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 (c) 2007, 2010, Oracle and/or its affiliates. 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 int
 714 kiblnd_setup_rd_kiov (lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 715                       int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 716 {
 717         kib_net_t          *net = ni->ni_data;
 718         struct scatterlist *sg;
 719         int                 fragnob;
 720
 721         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 722
 723         LASSERT (nob > 0);
 724         LASSERT (nkiov > 0);
 725         LASSERT (net != NULL);
 726
 727         while (offset >= kiov->kiov_len) {
 728                 offset -= kiov->kiov_len;
 729                 nkiov--;
 730                 kiov++;
 731                 LASSERT (nkiov > 0);
 732         }
 733
 734         sg = tx->tx_frags;
 735         do {
 736                 LASSERT (nkiov > 0);
 737
 738                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 739
 740                 memset(sg, 0, sizeof(*sg));
 741                 sg_set_page(sg, kiov->kiov_page, fragnob,
 742                             kiov->kiov_offset + offset);
 743                 sg++;
 744
 745                 offset = 0;
 746                 kiov++;
 747                 nkiov--;
 748                 nob -= fragnob;
 749         } while (nob > 0);
 750
 751         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 752 }
 753
 754 int
 755 kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
 756 {
 757         kib_msg_t         *msg = tx->tx_msg;
 758         kib_peer_t        *peer = conn->ibc_peer;
 759         int                ver = conn->ibc_version;
 760         int                rc;
 761         int                done;
 762         struct ib_send_wr *bad_wrq;
 763
 764         LASSERT (tx->tx_queued);
 765         /* We rely on this for QP sizing */
 766         LASSERT (tx->tx_nwrq > 0);
 767         LASSERT (tx->tx_nwrq <= 1 + IBLND_RDMA_FRAGS(ver));
 768
 769         LASSERT (credit == 0 || credit == 1);
 770         LASSERT (conn->ibc_outstanding_credits >= 0);
 771         LASSERT (conn->ibc_outstanding_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 772         LASSERT (conn->ibc_credits >= 0);
 773         LASSERT (conn->ibc_credits <= IBLND_MSG_QUEUE_SIZE(ver));
 774
 775         if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
 776                 /* tx completions outstanding... */
 777                 CDEBUG(D_NET, "%s: posted enough\n",
 778                        libcfs_nid2str(peer->ibp_nid));
 779                 return -EAGAIN;
 780         }
 781
 782         if (credit != 0 && conn->ibc_credits == 0) {   /* no credits */
 783                 CDEBUG(D_NET, "%s: no credits\n",
 784                        libcfs_nid2str(peer->ibp_nid));
 785                 return -EAGAIN;
 786         }
 787
 788         if (credit != 0 && !IBLND_OOB_CAPABLE(ver) &&
 789             conn->ibc_credits == 1 &&   /* last credit reserved for */
 790             conn->ibc_outstanding_credits == 0) { /* giving back credits */
 791                 CDEBUG(D_NET, "%s: not using last credit\n",
 792                        libcfs_nid2str(peer->ibp_nid));
 793                 return -EAGAIN;
 794         }
 795
 796         /* NB don't drop ibc_lock before bumping tx_sending */
 797         list_del(&tx->tx_list);
 798         tx->tx_queued = 0;
 799
 800         if (msg->ibm_type == IBLND_MSG_NOOP &&
 801             (!kiblnd_send_noop(conn) ||     /* redundant NOOP */
 802              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 803               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 804                 /* OK to drop when posted enough NOOPs, since
 805                  * kiblnd_check_sends will queue NOOP again when
 806                  * posted NOOPs complete */
 807                 spin_unlock(&conn->ibc_lock);
 808                 kiblnd_tx_done(peer->ibp_ni, tx);
 809                 spin_lock(&conn->ibc_lock);
 810                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 811                        libcfs_nid2str(peer->ibp_nid),
 812                        conn->ibc_noops_posted);
 813                 return 0;
 814         }
 815
 816         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 817                         peer->ibp_nid, conn->ibc_incarnation);
 818
 819         conn->ibc_credits -= credit;
 820         conn->ibc_outstanding_credits = 0;
 821         conn->ibc_nsends_posted++;
 822         if (msg->ibm_type == IBLND_MSG_NOOP)
 823                 conn->ibc_noops_posted++;
 824
 825         /* CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 826          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 827          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 828          * and then re-queued here.  It's (just) possible that
 829          * tx_sending is non-zero if we've not done the tx_complete()
 830          * from the first send; hence the ++ rather than = below. */
 831         tx->tx_sending++;
 832         list_add(&tx->tx_list, &conn->ibc_active_txs);
 833
 834         /* I'm still holding ibc_lock! */
 835         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
 836                 rc = -ECONNABORTED;
 837         else
 838                 rc = ib_post_send(conn->ibc_cmid->qp,
 839                                   tx->tx_wrq, &bad_wrq);
 840         conn->ibc_last_send = jiffies;
 841
 842         if (rc == 0)
 843                 return 0;
 844
 845         /* NB credits are transferred in the actual
 846          * message, which can only be the last work item */
 847         conn->ibc_credits += credit;
 848         conn->ibc_outstanding_credits += msg->ibm_credits;
 849         conn->ibc_nsends_posted--;
 850         if (msg->ibm_type == IBLND_MSG_NOOP)
 851                 conn->ibc_noops_posted--;
 852
 853         tx->tx_status = rc;
 854         tx->tx_waiting = 0;
 855         tx->tx_sending--;
 856
 857         done = (tx->tx_sending == 0);
 858         if (done)
 859                 list_del(&tx->tx_list);
 860
 861         spin_unlock(&conn->ibc_lock);
 862
 863         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 864                 CERROR("Error %d posting transmit to %s\n",
 865                        rc, libcfs_nid2str(peer->ibp_nid));
 866         else
 867                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 868                        rc, libcfs_nid2str(peer->ibp_nid));
 869
 870         kiblnd_close_conn(conn, rc);
 871
 872         if (done)
 873                 kiblnd_tx_done(peer->ibp_ni, tx);
 874
 875         spin_lock(&conn->ibc_lock);
 876
 877         return -EIO;
 878 }
 879
 880 void
 881 kiblnd_check_sends (kib_conn_t *conn)
 882 {
 883         int        ver = conn->ibc_version;
 884         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 885         kib_tx_t  *tx;
 886
 887         /* Don't send anything until after the connection is established */
 888         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 889                 CDEBUG(D_NET, "%s too soon\n",
 890                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 891                 return;
 892         }
 893
 894         spin_lock(&conn->ibc_lock);
 895
 896         LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
 897         LASSERT (!IBLND_OOB_CAPABLE(ver) ||
 898                  conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 899         LASSERT (conn->ibc_reserved_credits >= 0);
 900
 901         while (conn->ibc_reserved_credits > 0 &&
 902                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 903                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 904                                 kib_tx_t, tx_list);
 905                 list_del(&tx->tx_list);
 906                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 907                 conn->ibc_reserved_credits--;
 908         }
 909
 910         if (kiblnd_send_noop(conn)) {
 911                 spin_unlock(&conn->ibc_lock);
 912
 913                 tx = kiblnd_get_idle_tx(ni);
 914                 if (tx != NULL)
 915                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 916
 917                 spin_lock(&conn->ibc_lock);
 918                 if (tx != NULL)
 919                         kiblnd_queue_tx_locked(tx, conn);
 920         }
 921
 922         kiblnd_conn_addref(conn); /* 1 ref for me.... (see b21911) */
 923
 924         for (;;) {
 925                 int credit;
 926
 927                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 928                         credit = 0;
 929                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
 930                                         kib_tx_t, tx_list);
 931                 } else if (!list_empty(&conn->ibc_tx_queue)) {
 932                         credit = 1;
 933                         tx = list_entry(conn->ibc_tx_queue.next,
 934                                         kib_tx_t, tx_list);
 935                 } else
 936                         break;
 937
 938                 if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
 939                         break;
 940         }
 941
 942         spin_unlock(&conn->ibc_lock);
 943
 944         kiblnd_conn_decref(conn); /* ...until here */
 945 }
 946
 947 void
 948 kiblnd_tx_complete (kib_tx_t *tx, int status)
 949 {
 950         int           failed = (status != IB_WC_SUCCESS);
 951         kib_conn_t   *conn = tx->tx_conn;
 952         int           idle;
 953
 954         LASSERT (tx->tx_sending > 0);
 955
 956         if (failed) {
 957                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 958                         CNETERR("Tx -> %s cookie "LPX64
 959                                 " sending %d waiting %d: failed %d\n",
 960                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
 961                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
 962                                 status);
 963
 964                 kiblnd_close_conn(conn, -EIO);
 965         } else {
 966                 kiblnd_peer_alive(conn->ibc_peer);
 967         }
 968
 969         spin_lock(&conn->ibc_lock);
 970
 971         /* I could be racing with rdma completion.  Whoever makes 'tx' idle
 972          * gets to free it, which also drops its ref on 'conn'. */
 973
 974         tx->tx_sending--;
 975         conn->ibc_nsends_posted--;
 976         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
 977                 conn->ibc_noops_posted--;
 978
 979         if (failed) {
 980                 tx->tx_waiting = 0;             /* don't wait for peer */
 981                 tx->tx_status = -EIO;
 982         }
 983
 984         idle = (tx->tx_sending == 0) &&         /* This is the final callback */
 985                !tx->tx_waiting &&               /* Not waiting for peer */
 986                !tx->tx_queued;                  /* Not re-queued (PUT_DONE) */
 987         if (idle)
 988                 list_del(&tx->tx_list);
 989
 990         kiblnd_conn_addref(conn);               /* 1 ref for me.... */
 991
 992         spin_unlock(&conn->ibc_lock);
 993
 994         if (idle)
 995                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
 996
 997         kiblnd_check_sends(conn);
 998
 999         kiblnd_conn_decref(conn);               /* ...until here */
1000 }
1001
1002 void
1003 kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1004 {
1005         kib_net_t         *net = ni->ni_data;
1006         struct ib_sge     *sge = &tx->tx_sge[tx->tx_nwrq];
1007         struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1008         int                nob = offsetof (kib_msg_t, ibm_u) + body_nob;
1009         struct ib_mr      *mr;
1010
1011         LASSERT (net != NULL);
1012         LASSERT (tx->tx_nwrq >= 0);
1013         LASSERT (tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1014         LASSERT (nob <= IBLND_MSG_SIZE);
1015
1016         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1017
1018         mr = kiblnd_find_dma_mr(net, tx->tx_msgaddr, nob);
1019         LASSERT (mr != NULL);
1020
1021         sge->lkey   = mr->lkey;
1022         sge->addr   = tx->tx_msgaddr;
1023         sge->length = nob;
1024
1025         memset(wrq, 0, sizeof(*wrq));
1026
1027         wrq->next       = NULL;
1028         wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1029         wrq->sg_list    = sge;
1030         wrq->num_sge    = 1;
1031         wrq->opcode     = IB_WR_SEND;
1032         wrq->send_flags = IB_SEND_SIGNALED;
1033
1034         tx->tx_nwrq++;
1035 }
1036
1037 int
1038 kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1039                   int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1040 {
1041         kib_msg_t         *ibmsg = tx->tx_msg;
1042         kib_rdma_desc_t   *srcrd = tx->tx_rd;
1043         struct ib_sge     *sge = &tx->tx_sge[0];
1044         struct ib_send_wr *wrq = &tx->tx_wrq[0];
1045         int                rc  = resid;
1046         int                srcidx;
1047         int                dstidx;
1048         int                wrknob;
1049
1050         LASSERT (!in_interrupt());
1051         LASSERT (tx->tx_nwrq == 0);
1052         LASSERT (type == IBLND_MSG_GET_DONE ||
1053                  type == IBLND_MSG_PUT_DONE);
1054
1055         srcidx = dstidx = 0;
1056
1057         while (resid > 0) {
1058                 if (srcidx >= srcrd->rd_nfrags) {
1059                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1060                         rc = -EPROTO;
1061                         break;
1062                 }
1063
1064                 if (dstidx == dstrd->rd_nfrags) {
1065                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1066                         rc = -EPROTO;
1067                         break;
1068                 }
1069
1070                 if (tx->tx_nwrq == IBLND_RDMA_FRAGS(conn->ibc_version)) {
1071                         CERROR("RDMA too fragmented for %s (%d): "
1072                                "%d/%d src %d/%d dst frags\n",
1073                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1074                                IBLND_RDMA_FRAGS(conn->ibc_version),
1075                                srcidx, srcrd->rd_nfrags,
1076                                dstidx, dstrd->rd_nfrags);
1077                         rc = -EMSGSIZE;
1078                         break;
1079                 }
1080
1081                 wrknob = MIN(MIN(kiblnd_rd_frag_size(srcrd, srcidx),
1082                                  kiblnd_rd_frag_size(dstrd, dstidx)), resid);
1083
1084                 sge = &tx->tx_sge[tx->tx_nwrq];
1085                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1086                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1087                 sge->length = wrknob;
1088
1089                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1090
1091                 wrq->next       = wrq + 1;
1092                 wrq->wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1093                 wrq->sg_list    = sge;
1094                 wrq->num_sge    = 1;
1095                 wrq->opcode     = IB_WR_RDMA_WRITE;
1096                 wrq->send_flags = 0;
1097
1098                 wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1099                 wrq->wr.rdma.rkey        = kiblnd_rd_frag_key(dstrd, dstidx);
1100
1101                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1102                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1103
1104                 resid -= wrknob;
1105
1106                 tx->tx_nwrq++;
1107                 wrq++;
1108                 sge++;
1109         }
1110
1111         if (rc < 0)                             /* no RDMA if completing with failure */
1112                 tx->tx_nwrq = 0;
1113
1114         ibmsg->ibm_u.completion.ibcm_status = rc;
1115         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1116         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1117                            type, sizeof (kib_completion_msg_t));
1118
1119         return rc;
1120 }
1121
1122 void
1123 kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn)
1124 {
1125         struct list_head   *q;
1126
1127         LASSERT (tx->tx_nwrq > 0);              /* work items set up */
1128         LASSERT (!tx->tx_queued);               /* not queued for sending already */
1129         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1130
1131         tx->tx_queued = 1;
1132         tx->tx_deadline = jiffies + (*kiblnd_tunables.kib_timeout * HZ);
1133
1134         if (tx->tx_conn == NULL) {
1135                 kiblnd_conn_addref(conn);
1136                 tx->tx_conn = conn;
1137                 LASSERT (tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1138         } else {
1139                 /* PUT_DONE first attached to conn as a PUT_REQ */
1140                 LASSERT (tx->tx_conn == conn);
1141                 LASSERT (tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1142         }
1143
1144         switch (tx->tx_msg->ibm_type) {
1145         default:
1146                 LBUG();
1147
1148         case IBLND_MSG_PUT_REQ:
1149         case IBLND_MSG_GET_REQ:
1150                 q = &conn->ibc_tx_queue_rsrvd;
1151                 break;
1152
1153         case IBLND_MSG_PUT_NAK:
1154         case IBLND_MSG_PUT_ACK:
1155         case IBLND_MSG_PUT_DONE:
1156         case IBLND_MSG_GET_DONE:
1157                 q = &conn->ibc_tx_queue_nocred;
1158                 break;
1159
1160         case IBLND_MSG_NOOP:
1161                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1162                         q = &conn->ibc_tx_queue_nocred;
1163                 else
1164                         q = &conn->ibc_tx_queue;
1165                 break;
1166
1167         case IBLND_MSG_IMMEDIATE:
1168                 q = &conn->ibc_tx_queue;
1169                 break;
1170         }
1171
1172         list_add_tail(&tx->tx_list, q);
1173 }
1174
1175 void
1176 kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
1177 {
1178         spin_lock(&conn->ibc_lock);
1179         kiblnd_queue_tx_locked(tx, conn);
1180         spin_unlock(&conn->ibc_lock);
1181
1182         kiblnd_check_sends(conn);
1183 }
1184
1185 void
1186 kiblnd_connect_peer (kib_peer_t *peer)
1187 {
1188         struct rdma_cm_id *cmid;
1189         kib_dev_t         *dev;
1190         kib_net_t         *net = peer->ibp_ni->ni_data;
1191         struct sockaddr_in srcaddr;
1192         struct sockaddr_in dstaddr;
1193         int                rc;
1194
1195         LASSERT (net != NULL);
1196         LASSERT (peer->ibp_connecting > 0);
1197
1198         cmid = rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP);
1199         if (IS_ERR(cmid)) {
1200                 CERROR("Can't create CMID for %s: %ld\n",
1201                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1202                 rc = PTR_ERR(cmid);
1203                 goto failed;
1204         }
1205
1206         dev = net->ibn_dev;
1207         memset(&srcaddr, 0, sizeof(srcaddr));
1208         srcaddr.sin_family = AF_INET;
1209         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1210
1211         memset(&dstaddr, 0, sizeof(dstaddr));
1212         dstaddr.sin_family = AF_INET;
1213         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1214         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1215
1216         kiblnd_peer_addref(peer);               /* cmid's ref */
1217
1218         rc = rdma_resolve_addr(cmid,
1219                                (struct sockaddr *)&srcaddr,
1220                                (struct sockaddr *)&dstaddr,
1221                                *kiblnd_tunables.kib_timeout * 1000);
1222         if (rc == 0) {
1223                 LASSERT (cmid->device != NULL);
1224                 CDEBUG(D_NET, "%s: connection bound to %s:%u.%u.%u.%u:%s\n",
1225                        libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1226                        HIPQUAD(dev->ibd_ifip), cmid->device->name);
1227                 return;
1228         }
1229
1230         /* Can't initiate address resolution:  */
1231         CERROR("Can't resolve addr for %s: %d\n",
1232                libcfs_nid2str(peer->ibp_nid), rc);
1233
1234         kiblnd_peer_decref(peer);               /* cmid's ref */
1235         rdma_destroy_id(cmid);
1236  failed:
1237         kiblnd_peer_connect_failed(peer, 1, rc);
1238 }
1239
1240 void
1241 kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1242 {
1243         kib_peer_t        *peer;
1244         kib_peer_t        *peer2;
1245         kib_conn_t        *conn;
1246         rwlock_t          *g_lock = &kiblnd_data.kib_global_lock;
1247         unsigned long      flags;
1248         int                rc;
1249
1250         /* If I get here, I've committed to send, so I complete the tx with
1251          * failure on any problems */
1252
1253         LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
1254         LASSERT (tx == NULL || tx->tx_nwrq > 0);     /* work items have been set up */
1255
1256         /* First time, just use a read lock since I expect to find my peer
1257          * connected */
1258         read_lock_irqsave(g_lock, flags);
1259
1260         peer = kiblnd_find_peer_locked(nid);
1261         if (peer != NULL && !list_empty(&peer->ibp_conns)) {
1262                 /* Found a peer with an established connection */
1263                 conn = kiblnd_get_conn_locked(peer);
1264                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1265
1266                 read_unlock_irqrestore(g_lock, flags);
1267
1268                 if (tx != NULL)
1269                         kiblnd_queue_tx(tx, conn);
1270                 kiblnd_conn_decref(conn); /* ...to here */
1271                 return;
1272         }
1273
1274         read_unlock(g_lock);
1275         /* Re-try with a write lock */
1276         write_lock(g_lock);
1277
1278         peer = kiblnd_find_peer_locked(nid);
1279         if (peer != NULL) {
1280                 if (list_empty(&peer->ibp_conns)) {
1281                         /* found a peer, but it's still connecting... */
1282                         LASSERT (peer->ibp_connecting != 0 ||
1283                                  peer->ibp_accepting != 0);
1284                         if (tx != NULL)
1285                                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1286                         write_unlock_irqrestore(g_lock, flags);
1287                 } else {
1288                         conn = kiblnd_get_conn_locked(peer);
1289                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1290
1291                         write_unlock_irqrestore(g_lock, flags);
1292
1293                         if (tx != NULL)
1294                                 kiblnd_queue_tx(tx, conn);
1295                         kiblnd_conn_decref(conn); /* ...to here */
1296                 }
1297                 return;
1298         }
1299
1300         write_unlock_irqrestore(g_lock, flags);
1301
1302         /* Allocate a peer ready to add to the peer table and retry */
1303         rc = kiblnd_create_peer(ni, &peer, nid);
1304         if (rc != 0) {
1305                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1306                 if (tx != NULL) {
1307                         tx->tx_status = -EHOSTUNREACH;
1308                         tx->tx_waiting = 0;
1309                         kiblnd_tx_done(ni, tx);
1310                 }
1311                 return;
1312         }
1313
1314         write_lock_irqsave(g_lock, flags);
1315
1316         peer2 = kiblnd_find_peer_locked(nid);
1317         if (peer2 != NULL) {
1318                 if (list_empty(&peer2->ibp_conns)) {
1319                         /* found a peer, but it's still connecting... */
1320                         LASSERT (peer2->ibp_connecting != 0 ||
1321                                  peer2->ibp_accepting != 0);
1322                         if (tx != NULL)
1323                                 list_add_tail(&tx->tx_list, &peer2->ibp_tx_queue);
1324                         write_unlock_irqrestore(g_lock, flags);
1325                 } else {
1326                         conn = kiblnd_get_conn_locked(peer2);
1327                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1328
1329                         write_unlock_irqrestore(g_lock, flags);
1330
1331                         if (tx != NULL)
1332                                 kiblnd_queue_tx(tx, conn);
1333                         kiblnd_conn_decref(conn); /* ...to here */
1334                 }
1335
1336                 kiblnd_peer_decref(peer);
1337                 return;
1338         }
1339
1340         /* Brand new peer */
1341         LASSERT (peer->ibp_connecting == 0);
1342         peer->ibp_connecting = 1;
1343
1344         /* always called with a ref on ni, which prevents ni being shutdown */
1345         LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
1346
1347         if (tx != NULL)
1348                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1349
1350         kiblnd_peer_addref(peer);
1351         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1352
1353         write_unlock_irqrestore(g_lock, flags);
1354
1355         kiblnd_connect_peer(peer);
1356         kiblnd_peer_decref(peer);
1357 }
1358
1359 int
1360 kiblnd_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1361 {
1362         lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
1363         int               type = lntmsg->msg_type;
1364         lnet_process_id_t target = lntmsg->msg_target;
1365         int               target_is_router = lntmsg->msg_target_is_router;
1366         int               routing = lntmsg->msg_routing;
1367         unsigned int      payload_niov = lntmsg->msg_niov;
1368         struct iovec     *payload_iov = lntmsg->msg_iov;
1369         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
1370         unsigned int      payload_offset = lntmsg->msg_offset;
1371         unsigned int      payload_nob = lntmsg->msg_len;
1372         kib_msg_t        *ibmsg;
1373         kib_tx_t         *tx;
1374         int               nob;
1375         int               rc;
1376
1377         /* NB 'private' is different depending on what we're sending.... */
1378
1379         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1380                payload_nob, payload_niov, libcfs_id2str(target));
1381
1382         LASSERT (payload_nob == 0 || payload_niov > 0);
1383         LASSERT (payload_niov <= LNET_MAX_IOV);
1384
1385         /* Thread context */
1386         LASSERT (!in_interrupt());
1387         /* payload is either all vaddrs or all pages */
1388         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1389
1390         switch (type) {
1391         default:
1392                 LBUG();
1393                 return (-EIO);
1394
1395         case LNET_MSG_ACK:
1396                 LASSERT (payload_nob == 0);
1397                 break;
1398
1399         case LNET_MSG_GET:
1400                 if (routing || target_is_router)
1401                         break;                  /* send IMMEDIATE */
1402
1403                 /* is the REPLY message too small for RDMA? */
1404                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1405                 if (nob <= IBLND_MSG_SIZE)
1406                         break;                  /* send IMMEDIATE */
1407
1408                 tx = kiblnd_get_idle_tx(ni);
1409                 if (tx == NULL) {
1410                         CERROR("Can't allocate txd for GET to %s: \n",
1411                                libcfs_nid2str(target.nid));
1412                         return -ENOMEM;
1413                 }
1414
1415                 ibmsg = tx->tx_msg;
1416
1417                 if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
1418                         rc = kiblnd_setup_rd_iov(ni, tx,
1419                                                  &ibmsg->ibm_u.get.ibgm_rd,
1420                                                  lntmsg->msg_md->md_niov,
1421                                                  lntmsg->msg_md->md_iov.iov,
1422                                                  0, lntmsg->msg_md->md_length);
1423                 else
1424                         rc = kiblnd_setup_rd_kiov(ni, tx,
1425                                                   &ibmsg->ibm_u.get.ibgm_rd,
1426                                                   lntmsg->msg_md->md_niov,
1427                                                   lntmsg->msg_md->md_iov.kiov,
1428                                                   0, lntmsg->msg_md->md_length);
1429                 if (rc != 0) {
1430                         CERROR("Can't setup GET sink for %s: %d\n",
1431                                libcfs_nid2str(target.nid), rc);
1432                         kiblnd_tx_done(ni, tx);
1433                         return -EIO;
1434                 }
1435
1436                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[tx->tx_nfrags]);
1437                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1438                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1439
1440                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1441
1442                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1443                 if (tx->tx_lntmsg[1] == NULL) {
1444                         CERROR("Can't create reply for GET -> %s\n",
1445                                libcfs_nid2str(target.nid));
1446                         kiblnd_tx_done(ni, tx);
1447                         return -EIO;
1448                 }
1449
1450                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1451                 tx->tx_waiting = 1;             /* waiting for GET_DONE */
1452                 kiblnd_launch_tx(ni, tx, target.nid);
1453                 return 0;
1454
1455         case LNET_MSG_REPLY:
1456         case LNET_MSG_PUT:
1457                 /* Is the payload small enough not to need RDMA? */
1458                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1459                 if (nob <= IBLND_MSG_SIZE)
1460                         break;                  /* send IMMEDIATE */
1461
1462                 tx = kiblnd_get_idle_tx(ni);
1463                 if (tx == NULL) {
1464                         CERROR("Can't allocate %s txd for %s\n",
1465                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1466                                libcfs_nid2str(target.nid));
1467                         return -ENOMEM;
1468                 }
1469
1470                 if (payload_kiov == NULL)
1471                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1472                                                  payload_niov, payload_iov,
1473                                                  payload_offset, payload_nob);
1474                 else
1475                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1476                                                   payload_niov, payload_kiov,
1477                                                   payload_offset, payload_nob);
1478                 if (rc != 0) {
1479                         CERROR("Can't setup PUT src for %s: %d\n",
1480                                libcfs_nid2str(target.nid), rc);
1481                         kiblnd_tx_done(ni, tx);
1482                         return -EIO;
1483                 }
1484
1485                 ibmsg = tx->tx_msg;
1486                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1487                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1488                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1489
1490                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1491                 tx->tx_waiting = 1;             /* waiting for PUT_{ACK,NAK} */
1492                 kiblnd_launch_tx(ni, tx, target.nid);
1493                 return 0;
1494         }
1495
1496         /* send IMMEDIATE */
1497
1498         LASSERT (offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1499                  <= IBLND_MSG_SIZE);
1500
1501         tx = kiblnd_get_idle_tx(ni);
1502         if (tx == NULL) {
1503                 CERROR ("Can't send %d to %s: tx descs exhausted\n",
1504                         type, libcfs_nid2str(target.nid));
1505                 return -ENOMEM;
1506         }
1507
1508         ibmsg = tx->tx_msg;
1509         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1510
1511         if (payload_kiov != NULL)
1512                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1513                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1514                                     payload_niov, payload_kiov,
1515                                     payload_offset, payload_nob);
1516         else
1517                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1518                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1519                                    payload_niov, payload_iov,
1520                                    payload_offset, payload_nob);
1521
1522         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1523         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1524
1525         tx->tx_lntmsg[0] = lntmsg;              /* finalise lntmsg on completion */
1526         kiblnd_launch_tx(ni, tx, target.nid);
1527         return 0;
1528 }
1529
1530 void
1531 kiblnd_reply (lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1532 {
1533         lnet_process_id_t target = lntmsg->msg_target;
1534         unsigned int      niov = lntmsg->msg_niov;
1535         struct iovec     *iov = lntmsg->msg_iov;
1536         lnet_kiov_t      *kiov = lntmsg->msg_kiov;
1537         unsigned int      offset = lntmsg->msg_offset;
1538         unsigned int      nob = lntmsg->msg_len;
1539         kib_tx_t         *tx;
1540         int               rc;
1541
1542         tx = kiblnd_get_idle_tx(ni);
1543         if (tx == NULL) {
1544                 CERROR("Can't get tx for REPLY to %s\n",
1545                        libcfs_nid2str(target.nid));
1546                 goto failed_0;
1547         }
1548
1549         if (nob == 0)
1550                 rc = 0;
1551         else if (kiov == NULL)
1552                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1553                                          niov, iov, offset, nob);
1554         else
1555                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1556                                           niov, kiov, offset, nob);
1557
1558         if (rc != 0) {
1559                 CERROR("Can't setup GET src for %s: %d\n",
1560                        libcfs_nid2str(target.nid), rc);
1561                 goto failed_1;
1562         }
1563
1564         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1565                               IBLND_MSG_GET_DONE, nob,
1566                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1567                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1568         if (rc < 0) {
1569                 CERROR("Can't setup rdma for GET from %s: %d\n",
1570                        libcfs_nid2str(target.nid), rc);
1571                 goto failed_1;
1572         }
1573
1574         if (nob == 0) {
1575                 /* No RDMA: local completion may happen now! */
1576                 lnet_finalize(ni, lntmsg, 0);
1577         } else {
1578                 /* RDMA: lnet_finalize(lntmsg) when it
1579                  * completes */
1580                 tx->tx_lntmsg[0] = lntmsg;
1581         }
1582
1583         kiblnd_queue_tx(tx, rx->rx_conn);
1584         return;
1585
1586  failed_1:
1587         kiblnd_tx_done(ni, tx);
1588  failed_0:
1589         lnet_finalize(ni, lntmsg, -EIO);
1590 }
1591
1592 int
1593 kiblnd_recv (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1594              unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1595              unsigned int offset, unsigned int mlen, unsigned int rlen)
1596 {
1597         kib_rx_t    *rx = private;
1598         kib_msg_t   *rxmsg = rx->rx_msg;
1599         kib_conn_t  *conn = rx->rx_conn;
1600         kib_tx_t    *tx;
1601         kib_msg_t   *txmsg;
1602         int          nob;
1603         int          post_credit = IBLND_POSTRX_PEER_CREDIT;
1604         int          rc = 0;
1605
1606         LASSERT (mlen <= rlen);
1607         LASSERT (!in_interrupt());
1608         /* Either all pages or all vaddrs */
1609         LASSERT (!(kiov != NULL && iov != NULL));
1610
1611         switch (rxmsg->ibm_type) {
1612         default:
1613                 LBUG();
1614
1615         case IBLND_MSG_IMMEDIATE:
1616                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1617                 if (nob > rx->rx_nob) {
1618                         CERROR ("Immediate message from %s too big: %d(%d)\n",
1619                                 libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1620                                 nob, rx->rx_nob);
1621                         rc = -EPROTO;
1622                         break;
1623                 }
1624
1625                 if (kiov != NULL)
1626                         lnet_copy_flat2kiov(niov, kiov, offset,
1627                                             IBLND_MSG_SIZE, rxmsg,
1628                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1629                                             mlen);
1630                 else
1631                         lnet_copy_flat2iov(niov, iov, offset,
1632                                            IBLND_MSG_SIZE, rxmsg,
1633                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1634                                            mlen);
1635                 lnet_finalize (ni, lntmsg, 0);
1636                 break;
1637
1638         case IBLND_MSG_PUT_REQ:
1639                 if (mlen == 0) {
1640                         lnet_finalize(ni, lntmsg, 0);
1641                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1642                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1643                         break;
1644                 }
1645
1646                 tx = kiblnd_get_idle_tx(ni);
1647                 if (tx == NULL) {
1648                         CERROR("Can't allocate tx for %s\n",
1649                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1650                         /* Not replying will break the connection */
1651                         rc = -ENOMEM;
1652                         break;
1653                 }
1654
1655                 txmsg = tx->tx_msg;
1656                 if (kiov == NULL)
1657                         rc = kiblnd_setup_rd_iov(ni, tx,
1658                                                  &txmsg->ibm_u.putack.ibpam_rd,
1659                                                  niov, iov, offset, mlen);
1660                 else
1661                         rc = kiblnd_setup_rd_kiov(ni, tx,
1662                                                   &txmsg->ibm_u.putack.ibpam_rd,
1663                                                   niov, kiov, offset, mlen);
1664                 if (rc != 0) {
1665                         CERROR("Can't setup PUT sink for %s: %d\n",
1666                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1667                         kiblnd_tx_done(ni, tx);
1668                         /* tell peer it's over */
1669                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1670                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1671                         break;
1672                 }
1673
1674                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[tx->tx_nfrags]);
1675                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1676                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1677
1678                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1679
1680                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1681                 tx->tx_waiting = 1;             /* waiting for PUT_DONE */
1682                 kiblnd_queue_tx(tx, conn);
1683
1684                 /* reposted buffer reserved for PUT_DONE */
1685                 post_credit = IBLND_POSTRX_NO_CREDIT;
1686                 break;
1687
1688         case IBLND_MSG_GET_REQ:
1689                 if (lntmsg != NULL) {
1690                         /* Optimized GET; RDMA lntmsg's payload */
1691                         kiblnd_reply(ni, rx, lntmsg);
1692                 } else {
1693                         /* GET didn't match anything */
1694                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1695                                                -ENODATA,
1696                                                rxmsg->ibm_u.get.ibgm_cookie);
1697                 }
1698                 break;
1699         }
1700
1701         kiblnd_post_rx(rx, post_credit);
1702         return rc;
1703 }
1704
1705 int
1706 kiblnd_thread_start (int (*fn)(void *arg), void *arg)
1707 {
1708         long    pid = kernel_thread (fn, arg, 0);
1709
1710         if (pid < 0)
1711                 return ((int)pid);
1712
1713         atomic_inc (&kiblnd_data.kib_nthreads);
1714         return (0);
1715 }
1716
1717 void
1718 kiblnd_thread_fini (void)
1719 {
1720         atomic_dec (&kiblnd_data.kib_nthreads);
1721 }
1722
1723 void
1724 kiblnd_peer_alive (kib_peer_t *peer)
1725 {
1726         /* This is racy, but everyone's only writing cfs_time_current() */
1727         peer->ibp_last_alive = cfs_time_current();
1728         mb();
1729 }
1730
1731 void
1732 kiblnd_peer_notify (kib_peer_t *peer)
1733 {
1734         int           error = 0;
1735         cfs_time_t    last_alive = 0;
1736         unsigned long flags;
1737
1738         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1739
1740         if (list_empty(&peer->ibp_conns) &&
1741             peer->ibp_accepting == 0 &&
1742             peer->ibp_connecting == 0 &&
1743             peer->ibp_error != 0) {
1744                 error = peer->ibp_error;
1745                 peer->ibp_error = 0;
1746
1747                 last_alive = peer->ibp_last_alive;
1748         }
1749
1750         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1751
1752         if (error != 0)
1753                 lnet_notify(peer->ibp_ni,
1754                             peer->ibp_nid, 0, last_alive);
1755 }
1756
1757 void
1758 kiblnd_close_conn_locked (kib_conn_t *conn, int error)
1759 {
1760         /* This just does the immediate housekeeping.  'error' is zero for a
1761          * normal shutdown which can happen only after the connection has been
1762          * established.  If the connection is established, schedule the
1763          * connection to be finished off by the connd.  Otherwise the connd is
1764          * already dealing with it (either to set it up or tear it down).
1765          * Caller holds kib_global_lock exclusively in irq context */
1766         unsigned long     flags;
1767         kib_peer_t       *peer = conn->ibc_peer;
1768
1769         LASSERT (error != 0 || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1770
1771         if (error != 0 && conn->ibc_comms_error == 0)
1772                 conn->ibc_comms_error = error;
1773
1774         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1775                 return; /* already being handled  */
1776
1777         if (error == 0 &&
1778             list_empty(&conn->ibc_tx_queue) &&
1779             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1780             list_empty(&conn->ibc_tx_queue_nocred) &&
1781             list_empty(&conn->ibc_active_txs)) {
1782                 CDEBUG(D_NET, "closing conn to %s\n",
1783                        libcfs_nid2str(peer->ibp_nid));
1784         } else {
1785                 CNETERR("Closing conn to %s: error %d%s%s%s%s\n",
1786                         libcfs_nid2str(peer->ibp_nid), error,
1787                         list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1788                         list_empty(&conn->ibc_tx_queue_rsrvd) ?
1789                                 "" : "(sending_rsrvd)",
1790                         list_empty(&conn->ibc_tx_queue_nocred) ?
1791                                 "" : "(sending_nocred)",
1792                         list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1793         }
1794
1795         list_del(&conn->ibc_list);
1796         /* connd (see below) takes over ibc_list's ref */
1797
1798         if (list_empty (&peer->ibp_conns) &&    /* no more conns */
1799             kiblnd_peer_active(peer)) {         /* still in peer table */
1800                 kiblnd_unlink_peer_locked(peer);
1801
1802                 /* set/clear error on last conn */
1803                 peer->ibp_error = conn->ibc_comms_error;
1804         }
1805
1806         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1807
1808         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1809
1810         list_add_tail (&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1811         wake_up (&kiblnd_data.kib_connd_waitq);
1812
1813         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1814 }
1815
1816 void
1817 kiblnd_close_conn (kib_conn_t *conn, int error)
1818 {
1819         unsigned long flags;
1820
1821         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1822
1823         kiblnd_close_conn_locked(conn, error);
1824
1825         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1826 }
1827
1828 void
1829 kiblnd_handle_early_rxs(kib_conn_t *conn)
1830 {
1831         unsigned long    flags;
1832         kib_rx_t        *rx;
1833
1834         LASSERT (!in_interrupt());
1835         LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1836
1837         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1838         while (!list_empty(&conn->ibc_early_rxs)) {
1839                 rx = list_entry(conn->ibc_early_rxs.next,
1840                                 kib_rx_t, rx_list);
1841                 list_del(&rx->rx_list);
1842                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1843
1844                 kiblnd_handle_rx(rx);
1845
1846                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1847         }
1848         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1849 }
1850
1851 void
1852 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1853 {
1854         LIST_HEAD           (zombies);
1855         struct list_head    *tmp;
1856         struct list_head    *nxt;
1857         kib_tx_t            *tx;
1858
1859         spin_lock(&conn->ibc_lock);
1860
1861         list_for_each_safe (tmp, nxt, txs) {
1862                 tx = list_entry (tmp, kib_tx_t, tx_list);
1863
1864                 if (txs == &conn->ibc_active_txs) {
1865                         LASSERT (!tx->tx_queued);
1866                         LASSERT (tx->tx_waiting ||
1867                                  tx->tx_sending != 0);
1868                 } else {
1869                         LASSERT (tx->tx_queued);
1870                 }
1871
1872                 tx->tx_status = -ECONNABORTED;
1873                 tx->tx_waiting = 0;
1874
1875                 if (tx->tx_sending == 0) {
1876                         tx->tx_queued = 0;
1877                         list_del (&tx->tx_list);
1878                         list_add (&tx->tx_list, &zombies);
1879                 }
1880         }
1881
1882         spin_unlock(&conn->ibc_lock);
1883
1884         kiblnd_txlist_done(conn->ibc_peer->ibp_ni,
1885                            &zombies, -ECONNABORTED);
1886 }
1887
1888 void
1889 kiblnd_finalise_conn (kib_conn_t *conn)
1890 {
1891         LASSERT (!in_interrupt());
1892         LASSERT (conn->ibc_state > IBLND_CONN_INIT);
1893
1894         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
1895
1896         /* abort_receives moves QP state to IB_QPS_ERR.  This is only required
1897          * for connections that didn't get as far as being connected, because
1898          * rdma_disconnect() does this for free. */
1899         kiblnd_abort_receives(conn);
1900
1901         /* Complete all tx descs not waiting for sends to complete.
1902          * NB we should be safe from RDMA now that the QP has changed state */
1903
1904         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
1905         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
1906         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
1907         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
1908
1909         kiblnd_handle_early_rxs(conn);
1910 }
1911
1912 void
1913 kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error)
1914 {
1915         LIST_HEAD        (zombies);
1916         unsigned long     flags;
1917
1918         LASSERT (error != 0);
1919         LASSERT (!in_interrupt());
1920
1921         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1922
1923         if (active) {
1924                 LASSERT (peer->ibp_connecting > 0);
1925                 peer->ibp_connecting--;
1926         } else {
1927                 LASSERT (peer->ibp_accepting > 0);
1928                 peer->ibp_accepting--;
1929         }
1930
1931         if (peer->ibp_connecting != 0 ||
1932             peer->ibp_accepting != 0) {
1933                 /* another connection attempt under way... */
1934                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1935                 return;
1936         }
1937
1938         if (list_empty(&peer->ibp_conns)) {
1939                 /* Take peer's blocked transmits to complete with error */
1940                 list_add(&zombies, &peer->ibp_tx_queue);
1941                 list_del_init(&peer->ibp_tx_queue);
1942
1943                 if (kiblnd_peer_active(peer))
1944                         kiblnd_unlink_peer_locked(peer);
1945
1946                 peer->ibp_error = error;
1947         } else {
1948                 /* Can't have blocked transmits if there are connections */
1949                 LASSERT (list_empty(&peer->ibp_tx_queue));
1950         }
1951
1952         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1953
1954         kiblnd_peer_notify(peer);
1955
1956         if (list_empty (&zombies))
1957                 return;
1958
1959         CNETERR("Deleting messages for %s: connection failed\n",
1960                 libcfs_nid2str(peer->ibp_nid));
1961
1962         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
1963 }
1964
1965 void
1966 kiblnd_connreq_done(kib_conn_t *conn, int status)
1967 {
1968         kib_peer_t        *peer = conn->ibc_peer;
1969         kib_tx_t          *tx;
1970         struct list_head   txs;
1971         unsigned long      flags;
1972         int                active;
1973
1974         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
1975
1976         CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
1977                libcfs_nid2str(peer->ibp_nid), active,
1978                conn->ibc_version, status);
1979
1980         LASSERT (!in_interrupt());
1981         LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
1982                   peer->ibp_connecting > 0) ||
1983                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
1984                   peer->ibp_accepting > 0));
1985
1986         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
1987         conn->ibc_connvars = NULL;
1988
1989         if (status != 0) {
1990                 /* failed to establish connection */
1991                 kiblnd_peer_connect_failed(peer, active, status);
1992                 kiblnd_finalise_conn(conn);
1993                 return;
1994         }
1995
1996         /* connection established */
1997         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1998
1999         conn->ibc_last_send = jiffies;
2000         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2001         kiblnd_peer_alive(peer);
2002
2003         /* Add conn to peer's list and nuke any dangling conns from a different
2004          * peer instance... */
2005         kiblnd_conn_addref(conn);               /* +1 ref for ibc_list */
2006         list_add(&conn->ibc_list, &peer->ibp_conns);
2007         if (active)
2008                 peer->ibp_connecting--;
2009         else
2010                 peer->ibp_accepting--;
2011
2012         if (peer->ibp_version == 0) {
2013                 peer->ibp_version     = conn->ibc_version;
2014                 peer->ibp_incarnation = conn->ibc_incarnation;
2015         }
2016
2017         if (peer->ibp_version     != conn->ibc_version ||
2018             peer->ibp_incarnation != conn->ibc_incarnation) {
2019                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2020                                                 conn->ibc_incarnation);
2021                 peer->ibp_version     = conn->ibc_version;
2022                 peer->ibp_incarnation = conn->ibc_incarnation;
2023         }
2024
2025         /* grab pending txs while I have the lock */
2026         list_add(&txs, &peer->ibp_tx_queue);
2027         list_del_init(&peer->ibp_tx_queue);
2028
2029         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2030             conn->ibc_comms_error != 0) {       /* error has happened already */
2031                 lnet_ni_t *ni = peer->ibp_ni;
2032
2033                 /* start to shut down connection */
2034                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2035                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2036
2037                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2038
2039                 return;
2040         }
2041
2042         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2043
2044         /* Schedule blocked txs */
2045         spin_lock (&conn->ibc_lock);
2046         while (!list_empty (&txs)) {
2047                 tx = list_entry (txs.next, kib_tx_t, tx_list);
2048                 list_del(&tx->tx_list);
2049
2050                 kiblnd_queue_tx_locked(tx, conn);
2051         }
2052         spin_unlock (&conn->ibc_lock);
2053
2054         kiblnd_check_sends(conn);
2055
2056         /* schedule blocked rxs */
2057         kiblnd_handle_early_rxs(conn);
2058 }
2059
2060 void
2061 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2062 {
2063         int          rc;
2064
2065         rc = rdma_reject(cmid, rej, sizeof(*rej));
2066
2067         if (rc != 0)
2068                 CWARN("Error %d sending reject\n", rc);
2069 }
2070
2071 int
2072 kiblnd_passive_connect (struct rdma_cm_id *cmid, void *priv, int priv_nob)
2073 {
2074         rwlock_t              *g_lock = &kiblnd_data.kib_global_lock;
2075         kib_msg_t             *reqmsg = priv;
2076         kib_msg_t             *ackmsg;
2077         kib_dev_t             *ibdev;
2078         kib_peer_t            *peer;
2079         kib_peer_t            *peer2;
2080         kib_conn_t            *conn;
2081         lnet_ni_t             *ni  = NULL;
2082         kib_net_t             *net = NULL;
2083         lnet_nid_t             nid;
2084         struct rdma_conn_param cp;
2085         kib_rej_t              rej;
2086         int                    version = IBLND_MSG_VERSION;
2087         unsigned long          flags;
2088         int                    rc;
2089
2090         LASSERT (!in_interrupt());
2091
2092         /* cmid inherits 'context' from the corresponding listener id */
2093         ibdev = (kib_dev_t *)cmid->context;
2094         LASSERT (ibdev != NULL);
2095
2096         memset(&rej, 0, sizeof(rej));
2097         rej.ibr_magic                = IBLND_MSG_MAGIC;
2098         rej.ibr_why                  = IBLND_REJECT_FATAL;
2099         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2100
2101         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2102                 CERROR("Short connection request\n");
2103                 goto failed;
2104         }
2105
2106         /* Future protocol version compatibility support!  If the
2107          * o2iblnd-specific protocol changes, or when LNET unifies
2108          * protocols over all LNDs, the initial connection will
2109          * negotiate a protocol version.  I trap this here to avoid
2110          * console errors; the reject tells the peer which protocol I
2111          * speak. */
2112         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2113             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2114                 goto failed;
2115         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2116             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2117             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2118                 goto failed;
2119         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2120             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2121             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2122                 goto failed;
2123
2124         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2125         if (rc != 0) {
2126                 CERROR("Can't parse connection request: %d\n", rc);
2127                 goto failed;
2128         }
2129
2130         nid = reqmsg->ibm_srcnid;
2131         ni  = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2132
2133         if (ni != NULL) {
2134                 net = (kib_net_t *)ni->ni_data;
2135                 rej.ibr_incarnation = net->ibn_incarnation;
2136         }
2137
2138         if (ni == NULL ||                         /* no matching net */
2139             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2140             net->ibn_dev != ibdev) {              /* wrong device */
2141                 CERROR("Can't accept %s on %s (%s:%d:%u.%u.%u.%u): "
2142                        "bad dst nid %s\n", libcfs_nid2str(nid),
2143                        ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
2144                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2145                        HIPQUAD(ibdev->ibd_ifip),
2146                        libcfs_nid2str(reqmsg->ibm_dstnid));
2147
2148                 goto failed;
2149         }
2150
2151        /* check time stamp as soon as possible */
2152         if (reqmsg->ibm_dststamp != 0 &&
2153             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2154                 CWARN("Stale connection request\n");
2155                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2156                 goto failed;
2157         }
2158
2159         /* I can accept peer's version */
2160         version = reqmsg->ibm_version;
2161
2162         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2163                 CERROR("Unexpected connreq msg type: %x from %s\n",
2164                        reqmsg->ibm_type, libcfs_nid2str(nid));
2165                 goto failed;
2166         }
2167
2168         if (reqmsg->ibm_u.connparams.ibcp_queue_depth !=
2169             IBLND_MSG_QUEUE_SIZE(version)) {
2170                 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2171                        libcfs_nid2str(nid), reqmsg->ibm_u.connparams.ibcp_queue_depth,
2172                        IBLND_MSG_QUEUE_SIZE(version));
2173
2174                 if (version == IBLND_MSG_VERSION)
2175                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2176
2177                 goto failed;
2178         }
2179
2180         if (reqmsg->ibm_u.connparams.ibcp_max_frags !=
2181             IBLND_RDMA_FRAGS(version)) {
2182                 CERROR("Can't accept %s(version %x): "
2183                        "incompatible max_frags %d (%d wanted)\n",
2184                        libcfs_nid2str(nid), version,
2185                        reqmsg->ibm_u.connparams.ibcp_max_frags,
2186                        IBLND_RDMA_FRAGS(version));
2187
2188                 if (version == IBLND_MSG_VERSION)
2189                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2190
2191                 goto failed;
2192
2193         }
2194
2195         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2196                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2197                        libcfs_nid2str(nid),
2198                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2199                        IBLND_MSG_SIZE);
2200                 goto failed;
2201         }
2202
2203         /* assume 'nid' is a new peer; create  */
2204         rc = kiblnd_create_peer(ni, &peer, nid);
2205         if (rc != 0) {
2206                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2207                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2208                 goto failed;
2209         }
2210
2211         write_lock_irqsave(g_lock, flags);
2212
2213         peer2 = kiblnd_find_peer_locked(nid);
2214         if (peer2 != NULL) {
2215                 if (peer2->ibp_version == 0) {
2216                         peer2->ibp_version     = version;
2217                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2218                 }
2219
2220                 /* not the guy I've talked with */
2221                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2222                     peer2->ibp_version     != version) {
2223                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2224                         write_unlock_irqrestore(g_lock, flags);
2225
2226                         CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2227                               libcfs_nid2str(nid), peer2->ibp_version, version);
2228
2229                         kiblnd_peer_decref(peer);
2230                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2231                         goto failed;
2232                 }
2233
2234                 /* tie-break connection race in favour of the higher NID */
2235                 if (peer2->ibp_connecting != 0 &&
2236                     nid < ni->ni_nid) {
2237                         write_unlock_irqrestore(g_lock, flags);
2238
2239                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2240
2241                         kiblnd_peer_decref(peer);
2242                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2243                         goto failed;
2244                 }
2245
2246                 peer2->ibp_accepting++;
2247                 kiblnd_peer_addref(peer2);
2248
2249                 write_unlock_irqrestore(g_lock, flags);
2250                 kiblnd_peer_decref(peer);
2251                 peer = peer2;
2252         } else {
2253                 /* Brand new peer */
2254                 LASSERT (peer->ibp_accepting == 0);
2255                 LASSERT (peer->ibp_version == 0 &&
2256                          peer->ibp_incarnation == 0);
2257
2258                 peer->ibp_accepting   = 1;
2259                 peer->ibp_version     = version;
2260                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2261
2262                 /* I have a ref on ni that prevents it being shutdown */
2263                 LASSERT (net->ibn_shutdown == 0);
2264
2265                 kiblnd_peer_addref(peer);
2266                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2267
2268                 write_unlock_irqrestore(g_lock, flags);
2269         }
2270
2271         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
2272         if (conn == NULL) {
2273                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2274                 kiblnd_peer_decref(peer);
2275                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2276                 goto failed;
2277         }
2278
2279         /* conn now "owns" cmid, so I return success from here on to ensure the
2280          * CM callback doesn't destroy cmid. */
2281
2282         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2283         conn->ibc_credits          = IBLND_MSG_QUEUE_SIZE(version);
2284         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(version);
2285         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(version)
2286                  <= IBLND_RX_MSGS(version));
2287
2288         ackmsg = &conn->ibc_connvars->cv_msg;
2289         memset(ackmsg, 0, sizeof(*ackmsg));
2290
2291         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2292                         sizeof(ackmsg->ibm_u.connparams));
2293         ackmsg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2294         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2295         ackmsg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2296
2297         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2298
2299         memset(&cp, 0, sizeof(cp));
2300         cp.private_data        = ackmsg;
2301         cp.private_data_len    = ackmsg->ibm_nob;
2302         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2303         cp.initiator_depth     = 0;
2304         cp.flow_control        = 1;
2305         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2306         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2307
2308         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2309
2310         rc = rdma_accept(cmid, &cp);
2311         if (rc != 0) {
2312                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2313                 rej.ibr_version = version;
2314                 rej.ibr_why     = IBLND_REJECT_FATAL;
2315
2316                 kiblnd_reject(cmid, &rej);
2317                 kiblnd_connreq_done(conn, rc);
2318                 kiblnd_conn_decref(conn);
2319         }
2320
2321         lnet_ni_decref(ni);
2322         return 0;
2323
2324  failed:
2325         if (ni != NULL)
2326                 lnet_ni_decref(ni);
2327
2328         rej.ibr_version = version;
2329         rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2330         rej.ibr_cp.ibcp_max_frags   = IBLND_RDMA_FRAGS(version);
2331         kiblnd_reject(cmid, &rej);
2332
2333         return -ECONNREFUSED;
2334 }
2335
2336 void
2337 kiblnd_reconnect (kib_conn_t *conn, int version,
2338                   __u64 incarnation, int why, kib_connparams_t *cp)
2339 {
2340         kib_peer_t    *peer = conn->ibc_peer;
2341         char          *reason;
2342         int            retry = 0;
2343         unsigned long  flags;
2344
2345         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2346         LASSERT (peer->ibp_connecting > 0);     /* 'conn' at least */
2347
2348         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2349
2350         /* retry connection if it's still needed and no other connection
2351          * attempts (active or passive) are in progress
2352          * NB: reconnect is still needed even when ibp_tx_queue is
2353          * empty if ibp_version != version because reconnect may be
2354          * initiated by kiblnd_query() */
2355         if ((!list_empty(&peer->ibp_tx_queue) ||
2356              peer->ibp_version != version) &&
2357             peer->ibp_connecting == 1 &&
2358             peer->ibp_accepting == 0) {
2359                 retry = 1;
2360                 peer->ibp_connecting++;
2361
2362                 peer->ibp_version     = version;
2363                 peer->ibp_incarnation = incarnation;
2364         }
2365
2366         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2367
2368         if (!retry)
2369                 return;
2370
2371         switch (why) {
2372         default:
2373                 reason = "Unknown";
2374                 break;
2375
2376         case IBLND_REJECT_CONN_STALE:
2377                 reason = "stale";
2378                 break;
2379
2380         case IBLND_REJECT_CONN_RACE:
2381                 reason = "conn race";
2382                 break;
2383
2384         case IBLND_REJECT_CONN_UNCOMPAT:
2385                 reason = "version negotiation";
2386                 break;
2387         }
2388
2389         CNETERR("%s: retrying (%s), %x, %x, "
2390                 "queue_dep: %d, max_frag: %d, msg_size: %d\n",
2391                 libcfs_nid2str(peer->ibp_nid),
2392                 reason, IBLND_MSG_VERSION, version,
2393                 cp != NULL? cp->ibcp_queue_depth :IBLND_MSG_QUEUE_SIZE(version),
2394                 cp != NULL? cp->ibcp_max_frags   : IBLND_RDMA_FRAGS(version),
2395                 cp != NULL? cp->ibcp_max_msg_size: IBLND_MSG_SIZE);
2396
2397         kiblnd_connect_peer(peer);
2398 }
2399
2400 void
2401 kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
2402 {
2403         kib_peer_t    *peer = conn->ibc_peer;
2404
2405         LASSERT (!in_interrupt());
2406         LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2407
2408         switch (reason) {
2409         case IB_CM_REJ_STALE_CONN:
2410                 kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
2411                                  IBLND_REJECT_CONN_STALE, NULL);
2412                 break;
2413
2414         case IB_CM_REJ_INVALID_SERVICE_ID:
2415                 CNETERR("%s rejected: no listener at %d\n",
2416                         libcfs_nid2str(peer->ibp_nid),
2417                         *kiblnd_tunables.kib_service);
2418                 break;
2419
2420         case IB_CM_REJ_CONSUMER_DEFINED:
2421                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2422                         kib_rej_t        *rej         = priv;
2423                         kib_connparams_t *cp          = NULL;
2424                         int               flip        = 0;
2425                         __u64             incarnation = -1;
2426
2427                         /* NB. default incarnation is -1 because:
2428                          * a) V1 will ignore dst incarnation in connreq.
2429                          * b) V2 will provide incarnation while rejecting me,
2430                          *    -1 will be overwrote.
2431                          *
2432                          * if I try to connect to a V1 peer with V2 protocol,
2433                          * it rejected me then upgrade to V2, I have no idea
2434                          * about the upgrading and try to reconnect with V1,
2435                          * in this case upgraded V2 can find out I'm trying to
2436                          * talk to the old guy and reject me(incarnation is -1).
2437                          */
2438
2439                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2440                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2441                                 __swab32s(&rej->ibr_magic);
2442                                 __swab16s(&rej->ibr_version);
2443                                 flip = 1;
2444                         }
2445
2446                         if (priv_nob >= sizeof(kib_rej_t) &&
2447                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2448                                 /* priv_nob is always 148 in current version
2449                                  * of OFED, so we still need to check version.
2450                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2451                                 cp = &rej->ibr_cp;
2452
2453                                 if (flip) {
2454                                         __swab64s(&rej->ibr_incarnation);
2455                                         __swab16s(&cp->ibcp_queue_depth);
2456                                         __swab16s(&cp->ibcp_max_frags);
2457                                         __swab32s(&cp->ibcp_max_msg_size);
2458                                 }
2459
2460                                 incarnation = rej->ibr_incarnation;
2461                         }
2462
2463                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2464                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2465                                 CERROR("%s rejected: consumer defined fatal error\n",
2466                                        libcfs_nid2str(peer->ibp_nid));
2467                                 break;
2468                         }
2469
2470                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2471                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2472                                 CERROR("%s rejected: o2iblnd version %x error\n",
2473                                        libcfs_nid2str(peer->ibp_nid),
2474                                        rej->ibr_version);
2475                                 break;
2476                         }
2477
2478                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2479                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2480                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2481                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2482
2483                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2484                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2485                         }
2486
2487                         switch (rej->ibr_why) {
2488                         case IBLND_REJECT_CONN_RACE:
2489                         case IBLND_REJECT_CONN_STALE:
2490                         case IBLND_REJECT_CONN_UNCOMPAT:
2491                                 kiblnd_reconnect(conn, rej->ibr_version,
2492                                                  incarnation, rej->ibr_why, cp);
2493                                 break;
2494
2495                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2496                                 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2497                                        libcfs_nid2str(peer->ibp_nid), cp->ibcp_queue_depth,
2498                                        IBLND_MSG_QUEUE_SIZE(conn->ibc_version));
2499                                 break;
2500
2501                         case IBLND_REJECT_RDMA_FRAGS:
2502                                 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2503                                        libcfs_nid2str(peer->ibp_nid), cp->ibcp_max_frags,
2504                                        IBLND_RDMA_FRAGS(conn->ibc_version));
2505                                 break;
2506
2507                         case IBLND_REJECT_NO_RESOURCES:
2508                                 CERROR("%s rejected: o2iblnd no resources\n",
2509                                        libcfs_nid2str(peer->ibp_nid));
2510                                 break;
2511
2512                         case IBLND_REJECT_FATAL:
2513                                 CERROR("%s rejected: o2iblnd fatal error\n",
2514                                        libcfs_nid2str(peer->ibp_nid));
2515                                 break;
2516
2517                         default:
2518                                 CERROR("%s rejected: o2iblnd reason %d\n",
2519                                        libcfs_nid2str(peer->ibp_nid),
2520                                        rej->ibr_why);
2521                                 break;
2522                         }
2523                         break;
2524                 }
2525                 /* fall through */
2526         default:
2527                 CNETERR("%s rejected: reason %d, size %d\n",
2528                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2529                 break;
2530         }
2531
2532         kiblnd_connreq_done(conn, -ECONNREFUSED);
2533 }
2534
2535 void
2536 kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
2537 {
2538         kib_peer_t    *peer = conn->ibc_peer;
2539         lnet_ni_t     *ni   = peer->ibp_ni;
2540         kib_net_t     *net  = ni->ni_data;
2541         kib_msg_t     *msg  = priv;
2542         int            ver  = conn->ibc_version;
2543         int            rc   = kiblnd_unpack_msg(msg, priv_nob);
2544         unsigned long  flags;
2545
2546         LASSERT (net != NULL);
2547
2548         if (rc != 0) {
2549                 CERROR("Can't unpack connack from %s: %d\n",
2550                        libcfs_nid2str(peer->ibp_nid), rc);
2551                 goto failed;
2552         }
2553
2554         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2555                 CERROR("Unexpected message %d from %s\n",
2556                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2557                 rc = -EPROTO;
2558                 goto failed;
2559         }
2560
2561         if (ver != msg->ibm_version) {
2562                 CERROR("%s replied version %x is different with "
2563                        "requested version %x\n",
2564                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2565                 rc = -EPROTO;
2566                 goto failed;
2567         }
2568
2569         if (msg->ibm_u.connparams.ibcp_queue_depth !=
2570             IBLND_MSG_QUEUE_SIZE(ver)) {
2571                 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2572                        libcfs_nid2str(peer->ibp_nid),
2573                        msg->ibm_u.connparams.ibcp_queue_depth,
2574                        IBLND_MSG_QUEUE_SIZE(ver));
2575                 rc = -EPROTO;
2576                 goto failed;
2577         }
2578
2579         if (msg->ibm_u.connparams.ibcp_max_frags !=
2580             IBLND_RDMA_FRAGS(ver)) {
2581                 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2582                        libcfs_nid2str(peer->ibp_nid),
2583                        msg->ibm_u.connparams.ibcp_max_frags,
2584                        IBLND_RDMA_FRAGS(ver));
2585                 rc = -EPROTO;
2586                 goto failed;
2587         }
2588
2589         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2590                 CERROR("%s max message size %d too big (%d max)\n",
2591                        libcfs_nid2str(peer->ibp_nid),
2592                        msg->ibm_u.connparams.ibcp_max_msg_size,
2593                        IBLND_MSG_SIZE);
2594                 rc = -EPROTO;
2595                 goto failed;
2596         }
2597
2598         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2599         if (msg->ibm_dstnid == ni->ni_nid &&
2600             msg->ibm_dststamp == net->ibn_incarnation)
2601                 rc = 0;
2602         else
2603                 rc = -ESTALE;
2604         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2605
2606         if (rc != 0) {
2607                 CERROR("Bad connection reply from %s, rc = %d, "
2608                        "version: %x max_frags: %d\n",
2609                        libcfs_nid2str(peer->ibp_nid), rc,
2610                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2611                 goto failed;
2612         }
2613
2614         conn->ibc_incarnation      = msg->ibm_srcstamp;
2615         conn->ibc_credits          =
2616         conn->ibc_reserved_credits = IBLND_MSG_QUEUE_SIZE(ver);
2617         LASSERT (conn->ibc_credits + conn->ibc_reserved_credits + IBLND_OOB_MSGS(ver)
2618                  <= IBLND_RX_MSGS(ver));
2619
2620         kiblnd_connreq_done(conn, 0);
2621         return;
2622
2623  failed:
2624         /* NB My QP has already established itself, so I handle anything going
2625          * wrong here by setting ibc_comms_error.
2626          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2627          * immediately tears it down. */
2628
2629         LASSERT (rc != 0);
2630         conn->ibc_comms_error = rc;
2631         kiblnd_connreq_done(conn, 0);
2632 }
2633
2634 int
2635 kiblnd_active_connect (struct rdma_cm_id *cmid)
2636 {
2637         kib_peer_t              *peer = (kib_peer_t *)cmid->context;
2638         kib_conn_t              *conn;
2639         kib_msg_t               *msg;
2640         struct rdma_conn_param   cp;
2641         int                      version;
2642         __u64                    incarnation;
2643         unsigned long            flags;
2644         int                      rc;
2645
2646         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2647
2648         incarnation = peer->ibp_incarnation;
2649         version     = (peer->ibp_version == 0) ? IBLND_MSG_VERSION : peer->ibp_version;
2650
2651         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2652
2653         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT, version);
2654         if (conn == NULL) {
2655                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2656                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2657                 return -ENOMEM;
2658         }
2659
2660         /* conn "owns" cmid now, so I return success from here on to ensure the
2661          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2662          * on peer */
2663
2664         msg = &conn->ibc_connvars->cv_msg;
2665
2666         memset(msg, 0, sizeof(*msg));
2667         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2668         msg->ibm_u.connparams.ibcp_queue_depth  = IBLND_MSG_QUEUE_SIZE(version);
2669         msg->ibm_u.connparams.ibcp_max_frags    = IBLND_RDMA_FRAGS(version);
2670         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2671
2672         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2673                         0, peer->ibp_nid, incarnation);
2674
2675         memset(&cp, 0, sizeof(cp));
2676         cp.private_data        = msg;
2677         cp.private_data_len    = msg->ibm_nob;
2678         cp.responder_resources = 0;             /* No atomic ops or RDMA reads */
2679         cp.initiator_depth     = 0;
2680         cp.flow_control        = 1;
2681         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2682         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2683
2684         LASSERT(cmid->context == (void *)conn);
2685         LASSERT(conn->ibc_cmid == cmid);
2686
2687         rc = rdma_connect(cmid, &cp);
2688         if (rc != 0) {
2689                 CERROR("Can't connect to %s: %d\n",
2690                        libcfs_nid2str(peer->ibp_nid), rc);
2691                 kiblnd_connreq_done(conn, rc);
2692                 kiblnd_conn_decref(conn);
2693         }
2694
2695         return 0;
2696 }
2697
2698 int
2699 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2700 {
2701         kib_peer_t  *peer;
2702         kib_conn_t  *conn;
2703         int          rc;
2704
2705         switch (event->event) {
2706         default:
2707                 CERROR("Unexpected event: %d, status: %d\n",
2708                        event->event, event->status);
2709                 LBUG();
2710
2711         case RDMA_CM_EVENT_CONNECT_REQUEST:
2712                 /* destroy cmid on failure */
2713                 rc = kiblnd_passive_connect(cmid,
2714                                             (void *)KIBLND_CONN_PARAM(event),
2715                                             KIBLND_CONN_PARAM_LEN(event));
2716                 CDEBUG(D_NET, "connreq: %d\n", rc);
2717                 return rc;
2718
2719         case RDMA_CM_EVENT_ADDR_ERROR:
2720                 peer = (kib_peer_t *)cmid->context;
2721                 CNETERR("%s: ADDR ERROR %d\n",
2722                         libcfs_nid2str(peer->ibp_nid), event->status);
2723                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2724                 kiblnd_peer_decref(peer);
2725                 return -EHOSTUNREACH;      /* rc != 0 destroys cmid */
2726
2727         case RDMA_CM_EVENT_ADDR_RESOLVED:
2728                 peer = (kib_peer_t *)cmid->context;
2729
2730                 CDEBUG(D_NET,"%s Addr resolved: %d\n",
2731                        libcfs_nid2str(peer->ibp_nid), event->status);
2732
2733                 if (event->status != 0) {
2734                         CNETERR("Can't resolve address for %s: %d\n",
2735                                 libcfs_nid2str(peer->ibp_nid), event->status);
2736                         rc = event->status;
2737                 } else {
2738                         rc = rdma_resolve_route(
2739                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2740                         if (rc == 0)
2741                                 return 0;
2742                         /* Can't initiate route resolution */
2743                         CERROR("Can't resolve route for %s: %d\n",
2744                                libcfs_nid2str(peer->ibp_nid), rc);
2745                 }
2746                 kiblnd_peer_connect_failed(peer, 1, rc);
2747                 kiblnd_peer_decref(peer);
2748                 return rc;                      /* rc != 0 destroys cmid */
2749
2750         case RDMA_CM_EVENT_ROUTE_ERROR:
2751                 peer = (kib_peer_t *)cmid->context;
2752                 CNETERR("%s: ROUTE ERROR %d\n",
2753                         libcfs_nid2str(peer->ibp_nid), event->status);
2754                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2755                 kiblnd_peer_decref(peer);
2756                 return -EHOSTUNREACH;           /* rc != 0 destroys cmid */
2757
2758         case RDMA_CM_EVENT_ROUTE_RESOLVED:
2759                 peer = (kib_peer_t *)cmid->context;
2760                 CDEBUG(D_NET,"%s Route resolved: %d\n",
2761                        libcfs_nid2str(peer->ibp_nid), event->status);
2762
2763                 if (event->status == 0)
2764                         return kiblnd_active_connect(cmid);
2765
2766                 CNETERR("Can't resolve route for %s: %d\n",
2767                         libcfs_nid2str(peer->ibp_nid), event->status);
2768                 kiblnd_peer_connect_failed(peer, 1, event->status);
2769                 kiblnd_peer_decref(peer);
2770                 return event->status;           /* rc != 0 destroys cmid */
2771
2772         case RDMA_CM_EVENT_UNREACHABLE:
2773                 conn = (kib_conn_t *)cmid->context;
2774                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2775                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2776                 CNETERR("%s: UNREACHABLE %d\n",
2777                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2778                 kiblnd_connreq_done(conn, -ENETDOWN);
2779                 kiblnd_conn_decref(conn);
2780                 return 0;
2781
2782         case RDMA_CM_EVENT_CONNECT_ERROR:
2783                 conn = (kib_conn_t *)cmid->context;
2784                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
2785                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
2786                 CNETERR("%s: CONNECT ERROR %d\n",
2787                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
2788                 kiblnd_connreq_done(conn, -ENOTCONN);
2789                 kiblnd_conn_decref(conn);
2790                 return 0;
2791
2792         case RDMA_CM_EVENT_REJECTED:
2793                 conn = (kib_conn_t *)cmid->context;
2794                 switch (conn->ibc_state) {
2795                 default:
2796                         LBUG();
2797
2798                 case IBLND_CONN_PASSIVE_WAIT:
2799                         CERROR ("%s: REJECTED %d\n",
2800                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
2801                                 event->status);
2802                         kiblnd_connreq_done(conn, -ECONNRESET);
2803                         break;
2804
2805                 case IBLND_CONN_ACTIVE_CONNECT:
2806                         kiblnd_rejected(conn, event->status,
2807                                         (void *)KIBLND_CONN_PARAM(event),
2808                                         KIBLND_CONN_PARAM_LEN(event));
2809                         break;
2810                 }
2811                 kiblnd_conn_decref(conn);
2812                 return 0;
2813
2814         case RDMA_CM_EVENT_ESTABLISHED:
2815                 conn = (kib_conn_t *)cmid->context;
2816                 switch (conn->ibc_state) {
2817                 default:
2818                         LBUG();
2819
2820                 case IBLND_CONN_PASSIVE_WAIT:
2821                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
2822                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2823                         kiblnd_connreq_done(conn, 0);
2824                         break;
2825
2826                 case IBLND_CONN_ACTIVE_CONNECT:
2827                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
2828                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2829                         kiblnd_check_connreply(conn,
2830                                                (void *)KIBLND_CONN_PARAM(event),
2831                                                KIBLND_CONN_PARAM_LEN(event));
2832                         break;
2833                 }
2834                 /* net keeps its ref on conn! */
2835                 return 0;
2836
2837 #ifdef HAVE_OFED_RDMA_CMEV_TIMEWAIT_EXIT
2838         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2839                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
2840                 return 0;
2841 #endif
2842         case RDMA_CM_EVENT_DISCONNECTED:
2843                 conn = (kib_conn_t *)cmid->context;
2844                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
2845                         CERROR("%s DISCONNECTED\n",
2846                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
2847                         kiblnd_connreq_done(conn, -ECONNRESET);
2848                 } else {
2849                         kiblnd_close_conn(conn, 0);
2850                 }
2851                 kiblnd_conn_decref(conn);
2852                 cmid->context = NULL;
2853                 return 0;
2854
2855         case RDMA_CM_EVENT_DEVICE_REMOVAL:
2856                 LCONSOLE_ERROR_MSG(0x131,
2857                                    "Received notification of device removal\n"
2858                                    "Please shutdown LNET to allow this to proceed\n");
2859                 /* Can't remove network from underneath LNET for now, so I have
2860                  * to ignore this */
2861                 return 0;
2862
2863 #ifdef HAVE_OFED_RDMA_CMEV_ADDRCHANGE
2864         case RDMA_CM_EVENT_ADDR_CHANGE:
2865                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2866                 return 0;
2867 #endif
2868         }
2869 }
2870
2871 int
2872 kiblnd_check_txs (kib_conn_t *conn, struct list_head *txs)
2873 {
2874         kib_tx_t          *tx;
2875         struct list_head  *ttmp;
2876         int                timed_out = 0;
2877
2878         spin_lock(&conn->ibc_lock);
2879
2880         list_for_each (ttmp, txs) {
2881                 tx = list_entry (ttmp, kib_tx_t, tx_list);
2882
2883                 if (txs != &conn->ibc_active_txs) {
2884                         LASSERT (tx->tx_queued);
2885                 } else {
2886                         LASSERT (!tx->tx_queued);
2887                         LASSERT (tx->tx_waiting || tx->tx_sending != 0);
2888                 }
2889
2890                 if (time_after_eq (jiffies, tx->tx_deadline)) {
2891                         timed_out = 1;
2892                         CERROR("Timed out tx: %s, %lu seconds\n",
2893                                kiblnd_queue2str(conn, txs),
2894                                cfs_duration_sec(jiffies - tx->tx_deadline));
2895                         break;
2896                 }
2897         }
2898
2899         spin_unlock(&conn->ibc_lock);
2900         return timed_out;
2901 }
2902
2903 int
2904 kiblnd_conn_timed_out (kib_conn_t *conn)
2905 {
2906         return  kiblnd_check_txs(conn, &conn->ibc_tx_queue) ||
2907                 kiblnd_check_txs(conn, &conn->ibc_tx_queue_rsrvd) ||
2908                 kiblnd_check_txs(conn, &conn->ibc_tx_queue_nocred) ||
2909                 kiblnd_check_txs(conn, &conn->ibc_active_txs);
2910 }
2911
2912 void
2913 kiblnd_check_conns (int idx)
2914 {
2915         struct list_head  *peers = &kiblnd_data.kib_peers[idx];
2916         struct list_head  *ptmp;
2917         kib_peer_t        *peer;
2918         kib_conn_t        *conn;
2919         struct list_head  *ctmp;
2920         unsigned long      flags;
2921
2922  again:
2923         /* NB. We expect to have a look at all the peers and not find any
2924          * rdmas to time out, so we just use a shared lock while we
2925          * take a look... */
2926         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2927
2928         list_for_each (ptmp, peers) {
2929                 peer = list_entry (ptmp, kib_peer_t, ibp_list);
2930
2931                 list_for_each (ctmp, &peer->ibp_conns) {
2932                         conn = list_entry (ctmp, kib_conn_t, ibc_list);
2933
2934                         LASSERT (conn->ibc_state == IBLND_CONN_ESTABLISHED);
2935
2936                         /* In case we have enough credits to return via a
2937                          * NOOP, but there were no non-blocking tx descs
2938                          * free to do it last time... */
2939                         kiblnd_check_sends(conn);
2940
2941                         if (!kiblnd_conn_timed_out(conn))
2942                                 continue;
2943
2944                         /* Handle timeout by closing the whole connection.  We
2945                          * can only be sure RDMA activity has ceased once the
2946                          * QP has been modified. */
2947
2948                         kiblnd_conn_addref(conn); /* 1 ref for me... */
2949
2950                         read_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2951                                                flags);
2952
2953                         CERROR("Timed out RDMA with %s (%lu)\n",
2954                                libcfs_nid2str(peer->ibp_nid),
2955                                cfs_duration_sec(cfs_time_current() -
2956                                                 peer->ibp_last_alive));
2957
2958                         kiblnd_close_conn(conn, -ETIMEDOUT);
2959                         kiblnd_conn_decref(conn); /* ...until here */
2960
2961                         /* start again now I've dropped the lock */
2962                         goto again;
2963                 }
2964         }
2965
2966         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2967 }
2968
2969 void
2970 kiblnd_disconnect_conn (kib_conn_t *conn)
2971 {
2972         LASSERT (!in_interrupt());
2973         LASSERT (current == kiblnd_data.kib_connd);
2974         LASSERT (conn->ibc_state == IBLND_CONN_CLOSING);
2975
2976         rdma_disconnect(conn->ibc_cmid);
2977         kiblnd_finalise_conn(conn);
2978
2979         kiblnd_peer_notify(conn->ibc_peer);
2980 }
2981
2982 int
2983 kiblnd_connd (void *arg)
2984 {
2985         wait_queue_t       wait;
2986         unsigned long      flags;
2987         kib_conn_t        *conn;
2988         int                timeout;
2989         int                i;
2990         int                dropped_lock;
2991         int                peer_index = 0;
2992         unsigned long      deadline = jiffies;
2993
2994         cfs_daemonize ("kiblnd_connd");
2995         cfs_block_allsigs ();
2996
2997         init_waitqueue_entry (&wait, current);
2998         kiblnd_data.kib_connd = current;
2999
3000         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3001
3002         while (!kiblnd_data.kib_shutdown) {
3003
3004                 dropped_lock = 0;
3005
3006                 if (!list_empty (&kiblnd_data.kib_connd_zombies)) {
3007                         conn = list_entry (kiblnd_data.kib_connd_zombies.next,
3008                                            kib_conn_t, ibc_list);
3009                         list_del(&conn->ibc_list);
3010
3011                         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3012                         dropped_lock = 1;
3013
3014                         kiblnd_destroy_conn(conn);
3015
3016                         spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3017                 }
3018
3019                 if (!list_empty (&kiblnd_data.kib_connd_conns)) {
3020                         conn = list_entry (kiblnd_data.kib_connd_conns.next,
3021                                            kib_conn_t, ibc_list);
3022                         list_del(&conn->ibc_list);
3023
3024                         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3025                         dropped_lock = 1;
3026
3027                         kiblnd_disconnect_conn(conn);
3028                         kiblnd_conn_decref(conn);
3029
3030                         spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3031                 }
3032
3033                 /* careful with the jiffy wrap... */
3034                 timeout = (int)(deadline - jiffies);
3035                 if (timeout <= 0) {
3036                         const int n = 4;
3037                         const int p = 1;
3038                         int       chunk = kiblnd_data.kib_peer_hash_size;
3039
3040                         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
3041                         dropped_lock = 1;
3042
3043                         /* Time to check for RDMA timeouts on a few more
3044                          * peers: I do checks every 'p' seconds on a
3045                          * proportion of the peer table and I need to check
3046                          * every connection 'n' times within a timeout
3047                          * interval, to ensure I detect a timeout on any
3048                          * connection within (n+1)/n times the timeout
3049                          * interval. */
3050
3051                         if (*kiblnd_tunables.kib_timeout > n * p)
3052                                 chunk = (chunk * n * p) /
3053                                         *kiblnd_tunables.kib_timeout;
3054                         if (chunk == 0)
3055                                 chunk = 1;
3056
3057                         for (i = 0; i < chunk; i++) {
3058                                 kiblnd_check_conns(peer_index);
3059                                 peer_index = (peer_index + 1) %
3060                                              kiblnd_data.kib_peer_hash_size;
3061                         }
3062
3063                         deadline += p * HZ;
3064                         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
3065                 }
3066
3067                 if (dropped_lock)
3068                         continue;
3069
3070                 /* Nothing to do for 'timeout'  */
3071                 set_current_state (TASK_INTERRUPTIBLE);
3072                 add_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
3073                 spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3074
3075                 schedule_timeout (timeout);
3076
3077                 set_current_state (TASK_RUNNING);
3078                 remove_wait_queue (&kiblnd_data.kib_connd_waitq, &wait);
3079                 spin_lock_irqsave (&kiblnd_data.kib_connd_lock, flags);
3080         }
3081
3082         spin_unlock_irqrestore (&kiblnd_data.kib_connd_lock, flags);
3083
3084         kiblnd_thread_fini();
3085         return (0);
3086 }
3087
3088 void
3089 kiblnd_qp_event(struct ib_event *event, void *arg)
3090 {
3091         kib_conn_t *conn = arg;
3092
3093         switch (event->event) {
3094         case IB_EVENT_COMM_EST:
3095                 CDEBUG(D_NET, "%s established\n",
3096                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3097                 return;
3098
3099         default:
3100                 CERROR("%s: Async QP event type %d\n",
3101                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3102                 return;
3103         }
3104 }
3105
3106 void
3107 kiblnd_complete (struct ib_wc *wc)
3108 {
3109         switch (kiblnd_wreqid2type(wc->wr_id)) {
3110         default:
3111                 LBUG();
3112
3113         case IBLND_WID_RDMA:
3114                 /* We only get RDMA completion notification if it fails.  All
3115                  * subsequent work items, including the final SEND will fail
3116                  * too.  However we can't print out any more info about the
3117                  * failing RDMA because 'tx' might be back on the idle list or
3118                  * even reused already if we didn't manage to post all our work
3119                  * items */
3120                 CNETERR("RDMA (tx: %p) failed: %d\n",
3121                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3122                 return;
3123
3124         case IBLND_WID_TX:
3125                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3126                 return;
3127
3128         case IBLND_WID_RX:
3129                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3130                                    wc->byte_len);
3131                 return;
3132         }
3133 }
3134
3135 void
3136 kiblnd_cq_completion (struct ib_cq *cq, void *arg)
3137 {
3138         /* NB I'm not allowed to schedule this conn once its refcount has
3139          * reached 0.  Since fundamentally I'm racing with scheduler threads
3140          * consuming my CQ I could be called after all completions have
3141          * occurred.  But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3142          * and this CQ is about to be destroyed so I NOOP. */
3143         kib_conn_t     *conn = (kib_conn_t *)arg;
3144         unsigned long   flags;
3145
3146         LASSERT (cq == conn->ibc_cq);
3147
3148         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3149
3150         conn->ibc_ready = 1;
3151
3152         if (!conn->ibc_scheduled &&
3153             (conn->ibc_nrx > 0 ||
3154              conn->ibc_nsends_posted > 0)) {
3155                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3156                 conn->ibc_scheduled = 1;
3157                 list_add_tail(&conn->ibc_sched_list,
3158                               &kiblnd_data.kib_sched_conns);
3159                 wake_up(&kiblnd_data.kib_sched_waitq);
3160         }
3161
3162         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3163 }
3164
3165 void
3166 kiblnd_cq_event(struct ib_event *event, void *arg)
3167 {
3168         kib_conn_t *conn = arg;
3169
3170         CERROR("%s: async CQ event type %d\n",
3171                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3172 }
3173
3174 int
3175 kiblnd_scheduler(void *arg)
3176 {
3177         long            id = (long)arg;
3178         wait_queue_t    wait;
3179         char            name[16];
3180         unsigned long   flags;
3181         kib_conn_t     *conn;
3182         struct ib_wc    wc;
3183         int             rc;
3184         int             did_something;
3185         int             busy_loops = 0;
3186
3187         snprintf(name, sizeof(name), "kiblnd_sd_%02ld", id);
3188         cfs_daemonize(name);
3189         cfs_block_allsigs();
3190
3191         init_waitqueue_entry(&wait, current);
3192
3193         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3194
3195         while (!kiblnd_data.kib_shutdown) {
3196                 if (busy_loops++ >= IBLND_RESCHED) {
3197                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3198                                                flags);
3199
3200                         cfs_cond_resched();
3201                         busy_loops = 0;
3202
3203                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3204                 }
3205
3206                 did_something = 0;
3207
3208                 if (!list_empty(&kiblnd_data.kib_sched_conns)) {
3209                         conn = list_entry(kiblnd_data.kib_sched_conns.next,
3210                                           kib_conn_t, ibc_sched_list);
3211                         /* take over kib_sched_conns' ref on conn... */
3212                         LASSERT(conn->ibc_scheduled);
3213                         list_del(&conn->ibc_sched_list);
3214                         conn->ibc_ready = 0;
3215
3216                         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3217                                                flags);
3218
3219                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3220                         if (rc == 0) {
3221                                 rc = ib_req_notify_cq(conn->ibc_cq,
3222                                                       IB_CQ_NEXT_COMP);
3223                                 if (rc < 0) {
3224                                         CWARN("%s: ib_req_notify_cq failed: %d, "
3225                                               "closing connection\n",
3226                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3227                                         kiblnd_close_conn(conn, -EIO);
3228                                         kiblnd_conn_decref(conn);
3229                                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3230                                         continue;
3231                                 }
3232
3233                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3234                         }
3235
3236                         if (rc < 0) {
3237                                 CWARN("%s: ib_poll_cq failed: %d, "
3238                                       "closing connection\n",
3239                                       libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3240                                 kiblnd_close_conn(conn, -EIO);
3241                                 kiblnd_conn_decref(conn);
3242                                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3243                                 continue;
3244                         }
3245
3246                         spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3247                                           flags);
3248
3249                         if (rc != 0 || conn->ibc_ready) {
3250                                 /* There may be another completion waiting; get
3251                                  * another scheduler to check while I handle
3252                                  * this one... */
3253                                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3254                                 list_add_tail(&conn->ibc_sched_list,
3255                                               &kiblnd_data.kib_sched_conns);
3256                                 wake_up(&kiblnd_data.kib_sched_waitq);
3257                         } else {
3258                                 conn->ibc_scheduled = 0;
3259                         }
3260
3261                         if (rc != 0) {
3262                                 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock,
3263                                                        flags);
3264
3265                                 kiblnd_complete(&wc);
3266
3267                                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock,
3268                                                   flags);
3269                         }
3270
3271                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3272                         did_something = 1;
3273                 }
3274
3275                 if (did_something)
3276                         continue;
3277
3278                 set_current_state(TASK_INTERRUPTIBLE);
3279                 add_wait_queue_exclusive(&kiblnd_data.kib_sched_waitq, &wait);
3280                 spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3281
3282                 schedule();
3283                 busy_loops = 0;
3284
3285                 remove_wait_queue(&kiblnd_data.kib_sched_waitq, &wait);
3286                 set_current_state(TASK_RUNNING);
3287                 spin_lock_irqsave(&kiblnd_data.kib_sched_lock, flags);
3288         }
3289
3290         spin_unlock_irqrestore(&kiblnd_data.kib_sched_lock, flags);
3291
3292         kiblnd_thread_fini();
3293         return (0);
3294 }