connect to %s: %d\n",
2709                        libcfs_nid2str(peer->ibp_nid), rc);
2710                 kiblnd_connreq_done(conn, rc);
2711                 kiblnd_conn_decref(conn);
2712         }
2713
2714         return 0;
2715 }
2716
2717 int
2718 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2719 {
2720         kib_peer_t  *peer;
2721         kib_conn_t  *conn;
2722         int          rc;
2723
2724         switch (event->event) {
2725         default:
2726                 CERROR("Unexpected event: %d, status: %d\n",
2727                        event->event, event->status);
2728                 LBUG();
2729
2730         case RDMA_CM_EVENT_CONNECT_REQUEST:
2731                 /* destroy cmid on failure */
2732                 rc = kiblnd_passive_connect(cmid,
2733                                             (void *)KIBLND_CONN_PARAM(event),
2734                                             KIBLND_CONN_PARAM_LEN(event));
2735                 CDEBUG(D_NET, "connreq: %d\n", rc);
2736                 return rc;
2737
2738         case RDMA_CM_EVENT_ADDR_ERROR:
2739                 peer = (kib_peer_t *)cmid->context;
2740                 CNETERR("%s: ADDR ERROR %d\n",
2741                         libcfs_nid2str(peer->ibp_nid), event->status);
2742                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2743                 kiblnd_peer_decref(peer);
2744                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
2745
2746         case RDMA_CM_EVENT_ADDR_RESOLVED:
2747                 peer = (kib_peer_t *)cmid->context;
2748
2749                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2750                        libcfs_nid2str(peer->ibp_nid), event->status);
2751
2752                 if (event->status != 0) {
2753                         CNETERR("Can't resolve address for %s: %d\n",
2754                                 libcfs_nid2str(peer->ibp_nid), event->status);
2755                         rc = event->status;
2756                 } else {
2757                         rc = rdma_resolve_route(
2758                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2759                         if (rc == 0)
2760                                 return 0;
2761                         /* Can't initiate route resolution */
2762                         CERROR("Can't resolve route for %s: %d\n",
2763                                libcfs_nid2str(peer->ibp_nid), rc);
2764                 }
2765                 kiblnd_peer_connect_failed(peer, 1, rc);
2766                 kiblnd_peer_decref(peer);
2767                 return rc;                      /* rc != 0 destroys cmid */
2768
2769         case RDMA_CM_EVENT_ROUTE_ERROR:
2770                 peer = (kib_peer_t *)cmid->context;
2771                 CNETERR("%s: ROUTE ERROR %d\n",
2772                         libcfs_nid2str(peer->ibp_nid), event->status);
2773                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2774                 kiblnd_peer_decref(peer);
2775                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
2776
2777         case RDMA_CM_EVENT_ROUTE_RESOLVED:
2778                 peer = (kib_peer_t *)cmid->context;
2779                 CDEBUG(D_NET,"%s Route resolved: %d\n",
2780                        libcfs_nid2str(peer->ibp_nid), event->status);
2781
2782                 if (event->status == 0)
2783                         return kiblnd_active_connect(cmid);
2784
2785                 CNETERR("Can't resolve route for %s: %d\n",
2786                         libcfs_nid2str(peer->ibp_nid), event->status);
2787                 kiblnd_peer_connect_failed(peer, 1, event->status);
2788                 kiblnd_peer_decref(peer);
2789                 return event->status;           /* rc != 0 destroys cmid */
2790
2791         case RDMA_CM_EVENT_UNREACHABLE:
2792                 conn = (kib_conn_t *)cmid->context;
2793                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2794                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2795                 CNETERR("%s: UNREACHABLE %d\n",
2796                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2797                 kiblnd_connreq_done(conn, -ENETDOWN);
2798                 kiblnd_conn_decref(conn);
2799                 return 0;
2800
2801         case RDMA_CM_EVENT_CONNECT_ERROR:
2802                 conn = (kib_conn_t *)cmid->context;
2803                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2804                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2805                 CNETERR("%s: CONNECT ERROR %d\n",
2806                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2807                 kiblnd_connreq_done(conn, -ENOTCONN);
2808                 kiblnd_conn_decref(conn);
2809                 return 0;
2810
2811         case RDMA_CM_EVENT_REJECTED:
2812                 conn = (kib_conn_t *)cmid->context;
2813                 switch (conn->ibc_state) {
2814                 default:
2815                         LBUG();
2816
2817                 case IBLND_CONN_PASSIVE_WAIT:
2818                         CERROR ("%s: REJECTED %d\n",
2819                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2820                                 event->status);
2821                         kiblnd_connreq_done(conn, -ECONNRESET);
2822                         break;
2823
2824                 case IBLND_CONN_ACTIVE_CONNECT:
2825                         kiblnd_rejected(conn, event->status,
2826                                         (void *)KIBLND_CONN_PARAM(event),
2827                                         KIBLND_CONN_PARAM_LEN(event));
2828                         break;
2829                 }
2830                 kiblnd_conn_decref(conn);
2831                 return 0;
2832
2833         case RDMA_CM_EVENT_ESTABLISHED:
2834                 conn = (kib_conn_t *)cmid->context;
2835                 switch (conn->ibc_state) {
2836                 default:
2837                         LBUG();
2838
2839                 case IBLND_CONN_PASSIVE_WAIT:
2840                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2841                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2842                         kiblnd_connreq_done(conn, 0);
2843                         break;
2844
2845                 case IBLND_CONN_ACTIVE_CONNECT:
2846                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2847                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2848                         kiblnd_check_connreply(conn,
2849                                                (void *)KIBLND_CONN_PARAM(event),
2850                                                KIBLND_CONN_PARAM_LEN(event));
2851                         break;
2852                 }
2853                 /* net keeps its ref on conn! */
2854                 return 0;
2855
2856 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2857         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2858                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2859                 return 0;
2860 #endif
2861         case RDMA_CM_EVENT_DISCONNECTED:
2862                 conn = (kib_conn_t *)cmid->context;
2863                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2864                         CERROR("%s DISCONNECTED\n",
2865                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2866                         kiblnd_connreq_done(conn, -ECONNRESET);
2867                 } else {
2868                         kiblnd_close_conn(conn, 0);
2869                 }
2870                 kiblnd_conn_decref(conn);
2871                 cmid->context = NULL;
2872                 return 0;
2873
2874         case RDMA_CM_EVENT_DEVICE_REMOVAL:
2875                 LCONSOLE_ERROR_MSG(0x131,
2876                                    "Received notification of device removal\n"
2877                                    "Please shutdown LNET to allow this to proceed\n");
2878                 /* Can't remove network from underneath LNET for now, so I have
2879                  * to ignore this */
2880                 return 0;
2881
2882 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2883         case RDMA_CM_EVENT_ADDR_CHANGE:
2884                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2885                 return 0;
2886 #endif
2887         }
2888 }
2889
2890 int
2891 kiblnd_check_txs (kib_conn_t *conn, struct list_head *txs)
2892 {
2893         kib_tx_t          *tx;
2894         struct list_head  *ttmp;
2895         int                timed_out = 0;
2896
2897         spin_lock(&conn->ibc_lock);
2898
2899         list_for_each (ttmp, txs) {
2900                 tx = list_entry (ttmp, kib_tx_t, tx_list);
2901
2902                 if (txs != &conn->ibc_active_txs) {
2903                         LASSERT (tx->tx_queued);
2904                 } else {
2905                         LASSERT (!tx->tx_queued);
2906                         LASSERT (tx->tx_waiting || tx->tx_sending != 0);
2907                 }
2908
2909                 if (time_after_eq (jiffies, tx->tx_deadline)) {
2910                         timed_out = 1;
2911                         CERROR("Timed out tx: %s, %lu seconds\n",
2912                                kiblnd_queue2str(conn, txs),
2913                                cfs_duration_sec(jiffies - tx->tx_deadline));
2914                         break;
2915                 }
2916         }
2917
2918         spin_unlock(&conn->ibc_lock);
2919         return timed_out;
2920 }
2921
2922 int
2923 kiblnd_conn_timed_out (kib_conn_t *conn)
2924 {
2925         return  kiblnd_check_txs(conn, &conn->ibc_tx_queue) ||
2926                 kiblnd_check_txs(conn, &conn->ibc_tx_queue_rsrvd) ||
2927                 kiblnd_check_txs(conn, &conn->ibc_tx_queue_nocred) ||
2928                 kiblnd_check_txs(conn, &conn->ibc_active_txs);
2929 }
2930
2931 void
2932 kiblnd_check_conns (int idx)
2933 {
2934         struct list_head  *peers = &kiblnd_data.kib_peers[idx];
2935         struct list_head  *ptmp;
2936         kib_peer_t        *peer;
2937         kib_conn_t        *conn;
2938         struct list_head  *ctmp;
2939         unsigned long      flags;
2940
2941  again:
2942         /* NB. We expect to have a look at all the peers and not find any
2943          * rdmas to time out, so we just use a shared lock while we
2944          * take a look... */
2945         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2946
2947         list_for_each (ptmp, peers) {
2948                 peer = list_entry (ptmp, kib_peer_t, ibp_list);
2949
2950                 list_for_each (ctmp, &peer->ibp_conns) {
2951                         conn = list_entry (ctmp, kib_conn_t, ibc_list);
2952
2953                         LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
2954
2955                         /* In case we have enough credits to return via a
2956                          * NOOP, but there were no non-blocking tx descs
2957                          * free to do it last time... */
2958                         kiblnd_check_sends(conn);
2959
2960                         if (!kiblnd_conn_timed_out(conn))
2961                                 continue;
2962
2963                         /* Handle timeout by closing the whole connection.  We
2964                          * can only be sure RDMA activity has ceased once the
2965                          * QP has been modified. */
2966
2967                         kiblnd_conn_addref(conn); /* 1 ref for me... */
2968
2969                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2970                                                flags);
2971
2972                         CERROR("Timed out RDMA with %s (%lu)\n",
2973                                libcfs_nid2str(peer->ibp_nid),
2974                                cfs_duration_sec(cfs_time_current() -
2975                                                 peer->ibp_last_alive));
2976
2977                         kiblnd_close_conn(conn, -ETIMEDOUT);
2978                         kiblnd_conn_decref(conn); /* ...until here */
2979
2980                         /* start again now I've dropped the lock */
2981                         goto again;
2982                 }
2983         }
2984
2985         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2986 }
2987
2988 void
2989 kiblnd_disconnect_conn (kib_conn_t *conn)
2990 {
2991         LASSERT (!in_interrupt());
2992         LASSERT (current == kiblnd_data.kib_connd);
2993         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
2994
2995         rdma_disconnect(conn->ibc_cmid);
2996         kiblnd_finalise_conn(conn);
2997
2998         kiblnd_peer_notify(conn->ibc_peer);
2999 }
3000
3001 int
3002 kiblnd_connd (void *arg)
3003 {
3004         wait_queue_t       wait;
3005         unsigned long      flags;
3006         kib_conn_t        *conn;
3007         int                timeout;
3008         int                i;
3009         int                dropped_lock;
3010         int                peer_index = 0;
3011         unsigned long      deadline = jiffies;
3012
3013         cfs_daemonize ("kiblnd_connd");
3014         cfs_block_allsigs ();
3015
3016         init_waitqueue_entry (&wait, current);
3017         kiblnd_data.kib_connd = current;
3018
3019         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3020
3021         while (!kiblnd_data.kib_shutdown) {
3022
3023                 dropped_lock = 0;
3024
3025                 if (!list_empty (&kiblnd_data.kib_connd_zombies)) {
3026                         conn = list_entry (kiblnd_data.kib_connd_zombies.next,
3027                                            kib_conn_t, ibc_list);
3028                         list_del(&conn->ibc_list);
3029
3030                         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3031                         dropped_lock = 1;
3032
3033                         kiblnd_destroy_conn(conn);
3034
3035                         spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3036                 }
3037
3038                 if (!list_empty (&kiblnd_data.kib_connd_conns)) {
3039                         conn = list_entry (kiblnd_data.kib_connd_conns.next,
3040                                            kib_conn_t, ibc_list);
3041                         list_del(&conn->ibc_list);
3042
3043                         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3044                         dropped_lock = 1;
3045
3046                         kiblnd_disconnect_conn(conn);
3047                         kiblnd_conn_decref(conn);
3048
3049                         spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3050                 }
3051
3052                 /* careful with the jiffy wrap... */
3053                 timeout = (int)(deadline - jiffies);
3054                 if (timeout <= 0) {
3055                         const int n = 4;
3056                         const int p = 1;
3057                         int       chunk = kiblnd_data.kib_peer_hash_size;
3058
3059                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3060                         dropped_lock = 1;
3061
3062                         /* Time to check for RDMA timeouts on a few more
3063                          * peers: I do checks every 'p' seconds on a
3064                          * proportion of the peer table and I need to check
3065                          * every connection 'n' times within a timeout
3066                          * interval, to ensure I detect a timeout on any
3067                          * connection within (n+1)/n times the timeout
3068                          * interval. */
3069
3070                         if (*kiblnd_tunables.kib_timeout > n * p)
3071                                 chunk = (chunk * n * p) /
3072                                         *kiblnd_tunables.kib_timeout;
3073                         if (chunk == 0)
3074                                 chunk = 1;
3075
3076                         for (i = 0; i < chunk; i++) {
3077                                 kiblnd_check_conns(peer_index);
3078                                 peer_index = (peer_index + 1) %
3079                                              kiblnd_data.kib_peer_hash_size;
3080                         }
3081
3082                         deadline += p * HZ;
3083                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3084                 }
3085
3086                 if (dropped_lock)
3087                         continue;
3088
3089                 /* Nothing to do for 'timeout'  */
3090                 set_current_state (TASK_INTERRUPTIBLE);
3091                 add_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
3092                 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3093
3094                 schedule_timeout (timeout);
3095
3096                 set_current_state (TASK_RUNNING);
3097                 remove_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
3098                 spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3099         }
3100
3101         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3102
3103         kiblnd_thread_fini();
3104         return (0);
3105 }
3106
3107 void
3108 kiblnd_qp_event(struct ib_event *event, void *arg)
3109 {
3110         kib_conn_t *conn = arg;
3111
3112         switch (event->event) {
3113         case IB_EVENT_COMM_EST:
3114                 CDEBUG(D_NET, "%s established\n",
3115                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3116                 return;
3117
3118         default:
3119                 CERROR("%s: Async QP event type %d\n",
3120                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3121                 return;
3122         }
3123 }
3124
3125 void
3126 kiblnd_complete (struct ib_wc *wc)
3127 {
3128         switch (kiblnd_wreqid2type(wc->wr_id)) {
3129         default:
3130                 LBUG();
3131
3132         case IBLND_WID_RDMA:
3133                 /* We only get RDMA completion notification if it fails.  All
3134                  * subsequent work items, including the final SEND will fail
3135                  * too.  However we can't print out any more info about the
3136                  * failing RDMA because 'tx' might be back on the idle list or
3137                  * even reused already if we didn't manage to post all our work
3138                  * items */
3139                 CNETERR("RDMA (tx: %p) failed: %d\n",
3140                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3141                 return;
3142
3143         case IBLND_WID_TX:
3144                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3145                 return;
3146
3147         case IBLND_WID_RX:
3148                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3149                                    wc->byte_len);
3150                 return;
3151         }
3152 }
3153
3154 void
3155 kiblnd_cq_completion (struct ib_cq *cq, void *arg)
3156 {
3157         /* NB I'm not allowed to schedule this conn once its refcount has
3158          * reached 0.  Since fundamentally I'm racing with scheduler threads
3159          * consuming my CQ I could be called after all completions have
3160          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3161          * and this CQ is about to be destroyed so I NOOP. */
3162         kib_conn_t     *conn = (kib_conn_t *)arg;
3163         unsigned long   flags;
3164
3165         LASSERT (cq == conn->ibc_cq);
3166
3167         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3168
3169         conn->ibc_ready = 1;
3170
3171         if (!conn->ibc_scheduled &&
3172             (conn->ibc_nrx > 0 ||
3173              conn->ibc_nsends_posted > 0)) {
3174                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3175                 conn->ibc_scheduled = 1;
3176                 list_add_tail(&conn->ibc_sched_list,
3177                               &kiblnd_data.kib_sched_conns);
3178                 wake_up(&kiblnd_data.kib_sched_waitq);
3179         }
3180
3181         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3182 }
3183
3184 void
3185 kiblnd_cq_event(struct ib_event *event, void *arg)
3186 {
3187         kib_conn_t *conn = arg;
3188
3189         CERROR("%s: async CQ event type %d\n",
3190                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3191 }
3192
3193 int
3194 kiblnd_scheduler(void *arg)
3195 {
3196         long            id = (long)arg;
3197         wait_queue_t    wait;
3198         char            name[16];
3199         unsigned long   flags;
3200         kib_conn_t     *conn;
3201         struct ib_wc    wc;
3202         int             rc;
3203         int             did_something;
3204         int             busy_loops = 0;
3205
3206         snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id);
3207         cfs_daemonize(name);
3208         cfs_block_allsigs();
3209
3210         init_waitqueue_entry(&wait, current);
3211
3212         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3213
3214         while (!kiblnd_data.kib_shutdown) {
3215                 if (busy_loops++ >= IBLND_RESCHED) {
3216                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3217                                                flags);
3218
3219                         cfs_cond_resched();
3220                         busy_loops = 0;
3221
3222                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3223                 }
3224
3225                 did_something = 0;
3226
3227                 if (!list_empty(&kiblnd_data.kib_sched_conns)) {
3228                         conn = list_entry(kiblnd_data.kib_sched_conns.next,
3229                                           kib_conn_t, ibc_sched_list);
3230                         /* take over kib_sched_conns' ref on conn... */
3231                         LASSERT(conn->ibc_scheduled);
3232                         list_del(&conn->ibc_sched_list);
3233                         conn->ibc_ready = 0;
3234
3235                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3236                                                flags);
3237
3238                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3239                         if (rc == 0) {
3240                                 rc = ib_req_notify_cq(conn->ibc_cq,
3241                                                       IB_CQ_NEXT_COMP);
3242                                 if (rc < 0) {
3243                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3244                                               "closing connection\n",
3245                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3246                                         kiblnd_close_conn(conn, -EIO);
3247                                         kiblnd_conn_decref(conn);
3248                                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3249                                         continue;
3250                                 }
3251
3252                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3253                         }
3254
3255                         if (rc < 0) {
3256                                 CWARN("%s: ib_poll_cq failed: %d, "
3257                                       "closing connection\n",
3258                                       libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3259                                 kiblnd_close_conn(conn, -EIO);
3260                                 kiblnd_conn_decref(conn);
3261                                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3262                                 continue;
3263                         }
3264
3265                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3266                                           flags);
3267
3268                         if (rc != 0 || conn->ibc_ready) {
3269                                 /* There may be another completion waiting; get
3270                                  * another scheduler to check while I handle
3271                                  * this one... */
3272                                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3273                                 list_add_tail(&conn->ibc_sched_list,
3274                                               &kiblnd_data.kib_sched_conns);
3275                                 wake_up(&kiblnd_data.kib_sched_waitq);
3276                         } else {
3277                                 conn->ibc_scheduled = 0;
3278                         }
3279
3280                         if (rc != 0) {
3281                                 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3282                                                        flags);
3283
3284                                 kiblnd_complete(&wc);
3285
3286                                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3287                                                   flags);
3288                         }
3289
3290                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3291                         did_something = 1;
3292                 }
3293
3294                 if (did_something)
3295                         continue;
3296
3297                 set_current_state(TASK_INTERRUPTIBLE);
3298                 add_wait_queue_exclusive(&kiblnd_data.kib_sched_waitq, &wait);
3299                 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3300
3301                 schedule();
3302                 busy_loops = 0;
3303
3304                 remove_wait_queue(&kiblnd_data.kib_sched_waitq, &wait);
3305                 set_current_state(TASK_RUNNING);
3306                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3307         }
3308
3309         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3310
3311         kiblnd_thread_fini();
3312         return (0);
3313 }