drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd_cb.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/o2iblnd/o2iblnd_cb.c
  37  *
  38  * Author: Eric Barton <eric@bartonsoftware.com>
  39  */
  40
  41 #include "o2iblnd.h"
  42
  43 static void kiblnd_peer_alive(kib_peer_t *peer);
  44 static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
  45 static void kiblnd_check_sends(kib_conn_t *conn);
  46 static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
  47                                 int type, int body_nob);
  48 static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
  49                              int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie);
  50 static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
  51 static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
  52 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
  53
  54 static void
  55 kiblnd_tx_done(lnet_ni_t *ni, kib_tx_t *tx)
  56 {
  57         lnet_msg_t *lntmsg[2];
  58         kib_net_t *net = ni->ni_data;
  59         int rc;
  60         int i;
  61
  62         LASSERT(net);
  63         LASSERT(!in_interrupt());
  64         LASSERT(!tx->tx_queued);               /* mustn't be queued for sending */
  65         LASSERT(!tx->tx_sending);         /* mustn't be awaiting sent callback */
  66         LASSERT(!tx->tx_waiting);             /* mustn't be awaiting peer response */
  67         LASSERT(tx->tx_pool);
  68
  69         kiblnd_unmap_tx(ni, tx);
  70
  71         /* tx may have up to 2 lnet msgs to finalise */
  72         lntmsg[0] = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
  73         lntmsg[1] = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;
  74         rc = tx->tx_status;
  75
  76         if (tx->tx_conn) {
  77                 LASSERT(ni == tx->tx_conn->ibc_peer->ibp_ni);
  78
  79                 kiblnd_conn_decref(tx->tx_conn);
  80                 tx->tx_conn = NULL;
  81         }
  82
  83         tx->tx_nwrq = 0;
  84         tx->tx_status = 0;
  85
  86         kiblnd_pool_free_node(&tx->tx_pool->tpo_pool, &tx->tx_list);
  87
  88         /* delay finalize until my descs have been freed */
  89         for (i = 0; i < 2; i++) {
  90                 if (!lntmsg[i])
  91                         continue;
  92
  93                 lnet_finalize(ni, lntmsg[i], rc);
  94         }
  95 }
  96
  97 void
  98 kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status)
  99 {
 100         kib_tx_t *tx;
 101
 102         while (!list_empty(txlist)) {
 103                 tx = list_entry(txlist->next, kib_tx_t, tx_list);
 104
 105                 list_del(&tx->tx_list);
 106                 /* complete now */
 107                 tx->tx_waiting = 0;
 108                 tx->tx_status = status;
 109                 kiblnd_tx_done(ni, tx);
 110         }
 111 }
 112
 113 static kib_tx_t *
 114 kiblnd_get_idle_tx(lnet_ni_t *ni, lnet_nid_t target)
 115 {
 116         kib_net_t *net = (kib_net_t *)ni->ni_data;
 117         struct list_head *node;
 118         kib_tx_t *tx;
 119         kib_tx_poolset_t *tps;
 120
 121         tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
 122         node = kiblnd_pool_alloc_node(&tps->tps_poolset);
 123         if (!node)
 124                 return NULL;
 125         tx = list_entry(node, kib_tx_t, tx_list);
 126
 127         LASSERT(!tx->tx_nwrq);
 128         LASSERT(!tx->tx_queued);
 129         LASSERT(!tx->tx_sending);
 130         LASSERT(!tx->tx_waiting);
 131         LASSERT(!tx->tx_status);
 132         LASSERT(!tx->tx_conn);
 133         LASSERT(!tx->tx_lntmsg[0]);
 134         LASSERT(!tx->tx_lntmsg[1]);
 135         LASSERT(!tx->tx_nfrags);
 136
 137         return tx;
 138 }
 139
 140 static void
 141 kiblnd_drop_rx(kib_rx_t *rx)
 142 {
 143         kib_conn_t *conn = rx->rx_conn;
 144         struct kib_sched_info *sched = conn->ibc_sched;
 145         unsigned long flags;
 146
 147         spin_lock_irqsave(&sched->ibs_lock, flags);
 148         LASSERT(conn->ibc_nrx > 0);
 149         conn->ibc_nrx--;
 150         spin_unlock_irqrestore(&sched->ibs_lock, flags);
 151
 152         kiblnd_conn_decref(conn);
 153 }
 154
 155 int
 156 kiblnd_post_rx(kib_rx_t *rx, int credit)
 157 {
 158         kib_conn_t *conn = rx->rx_conn;
 159         kib_net_t *net = conn->ibc_peer->ibp_ni->ni_data;
 160         struct ib_recv_wr *bad_wrq = NULL;
 161         struct ib_mr *mr = conn->ibc_hdev->ibh_mrs;
 162         int rc;
 163
 164         LASSERT(net);
 165         LASSERT(!in_interrupt());
 166         LASSERT(credit == IBLND_POSTRX_NO_CREDIT ||
 167                 credit == IBLND_POSTRX_PEER_CREDIT ||
 168                 credit == IBLND_POSTRX_RSRVD_CREDIT);
 169         LASSERT(mr);
 170
 171         rx->rx_sge.lkey   = mr->lkey;
 172         rx->rx_sge.addr   = rx->rx_msgaddr;
 173         rx->rx_sge.length = IBLND_MSG_SIZE;
 174
 175         rx->rx_wrq.next    = NULL;
 176         rx->rx_wrq.sg_list = &rx->rx_sge;
 177         rx->rx_wrq.num_sge = 1;
 178         rx->rx_wrq.wr_id   = kiblnd_ptr2wreqid(rx, IBLND_WID_RX);
 179
 180         LASSERT(conn->ibc_state >= IBLND_CONN_INIT);
 181         LASSERT(rx->rx_nob >= 0);             /* not posted */
 182
 183         if (conn->ibc_state > IBLND_CONN_ESTABLISHED) {
 184                 kiblnd_drop_rx(rx);          /* No more posts for this rx */
 185                 return 0;
 186         }
 187
 188         rx->rx_nob = -1;                        /* flag posted */
 189
 190         /* NB: need an extra reference after ib_post_recv because we don't
 191          * own this rx (and rx::rx_conn) anymore, LU-5678.
 192          */
 193         kiblnd_conn_addref(conn);
 194         rc = ib_post_recv(conn->ibc_cmid->qp, &rx->rx_wrq, &bad_wrq);
 195         if (unlikely(rc)) {
 196                 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
 197                        libcfs_nid2str(conn->ibc_peer->ibp_nid), rc, bad_wrq);
 198                 rx->rx_nob = 0;
 199         }
 200
 201         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) /* Initial post */
 202                 goto out;
 203
 204         if (unlikely(rc)) {
 205                 kiblnd_close_conn(conn, rc);
 206                 kiblnd_drop_rx(rx);          /* No more posts for this rx */
 207                 goto out;
 208         }
 209
 210         if (credit == IBLND_POSTRX_NO_CREDIT)
 211                 goto out;
 212
 213         spin_lock(&conn->ibc_lock);
 214         if (credit == IBLND_POSTRX_PEER_CREDIT)
 215                 conn->ibc_outstanding_credits++;
 216         else
 217                 conn->ibc_reserved_credits++;
 218         spin_unlock(&conn->ibc_lock);
 219
 220         kiblnd_check_sends(conn);
 221 out:
 222         kiblnd_conn_decref(conn);
 223         return rc;
 224 }
 225
 226 static kib_tx_t *
 227 kiblnd_find_waiting_tx_locked(kib_conn_t *conn, int txtype, __u64 cookie)
 228 {
 229         struct list_head *tmp;
 230
 231         list_for_each(tmp, &conn->ibc_active_txs) {
 232                 kib_tx_t *tx = list_entry(tmp, kib_tx_t, tx_list);
 233
 234                 LASSERT(!tx->tx_queued);
 235                 LASSERT(tx->tx_sending || tx->tx_waiting);
 236
 237                 if (tx->tx_cookie != cookie)
 238                         continue;
 239
 240                 if (tx->tx_waiting &&
 241                     tx->tx_msg->ibm_type == txtype)
 242                         return tx;
 243
 244                 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
 245                       tx->tx_waiting ? "" : "NOT ",
 246                       tx->tx_msg->ibm_type, txtype);
 247         }
 248         return NULL;
 249 }
 250
 251 static void
 252 kiblnd_handle_completion(kib_conn_t *conn, int txtype, int status, __u64 cookie)
 253 {
 254         kib_tx_t *tx;
 255         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 256         int idle;
 257
 258         spin_lock(&conn->ibc_lock);
 259
 260         tx = kiblnd_find_waiting_tx_locked(conn, txtype, cookie);
 261         if (!tx) {
 262                 spin_unlock(&conn->ibc_lock);
 263
 264                 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
 265                       txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 266                 kiblnd_close_conn(conn, -EPROTO);
 267                 return;
 268         }
 269
 270         if (!tx->tx_status) {          /* success so far */
 271                 if (status < 0) /* failed? */
 272                         tx->tx_status = status;
 273                 else if (txtype == IBLND_MSG_GET_REQ)
 274                         lnet_set_reply_msg_len(ni, tx->tx_lntmsg[1], status);
 275         }
 276
 277         tx->tx_waiting = 0;
 278
 279         idle = !tx->tx_queued && !tx->tx_sending;
 280         if (idle)
 281                 list_del(&tx->tx_list);
 282
 283         spin_unlock(&conn->ibc_lock);
 284
 285         if (idle)
 286                 kiblnd_tx_done(ni, tx);
 287 }
 288
 289 static void
 290 kiblnd_send_completion(kib_conn_t *conn, int type, int status, __u64 cookie)
 291 {
 292         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 293         kib_tx_t *tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 294
 295         if (!tx) {
 296                 CERROR("Can't get tx for completion %x for %s\n",
 297                        type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 298                 return;
 299         }
 300
 301         tx->tx_msg->ibm_u.completion.ibcm_status = status;
 302         tx->tx_msg->ibm_u.completion.ibcm_cookie = cookie;
 303         kiblnd_init_tx_msg(ni, tx, type, sizeof(kib_completion_msg_t));
 304
 305         kiblnd_queue_tx(tx, conn);
 306 }
 307
 308 static void
 309 kiblnd_handle_rx(kib_rx_t *rx)
 310 {
 311         kib_msg_t *msg = rx->rx_msg;
 312         kib_conn_t *conn = rx->rx_conn;
 313         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 314         int credits = msg->ibm_credits;
 315         kib_tx_t *tx;
 316         int rc = 0;
 317         int rc2;
 318         int post_credit;
 319
 320         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
 321
 322         CDEBUG(D_NET, "Received %x[%d] from %s\n",
 323                msg->ibm_type, credits,
 324                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 325
 326         if (credits) {
 327                 /* Have I received credits that will let me send? */
 328                 spin_lock(&conn->ibc_lock);
 329
 330                 if (conn->ibc_credits + credits >
 331                     conn->ibc_queue_depth) {
 332                         rc2 = conn->ibc_credits;
 333                         spin_unlock(&conn->ibc_lock);
 334
 335                         CERROR("Bad credits from %s: %d + %d > %d\n",
 336                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
 337                                rc2, credits, conn->ibc_queue_depth);
 338
 339                         kiblnd_close_conn(conn, -EPROTO);
 340                         kiblnd_post_rx(rx, IBLND_POSTRX_NO_CREDIT);
 341                         return;
 342                 }
 343
 344                 conn->ibc_credits += credits;
 345
 346                 /* This ensures the credit taken by NOOP can be returned */
 347                 if (msg->ibm_type == IBLND_MSG_NOOP &&
 348                     !IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
 349                         conn->ibc_outstanding_credits++;
 350
 351                 spin_unlock(&conn->ibc_lock);
 352                 kiblnd_check_sends(conn);
 353         }
 354
 355         switch (msg->ibm_type) {
 356         default:
 357                 CERROR("Bad IBLND message type %x from %s\n",
 358                        msg->ibm_type, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 359                 post_credit = IBLND_POSTRX_NO_CREDIT;
 360                 rc = -EPROTO;
 361                 break;
 362
 363         case IBLND_MSG_NOOP:
 364                 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
 365                         post_credit = IBLND_POSTRX_NO_CREDIT;
 366                         break;
 367                 }
 368
 369                 if (credits) /* credit already posted */
 370                         post_credit = IBLND_POSTRX_NO_CREDIT;
 371                 else          /* a keepalive NOOP */
 372                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 373                 break;
 374
 375         case IBLND_MSG_IMMEDIATE:
 376                 post_credit = IBLND_POSTRX_DONT_POST;
 377                 rc = lnet_parse(ni, &msg->ibm_u.immediate.ibim_hdr,
 378                                 msg->ibm_srcnid, rx, 0);
 379                 if (rc < 0)                  /* repost on error */
 380                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 381                 break;
 382
 383         case IBLND_MSG_PUT_REQ:
 384                 post_credit = IBLND_POSTRX_DONT_POST;
 385                 rc = lnet_parse(ni, &msg->ibm_u.putreq.ibprm_hdr,
 386                                 msg->ibm_srcnid, rx, 1);
 387                 if (rc < 0)                  /* repost on error */
 388                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 389                 break;
 390
 391         case IBLND_MSG_PUT_NAK:
 392                 CWARN("PUT_NACK from %s\n",
 393                       libcfs_nid2str(conn->ibc_peer->ibp_nid));
 394                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 395                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_REQ,
 396                                          msg->ibm_u.completion.ibcm_status,
 397                                          msg->ibm_u.completion.ibcm_cookie);
 398                 break;
 399
 400         case IBLND_MSG_PUT_ACK:
 401                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 402
 403                 spin_lock(&conn->ibc_lock);
 404                 tx = kiblnd_find_waiting_tx_locked(conn, IBLND_MSG_PUT_REQ,
 405                                                    msg->ibm_u.putack.ibpam_src_cookie);
 406                 if (tx)
 407                         list_del(&tx->tx_list);
 408                 spin_unlock(&conn->ibc_lock);
 409
 410                 if (!tx) {
 411                         CERROR("Unmatched PUT_ACK from %s\n",
 412                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
 413                         rc = -EPROTO;
 414                         break;
 415                 }
 416
 417                 LASSERT(tx->tx_waiting);
 418                 /*
 419                  * CAVEAT EMPTOR: I could be racing with tx_complete, but...
 420                  * (a) I can overwrite tx_msg since my peer has received it!
 421                  * (b) tx_waiting set tells tx_complete() it's not done.
 422                  */
 423                 tx->tx_nwrq = 0;                /* overwrite PUT_REQ */
 424
 425                 rc2 = kiblnd_init_rdma(conn, tx, IBLND_MSG_PUT_DONE,
 426                                        kiblnd_rd_size(&msg->ibm_u.putack.ibpam_rd),
 427                                        &msg->ibm_u.putack.ibpam_rd,
 428                                        msg->ibm_u.putack.ibpam_dst_cookie);
 429                 if (rc2 < 0)
 430                         CERROR("Can't setup rdma for PUT to %s: %d\n",
 431                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc2);
 432
 433                 spin_lock(&conn->ibc_lock);
 434                 tx->tx_waiting = 0;     /* clear waiting and queue atomically */
 435                 kiblnd_queue_tx_locked(tx, conn);
 436                 spin_unlock(&conn->ibc_lock);
 437                 break;
 438
 439         case IBLND_MSG_PUT_DONE:
 440                 post_credit = IBLND_POSTRX_PEER_CREDIT;
 441                 kiblnd_handle_completion(conn, IBLND_MSG_PUT_ACK,
 442                                          msg->ibm_u.completion.ibcm_status,
 443                                          msg->ibm_u.completion.ibcm_cookie);
 444                 break;
 445
 446         case IBLND_MSG_GET_REQ:
 447                 post_credit = IBLND_POSTRX_DONT_POST;
 448                 rc = lnet_parse(ni, &msg->ibm_u.get.ibgm_hdr,
 449                                 msg->ibm_srcnid, rx, 1);
 450                 if (rc < 0)                  /* repost on error */
 451                         post_credit = IBLND_POSTRX_PEER_CREDIT;
 452                 break;
 453
 454         case IBLND_MSG_GET_DONE:
 455                 post_credit = IBLND_POSTRX_RSRVD_CREDIT;
 456                 kiblnd_handle_completion(conn, IBLND_MSG_GET_REQ,
 457                                          msg->ibm_u.completion.ibcm_status,
 458                                          msg->ibm_u.completion.ibcm_cookie);
 459                 break;
 460         }
 461
 462         if (rc < 0)                          /* protocol error */
 463                 kiblnd_close_conn(conn, rc);
 464
 465         if (post_credit != IBLND_POSTRX_DONT_POST)
 466                 kiblnd_post_rx(rx, post_credit);
 467 }
 468
 469 static void
 470 kiblnd_rx_complete(kib_rx_t *rx, int status, int nob)
 471 {
 472         kib_msg_t *msg = rx->rx_msg;
 473         kib_conn_t *conn = rx->rx_conn;
 474         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 475         kib_net_t *net = ni->ni_data;
 476         int rc;
 477         int err = -EIO;
 478
 479         LASSERT(net);
 480         LASSERT(rx->rx_nob < 0);               /* was posted */
 481         rx->rx_nob = 0;                  /* isn't now */
 482
 483         if (conn->ibc_state > IBLND_CONN_ESTABLISHED)
 484                 goto ignore;
 485
 486         if (status != IB_WC_SUCCESS) {
 487                 CNETERR("Rx from %s failed: %d\n",
 488                         libcfs_nid2str(conn->ibc_peer->ibp_nid), status);
 489                 goto failed;
 490         }
 491
 492         LASSERT(nob >= 0);
 493         rx->rx_nob = nob;
 494
 495         rc = kiblnd_unpack_msg(msg, rx->rx_nob);
 496         if (rc) {
 497                 CERROR("Error %d unpacking rx from %s\n",
 498                        rc, libcfs_nid2str(conn->ibc_peer->ibp_nid));
 499                 goto failed;
 500         }
 501
 502         if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
 503             msg->ibm_dstnid != ni->ni_nid ||
 504             msg->ibm_srcstamp != conn->ibc_incarnation ||
 505             msg->ibm_dststamp != net->ibn_incarnation) {
 506                 CERROR("Stale rx from %s\n",
 507                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 508                 err = -ESTALE;
 509                 goto failed;
 510         }
 511
 512         /* set time last known alive */
 513         kiblnd_peer_alive(conn->ibc_peer);
 514
 515         /* racing with connection establishment/teardown! */
 516
 517         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 518                 rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
 519                 unsigned long flags;
 520
 521                 write_lock_irqsave(g_lock, flags);
 522                 /* must check holding global lock to eliminate race */
 523                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 524                         list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
 525                         write_unlock_irqrestore(g_lock, flags);
 526                         return;
 527                 }
 528                 write_unlock_irqrestore(g_lock, flags);
 529         }
 530         kiblnd_handle_rx(rx);
 531         return;
 532
 533  failed:
 534         CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
 535         kiblnd_close_conn(conn, err);
 536  ignore:
 537         kiblnd_drop_rx(rx);                  /* Don't re-post rx. */
 538 }
 539
 540 static struct page *
 541 kiblnd_kvaddr_to_page(unsigned long vaddr)
 542 {
 543         struct page *page;
 544
 545         if (is_vmalloc_addr((void *)vaddr)) {
 546                 page = vmalloc_to_page((void *)vaddr);
 547                 LASSERT(page);
 548                 return page;
 549         }
 550 #ifdef CONFIG_HIGHMEM
 551         if (vaddr >= PKMAP_BASE &&
 552             vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE)) {
 553                 /* No highmem pages only used for bulk (kiov) I/O */
 554                 CERROR("find page for address in highmem\n");
 555                 LBUG();
 556         }
 557 #endif
 558         page = virt_to_page(vaddr);
 559         LASSERT(page);
 560         return page;
 561 }
 562
 563 static int
 564 kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
 565 {
 566         kib_hca_dev_t *hdev;
 567         __u64 *pages = tx->tx_pages;
 568         kib_fmr_poolset_t *fps;
 569         int npages;
 570         int size;
 571         int cpt;
 572         int rc;
 573         int i;
 574
 575         LASSERT(tx->tx_pool);
 576         LASSERT(tx->tx_pool->tpo_pool.po_owner);
 577
 578         hdev = tx->tx_pool->tpo_hdev;
 579
 580         for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
 581                 for (size = 0; size <  rd->rd_frags[i].rf_nob;
 582                                size += hdev->ibh_page_size) {
 583                         pages[npages++] = (rd->rd_frags[i].rf_addr &
 584                                             hdev->ibh_page_mask) + size;
 585                 }
 586         }
 587
 588         cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 589
 590         fps = net->ibn_fmr_ps[cpt];
 591         rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
 592         if (rc) {
 593                 CERROR("Can't map %d pages: %d\n", npages, rc);
 594                 return rc;
 595         }
 596
 597         /*
 598          * If rd is not tx_rd, it's going to get sent to a peer, who will need
 599          * the rkey
 600          */
 601         rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
 602                                          tx->fmr.fmr_pfmr->fmr->lkey;
 603         rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
 604         rd->rd_frags[0].rf_nob = nob;
 605         rd->rd_nfrags = 1;
 606
 607         return 0;
 608 }
 609
 610 static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx)
 611 {
 612         kib_net_t *net = ni->ni_data;
 613
 614         LASSERT(net);
 615
 616         if (net->ibn_fmr_ps && tx->fmr.fmr_pfmr) {
 617                 kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
 618                 tx->fmr.fmr_pfmr = NULL;
 619         }
 620
 621         if (tx->tx_nfrags) {
 622                 kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
 623                                     tx->tx_frags, tx->tx_nfrags, tx->tx_dmadir);
 624                 tx->tx_nfrags = 0;
 625         }
 626 }
 627
 628 static int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 629                          int nfrags)
 630 {
 631         kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
 632         kib_net_t *net = ni->ni_data;
 633         struct ib_mr *mr    = NULL;
 634         __u32 nob;
 635         int i;
 636
 637         /*
 638          * If rd is not tx_rd, it's going to get sent to a peer and I'm the
 639          * RDMA sink
 640          */
 641         tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
 642         tx->tx_nfrags = nfrags;
 643
 644         rd->rd_nfrags = kiblnd_dma_map_sg(hdev->ibh_ibdev, tx->tx_frags,
 645                                           tx->tx_nfrags, tx->tx_dmadir);
 646
 647         for (i = 0, nob = 0; i < rd->rd_nfrags; i++) {
 648                 rd->rd_frags[i].rf_nob  = kiblnd_sg_dma_len(
 649                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 650                 rd->rd_frags[i].rf_addr = kiblnd_sg_dma_address(
 651                         hdev->ibh_ibdev, &tx->tx_frags[i]);
 652                 nob += rd->rd_frags[i].rf_nob;
 653         }
 654
 655         mr = kiblnd_find_rd_dma_mr(hdev, rd, tx->tx_conn ?
 656                                    tx->tx_conn->ibc_max_frags : -1);
 657         if (mr) {
 658                 /* found pre-mapping MR */
 659                 rd->rd_key = (rd != tx->tx_rd) ? mr->rkey : mr->lkey;
 660                 return 0;
 661         }
 662
 663         if (net->ibn_fmr_ps)
 664                 return kiblnd_fmr_map_tx(net, tx, rd, nob);
 665
 666         return -EINVAL;
 667 }
 668
 669 static int
 670 kiblnd_setup_rd_iov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 671                     unsigned int niov, struct kvec *iov, int offset, int nob)
 672 {
 673         kib_net_t *net = ni->ni_data;
 674         struct page *page;
 675         struct scatterlist *sg;
 676         unsigned long vaddr;
 677         int fragnob;
 678         int page_offset;
 679
 680         LASSERT(nob > 0);
 681         LASSERT(niov > 0);
 682         LASSERT(net);
 683
 684         while (offset >= iov->iov_len) {
 685                 offset -= iov->iov_len;
 686                 niov--;
 687                 iov++;
 688                 LASSERT(niov > 0);
 689         }
 690
 691         sg = tx->tx_frags;
 692         do {
 693                 LASSERT(niov > 0);
 694
 695                 vaddr = ((unsigned long)iov->iov_base) + offset;
 696                 page_offset = vaddr & (PAGE_SIZE - 1);
 697                 page = kiblnd_kvaddr_to_page(vaddr);
 698                 if (!page) {
 699                         CERROR("Can't find page\n");
 700                         return -EFAULT;
 701                 }
 702
 703                 fragnob = min((int)(iov->iov_len - offset), nob);
 704                 fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
 705
 706                 sg_set_page(sg, page, fragnob, page_offset);
 707                 sg++;
 708
 709                 if (offset + fragnob < iov->iov_len) {
 710                         offset += fragnob;
 711                 } else {
 712                         offset = 0;
 713                         iov++;
 714                         niov--;
 715                 }
 716                 nob -= fragnob;
 717         } while (nob > 0);
 718
 719         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 720 }
 721
 722 static int
 723 kiblnd_setup_rd_kiov(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd,
 724                      int nkiov, lnet_kiov_t *kiov, int offset, int nob)
 725 {
 726         kib_net_t *net = ni->ni_data;
 727         struct scatterlist *sg;
 728         int fragnob;
 729
 730         CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
 731
 732         LASSERT(nob > 0);
 733         LASSERT(nkiov > 0);
 734         LASSERT(net);
 735
 736         while (offset >= kiov->kiov_len) {
 737                 offset -= kiov->kiov_len;
 738                 nkiov--;
 739                 kiov++;
 740                 LASSERT(nkiov > 0);
 741         }
 742
 743         sg = tx->tx_frags;
 744         do {
 745                 LASSERT(nkiov > 0);
 746
 747                 fragnob = min((int)(kiov->kiov_len - offset), nob);
 748
 749                 sg_set_page(sg, kiov->kiov_page, fragnob,
 750                             kiov->kiov_offset + offset);
 751                 sg++;
 752
 753                 offset = 0;
 754                 kiov++;
 755                 nkiov--;
 756                 nob -= fragnob;
 757         } while (nob > 0);
 758
 759         return kiblnd_map_tx(ni, tx, rd, sg - tx->tx_frags);
 760 }
 761
 762 static int
 763 kiblnd_post_tx_locked(kib_conn_t *conn, kib_tx_t *tx, int credit)
 764         __must_hold(&conn->ibc_lock)
 765 {
 766         kib_msg_t *msg = tx->tx_msg;
 767         kib_peer_t *peer = conn->ibc_peer;
 768         int ver = conn->ibc_version;
 769         int rc;
 770         int done;
 771
 772         LASSERT(tx->tx_queued);
 773         /* We rely on this for QP sizing */
 774         LASSERT(tx->tx_nwrq > 0);
 775         LASSERT(tx->tx_nwrq <= 1 + conn->ibc_max_frags);
 776
 777         LASSERT(!credit || credit == 1);
 778         LASSERT(conn->ibc_outstanding_credits >= 0);
 779         LASSERT(conn->ibc_outstanding_credits <= conn->ibc_queue_depth);
 780         LASSERT(conn->ibc_credits >= 0);
 781         LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
 782
 783         if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
 784                 /* tx completions outstanding... */
 785                 CDEBUG(D_NET, "%s: posted enough\n",
 786                        libcfs_nid2str(peer->ibp_nid));
 787                 return -EAGAIN;
 788         }
 789
 790         if (credit && !conn->ibc_credits) {   /* no credits */
 791                 CDEBUG(D_NET, "%s: no credits\n",
 792                        libcfs_nid2str(peer->ibp_nid));
 793                 return -EAGAIN;
 794         }
 795
 796         if (credit && !IBLND_OOB_CAPABLE(ver) &&
 797             conn->ibc_credits == 1 &&   /* last credit reserved */
 798             msg->ibm_type != IBLND_MSG_NOOP) {      /* for NOOP */
 799                 CDEBUG(D_NET, "%s: not using last credit\n",
 800                        libcfs_nid2str(peer->ibp_nid));
 801                 return -EAGAIN;
 802         }
 803
 804         /* NB don't drop ibc_lock before bumping tx_sending */
 805         list_del(&tx->tx_list);
 806         tx->tx_queued = 0;
 807
 808         if (msg->ibm_type == IBLND_MSG_NOOP &&
 809             (!kiblnd_need_noop(conn) ||     /* redundant NOOP */
 810              (IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
 811               conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
 812                 /*
 813                  * OK to drop when posted enough NOOPs, since
 814                  * kiblnd_check_sends will queue NOOP again when
 815                  * posted NOOPs complete
 816                  */
 817                 spin_unlock(&conn->ibc_lock);
 818                 kiblnd_tx_done(peer->ibp_ni, tx);
 819                 spin_lock(&conn->ibc_lock);
 820                 CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
 821                        libcfs_nid2str(peer->ibp_nid),
 822                        conn->ibc_noops_posted);
 823                 return 0;
 824         }
 825
 826         kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
 827                         peer->ibp_nid, conn->ibc_incarnation);
 828
 829         conn->ibc_credits -= credit;
 830         conn->ibc_outstanding_credits = 0;
 831         conn->ibc_nsends_posted++;
 832         if (msg->ibm_type == IBLND_MSG_NOOP)
 833                 conn->ibc_noops_posted++;
 834
 835         /*
 836          * CAVEAT EMPTOR!  This tx could be the PUT_DONE of an RDMA
 837          * PUT.  If so, it was first queued here as a PUT_REQ, sent and
 838          * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
 839          * and then re-queued here.  It's (just) possible that
 840          * tx_sending is non-zero if we've not done the tx_complete()
 841          * from the first send; hence the ++ rather than = below.
 842          */
 843         tx->tx_sending++;
 844         list_add(&tx->tx_list, &conn->ibc_active_txs);
 845
 846         /* I'm still holding ibc_lock! */
 847         if (conn->ibc_state != IBLND_CONN_ESTABLISHED) {
 848                 rc = -ECONNABORTED;
 849         } else if (tx->tx_pool->tpo_pool.po_failed ||
 850                  conn->ibc_hdev != tx->tx_pool->tpo_hdev) {
 851                 /* close_conn will launch failover */
 852                 rc = -ENETDOWN;
 853         } else {
 854                 struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
 855
 856                 LASSERTF(wrq->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
 857                          "bad wr_id %llx, opc %d, flags %d, peer: %s\n",
 858                          wrq->wr_id, wrq->opcode, wrq->send_flags,
 859                 libcfs_nid2str(conn->ibc_peer->ibp_nid));
 860                 wrq = NULL;
 861                 rc = ib_post_send(conn->ibc_cmid->qp, &tx->tx_wrq->wr, &wrq);
 862         }
 863
 864         conn->ibc_last_send = jiffies;
 865
 866         if (!rc)
 867                 return 0;
 868
 869         /*
 870          * NB credits are transferred in the actual
 871          * message, which can only be the last work item
 872          */
 873         conn->ibc_credits += credit;
 874         conn->ibc_outstanding_credits += msg->ibm_credits;
 875         conn->ibc_nsends_posted--;
 876         if (msg->ibm_type == IBLND_MSG_NOOP)
 877                 conn->ibc_noops_posted--;
 878
 879         tx->tx_status = rc;
 880         tx->tx_waiting = 0;
 881         tx->tx_sending--;
 882
 883         done = !tx->tx_sending;
 884         if (done)
 885                 list_del(&tx->tx_list);
 886
 887         spin_unlock(&conn->ibc_lock);
 888
 889         if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 890                 CERROR("Error %d posting transmit to %s\n",
 891                        rc, libcfs_nid2str(peer->ibp_nid));
 892         else
 893                 CDEBUG(D_NET, "Error %d posting transmit to %s\n",
 894                        rc, libcfs_nid2str(peer->ibp_nid));
 895
 896         kiblnd_close_conn(conn, rc);
 897
 898         if (done)
 899                 kiblnd_tx_done(peer->ibp_ni, tx);
 900
 901         spin_lock(&conn->ibc_lock);
 902
 903         return -EIO;
 904 }
 905
 906 static void
 907 kiblnd_check_sends(kib_conn_t *conn)
 908 {
 909         int ver = conn->ibc_version;
 910         lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
 911         kib_tx_t *tx;
 912
 913         /* Don't send anything until after the connection is established */
 914         if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
 915                 CDEBUG(D_NET, "%s too soon\n",
 916                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
 917                 return;
 918         }
 919
 920         spin_lock(&conn->ibc_lock);
 921
 922         LASSERT(conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
 923         LASSERT(!IBLND_OOB_CAPABLE(ver) ||
 924                 conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
 925         LASSERT(conn->ibc_reserved_credits >= 0);
 926
 927         while (conn->ibc_reserved_credits > 0 &&
 928                !list_empty(&conn->ibc_tx_queue_rsrvd)) {
 929                 tx = list_entry(conn->ibc_tx_queue_rsrvd.next,
 930                                 kib_tx_t, tx_list);
 931                 list_del(&tx->tx_list);
 932                 list_add_tail(&tx->tx_list, &conn->ibc_tx_queue);
 933                 conn->ibc_reserved_credits--;
 934         }
 935
 936         if (kiblnd_need_noop(conn)) {
 937                 spin_unlock(&conn->ibc_lock);
 938
 939                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
 940                 if (tx)
 941                         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_NOOP, 0);
 942
 943                 spin_lock(&conn->ibc_lock);
 944                 if (tx)
 945                         kiblnd_queue_tx_locked(tx, conn);
 946         }
 947
 948         for (;;) {
 949                 int credit;
 950
 951                 if (!list_empty(&conn->ibc_tx_queue_nocred)) {
 952                         credit = 0;
 953                         tx = list_entry(conn->ibc_tx_queue_nocred.next,
 954                                         kib_tx_t, tx_list);
 955                 } else if (!list_empty(&conn->ibc_tx_noops)) {
 956                         LASSERT(!IBLND_OOB_CAPABLE(ver));
 957                         credit = 1;
 958                         tx = list_entry(conn->ibc_tx_noops.next,
 959                                         kib_tx_t, tx_list);
 960                 } else if (!list_empty(&conn->ibc_tx_queue)) {
 961                         credit = 1;
 962                         tx = list_entry(conn->ibc_tx_queue.next,
 963                                         kib_tx_t, tx_list);
 964                 } else {
 965                         break;
 966                 }
 967
 968                 if (kiblnd_post_tx_locked(conn, tx, credit))
 969                         break;
 970         }
 971
 972         spin_unlock(&conn->ibc_lock);
 973 }
 974
 975 static void
 976 kiblnd_tx_complete(kib_tx_t *tx, int status)
 977 {
 978         int failed = (status != IB_WC_SUCCESS);
 979         kib_conn_t *conn = tx->tx_conn;
 980         int idle;
 981
 982         LASSERT(tx->tx_sending > 0);
 983
 984         if (failed) {
 985                 if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
 986                         CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
 987                                 libcfs_nid2str(conn->ibc_peer->ibp_nid),
 988                                 tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
 989                                 status);
 990
 991                 kiblnd_close_conn(conn, -EIO);
 992         } else {
 993                 kiblnd_peer_alive(conn->ibc_peer);
 994         }
 995
 996         spin_lock(&conn->ibc_lock);
 997
 998         /*
 999          * I could be racing with rdma completion.  Whoever makes 'tx' idle
1000          * gets to free it, which also drops its ref on 'conn'.
1001          */
1002         tx->tx_sending--;
1003         conn->ibc_nsends_posted--;
1004         if (tx->tx_msg->ibm_type == IBLND_MSG_NOOP)
1005                 conn->ibc_noops_posted--;
1006
1007         if (failed) {
1008                 tx->tx_waiting = 0;          /* don't wait for peer */
1009                 tx->tx_status = -EIO;
1010         }
1011
1012         idle = !tx->tx_sending &&        /* This is the final callback */
1013                !tx->tx_waiting &&              /* Not waiting for peer */
1014                !tx->tx_queued;            /* Not re-queued (PUT_DONE) */
1015         if (idle)
1016                 list_del(&tx->tx_list);
1017
1018         kiblnd_conn_addref(conn);              /* 1 ref for me.... */
1019
1020         spin_unlock(&conn->ibc_lock);
1021
1022         if (idle)
1023                 kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
1024
1025         kiblnd_check_sends(conn);
1026
1027         kiblnd_conn_decref(conn);              /* ...until here */
1028 }
1029
1030 static void
1031 kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
1032 {
1033         kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
1034         struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
1035         struct ib_rdma_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
1036         int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1037         struct ib_mr *mr = hdev->ibh_mrs;
1038
1039         LASSERT(tx->tx_nwrq >= 0);
1040         LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
1041         LASSERT(nob <= IBLND_MSG_SIZE);
1042         LASSERT(mr);
1043
1044         kiblnd_init_msg(tx->tx_msg, type, body_nob);
1045
1046         sge->lkey   = mr->lkey;
1047         sge->addr   = tx->tx_msgaddr;
1048         sge->length = nob;
1049
1050         memset(wrq, 0, sizeof(*wrq));
1051
1052         wrq->wr.next       = NULL;
1053         wrq->wr.wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
1054         wrq->wr.sg_list    = sge;
1055         wrq->wr.num_sge    = 1;
1056         wrq->wr.opcode     = IB_WR_SEND;
1057         wrq->wr.send_flags = IB_SEND_SIGNALED;
1058
1059         tx->tx_nwrq++;
1060 }
1061
1062 static int
1063 kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
1064                  int resid, kib_rdma_desc_t *dstrd, __u64 dstcookie)
1065 {
1066         kib_msg_t *ibmsg = tx->tx_msg;
1067         kib_rdma_desc_t *srcrd = tx->tx_rd;
1068         struct ib_sge *sge = &tx->tx_sge[0];
1069         struct ib_rdma_wr *wrq = &tx->tx_wrq[0], *next;
1070         int rc  = resid;
1071         int srcidx = 0;
1072         int dstidx = 0;
1073         int wrknob;
1074
1075         LASSERT(!in_interrupt());
1076         LASSERT(!tx->tx_nwrq);
1077         LASSERT(type == IBLND_MSG_GET_DONE ||
1078                 type == IBLND_MSG_PUT_DONE);
1079
1080         while (resid > 0) {
1081                 if (srcidx >= srcrd->rd_nfrags) {
1082                         CERROR("Src buffer exhausted: %d frags\n", srcidx);
1083                         rc = -EPROTO;
1084                         break;
1085                 }
1086
1087                 if (dstidx == dstrd->rd_nfrags) {
1088                         CERROR("Dst buffer exhausted: %d frags\n", dstidx);
1089                         rc = -EPROTO;
1090                         break;
1091                 }
1092
1093                 if (tx->tx_nwrq >= conn->ibc_max_frags) {
1094                         CERROR("RDMA has too many fragments for peer %s (%d), src idx/frags: %d/%d dst idx/frags: %d/%d\n",
1095                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
1096                                conn->ibc_max_frags,
1097                                srcidx, srcrd->rd_nfrags,
1098                                dstidx, dstrd->rd_nfrags);
1099                         rc = -EMSGSIZE;
1100                         break;
1101                 }
1102
1103                 wrknob = min(min(kiblnd_rd_frag_size(srcrd, srcidx),
1104                                  kiblnd_rd_frag_size(dstrd, dstidx)),
1105                              (__u32) resid);
1106
1107                 sge = &tx->tx_sge[tx->tx_nwrq];
1108                 sge->addr   = kiblnd_rd_frag_addr(srcrd, srcidx);
1109                 sge->lkey   = kiblnd_rd_frag_key(srcrd, srcidx);
1110                 sge->length = wrknob;
1111
1112                 wrq = &tx->tx_wrq[tx->tx_nwrq];
1113                 next = wrq + 1;
1114
1115                 wrq->wr.next       = &next->wr;
1116                 wrq->wr.wr_id      = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
1117                 wrq->wr.sg_list    = sge;
1118                 wrq->wr.num_sge    = 1;
1119                 wrq->wr.opcode     = IB_WR_RDMA_WRITE;
1120                 wrq->wr.send_flags = 0;
1121
1122                 wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
1123                 wrq->rkey        = kiblnd_rd_frag_key(dstrd, dstidx);
1124
1125                 srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
1126                 dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
1127
1128                 resid -= wrknob;
1129
1130                 tx->tx_nwrq++;
1131                 wrq++;
1132                 sge++;
1133         }
1134
1135         if (rc < 0)                          /* no RDMA if completing with failure */
1136                 tx->tx_nwrq = 0;
1137
1138         ibmsg->ibm_u.completion.ibcm_status = rc;
1139         ibmsg->ibm_u.completion.ibcm_cookie = dstcookie;
1140         kiblnd_init_tx_msg(conn->ibc_peer->ibp_ni, tx,
1141                            type, sizeof(kib_completion_msg_t));
1142
1143         return rc;
1144 }
1145
1146 static void
1147 kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn)
1148 {
1149         struct list_head *q;
1150
1151         LASSERT(tx->tx_nwrq > 0);             /* work items set up */
1152         LASSERT(!tx->tx_queued);               /* not queued for sending already */
1153         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1154
1155         tx->tx_queued = 1;
1156         tx->tx_deadline = jiffies +
1157                           msecs_to_jiffies(*kiblnd_tunables.kib_timeout *
1158                                            MSEC_PER_SEC);
1159
1160         if (!tx->tx_conn) {
1161                 kiblnd_conn_addref(conn);
1162                 tx->tx_conn = conn;
1163                 LASSERT(tx->tx_msg->ibm_type != IBLND_MSG_PUT_DONE);
1164         } else {
1165                 /* PUT_DONE first attached to conn as a PUT_REQ */
1166                 LASSERT(tx->tx_conn == conn);
1167                 LASSERT(tx->tx_msg->ibm_type == IBLND_MSG_PUT_DONE);
1168         }
1169
1170         switch (tx->tx_msg->ibm_type) {
1171         default:
1172                 LBUG();
1173
1174         case IBLND_MSG_PUT_REQ:
1175         case IBLND_MSG_GET_REQ:
1176                 q = &conn->ibc_tx_queue_rsrvd;
1177                 break;
1178
1179         case IBLND_MSG_PUT_NAK:
1180         case IBLND_MSG_PUT_ACK:
1181         case IBLND_MSG_PUT_DONE:
1182         case IBLND_MSG_GET_DONE:
1183                 q = &conn->ibc_tx_queue_nocred;
1184                 break;
1185
1186         case IBLND_MSG_NOOP:
1187                 if (IBLND_OOB_CAPABLE(conn->ibc_version))
1188                         q = &conn->ibc_tx_queue_nocred;
1189                 else
1190                         q = &conn->ibc_tx_noops;
1191                 break;
1192
1193         case IBLND_MSG_IMMEDIATE:
1194                 q = &conn->ibc_tx_queue;
1195                 break;
1196         }
1197
1198         list_add_tail(&tx->tx_list, q);
1199 }
1200
1201 static void
1202 kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn)
1203 {
1204         spin_lock(&conn->ibc_lock);
1205         kiblnd_queue_tx_locked(tx, conn);
1206         spin_unlock(&conn->ibc_lock);
1207
1208         kiblnd_check_sends(conn);
1209 }
1210
1211 static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
1212                                struct sockaddr_in *srcaddr,
1213                                struct sockaddr_in *dstaddr,
1214                                int timeout_ms)
1215 {
1216         unsigned short port;
1217         int rc;
1218
1219         /* allow the port to be reused */
1220         rc = rdma_set_reuseaddr(cmid, 1);
1221         if (rc) {
1222                 CERROR("Unable to set reuse on cmid: %d\n", rc);
1223                 return rc;
1224         }
1225
1226         /* look for a free privileged port */
1227         for (port = PROT_SOCK - 1; port > 0; port--) {
1228                 srcaddr->sin_port = htons(port);
1229                 rc = rdma_resolve_addr(cmid,
1230                                        (struct sockaddr *)srcaddr,
1231                                        (struct sockaddr *)dstaddr,
1232                                        timeout_ms);
1233                 if (!rc) {
1234                         CDEBUG(D_NET, "bound to port %hu\n", port);
1235                         return 0;
1236                 } else if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL) {
1237                         CDEBUG(D_NET, "bind to port %hu failed: %d\n",
1238                                port, rc);
1239                 } else {
1240                         return rc;
1241                 }
1242         }
1243
1244         CERROR("Failed to bind to a free privileged port\n");
1245         return rc;
1246 }
1247
1248 static void
1249 kiblnd_connect_peer(kib_peer_t *peer)
1250 {
1251         struct rdma_cm_id *cmid;
1252         kib_dev_t *dev;
1253         kib_net_t *net = peer->ibp_ni->ni_data;
1254         struct sockaddr_in srcaddr;
1255         struct sockaddr_in dstaddr;
1256         int rc;
1257
1258         LASSERT(net);
1259         LASSERT(peer->ibp_connecting > 0);
1260         LASSERT(!peer->ibp_reconnecting);
1261
1262         cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
1263                                      IB_QPT_RC);
1264
1265         if (IS_ERR(cmid)) {
1266                 CERROR("Can't create CMID for %s: %ld\n",
1267                        libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
1268                 rc = PTR_ERR(cmid);
1269                 goto failed;
1270         }
1271
1272         dev = net->ibn_dev;
1273         memset(&srcaddr, 0, sizeof(srcaddr));
1274         srcaddr.sin_family = AF_INET;
1275         srcaddr.sin_addr.s_addr = htonl(dev->ibd_ifip);
1276
1277         memset(&dstaddr, 0, sizeof(dstaddr));
1278         dstaddr.sin_family = AF_INET;
1279         dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
1280         dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
1281
1282         kiblnd_peer_addref(peer);              /* cmid's ref */
1283
1284         if (*kiblnd_tunables.kib_use_priv_port) {
1285                 rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
1286                                          *kiblnd_tunables.kib_timeout * 1000);
1287         } else {
1288                 rc = rdma_resolve_addr(cmid,
1289                                        (struct sockaddr *)&srcaddr,
1290                                        (struct sockaddr *)&dstaddr,
1291                                        *kiblnd_tunables.kib_timeout * 1000);
1292         }
1293         if (rc) {
1294                 /* Can't initiate address resolution:  */
1295                 CERROR("Can't resolve addr for %s: %d\n",
1296                        libcfs_nid2str(peer->ibp_nid), rc);
1297                 goto failed2;
1298         }
1299
1300         LASSERT(cmid->device);
1301         CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
1302                libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
1303                &dev->ibd_ifip, cmid->device->name);
1304
1305         return;
1306
1307  failed2:
1308         kiblnd_peer_connect_failed(peer, 1, rc);
1309         kiblnd_peer_decref(peer);              /* cmid's ref */
1310         rdma_destroy_id(cmid);
1311         return;
1312  failed:
1313         kiblnd_peer_connect_failed(peer, 1, rc);
1314 }
1315
1316 bool
1317 kiblnd_reconnect_peer(kib_peer_t *peer)
1318 {
1319         rwlock_t *glock = &kiblnd_data.kib_global_lock;
1320         char *reason = NULL;
1321         struct list_head txs;
1322         unsigned long flags;
1323
1324         INIT_LIST_HEAD(&txs);
1325
1326         write_lock_irqsave(glock, flags);
1327         if (!peer->ibp_reconnecting) {
1328                 if (peer->ibp_accepting)
1329                         reason = "accepting";
1330                 else if (peer->ibp_connecting)
1331                         reason = "connecting";
1332                 else if (!list_empty(&peer->ibp_conns))
1333                         reason = "connected";
1334                 else /* connected then closed */
1335                         reason = "closed";
1336
1337                 goto no_reconnect;
1338         }
1339
1340         LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
1341                 list_empty(&peer->ibp_conns));
1342         peer->ibp_reconnecting = 0;
1343
1344         if (!kiblnd_peer_active(peer)) {
1345                 list_splice_init(&peer->ibp_tx_queue, &txs);
1346                 reason = "unlinked";
1347                 goto no_reconnect;
1348         }
1349
1350         peer->ibp_connecting++;
1351         peer->ibp_reconnected++;
1352         write_unlock_irqrestore(glock, flags);
1353
1354         kiblnd_connect_peer(peer);
1355         return true;
1356
1357 no_reconnect:
1358         write_unlock_irqrestore(glock, flags);
1359
1360         CWARN("Abort reconnection of %s: %s\n",
1361               libcfs_nid2str(peer->ibp_nid), reason);
1362         kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
1363         return false;
1364 }
1365
1366 void
1367 kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
1368 {
1369         kib_peer_t *peer;
1370         kib_peer_t *peer2;
1371         kib_conn_t *conn;
1372         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
1373         unsigned long flags;
1374         int rc;
1375
1376         /*
1377          * If I get here, I've committed to send, so I complete the tx with
1378          * failure on any problems
1379          */
1380         LASSERT(!tx || !tx->tx_conn); /* only set when assigned a conn */
1381         LASSERT(!tx || tx->tx_nwrq > 0);     /* work items have been set up */
1382
1383         /*
1384          * First time, just use a read lock since I expect to find my peer
1385          * connected
1386          */
1387         read_lock_irqsave(g_lock, flags);
1388
1389         peer = kiblnd_find_peer_locked(nid);
1390         if (peer && !list_empty(&peer->ibp_conns)) {
1391                 /* Found a peer with an established connection */
1392                 conn = kiblnd_get_conn_locked(peer);
1393                 kiblnd_conn_addref(conn); /* 1 ref for me... */
1394
1395                 read_unlock_irqrestore(g_lock, flags);
1396
1397                 if (tx)
1398                         kiblnd_queue_tx(tx, conn);
1399                 kiblnd_conn_decref(conn); /* ...to here */
1400                 return;
1401         }
1402
1403         read_unlock(g_lock);
1404         /* Re-try with a write lock */
1405         write_lock(g_lock);
1406
1407         peer = kiblnd_find_peer_locked(nid);
1408         if (peer) {
1409                 if (list_empty(&peer->ibp_conns)) {
1410                         /* found a peer, but it's still connecting... */
1411                         LASSERT(kiblnd_peer_connecting(peer));
1412                         if (tx)
1413                                 list_add_tail(&tx->tx_list,
1414                                               &peer->ibp_tx_queue);
1415                         write_unlock_irqrestore(g_lock, flags);
1416                 } else {
1417                         conn = kiblnd_get_conn_locked(peer);
1418                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1419
1420                         write_unlock_irqrestore(g_lock, flags);
1421
1422                         if (tx)
1423                                 kiblnd_queue_tx(tx, conn);
1424                         kiblnd_conn_decref(conn); /* ...to here */
1425                 }
1426                 return;
1427         }
1428
1429         write_unlock_irqrestore(g_lock, flags);
1430
1431         /* Allocate a peer ready to add to the peer table and retry */
1432         rc = kiblnd_create_peer(ni, &peer, nid);
1433         if (rc) {
1434                 CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
1435                 if (tx) {
1436                         tx->tx_status = -EHOSTUNREACH;
1437                         tx->tx_waiting = 0;
1438                         kiblnd_tx_done(ni, tx);
1439                 }
1440                 return;
1441         }
1442
1443         write_lock_irqsave(g_lock, flags);
1444
1445         peer2 = kiblnd_find_peer_locked(nid);
1446         if (peer2) {
1447                 if (list_empty(&peer2->ibp_conns)) {
1448                         /* found a peer, but it's still connecting... */
1449                         LASSERT(kiblnd_peer_connecting(peer2));
1450                         if (tx)
1451                                 list_add_tail(&tx->tx_list,
1452                                               &peer2->ibp_tx_queue);
1453                         write_unlock_irqrestore(g_lock, flags);
1454                 } else {
1455                         conn = kiblnd_get_conn_locked(peer2);
1456                         kiblnd_conn_addref(conn); /* 1 ref for me... */
1457
1458                         write_unlock_irqrestore(g_lock, flags);
1459
1460                         if (tx)
1461                                 kiblnd_queue_tx(tx, conn);
1462                         kiblnd_conn_decref(conn); /* ...to here */
1463                 }
1464
1465                 kiblnd_peer_decref(peer);
1466                 return;
1467         }
1468
1469         /* Brand new peer */
1470         LASSERT(!peer->ibp_connecting);
1471         peer->ibp_connecting = 1;
1472
1473         /* always called with a ref on ni, which prevents ni being shutdown */
1474         LASSERT(!((kib_net_t *)ni->ni_data)->ibn_shutdown);
1475
1476         if (tx)
1477                 list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
1478
1479         kiblnd_peer_addref(peer);
1480         list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
1481
1482         write_unlock_irqrestore(g_lock, flags);
1483
1484         kiblnd_connect_peer(peer);
1485         kiblnd_peer_decref(peer);
1486 }
1487
1488 int
1489 kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
1490 {
1491         lnet_hdr_t *hdr = &lntmsg->msg_hdr;
1492         int type = lntmsg->msg_type;
1493         lnet_process_id_t target = lntmsg->msg_target;
1494         int target_is_router = lntmsg->msg_target_is_router;
1495         int routing = lntmsg->msg_routing;
1496         unsigned int payload_niov = lntmsg->msg_niov;
1497         struct kvec *payload_iov = lntmsg->msg_iov;
1498         lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
1499         unsigned int payload_offset = lntmsg->msg_offset;
1500         unsigned int payload_nob = lntmsg->msg_len;
1501         kib_msg_t *ibmsg;
1502         kib_rdma_desc_t  *rd;
1503         kib_tx_t *tx;
1504         int nob;
1505         int rc;
1506
1507         /* NB 'private' is different depending on what we're sending.... */
1508
1509         CDEBUG(D_NET, "sending %d bytes in %d frags to %s\n",
1510                payload_nob, payload_niov, libcfs_id2str(target));
1511
1512         LASSERT(!payload_nob || payload_niov > 0);
1513         LASSERT(payload_niov <= LNET_MAX_IOV);
1514
1515         /* Thread context */
1516         LASSERT(!in_interrupt());
1517         /* payload is either all vaddrs or all pages */
1518         LASSERT(!(payload_kiov && payload_iov));
1519
1520         switch (type) {
1521         default:
1522                 LBUG();
1523                 return -EIO;
1524
1525         case LNET_MSG_ACK:
1526                 LASSERT(!payload_nob);
1527                 break;
1528
1529         case LNET_MSG_GET:
1530                 if (routing || target_is_router)
1531                         break;            /* send IMMEDIATE */
1532
1533                 /* is the REPLY message too small for RDMA? */
1534                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[lntmsg->msg_md->md_length]);
1535                 if (nob <= IBLND_MSG_SIZE)
1536                         break;            /* send IMMEDIATE */
1537
1538                 tx = kiblnd_get_idle_tx(ni, target.nid);
1539                 if (!tx) {
1540                         CERROR("Can't allocate txd for GET to %s\n",
1541                                libcfs_nid2str(target.nid));
1542                         return -ENOMEM;
1543                 }
1544
1545                 ibmsg = tx->tx_msg;
1546                 rd = &ibmsg->ibm_u.get.ibgm_rd;
1547                 if (!(lntmsg->msg_md->md_options & LNET_MD_KIOV))
1548                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1549                                                  lntmsg->msg_md->md_niov,
1550                                                  lntmsg->msg_md->md_iov.iov,
1551                                                  0, lntmsg->msg_md->md_length);
1552                 else
1553                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1554                                                   lntmsg->msg_md->md_niov,
1555                                                   lntmsg->msg_md->md_iov.kiov,
1556                                                   0, lntmsg->msg_md->md_length);
1557                 if (rc) {
1558                         CERROR("Can't setup GET sink for %s: %d\n",
1559                                libcfs_nid2str(target.nid), rc);
1560                         kiblnd_tx_done(ni, tx);
1561                         return -EIO;
1562                 }
1563
1564                 nob = offsetof(kib_get_msg_t, ibgm_rd.rd_frags[rd->rd_nfrags]);
1565                 ibmsg->ibm_u.get.ibgm_cookie = tx->tx_cookie;
1566                 ibmsg->ibm_u.get.ibgm_hdr = *hdr;
1567
1568                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_GET_REQ, nob);
1569
1570                 tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
1571                 if (!tx->tx_lntmsg[1]) {
1572                         CERROR("Can't create reply for GET -> %s\n",
1573                                libcfs_nid2str(target.nid));
1574                         kiblnd_tx_done(ni, tx);
1575                         return -EIO;
1576                 }
1577
1578                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg[0,1] on completion */
1579                 tx->tx_waiting = 1;          /* waiting for GET_DONE */
1580                 kiblnd_launch_tx(ni, tx, target.nid);
1581                 return 0;
1582
1583         case LNET_MSG_REPLY:
1584         case LNET_MSG_PUT:
1585                 /* Is the payload small enough not to need RDMA? */
1586                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1587                 if (nob <= IBLND_MSG_SIZE)
1588                         break;            /* send IMMEDIATE */
1589
1590                 tx = kiblnd_get_idle_tx(ni, target.nid);
1591                 if (!tx) {
1592                         CERROR("Can't allocate %s txd for %s\n",
1593                                type == LNET_MSG_PUT ? "PUT" : "REPLY",
1594                                libcfs_nid2str(target.nid));
1595                         return -ENOMEM;
1596                 }
1597
1598                 if (!payload_kiov)
1599                         rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1600                                                  payload_niov, payload_iov,
1601                                                  payload_offset, payload_nob);
1602                 else
1603                         rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1604                                                   payload_niov, payload_kiov,
1605                                                   payload_offset, payload_nob);
1606                 if (rc) {
1607                         CERROR("Can't setup PUT src for %s: %d\n",
1608                                libcfs_nid2str(target.nid), rc);
1609                         kiblnd_tx_done(ni, tx);
1610                         return -EIO;
1611                 }
1612
1613                 ibmsg = tx->tx_msg;
1614                 ibmsg->ibm_u.putreq.ibprm_hdr = *hdr;
1615                 ibmsg->ibm_u.putreq.ibprm_cookie = tx->tx_cookie;
1616                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_REQ, sizeof(kib_putreq_msg_t));
1617
1618                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1619                 tx->tx_waiting = 1;          /* waiting for PUT_{ACK,NAK} */
1620                 kiblnd_launch_tx(ni, tx, target.nid);
1621                 return 0;
1622         }
1623
1624         /* send IMMEDIATE */
1625
1626         LASSERT(offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob])
1627                  <= IBLND_MSG_SIZE);
1628
1629         tx = kiblnd_get_idle_tx(ni, target.nid);
1630         if (!tx) {
1631                 CERROR("Can't send %d to %s: tx descs exhausted\n",
1632                        type, libcfs_nid2str(target.nid));
1633                 return -ENOMEM;
1634         }
1635
1636         ibmsg = tx->tx_msg;
1637         ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1638
1639         if (payload_kiov)
1640                 lnet_copy_kiov2flat(IBLND_MSG_SIZE, ibmsg,
1641                                     offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1642                                     payload_niov, payload_kiov,
1643                                     payload_offset, payload_nob);
1644         else
1645                 lnet_copy_iov2flat(IBLND_MSG_SIZE, ibmsg,
1646                                    offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1647                                    payload_niov, payload_iov,
1648                                    payload_offset, payload_nob);
1649
1650         nob = offsetof(kib_immediate_msg_t, ibim_payload[payload_nob]);
1651         kiblnd_init_tx_msg(ni, tx, IBLND_MSG_IMMEDIATE, nob);
1652
1653         tx->tx_lntmsg[0] = lntmsg;            /* finalise lntmsg on completion */
1654         kiblnd_launch_tx(ni, tx, target.nid);
1655         return 0;
1656 }
1657
1658 static void
1659 kiblnd_reply(lnet_ni_t *ni, kib_rx_t *rx, lnet_msg_t *lntmsg)
1660 {
1661         lnet_process_id_t target = lntmsg->msg_target;
1662         unsigned int niov = lntmsg->msg_niov;
1663         struct kvec *iov = lntmsg->msg_iov;
1664         lnet_kiov_t *kiov = lntmsg->msg_kiov;
1665         unsigned int offset = lntmsg->msg_offset;
1666         unsigned int nob = lntmsg->msg_len;
1667         kib_tx_t *tx;
1668         int rc;
1669
1670         tx = kiblnd_get_idle_tx(ni, rx->rx_conn->ibc_peer->ibp_nid);
1671         if (!tx) {
1672                 CERROR("Can't get tx for REPLY to %s\n",
1673                        libcfs_nid2str(target.nid));
1674                 goto failed_0;
1675         }
1676
1677         if (!nob)
1678                 rc = 0;
1679         else if (!kiov)
1680                 rc = kiblnd_setup_rd_iov(ni, tx, tx->tx_rd,
1681                                          niov, iov, offset, nob);
1682         else
1683                 rc = kiblnd_setup_rd_kiov(ni, tx, tx->tx_rd,
1684                                           niov, kiov, offset, nob);
1685
1686         if (rc) {
1687                 CERROR("Can't setup GET src for %s: %d\n",
1688                        libcfs_nid2str(target.nid), rc);
1689                 goto failed_1;
1690         }
1691
1692         rc = kiblnd_init_rdma(rx->rx_conn, tx,
1693                               IBLND_MSG_GET_DONE, nob,
1694                               &rx->rx_msg->ibm_u.get.ibgm_rd,
1695                               rx->rx_msg->ibm_u.get.ibgm_cookie);
1696         if (rc < 0) {
1697                 CERROR("Can't setup rdma for GET from %s: %d\n",
1698                        libcfs_nid2str(target.nid), rc);
1699                 goto failed_1;
1700         }
1701
1702         if (!nob) {
1703                 /* No RDMA: local completion may happen now! */
1704                 lnet_finalize(ni, lntmsg, 0);
1705         } else {
1706                 /* RDMA: lnet_finalize(lntmsg) when it completes */
1707                 tx->tx_lntmsg[0] = lntmsg;
1708         }
1709
1710         kiblnd_queue_tx(tx, rx->rx_conn);
1711         return;
1712
1713  failed_1:
1714         kiblnd_tx_done(ni, tx);
1715  failed_0:
1716         lnet_finalize(ni, lntmsg, -EIO);
1717 }
1718
1719 int
1720 kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1721             unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1722             unsigned int offset, unsigned int mlen, unsigned int rlen)
1723 {
1724         kib_rx_t *rx = private;
1725         kib_msg_t *rxmsg = rx->rx_msg;
1726         kib_conn_t *conn = rx->rx_conn;
1727         kib_tx_t *tx;
1728         int nob;
1729         int post_credit = IBLND_POSTRX_PEER_CREDIT;
1730         int rc = 0;
1731
1732         LASSERT(mlen <= rlen);
1733         LASSERT(!in_interrupt());
1734         /* Either all pages or all vaddrs */
1735         LASSERT(!(kiov && iov));
1736
1737         switch (rxmsg->ibm_type) {
1738         default:
1739                 LBUG();
1740
1741         case IBLND_MSG_IMMEDIATE:
1742                 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1743                 if (nob > rx->rx_nob) {
1744                         CERROR("Immediate message from %s too big: %d(%d)\n",
1745                                libcfs_nid2str(rxmsg->ibm_u.immediate.ibim_hdr.src_nid),
1746                                nob, rx->rx_nob);
1747                         rc = -EPROTO;
1748                         break;
1749                 }
1750
1751                 if (kiov)
1752                         lnet_copy_flat2kiov(niov, kiov, offset,
1753                                             IBLND_MSG_SIZE, rxmsg,
1754                                             offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1755                                             mlen);
1756                 else
1757                         lnet_copy_flat2iov(niov, iov, offset,
1758                                            IBLND_MSG_SIZE, rxmsg,
1759                                            offsetof(kib_msg_t, ibm_u.immediate.ibim_payload),
1760                                            mlen);
1761                 lnet_finalize(ni, lntmsg, 0);
1762                 break;
1763
1764         case IBLND_MSG_PUT_REQ: {
1765                 kib_msg_t       *txmsg;
1766                 kib_rdma_desc_t *rd;
1767
1768                 if (!mlen) {
1769                         lnet_finalize(ni, lntmsg, 0);
1770                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, 0,
1771                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1772                         break;
1773                 }
1774
1775                 tx = kiblnd_get_idle_tx(ni, conn->ibc_peer->ibp_nid);
1776                 if (!tx) {
1777                         CERROR("Can't allocate tx for %s\n",
1778                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
1779                         /* Not replying will break the connection */
1780                         rc = -ENOMEM;
1781                         break;
1782                 }
1783
1784                 txmsg = tx->tx_msg;
1785                 rd = &txmsg->ibm_u.putack.ibpam_rd;
1786                 if (!kiov)
1787                         rc = kiblnd_setup_rd_iov(ni, tx, rd,
1788                                                  niov, iov, offset, mlen);
1789                 else
1790                         rc = kiblnd_setup_rd_kiov(ni, tx, rd,
1791                                                   niov, kiov, offset, mlen);
1792                 if (rc) {
1793                         CERROR("Can't setup PUT sink for %s: %d\n",
1794                                libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
1795                         kiblnd_tx_done(ni, tx);
1796                         /* tell peer it's over */
1797                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
1798                                                rxmsg->ibm_u.putreq.ibprm_cookie);
1799                         break;
1800                 }
1801
1802                 nob = offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[rd->rd_nfrags]);
1803                 txmsg->ibm_u.putack.ibpam_src_cookie = rxmsg->ibm_u.putreq.ibprm_cookie;
1804                 txmsg->ibm_u.putack.ibpam_dst_cookie = tx->tx_cookie;
1805
1806                 kiblnd_init_tx_msg(ni, tx, IBLND_MSG_PUT_ACK, nob);
1807
1808                 tx->tx_lntmsg[0] = lntmsg;      /* finalise lntmsg on completion */
1809                 tx->tx_waiting = 1;          /* waiting for PUT_DONE */
1810                 kiblnd_queue_tx(tx, conn);
1811
1812                 /* reposted buffer reserved for PUT_DONE */
1813                 post_credit = IBLND_POSTRX_NO_CREDIT;
1814                 break;
1815                 }
1816
1817         case IBLND_MSG_GET_REQ:
1818                 if (lntmsg) {
1819                         /* Optimized GET; RDMA lntmsg's payload */
1820                         kiblnd_reply(ni, rx, lntmsg);
1821                 } else {
1822                         /* GET didn't match anything */
1823                         kiblnd_send_completion(rx->rx_conn, IBLND_MSG_GET_DONE,
1824                                                -ENODATA,
1825                                                rxmsg->ibm_u.get.ibgm_cookie);
1826                 }
1827                 break;
1828         }
1829
1830         kiblnd_post_rx(rx, post_credit);
1831         return rc;
1832 }
1833
1834 int
1835 kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name)
1836 {
1837         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1838
1839         if (IS_ERR(task))
1840                 return PTR_ERR(task);
1841
1842         atomic_inc(&kiblnd_data.kib_nthreads);
1843         return 0;
1844 }
1845
1846 static void
1847 kiblnd_thread_fini(void)
1848 {
1849         atomic_dec(&kiblnd_data.kib_nthreads);
1850 }
1851
1852 static void
1853 kiblnd_peer_alive(kib_peer_t *peer)
1854 {
1855         /* This is racy, but everyone's only writing cfs_time_current() */
1856         peer->ibp_last_alive = cfs_time_current();
1857         mb();
1858 }
1859
1860 static void
1861 kiblnd_peer_notify(kib_peer_t *peer)
1862 {
1863         int error = 0;
1864         unsigned long last_alive = 0;
1865         unsigned long flags;
1866
1867         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1868
1869         if (kiblnd_peer_idle(peer) && peer->ibp_error) {
1870                 error = peer->ibp_error;
1871                 peer->ibp_error = 0;
1872
1873                 last_alive = peer->ibp_last_alive;
1874         }
1875
1876         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1877
1878         if (error)
1879                 lnet_notify(peer->ibp_ni,
1880                             peer->ibp_nid, 0, last_alive);
1881 }
1882
1883 void
1884 kiblnd_close_conn_locked(kib_conn_t *conn, int error)
1885 {
1886         /*
1887          * This just does the immediate housekeeping. 'error' is zero for a
1888          * normal shutdown which can happen only after the connection has been
1889          * established.  If the connection is established, schedule the
1890          * connection to be finished off by the connd. Otherwise the connd is
1891          * already dealing with it (either to set it up or tear it down).
1892          * Caller holds kib_global_lock exclusively in irq context
1893          */
1894         kib_peer_t *peer = conn->ibc_peer;
1895         kib_dev_t *dev;
1896         unsigned long flags;
1897
1898         LASSERT(error || conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1899
1900         if (error && !conn->ibc_comms_error)
1901                 conn->ibc_comms_error = error;
1902
1903         if (conn->ibc_state != IBLND_CONN_ESTABLISHED)
1904                 return; /* already being handled  */
1905
1906         if (!error &&
1907             list_empty(&conn->ibc_tx_noops) &&
1908             list_empty(&conn->ibc_tx_queue) &&
1909             list_empty(&conn->ibc_tx_queue_rsrvd) &&
1910             list_empty(&conn->ibc_tx_queue_nocred) &&
1911             list_empty(&conn->ibc_active_txs)) {
1912                 CDEBUG(D_NET, "closing conn to %s\n",
1913                        libcfs_nid2str(peer->ibp_nid));
1914         } else {
1915                 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1916                         libcfs_nid2str(peer->ibp_nid), error,
1917                         list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
1918                         list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
1919                         list_empty(&conn->ibc_tx_queue_rsrvd) ? "" : "(sending_rsrvd)",
1920                         list_empty(&conn->ibc_tx_queue_nocred) ? "" : "(sending_nocred)",
1921                         list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
1922         }
1923
1924         dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
1925         list_del(&conn->ibc_list);
1926         /* connd (see below) takes over ibc_list's ref */
1927
1928         if (list_empty(&peer->ibp_conns) &&    /* no more conns */
1929             kiblnd_peer_active(peer)) {  /* still in peer table */
1930                 kiblnd_unlink_peer_locked(peer);
1931
1932                 /* set/clear error on last conn */
1933                 peer->ibp_error = conn->ibc_comms_error;
1934         }
1935
1936         kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
1937
1938         if (error &&
1939             kiblnd_dev_can_failover(dev)) {
1940                 list_add_tail(&dev->ibd_fail_list,
1941                               &kiblnd_data.kib_failed_devs);
1942                 wake_up(&kiblnd_data.kib_failover_waitq);
1943         }
1944
1945         spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
1946
1947         list_add_tail(&conn->ibc_list, &kiblnd_data.kib_connd_conns);
1948         wake_up(&kiblnd_data.kib_connd_waitq);
1949
1950         spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
1951 }
1952
1953 void
1954 kiblnd_close_conn(kib_conn_t *conn, int error)
1955 {
1956         unsigned long flags;
1957
1958         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1959
1960         kiblnd_close_conn_locked(conn, error);
1961
1962         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1963 }
1964
1965 static void
1966 kiblnd_handle_early_rxs(kib_conn_t *conn)
1967 {
1968         unsigned long flags;
1969         kib_rx_t *rx;
1970         kib_rx_t *tmp;
1971
1972         LASSERT(!in_interrupt());
1973         LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
1974
1975         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1976         list_for_each_entry_safe(rx, tmp, &conn->ibc_early_rxs, rx_list) {
1977                 list_del(&rx->rx_list);
1978                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1979
1980                 kiblnd_handle_rx(rx);
1981
1982                 write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
1983         }
1984         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
1985 }
1986
1987 static void
1988 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
1989 {
1990         LIST_HEAD(zombies);
1991         struct list_head *tmp;
1992         struct list_head *nxt;
1993         kib_tx_t *tx;
1994
1995         spin_lock(&conn->ibc_lock);
1996
1997         list_for_each_safe(tmp, nxt, txs) {
1998                 tx = list_entry(tmp, kib_tx_t, tx_list);
1999
2000                 if (txs == &conn->ibc_active_txs) {
2001                         LASSERT(!tx->tx_queued);
2002                         LASSERT(tx->tx_waiting || tx->tx_sending);
2003                 } else {
2004                         LASSERT(tx->tx_queued);
2005                 }
2006
2007                 tx->tx_status = -ECONNABORTED;
2008                 tx->tx_waiting = 0;
2009
2010                 if (!tx->tx_sending) {
2011                         tx->tx_queued = 0;
2012                         list_del(&tx->tx_list);
2013                         list_add(&tx->tx_list, &zombies);
2014                 }
2015         }
2016
2017         spin_unlock(&conn->ibc_lock);
2018
2019         kiblnd_txlist_done(conn->ibc_peer->ibp_ni, &zombies, -ECONNABORTED);
2020 }
2021
2022 static void
2023 kiblnd_finalise_conn(kib_conn_t *conn)
2024 {
2025         LASSERT(!in_interrupt());
2026         LASSERT(conn->ibc_state > IBLND_CONN_INIT);
2027
2028         kiblnd_set_conn_state(conn, IBLND_CONN_DISCONNECTED);
2029
2030         /*
2031          * abort_receives moves QP state to IB_QPS_ERR.  This is only required
2032          * for connections that didn't get as far as being connected, because
2033          * rdma_disconnect() does this for free.
2034          */
2035         kiblnd_abort_receives(conn);
2036
2037         /*
2038          * Complete all tx descs not waiting for sends to complete.
2039          * NB we should be safe from RDMA now that the QP has changed state
2040          */
2041         kiblnd_abort_txs(conn, &conn->ibc_tx_noops);
2042         kiblnd_abort_txs(conn, &conn->ibc_tx_queue);
2043         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
2044         kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
2045         kiblnd_abort_txs(conn, &conn->ibc_active_txs);
2046
2047         kiblnd_handle_early_rxs(conn);
2048 }
2049
2050 static void
2051 kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
2052 {
2053         LIST_HEAD(zombies);
2054         unsigned long flags;
2055
2056         LASSERT(error);
2057         LASSERT(!in_interrupt());
2058
2059         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2060
2061         if (active) {
2062                 LASSERT(peer->ibp_connecting > 0);
2063                 peer->ibp_connecting--;
2064         } else {
2065                 LASSERT(peer->ibp_accepting > 0);
2066                 peer->ibp_accepting--;
2067         }
2068
2069         if (kiblnd_peer_connecting(peer)) {
2070                 /* another connection attempt under way... */
2071                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
2072                                         flags);
2073                 return;
2074         }
2075
2076         peer->ibp_reconnected = 0;
2077         if (list_empty(&peer->ibp_conns)) {
2078                 /* Take peer's blocked transmits to complete with error */
2079                 list_add(&zombies, &peer->ibp_tx_queue);
2080                 list_del_init(&peer->ibp_tx_queue);
2081
2082                 if (kiblnd_peer_active(peer))
2083                         kiblnd_unlink_peer_locked(peer);
2084
2085                 peer->ibp_error = error;
2086         } else {
2087                 /* Can't have blocked transmits if there are connections */
2088                 LASSERT(list_empty(&peer->ibp_tx_queue));
2089         }
2090
2091         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2092
2093         kiblnd_peer_notify(peer);
2094
2095         if (list_empty(&zombies))
2096                 return;
2097
2098         CNETERR("Deleting messages for %s: connection failed\n",
2099                 libcfs_nid2str(peer->ibp_nid));
2100
2101         kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
2102 }
2103
2104 static void
2105 kiblnd_connreq_done(kib_conn_t *conn, int status)
2106 {
2107         kib_peer_t *peer = conn->ibc_peer;
2108         kib_tx_t *tx;
2109         kib_tx_t *tmp;
2110         struct list_head txs;
2111         unsigned long flags;
2112         int active;
2113
2114         active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2115
2116         CDEBUG(D_NET, "%s: active(%d), version(%x), status(%d)\n",
2117                libcfs_nid2str(peer->ibp_nid), active,
2118                conn->ibc_version, status);
2119
2120         LASSERT(!in_interrupt());
2121         LASSERT((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
2122                  peer->ibp_connecting > 0) ||
2123                  (conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
2124                  peer->ibp_accepting > 0));
2125
2126         LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
2127         conn->ibc_connvars = NULL;
2128
2129         if (status) {
2130                 /* failed to establish connection */
2131                 kiblnd_peer_connect_failed(peer, active, status);
2132                 kiblnd_finalise_conn(conn);
2133                 return;
2134         }
2135
2136         /* connection established */
2137         write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2138
2139         conn->ibc_last_send = jiffies;
2140         kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
2141         kiblnd_peer_alive(peer);
2142
2143         /*
2144          * Add conn to peer's list and nuke any dangling conns from a different
2145          * peer instance...
2146          */
2147         kiblnd_conn_addref(conn);              /* +1 ref for ibc_list */
2148         list_add(&conn->ibc_list, &peer->ibp_conns);
2149         peer->ibp_reconnected = 0;
2150         if (active)
2151                 peer->ibp_connecting--;
2152         else
2153                 peer->ibp_accepting--;
2154
2155         if (!peer->ibp_version) {
2156                 peer->ibp_version     = conn->ibc_version;
2157                 peer->ibp_incarnation = conn->ibc_incarnation;
2158         }
2159
2160         if (peer->ibp_version     != conn->ibc_version ||
2161             peer->ibp_incarnation != conn->ibc_incarnation) {
2162                 kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
2163                                                 conn->ibc_incarnation);
2164                 peer->ibp_version     = conn->ibc_version;
2165                 peer->ibp_incarnation = conn->ibc_incarnation;
2166         }
2167
2168         /* grab pending txs while I have the lock */
2169         list_add(&txs, &peer->ibp_tx_queue);
2170         list_del_init(&peer->ibp_tx_queue);
2171
2172         if (!kiblnd_peer_active(peer) ||        /* peer has been deleted */
2173             conn->ibc_comms_error) {       /* error has happened already */
2174                 lnet_ni_t *ni = peer->ibp_ni;
2175
2176                 /* start to shut down connection */
2177                 kiblnd_close_conn_locked(conn, -ECONNABORTED);
2178                 write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2179
2180                 kiblnd_txlist_done(ni, &txs, -ECONNABORTED);
2181
2182                 return;
2183         }
2184
2185         /**
2186          * refcount taken by cmid is not reliable after I released the glock
2187          * because this connection is visible to other threads now, another
2188          * thread can find and close this connection right after I released
2189          * the glock, if kiblnd_cm_callback for RDMA_CM_EVENT_DISCONNECTED is
2190          * called, it can release the connection refcount taken by cmid.
2191          * It means the connection could be destroyed before I finish my
2192          * operations on it.
2193          */
2194         kiblnd_conn_addref(conn);
2195         write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2196
2197         /* Schedule blocked txs */
2198         spin_lock(&conn->ibc_lock);
2199         list_for_each_entry_safe(tx, tmp, &txs, tx_list) {
2200                 list_del(&tx->tx_list);
2201
2202                 kiblnd_queue_tx_locked(tx, conn);
2203         }
2204         spin_unlock(&conn->ibc_lock);
2205
2206         kiblnd_check_sends(conn);
2207
2208         /* schedule blocked rxs */
2209         kiblnd_handle_early_rxs(conn);
2210
2211         kiblnd_conn_decref(conn);
2212 }
2213
2214 static void
2215 kiblnd_reject(struct rdma_cm_id *cmid, kib_rej_t *rej)
2216 {
2217         int rc;
2218
2219         rc = rdma_reject(cmid, rej, sizeof(*rej));
2220
2221         if (rc)
2222                 CWARN("Error %d sending reject\n", rc);
2223 }
2224
2225 static int
2226 kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
2227 {
2228         rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
2229         kib_msg_t *reqmsg = priv;
2230         kib_msg_t *ackmsg;
2231         kib_dev_t *ibdev;
2232         kib_peer_t *peer;
2233         kib_peer_t *peer2;
2234         kib_conn_t *conn;
2235         lnet_ni_t *ni  = NULL;
2236         kib_net_t *net = NULL;
2237         lnet_nid_t nid;
2238         struct rdma_conn_param cp;
2239         kib_rej_t rej;
2240         int version = IBLND_MSG_VERSION;
2241         unsigned long flags;
2242         int rc;
2243         struct sockaddr_in *peer_addr;
2244
2245         LASSERT(!in_interrupt());
2246
2247         /* cmid inherits 'context' from the corresponding listener id */
2248         ibdev = (kib_dev_t *)cmid->context;
2249         LASSERT(ibdev);
2250
2251         memset(&rej, 0, sizeof(rej));
2252         rej.ibr_magic = IBLND_MSG_MAGIC;
2253         rej.ibr_why = IBLND_REJECT_FATAL;
2254         rej.ibr_cp.ibcp_max_msg_size = IBLND_MSG_SIZE;
2255
2256         peer_addr = (struct sockaddr_in *)&cmid->route.addr.dst_addr;
2257         if (*kiblnd_tunables.kib_require_priv_port &&
2258             ntohs(peer_addr->sin_port) >= PROT_SOCK) {
2259                 __u32 ip = ntohl(peer_addr->sin_addr.s_addr);
2260
2261                 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2262                        &ip, ntohs(peer_addr->sin_port));
2263                 goto failed;
2264         }
2265
2266         if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
2267                 CERROR("Short connection request\n");
2268                 goto failed;
2269         }
2270
2271         /*
2272          * Future protocol version compatibility support!  If the
2273          * o2iblnd-specific protocol changes, or when LNET unifies
2274          * protocols over all LNDs, the initial connection will
2275          * negotiate a protocol version.  I trap this here to avoid
2276          * console errors; the reject tells the peer which protocol I
2277          * speak.
2278          */
2279         if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
2280             reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
2281                 goto failed;
2282         if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
2283             reqmsg->ibm_version != IBLND_MSG_VERSION &&
2284             reqmsg->ibm_version != IBLND_MSG_VERSION_1)
2285                 goto failed;
2286         if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
2287             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
2288             reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
2289                 goto failed;
2290
2291         rc = kiblnd_unpack_msg(reqmsg, priv_nob);
2292         if (rc) {
2293                 CERROR("Can't parse connection request: %d\n", rc);
2294                 goto failed;
2295         }
2296
2297         nid = reqmsg->ibm_srcnid;
2298         ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
2299
2300         if (ni) {
2301                 net = (kib_net_t *)ni->ni_data;
2302                 rej.ibr_incarnation = net->ibn_incarnation;
2303         }
2304
2305         if (!ni ||                       /* no matching net */
2306             ni->ni_nid != reqmsg->ibm_dstnid ||   /* right NET, wrong NID! */
2307             net->ibn_dev != ibdev) {          /* wrong device */
2308                 CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2309                        libcfs_nid2str(nid),
2310                        !ni ? "NA" : libcfs_nid2str(ni->ni_nid),
2311                        ibdev->ibd_ifname, ibdev->ibd_nnets,
2312                        &ibdev->ibd_ifip,
2313                        libcfs_nid2str(reqmsg->ibm_dstnid));
2314
2315                 goto failed;
2316         }
2317
2318        /* check time stamp as soon as possible */
2319         if (reqmsg->ibm_dststamp &&
2320             reqmsg->ibm_dststamp != net->ibn_incarnation) {
2321                 CWARN("Stale connection request\n");
2322                 rej.ibr_why = IBLND_REJECT_CONN_STALE;
2323                 goto failed;
2324         }
2325
2326         /* I can accept peer's version */
2327         version = reqmsg->ibm_version;
2328
2329         if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
2330                 CERROR("Unexpected connreq msg type: %x from %s\n",
2331                        reqmsg->ibm_type, libcfs_nid2str(nid));
2332                 goto failed;
2333         }
2334
2335         if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
2336             IBLND_MSG_QUEUE_SIZE(version)) {
2337                 CERROR("Can't accept conn from %s, queue depth too large: %d (<=%d wanted)\n",
2338                        libcfs_nid2str(nid),
2339                        reqmsg->ibm_u.connparams.ibcp_queue_depth,
2340                        IBLND_MSG_QUEUE_SIZE(version));
2341
2342                 if (version == IBLND_MSG_VERSION)
2343                         rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
2344
2345                 goto failed;
2346         }
2347
2348         if (reqmsg->ibm_u.connparams.ibcp_max_frags >
2349             IBLND_RDMA_FRAGS(version)) {
2350                 CWARN("Can't accept conn from %s (version %x): max_frags %d too large (%d wanted)\n",
2351                       libcfs_nid2str(nid), version,
2352                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2353                       IBLND_RDMA_FRAGS(version));
2354
2355                 if (version >= IBLND_MSG_VERSION)
2356                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2357
2358                 goto failed;
2359         } else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
2360                    IBLND_RDMA_FRAGS(version) && !net->ibn_fmr_ps) {
2361                 CWARN("Can't accept conn from %s (version %x): max_frags %d incompatible without FMR pool (%d wanted)\n",
2362                       libcfs_nid2str(nid), version,
2363                       reqmsg->ibm_u.connparams.ibcp_max_frags,
2364                       IBLND_RDMA_FRAGS(version));
2365
2366                 if (version >= IBLND_MSG_VERSION)
2367                         rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
2368
2369                 goto failed;
2370         }
2371
2372         if (reqmsg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2373                 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2374                        libcfs_nid2str(nid),
2375                        reqmsg->ibm_u.connparams.ibcp_max_msg_size,
2376                        IBLND_MSG_SIZE);
2377                 goto failed;
2378         }
2379
2380         /* assume 'nid' is a new peer; create  */
2381         rc = kiblnd_create_peer(ni, &peer, nid);
2382         if (rc) {
2383                 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
2384                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2385                 goto failed;
2386         }
2387
2388         /* We have validated the peer's parameters so use those */
2389         peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
2390         peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
2391
2392         write_lock_irqsave(g_lock, flags);
2393
2394         peer2 = kiblnd_find_peer_locked(nid);
2395         if (peer2) {
2396                 if (!peer2->ibp_version) {
2397                         peer2->ibp_version     = version;
2398                         peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2399                 }
2400
2401                 /* not the guy I've talked with */
2402                 if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
2403                     peer2->ibp_version     != version) {
2404                         kiblnd_close_peer_conns_locked(peer2, -ESTALE);
2405
2406                         if (kiblnd_peer_active(peer2)) {
2407                                 peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
2408                                 peer2->ibp_version = version;
2409                         }
2410                         write_unlock_irqrestore(g_lock, flags);
2411
2412                         CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
2413                               libcfs_nid2str(nid), peer2->ibp_version, version,
2414                               peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
2415
2416                         kiblnd_peer_decref(peer);
2417                         rej.ibr_why = IBLND_REJECT_CONN_STALE;
2418                         goto failed;
2419                 }
2420
2421                 /* tie-break connection race in favour of the higher NID */
2422                 if (peer2->ibp_connecting &&
2423                     nid < ni->ni_nid) {
2424                         write_unlock_irqrestore(g_lock, flags);
2425
2426                         CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
2427
2428                         kiblnd_peer_decref(peer);
2429                         rej.ibr_why = IBLND_REJECT_CONN_RACE;
2430                         goto failed;
2431                 }
2432
2433                 /**
2434                  * passive connection is allowed even this peer is waiting for
2435                  * reconnection.
2436                  */
2437                 peer2->ibp_reconnecting = 0;
2438                 peer2->ibp_accepting++;
2439                 kiblnd_peer_addref(peer2);
2440
2441                 /**
2442                  * Race with kiblnd_launch_tx (active connect) to create peer
2443                  * so copy validated parameters since we now know what the
2444                  * peer's limits are
2445                  */
2446                 peer2->ibp_max_frags = peer->ibp_max_frags;
2447                 peer2->ibp_queue_depth = peer->ibp_queue_depth;
2448
2449                 write_unlock_irqrestore(g_lock, flags);
2450                 kiblnd_peer_decref(peer);
2451                 peer = peer2;
2452         } else {
2453                 /* Brand new peer */
2454                 LASSERT(!peer->ibp_accepting);
2455                 LASSERT(!peer->ibp_version &&
2456                         !peer->ibp_incarnation);
2457
2458                 peer->ibp_accepting   = 1;
2459                 peer->ibp_version     = version;
2460                 peer->ibp_incarnation = reqmsg->ibm_srcstamp;
2461
2462                 /* I have a ref on ni that prevents it being shutdown */
2463                 LASSERT(!net->ibn_shutdown);
2464
2465                 kiblnd_peer_addref(peer);
2466                 list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
2467
2468                 write_unlock_irqrestore(g_lock, flags);
2469         }
2470
2471         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT,
2472                                   version);
2473         if (!conn) {
2474                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
2475                 kiblnd_peer_decref(peer);
2476                 rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
2477                 goto failed;
2478         }
2479
2480         /*
2481          * conn now "owns" cmid, so I return success from here on to ensure the
2482          * CM callback doesn't destroy cmid.
2483          */
2484         conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
2485         conn->ibc_credits          = conn->ibc_queue_depth;
2486         conn->ibc_reserved_credits = conn->ibc_queue_depth;
2487         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2488                 IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
2489
2490         ackmsg = &conn->ibc_connvars->cv_msg;
2491         memset(ackmsg, 0, sizeof(*ackmsg));
2492
2493         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
2494                         sizeof(ackmsg->ibm_u.connparams));
2495         ackmsg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2496         ackmsg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2497         ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2498
2499         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
2500
2501         memset(&cp, 0, sizeof(cp));
2502         cp.private_data = ackmsg;
2503         cp.private_data_len = ackmsg->ibm_nob;
2504         cp.responder_resources = 0;          /* No atomic ops or RDMA reads */
2505         cp.initiator_depth = 0;
2506         cp.flow_control = 1;
2507         cp.retry_count = *kiblnd_tunables.kib_retry_count;
2508         cp.rnr_retry_count = *kiblnd_tunables.kib_rnr_retry_count;
2509
2510         CDEBUG(D_NET, "Accept %s\n", libcfs_nid2str(nid));
2511
2512         rc = rdma_accept(cmid, &cp);
2513         if (rc) {
2514                 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid), rc);
2515                 rej.ibr_version = version;
2516                 rej.ibr_why     = IBLND_REJECT_FATAL;
2517
2518                 kiblnd_reject(cmid, &rej);
2519                 kiblnd_connreq_done(conn, rc);
2520                 kiblnd_conn_decref(conn);
2521         }
2522
2523         lnet_ni_decref(ni);
2524         return 0;
2525
2526  failed:
2527         if (ni)
2528                 lnet_ni_decref(ni);
2529
2530         rej.ibr_version             = version;
2531         rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
2532         rej.ibr_cp.ibcp_max_frags   = IBLND_RDMA_FRAGS(version);
2533         kiblnd_reject(cmid, &rej);
2534
2535         return -ECONNREFUSED;
2536 }
2537
2538 static void
2539 kiblnd_check_reconnect(kib_conn_t *conn, int version,
2540                        __u64 incarnation, int why, kib_connparams_t *cp)
2541 {
2542         rwlock_t *glock = &kiblnd_data.kib_global_lock;
2543         kib_peer_t *peer = conn->ibc_peer;
2544         char *reason;
2545         int msg_size = IBLND_MSG_SIZE;
2546         int frag_num = -1;
2547         int queue_dep = -1;
2548         bool reconnect;
2549         unsigned long flags;
2550
2551         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2552         LASSERT(peer->ibp_connecting > 0);     /* 'conn' at least */
2553         LASSERT(!peer->ibp_reconnecting);
2554
2555         if (cp) {
2556                 msg_size = cp->ibcp_max_msg_size;
2557                 frag_num = cp->ibcp_max_frags;
2558                 queue_dep = cp->ibcp_queue_depth;
2559         }
2560
2561         write_lock_irqsave(glock, flags);
2562         /**
2563          * retry connection if it's still needed and no other connection
2564          * attempts (active or passive) are in progress
2565          * NB: reconnect is still needed even when ibp_tx_queue is
2566          * empty if ibp_version != version because reconnect may be
2567          * initiated by kiblnd_query()
2568          */
2569         reconnect = (!list_empty(&peer->ibp_tx_queue) ||
2570                      peer->ibp_version != version) &&
2571                     peer->ibp_connecting == 1 &&
2572                     !peer->ibp_accepting;
2573         if (!reconnect) {
2574                 reason = "no need";
2575                 goto out;
2576         }
2577
2578         switch (why) {
2579         default:
2580                 reason = "Unknown";
2581                 break;
2582
2583         case IBLND_REJECT_RDMA_FRAGS:
2584                 if (!cp) {
2585                         reason = "can't negotiate max frags";
2586                         goto out;
2587                 }
2588                 if (!*kiblnd_tunables.kib_map_on_demand) {
2589                         reason = "map_on_demand must be enabled";
2590                         goto out;
2591                 }
2592                 if (conn->ibc_max_frags <= frag_num) {
2593                         reason = "unsupported max frags";
2594                         goto out;
2595                 }
2596
2597                 peer->ibp_max_frags = frag_num;
2598                 reason = "rdma fragments";
2599                 break;
2600
2601         case IBLND_REJECT_MSG_QUEUE_SIZE:
2602                 if (!cp) {
2603                         reason = "can't negotiate queue depth";
2604                         goto out;
2605                 }
2606                 if (conn->ibc_queue_depth <= queue_dep) {
2607                         reason = "unsupported queue depth";
2608                         goto out;
2609                 }
2610
2611                 peer->ibp_queue_depth = queue_dep;
2612                 reason = "queue depth";
2613                 break;
2614
2615         case IBLND_REJECT_CONN_STALE:
2616                 reason = "stale";
2617                 break;
2618
2619         case IBLND_REJECT_CONN_RACE:
2620                 reason = "conn race";
2621                 break;
2622
2623         case IBLND_REJECT_CONN_UNCOMPAT:
2624                 reason = "version negotiation";
2625                 break;
2626         }
2627
2628         conn->ibc_reconnect = 1;
2629         peer->ibp_reconnecting = 1;
2630         peer->ibp_version = version;
2631         if (incarnation)
2632                 peer->ibp_incarnation = incarnation;
2633 out:
2634         write_unlock_irqrestore(glock, flags);
2635
2636         CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
2637                 libcfs_nid2str(peer->ibp_nid),
2638                 reconnect ? "reconnect" : "don't reconnect",
2639                 reason, IBLND_MSG_VERSION, version, msg_size,
2640                 conn->ibc_queue_depth, queue_dep,
2641                 conn->ibc_max_frags, frag_num);
2642         /**
2643          * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
2644          * while destroying the zombie
2645          */
2646 }
2647
2648 static void
2649 kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
2650 {
2651         kib_peer_t *peer = conn->ibc_peer;
2652
2653         LASSERT(!in_interrupt());
2654         LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
2655
2656         switch (reason) {
2657         case IB_CM_REJ_STALE_CONN:
2658                 kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
2659                                        IBLND_REJECT_CONN_STALE, NULL);
2660                 break;
2661
2662         case IB_CM_REJ_INVALID_SERVICE_ID:
2663                 CNETERR("%s rejected: no listener at %d\n",
2664                         libcfs_nid2str(peer->ibp_nid),
2665                         *kiblnd_tunables.kib_service);
2666                 break;
2667
2668         case IB_CM_REJ_CONSUMER_DEFINED:
2669                 if (priv_nob >= offsetof(kib_rej_t, ibr_padding)) {
2670                         kib_rej_t *rej = priv;
2671                         kib_connparams_t *cp = NULL;
2672                         int flip = 0;
2673                         __u64 incarnation = -1;
2674
2675                         /* NB. default incarnation is -1 because:
2676                          * a) V1 will ignore dst incarnation in connreq.
2677                          * b) V2 will provide incarnation while rejecting me,
2678                          *    -1 will be overwrote.
2679                          *
2680                          * if I try to connect to a V1 peer with V2 protocol,
2681                          * it rejected me then upgrade to V2, I have no idea
2682                          * about the upgrading and try to reconnect with V1,
2683                          * in this case upgraded V2 can find out I'm trying to
2684                          * talk to the old guy and reject me(incarnation is -1).
2685                          */
2686
2687                         if (rej->ibr_magic == __swab32(IBLND_MSG_MAGIC) ||
2688                             rej->ibr_magic == __swab32(LNET_PROTO_MAGIC)) {
2689                                 __swab32s(&rej->ibr_magic);
2690                                 __swab16s(&rej->ibr_version);
2691                                 flip = 1;
2692                         }
2693
2694                         if (priv_nob >= sizeof(kib_rej_t) &&
2695                             rej->ibr_version > IBLND_MSG_VERSION_1) {
2696                                 /*
2697                                  * priv_nob is always 148 in current version
2698                                  * of OFED, so we still need to check version.
2699                                  * (define of IB_CM_REJ_PRIVATE_DATA_SIZE)
2700                                  */
2701                                 cp = &rej->ibr_cp;
2702
2703                                 if (flip) {
2704                                         __swab64s(&rej->ibr_incarnation);
2705                                         __swab16s(&cp->ibcp_queue_depth);
2706                                         __swab16s(&cp->ibcp_max_frags);
2707                                         __swab32s(&cp->ibcp_max_msg_size);
2708                                 }
2709
2710                                 incarnation = rej->ibr_incarnation;
2711                         }
2712
2713                         if (rej->ibr_magic != IBLND_MSG_MAGIC &&
2714                             rej->ibr_magic != LNET_PROTO_MAGIC) {
2715                                 CERROR("%s rejected: consumer defined fatal error\n",
2716                                        libcfs_nid2str(peer->ibp_nid));
2717                                 break;
2718                         }
2719
2720                         if (rej->ibr_version != IBLND_MSG_VERSION &&
2721                             rej->ibr_version != IBLND_MSG_VERSION_1) {
2722                                 CERROR("%s rejected: o2iblnd version %x error\n",
2723                                        libcfs_nid2str(peer->ibp_nid),
2724                                        rej->ibr_version);
2725                                 break;
2726                         }
2727
2728                         if (rej->ibr_why     == IBLND_REJECT_FATAL &&
2729                             rej->ibr_version == IBLND_MSG_VERSION_1) {
2730                                 CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
2731                                        libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
2732
2733                                 if (conn->ibc_version != IBLND_MSG_VERSION_1)
2734                                         rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
2735                         }
2736
2737                         switch (rej->ibr_why) {
2738                         case IBLND_REJECT_CONN_RACE:
2739                         case IBLND_REJECT_CONN_STALE:
2740                         case IBLND_REJECT_CONN_UNCOMPAT:
2741                         case IBLND_REJECT_MSG_QUEUE_SIZE:
2742                         case IBLND_REJECT_RDMA_FRAGS:
2743                                 kiblnd_check_reconnect(conn, rej->ibr_version,
2744                                                        incarnation,
2745                                                        rej->ibr_why, cp);
2746                                 break;
2747
2748                         case IBLND_REJECT_NO_RESOURCES:
2749                                 CERROR("%s rejected: o2iblnd no resources\n",
2750                                        libcfs_nid2str(peer->ibp_nid));
2751                                 break;
2752
2753                         case IBLND_REJECT_FATAL:
2754                                 CERROR("%s rejected: o2iblnd fatal error\n",
2755                                        libcfs_nid2str(peer->ibp_nid));
2756                                 break;
2757
2758                         default:
2759                                 CERROR("%s rejected: o2iblnd reason %d\n",
2760                                        libcfs_nid2str(peer->ibp_nid),
2761                                        rej->ibr_why);
2762                                 break;
2763                         }
2764                         break;
2765                 }
2766                 /* fall through */
2767         default:
2768                 CNETERR("%s rejected: reason %d, size %d\n",
2769                         libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
2770                 break;
2771         }
2772
2773         kiblnd_connreq_done(conn, -ECONNREFUSED);
2774 }
2775
2776 static void
2777 kiblnd_check_connreply(kib_conn_t *conn, void *priv, int priv_nob)
2778 {
2779         kib_peer_t *peer = conn->ibc_peer;
2780         lnet_ni_t *ni = peer->ibp_ni;
2781         kib_net_t *net = ni->ni_data;
2782         kib_msg_t *msg = priv;
2783         int ver = conn->ibc_version;
2784         int rc = kiblnd_unpack_msg(msg, priv_nob);
2785         unsigned long flags;
2786
2787         LASSERT(net);
2788
2789         if (rc) {
2790                 CERROR("Can't unpack connack from %s: %d\n",
2791                        libcfs_nid2str(peer->ibp_nid), rc);
2792                 goto failed;
2793         }
2794
2795         if (msg->ibm_type != IBLND_MSG_CONNACK) {
2796                 CERROR("Unexpected message %d from %s\n",
2797                        msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
2798                 rc = -EPROTO;
2799                 goto failed;
2800         }
2801
2802         if (ver != msg->ibm_version) {
2803                 CERROR("%s replied version %x is different with requested version %x\n",
2804                        libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
2805                 rc = -EPROTO;
2806                 goto failed;
2807         }
2808
2809         if (msg->ibm_u.connparams.ibcp_queue_depth >
2810             conn->ibc_queue_depth) {
2811                 CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
2812                        libcfs_nid2str(peer->ibp_nid),
2813                        msg->ibm_u.connparams.ibcp_queue_depth,
2814                        conn->ibc_queue_depth);
2815                 rc = -EPROTO;
2816                 goto failed;
2817         }
2818
2819         if (msg->ibm_u.connparams.ibcp_max_frags >
2820             conn->ibc_max_frags) {
2821                 CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
2822                        libcfs_nid2str(peer->ibp_nid),
2823                        msg->ibm_u.connparams.ibcp_max_frags,
2824                        conn->ibc_max_frags);
2825                 rc = -EPROTO;
2826                 goto failed;
2827         }
2828
2829         if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
2830                 CERROR("%s max message size %d too big (%d max)\n",
2831                        libcfs_nid2str(peer->ibp_nid),
2832                        msg->ibm_u.connparams.ibcp_max_msg_size,
2833                        IBLND_MSG_SIZE);
2834                 rc = -EPROTO;
2835                 goto failed;
2836         }
2837
2838         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2839         if (msg->ibm_dstnid == ni->ni_nid &&
2840             msg->ibm_dststamp == net->ibn_incarnation)
2841                 rc = 0;
2842         else
2843                 rc = -ESTALE;
2844         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2845
2846         if (rc) {
2847                 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2848                        libcfs_nid2str(peer->ibp_nid), rc,
2849                        msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
2850                 goto failed;
2851         }
2852
2853         conn->ibc_incarnation = msg->ibm_srcstamp;
2854         conn->ibc_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2855         conn->ibc_reserved_credits = msg->ibm_u.connparams.ibcp_queue_depth;
2856         conn->ibc_queue_depth = msg->ibm_u.connparams.ibcp_queue_depth;
2857         conn->ibc_max_frags = msg->ibm_u.connparams.ibcp_max_frags;
2858         LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
2859                 IBLND_OOB_MSGS(ver) <= IBLND_RX_MSGS(conn));
2860
2861         kiblnd_connreq_done(conn, 0);
2862         return;
2863
2864  failed:
2865         /*
2866          * NB My QP has already established itself, so I handle anything going
2867          * wrong here by setting ibc_comms_error.
2868          * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2869          * immediately tears it down.
2870          */
2871         LASSERT(rc);
2872         conn->ibc_comms_error = rc;
2873         kiblnd_connreq_done(conn, 0);
2874 }
2875
2876 static int
2877 kiblnd_active_connect(struct rdma_cm_id *cmid)
2878 {
2879         kib_peer_t *peer = (kib_peer_t *)cmid->context;
2880         kib_conn_t *conn;
2881         kib_msg_t *msg;
2882         struct rdma_conn_param cp;
2883         int version;
2884         __u64 incarnation;
2885         unsigned long flags;
2886         int rc;
2887
2888         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
2889
2890         incarnation = peer->ibp_incarnation;
2891         version = !peer->ibp_version ? IBLND_MSG_VERSION :
2892                                        peer->ibp_version;
2893
2894         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
2895
2896         conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
2897                                   version);
2898         if (!conn) {
2899                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
2900                 kiblnd_peer_decref(peer); /* lose cmid's ref */
2901                 return -ENOMEM;
2902         }
2903
2904         /*
2905          * conn "owns" cmid now, so I return success from here on to ensure the
2906          * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2907          * on peer
2908          */
2909         msg = &conn->ibc_connvars->cv_msg;
2910
2911         memset(msg, 0, sizeof(*msg));
2912         kiblnd_init_msg(msg, IBLND_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2913         msg->ibm_u.connparams.ibcp_queue_depth = conn->ibc_queue_depth;
2914         msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
2915         msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
2916
2917         kiblnd_pack_msg(peer->ibp_ni, msg, version,
2918                         0, peer->ibp_nid, incarnation);
2919
2920         memset(&cp, 0, sizeof(cp));
2921         cp.private_data = msg;
2922         cp.private_data_len    = msg->ibm_nob;
2923         cp.responder_resources = 0;          /* No atomic ops or RDMA reads */
2924         cp.initiator_depth     = 0;
2925         cp.flow_control        = 1;
2926         cp.retry_count         = *kiblnd_tunables.kib_retry_count;
2927         cp.rnr_retry_count     = *kiblnd_tunables.kib_rnr_retry_count;
2928
2929         LASSERT(cmid->context == (void *)conn);
2930         LASSERT(conn->ibc_cmid == cmid);
2931
2932         rc = rdma_connect(cmid, &cp);
2933         if (rc) {
2934                 CERROR("Can't connect to %s: %d\n",
2935                        libcfs_nid2str(peer->ibp_nid), rc);
2936                 kiblnd_connreq_done(conn, rc);
2937                 kiblnd_conn_decref(conn);
2938         }
2939
2940         return 0;
2941 }
2942
2943 int
2944 kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
2945 {
2946         kib_peer_t *peer;
2947         kib_conn_t *conn;
2948         int rc;
2949
2950         switch (event->event) {
2951         default:
2952                 CERROR("Unexpected event: %d, status: %d\n",
2953                        event->event, event->status);
2954                 LBUG();
2955
2956         case RDMA_CM_EVENT_CONNECT_REQUEST:
2957                 /* destroy cmid on failure */
2958                 rc = kiblnd_passive_connect(cmid,
2959                                             (void *)KIBLND_CONN_PARAM(event),
2960                                             KIBLND_CONN_PARAM_LEN(event));
2961                 CDEBUG(D_NET, "connreq: %d\n", rc);
2962                 return rc;
2963
2964         case RDMA_CM_EVENT_ADDR_ERROR:
2965                 peer = (kib_peer_t *)cmid->context;
2966                 CNETERR("%s: ADDR ERROR %d\n",
2967                         libcfs_nid2str(peer->ibp_nid), event->status);
2968                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
2969                 kiblnd_peer_decref(peer);
2970                 return -EHOSTUNREACH;      /* rc destroys cmid */
2971
2972         case RDMA_CM_EVENT_ADDR_RESOLVED:
2973                 peer = (kib_peer_t *)cmid->context;
2974
2975                 CDEBUG(D_NET, "%s Addr resolved: %d\n",
2976                        libcfs_nid2str(peer->ibp_nid), event->status);
2977
2978                 if (event->status) {
2979                         CNETERR("Can't resolve address for %s: %d\n",
2980                                 libcfs_nid2str(peer->ibp_nid), event->status);
2981                         rc = event->status;
2982                 } else {
2983                         rc = rdma_resolve_route(
2984                                 cmid, *kiblnd_tunables.kib_timeout * 1000);
2985                         if (!rc)
2986                                 return 0;
2987                         /* Can't initiate route resolution */
2988                         CERROR("Can't resolve route for %s: %d\n",
2989                                libcfs_nid2str(peer->ibp_nid), rc);
2990                 }
2991                 kiblnd_peer_connect_failed(peer, 1, rc);
2992                 kiblnd_peer_decref(peer);
2993                 return rc;                    /* rc destroys cmid */
2994
2995         case RDMA_CM_EVENT_ROUTE_ERROR:
2996                 peer = (kib_peer_t *)cmid->context;
2997                 CNETERR("%s: ROUTE ERROR %d\n",
2998                         libcfs_nid2str(peer->ibp_nid), event->status);
2999                 kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
3000                 kiblnd_peer_decref(peer);
3001                 return -EHOSTUNREACH;      /* rc destroys cmid */
3002
3003         case RDMA_CM_EVENT_ROUTE_RESOLVED:
3004                 peer = (kib_peer_t *)cmid->context;
3005                 CDEBUG(D_NET, "%s Route resolved: %d\n",
3006                        libcfs_nid2str(peer->ibp_nid), event->status);
3007
3008                 if (!event->status)
3009                         return kiblnd_active_connect(cmid);
3010
3011                 CNETERR("Can't resolve route for %s: %d\n",
3012                         libcfs_nid2str(peer->ibp_nid), event->status);
3013                 kiblnd_peer_connect_failed(peer, 1, event->status);
3014                 kiblnd_peer_decref(peer);
3015                 return event->status;      /* rc destroys cmid */
3016
3017         case RDMA_CM_EVENT_UNREACHABLE:
3018                 conn = (kib_conn_t *)cmid->context;
3019                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3020                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3021                 CNETERR("%s: UNREACHABLE %d\n",
3022                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3023                 kiblnd_connreq_done(conn, -ENETDOWN);
3024                 kiblnd_conn_decref(conn);
3025                 return 0;
3026
3027         case RDMA_CM_EVENT_CONNECT_ERROR:
3028                 conn = (kib_conn_t *)cmid->context;
3029                 LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT ||
3030                         conn->ibc_state == IBLND_CONN_PASSIVE_WAIT);
3031                 CNETERR("%s: CONNECT ERROR %d\n",
3032                         libcfs_nid2str(conn->ibc_peer->ibp_nid), event->status);
3033                 kiblnd_connreq_done(conn, -ENOTCONN);
3034                 kiblnd_conn_decref(conn);
3035                 return 0;
3036
3037         case RDMA_CM_EVENT_REJECTED:
3038                 conn = (kib_conn_t *)cmid->context;
3039                 switch (conn->ibc_state) {
3040                 default:
3041                         LBUG();
3042
3043                 case IBLND_CONN_PASSIVE_WAIT:
3044                         CERROR("%s: REJECTED %d\n",
3045                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
3046                                event->status);
3047                         kiblnd_connreq_done(conn, -ECONNRESET);
3048                         break;
3049
3050                 case IBLND_CONN_ACTIVE_CONNECT:
3051                         kiblnd_rejected(conn, event->status,
3052                                         (void *)KIBLND_CONN_PARAM(event),
3053                                         KIBLND_CONN_PARAM_LEN(event));
3054                         break;
3055                 }
3056                 kiblnd_conn_decref(conn);
3057                 return 0;
3058
3059         case RDMA_CM_EVENT_ESTABLISHED:
3060                 conn = (kib_conn_t *)cmid->context;
3061                 switch (conn->ibc_state) {
3062                 default:
3063                         LBUG();
3064
3065                 case IBLND_CONN_PASSIVE_WAIT:
3066                         CDEBUG(D_NET, "ESTABLISHED (passive): %s\n",
3067                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3068                         kiblnd_connreq_done(conn, 0);
3069                         break;
3070
3071                 case IBLND_CONN_ACTIVE_CONNECT:
3072                         CDEBUG(D_NET, "ESTABLISHED(active): %s\n",
3073                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3074                         kiblnd_check_connreply(conn,
3075                                                (void *)KIBLND_CONN_PARAM(event),
3076                                                KIBLND_CONN_PARAM_LEN(event));
3077                         break;
3078                 }
3079                 /* net keeps its ref on conn! */
3080                 return 0;
3081
3082         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3083                 CDEBUG(D_NET, "Ignore TIMEWAIT_EXIT event\n");
3084                 return 0;
3085         case RDMA_CM_EVENT_DISCONNECTED:
3086                 conn = (kib_conn_t *)cmid->context;
3087                 if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
3088                         CERROR("%s DISCONNECTED\n",
3089                                libcfs_nid2str(conn->ibc_peer->ibp_nid));
3090                         kiblnd_connreq_done(conn, -ECONNRESET);
3091                 } else {
3092                         kiblnd_close_conn(conn, 0);
3093                 }
3094                 kiblnd_conn_decref(conn);
3095                 cmid->context = NULL;
3096                 return 0;
3097
3098         case RDMA_CM_EVENT_DEVICE_REMOVAL:
3099                 LCONSOLE_ERROR_MSG(0x131,
3100                                    "Received notification of device removal\n"
3101                                    "Please shutdown LNET to allow this to proceed\n");
3102                 /*
3103                  * Can't remove network from underneath LNET for now, so I have
3104                  * to ignore this
3105                  */
3106                 return 0;
3107
3108         case RDMA_CM_EVENT_ADDR_CHANGE:
3109                 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
3110                 return 0;
3111         }
3112 }
3113
3114 static int
3115 kiblnd_check_txs_locked(kib_conn_t *conn, struct list_head *txs)
3116 {
3117         kib_tx_t *tx;
3118         struct list_head *ttmp;
3119
3120         list_for_each(ttmp, txs) {
3121                 tx = list_entry(ttmp, kib_tx_t, tx_list);
3122
3123                 if (txs != &conn->ibc_active_txs) {
3124                         LASSERT(tx->tx_queued);
3125                 } else {
3126                         LASSERT(!tx->tx_queued);
3127                         LASSERT(tx->tx_waiting || tx->tx_sending);
3128                 }
3129
3130                 if (cfs_time_aftereq(jiffies, tx->tx_deadline)) {
3131                         CERROR("Timed out tx: %s, %lu seconds\n",
3132                                kiblnd_queue2str(conn, txs),
3133                                cfs_duration_sec(jiffies - tx->tx_deadline));
3134                         return 1;
3135                 }
3136         }
3137
3138         return 0;
3139 }
3140
3141 static int
3142 kiblnd_conn_timed_out_locked(kib_conn_t *conn)
3143 {
3144         return  kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue) ||
3145                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_noops) ||
3146                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_rsrvd) ||
3147                 kiblnd_check_txs_locked(conn, &conn->ibc_tx_queue_nocred) ||
3148                 kiblnd_check_txs_locked(conn, &conn->ibc_active_txs);
3149 }
3150
3151 static void
3152 kiblnd_check_conns(int idx)
3153 {
3154         LIST_HEAD(closes);
3155         LIST_HEAD(checksends);
3156         struct list_head *peers = &kiblnd_data.kib_peers[idx];
3157         struct list_head *ptmp;
3158         kib_peer_t *peer;
3159         kib_conn_t *conn;
3160         kib_conn_t *temp;
3161         kib_conn_t *tmp;
3162         struct list_head *ctmp;
3163         unsigned long flags;
3164
3165         /*
3166          * NB. We expect to have a look at all the peers and not find any
3167          * RDMAs to time out, so we just use a shared lock while we
3168          * take a look...
3169          */
3170         read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
3171
3172         list_for_each(ptmp, peers) {
3173                 peer = list_entry(ptmp, kib_peer_t, ibp_list);
3174
3175                 list_for_each(ctmp, &peer->ibp_conns) {
3176                         int timedout;
3177                         int sendnoop;
3178
3179                         conn = list_entry(ctmp, kib_conn_t, ibc_list);
3180
3181                         LASSERT(conn->ibc_state == IBLND_CONN_ESTABLISHED);
3182
3183                         spin_lock(&conn->ibc_lock);
3184
3185                         sendnoop = kiblnd_need_noop(conn);
3186                         timedout = kiblnd_conn_timed_out_locked(conn);
3187                         if (!sendnoop && !timedout) {
3188                                 spin_unlock(&conn->ibc_lock);
3189                                 continue;
3190                         }
3191
3192                         if (timedout) {
3193                                 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3194                                        libcfs_nid2str(peer->ibp_nid),
3195                                        cfs_duration_sec(cfs_time_current() -
3196                                                         peer->ibp_last_alive),
3197                                        conn->ibc_credits,
3198                                        conn->ibc_outstanding_credits,
3199                                        conn->ibc_reserved_credits);
3200                                 list_add(&conn->ibc_connd_list, &closes);
3201                         } else {
3202                                 list_add(&conn->ibc_connd_list, &checksends);
3203                         }
3204                         /* +ref for 'closes' or 'checksends' */
3205                         kiblnd_conn_addref(conn);
3206
3207                         spin_unlock(&conn->ibc_lock);
3208                 }
3209         }
3210
3211         read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
3212
3213         /*
3214          * Handle timeout by closing the whole
3215          * connection. We can only be sure RDMA activity
3216          * has ceased once the QP has been modified.
3217          */
3218         list_for_each_entry_safe(conn, tmp, &closes, ibc_connd_list) {
3219                 list_del(&conn->ibc_connd_list);
3220                 kiblnd_close_conn(conn, -ETIMEDOUT);
3221                 kiblnd_conn_decref(conn);
3222         }
3223
3224         /*
3225          * In case we have enough credits to return via a
3226          * NOOP, but there were no non-blocking tx descs
3227          * free to do it last time...
3228          */
3229         list_for_each_entry_safe(conn, temp, &checksends, ibc_connd_list) {
3230                 list_del(&conn->ibc_connd_list);
3231                 kiblnd_check_sends(conn);
3232                 kiblnd_conn_decref(conn);
3233         }
3234 }
3235
3236 static void
3237 kiblnd_disconnect_conn(kib_conn_t *conn)
3238 {
3239         LASSERT(!in_interrupt());
3240         LASSERT(current == kiblnd_data.kib_connd);
3241         LASSERT(conn->ibc_state == IBLND_CONN_CLOSING);
3242
3243         rdma_disconnect(conn->ibc_cmid);
3244         kiblnd_finalise_conn(conn);
3245
3246         kiblnd_peer_notify(conn->ibc_peer);
3247 }
3248
3249 /**
3250  * High-water for reconnection to the same peer, reconnection attempt should
3251  * be delayed after trying more than KIB_RECONN_HIGH_RACE.
3252  */
3253 #define KIB_RECONN_HIGH_RACE    10
3254 /**
3255  * Allow connd to take a break and handle other things after consecutive
3256  * reconnection attemps.
3257  */
3258 #define KIB_RECONN_BREAK        100
3259
3260 int
3261 kiblnd_connd(void *arg)
3262 {
3263         spinlock_t *lock= &kiblnd_data.kib_connd_lock;
3264         wait_queue_t wait;
3265         unsigned long flags;
3266         kib_conn_t *conn;
3267         int timeout;
3268         int i;
3269         int dropped_lock;
3270         int peer_index = 0;
3271         unsigned long deadline = jiffies;
3272
3273         cfs_block_allsigs();
3274
3275         init_waitqueue_entry(&wait, current);
3276         kiblnd_data.kib_connd = current;
3277
3278         spin_lock_irqsave(lock, flags);
3279
3280         while (!kiblnd_data.kib_shutdown) {
3281                 int reconn = 0;
3282
3283                 dropped_lock = 0;
3284
3285                 if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
3286                         kib_peer_t *peer = NULL;
3287
3288                         conn = list_entry(kiblnd_data.kib_connd_zombies.next,
3289                                           kib_conn_t, ibc_list);
3290                         list_del(&conn->ibc_list);
3291                         if (conn->ibc_reconnect) {
3292                                 peer = conn->ibc_peer;
3293                                 kiblnd_peer_addref(peer);
3294                         }
3295
3296                         spin_unlock_irqrestore(lock, flags);
3297                         dropped_lock = 1;
3298
3299                         kiblnd_destroy_conn(conn, !peer);
3300
3301                         spin_lock_irqsave(lock, flags);
3302                         if (!peer)
3303                                 continue;
3304
3305                         conn->ibc_peer = peer;
3306                         if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
3307                                 list_add_tail(&conn->ibc_list,
3308                                               &kiblnd_data.kib_reconn_list);
3309                         else
3310                                 list_add_tail(&conn->ibc_list,
3311                                               &kiblnd_data.kib_reconn_wait);
3312                 }
3313
3314                 if (!list_empty(&kiblnd_data.kib_connd_conns)) {
3315                         conn = list_entry(kiblnd_data.kib_connd_conns.next,
3316                                           kib_conn_t, ibc_list);
3317                         list_del(&conn->ibc_list);
3318
3319                         spin_unlock_irqrestore(lock, flags);
3320                         dropped_lock = 1;
3321
3322                         kiblnd_disconnect_conn(conn);
3323                         kiblnd_conn_decref(conn);
3324
3325                         spin_lock_irqsave(lock, flags);
3326                 }
3327
3328                 while (reconn < KIB_RECONN_BREAK) {
3329                         if (kiblnd_data.kib_reconn_sec !=
3330                             ktime_get_real_seconds()) {
3331                                 kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
3332                                 list_splice_init(&kiblnd_data.kib_reconn_wait,
3333                                                  &kiblnd_data.kib_reconn_list);
3334                         }
3335
3336                         if (list_empty(&kiblnd_data.kib_reconn_list))
3337                                 break;
3338
3339                         conn = list_entry(kiblnd_data.kib_reconn_list.next,
3340                                           kib_conn_t, ibc_list);
3341                         list_del(&conn->ibc_list);
3342
3343                         spin_unlock_irqrestore(lock, flags);
3344                         dropped_lock = 1;
3345
3346                         reconn += kiblnd_reconnect_peer(conn->ibc_peer);
3347                         kiblnd_peer_decref(conn->ibc_peer);
3348                         LIBCFS_FREE(conn, sizeof(*conn));
3349
3350                         spin_lock_irqsave(lock, flags);
3351                 }
3352
3353                 /* careful with the jiffy wrap... */
3354                 timeout = (int)(deadline - jiffies);
3355                 if (timeout <= 0) {
3356                         const int n = 4;
3357                         const int p = 1;
3358                         int chunk = kiblnd_data.kib_peer_hash_size;
3359
3360                         spin_unlock_irqrestore(lock, flags);
3361                         dropped_lock = 1;
3362
3363                         /*
3364                          * Time to check for RDMA timeouts on a few more
3365                          * peers: I do checks every 'p' seconds on a
3366                          * proportion of the peer table and I need to check
3367                          * every connection 'n' times within a timeout
3368                          * interval, to ensure I detect a timeout on any
3369                          * connection within (n+1)/n times the timeout
3370                          * interval.
3371                          */
3372                         if (*kiblnd_tunables.kib_timeout > n * p)
3373                                 chunk = (chunk * n * p) /
3374                                         *kiblnd_tunables.kib_timeout;
3375                         if (!chunk)
3376                                 chunk = 1;
3377
3378                         for (i = 0; i < chunk; i++) {
3379                                 kiblnd_check_conns(peer_index);
3380                                 peer_index = (peer_index + 1) %
3381                                              kiblnd_data.kib_peer_hash_size;
3382                         }
3383
3384                         deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
3385                         spin_lock_irqsave(lock, flags);
3386                 }
3387
3388                 if (dropped_lock)
3389                         continue;
3390
3391                 /* Nothing to do for 'timeout'  */
3392                 set_current_state(TASK_INTERRUPTIBLE);
3393                 add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3394                 spin_unlock_irqrestore(lock, flags);
3395
3396                 schedule_timeout(timeout);
3397
3398                 remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
3399                 spin_lock_irqsave(lock, flags);
3400         }
3401
3402         spin_unlock_irqrestore(lock, flags);
3403
3404         kiblnd_thread_fini();
3405         return 0;
3406 }
3407
3408 void
3409 kiblnd_qp_event(struct ib_event *event, void *arg)
3410 {
3411         kib_conn_t *conn = arg;
3412
3413         switch (event->event) {
3414         case IB_EVENT_COMM_EST:
3415                 CDEBUG(D_NET, "%s established\n",
3416                        libcfs_nid2str(conn->ibc_peer->ibp_nid));
3417                 return;
3418
3419         default:
3420                 CERROR("%s: Async QP event type %d\n",
3421                        libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3422                 return;
3423         }
3424 }
3425
3426 static void
3427 kiblnd_complete(struct ib_wc *wc)
3428 {
3429         switch (kiblnd_wreqid2type(wc->wr_id)) {
3430         default:
3431                 LBUG();
3432
3433         case IBLND_WID_RDMA:
3434                 /*
3435                  * We only get RDMA completion notification if it fails.  All
3436                  * subsequent work items, including the final SEND will fail
3437                  * too.  However we can't print out any more info about the
3438                  * failing RDMA because 'tx' might be back on the idle list or
3439                  * even reused already if we didn't manage to post all our work
3440                  * items
3441                  */
3442                 CNETERR("RDMA (tx: %p) failed: %d\n",
3443                         kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3444                 return;
3445
3446         case IBLND_WID_TX:
3447                 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status);
3448                 return;
3449
3450         case IBLND_WID_RX:
3451                 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc->wr_id), wc->status,
3452                                    wc->byte_len);
3453                 return;
3454         }
3455 }
3456
3457 void
3458 kiblnd_cq_completion(struct ib_cq *cq, void *arg)
3459 {
3460         /*
3461          * NB I'm not allowed to schedule this conn once its refcount has
3462          * reached 0.  Since fundamentally I'm racing with scheduler threads
3463          * consuming my CQ I could be called after all completions have
3464          * occurred.  But in this case, !ibc_nrx && !ibc_nsends_posted
3465          * and this CQ is about to be destroyed so I NOOP.
3466          */
3467         kib_conn_t *conn = arg;
3468         struct kib_sched_info *sched = conn->ibc_sched;
3469         unsigned long flags;
3470
3471         LASSERT(cq == conn->ibc_cq);
3472
3473         spin_lock_irqsave(&sched->ibs_lock, flags);
3474
3475         conn->ibc_ready = 1;
3476
3477         if (!conn->ibc_scheduled &&
3478             (conn->ibc_nrx > 0 ||
3479              conn->ibc_nsends_posted > 0)) {
3480                 kiblnd_conn_addref(conn); /* +1 ref for sched_conns */
3481                 conn->ibc_scheduled = 1;
3482                 list_add_tail(&conn->ibc_sched_list, &sched->ibs_conns);
3483
3484                 if (waitqueue_active(&sched->ibs_waitq))
3485                         wake_up(&sched->ibs_waitq);
3486         }
3487
3488         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3489 }
3490
3491 void
3492 kiblnd_cq_event(struct ib_event *event, void *arg)
3493 {
3494         kib_conn_t *conn = arg;
3495
3496         CERROR("%s: async CQ event type %d\n",
3497                libcfs_nid2str(conn->ibc_peer->ibp_nid), event->event);
3498 }
3499
3500 int
3501 kiblnd_scheduler(void *arg)
3502 {
3503         long id = (long)arg;
3504         struct kib_sched_info *sched;
3505         kib_conn_t *conn;
3506         wait_queue_t wait;
3507         unsigned long flags;
3508         struct ib_wc wc;
3509         int did_something;
3510         int busy_loops = 0;
3511         int rc;
3512
3513         cfs_block_allsigs();
3514
3515         init_waitqueue_entry(&wait, current);
3516
3517         sched = kiblnd_data.kib_scheds[KIB_THREAD_CPT(id)];
3518
3519         rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
3520         if (rc) {
3521                 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3522                       sched->ibs_cpt);
3523         }
3524
3525         spin_lock_irqsave(&sched->ibs_lock, flags);
3526
3527         while (!kiblnd_data.kib_shutdown) {
3528                 if (busy_loops++ >= IBLND_RESCHED) {
3529                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3530
3531                         cond_resched();
3532                         busy_loops = 0;
3533
3534                         spin_lock_irqsave(&sched->ibs_lock, flags);
3535                 }
3536
3537                 did_something = 0;
3538
3539                 if (!list_empty(&sched->ibs_conns)) {
3540                         conn = list_entry(sched->ibs_conns.next, kib_conn_t,
3541                                           ibc_sched_list);
3542                         /* take over kib_sched_conns' ref on conn... */
3543                         LASSERT(conn->ibc_scheduled);
3544                         list_del(&conn->ibc_sched_list);
3545                         conn->ibc_ready = 0;
3546
3547                         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3548
3549                         wc.wr_id = IBLND_WID_INVAL;
3550
3551                         rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3552                         if (!rc) {
3553                                 rc = ib_req_notify_cq(conn->ibc_cq,
3554                                                       IB_CQ_NEXT_COMP);
3555                                 if (rc < 0) {
3556                                         CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3557                                               libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
3558                                         kiblnd_close_conn(conn, -EIO);
3559                                         kiblnd_conn_decref(conn);
3560                                         spin_lock_irqsave(&sched->ibs_lock,
3561                                                           flags);
3562                                         continue;
3563                                 }
3564
3565                                 rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
3566                         }
3567
3568                         if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
3569                                 LCONSOLE_ERROR("ib_poll_cq (rc: %d) returned invalid wr_id, opcode %d, status: %d, vendor_err: %d, conn: %s status: %d\nplease upgrade firmware and OFED or contact vendor.\n",
3570                                                rc, wc.opcode, wc.status,
3571                                                wc.vendor_err,
3572                                                libcfs_nid2str(conn->ibc_peer->ibp_nid),
3573                                                conn->ibc_state);
3574                                 rc = -EINVAL;
3575                         }
3576
3577                         if (rc < 0) {
3578                                 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3579                                       libcfs_nid2str(conn->ibc_peer->ibp_nid),
3580                                       rc);
3581                                 kiblnd_close_conn(conn, -EIO);
3582                                 kiblnd_conn_decref(conn);
3583                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3584                                 continue;
3585                         }
3586
3587                         spin_lock_irqsave(&sched->ibs_lock, flags);
3588
3589                         if (rc || conn->ibc_ready) {
3590                                 /*
3591                                  * There may be another completion waiting; get
3592                                  * another scheduler to check while I handle
3593                                  * this one...
3594                                  */
3595                                 /* +1 ref for sched_conns */
3596                                 kiblnd_conn_addref(conn);
3597                                 list_add_tail(&conn->ibc_sched_list,
3598                                               &sched->ibs_conns);
3599                                 if (waitqueue_active(&sched->ibs_waitq))
3600                                         wake_up(&sched->ibs_waitq);
3601                         } else {
3602                                 conn->ibc_scheduled = 0;
3603                         }
3604
3605                         if (rc) {
3606                                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3607                                 kiblnd_complete(&wc);
3608
3609                                 spin_lock_irqsave(&sched->ibs_lock, flags);
3610                         }
3611
3612                         kiblnd_conn_decref(conn); /* ...drop my ref from above */
3613                         did_something = 1;
3614                 }
3615
3616                 if (did_something)
3617                         continue;
3618
3619                 set_current_state(TASK_INTERRUPTIBLE);
3620                 add_wait_queue_exclusive(&sched->ibs_waitq, &wait);
3621                 spin_unlock_irqrestore(&sched->ibs_lock, flags);
3622
3623                 schedule();
3624                 busy_loops = 0;
3625
3626                 remove_wait_queue(&sched->ibs_waitq, &wait);
3627                 spin_lock_irqsave(&sched->ibs_lock, flags);
3628         }
3629
3630         spin_unlock_irqrestore(&sched->ibs_lock, flags);
3631
3632         kiblnd_thread_fini();
3633         return 0;
3634 }
3635
3636 int
3637 kiblnd_failover_thread(void *arg)
3638 {
3639         rwlock_t *glock = &kiblnd_data.kib_global_lock;
3640         kib_dev_t *dev;
3641         wait_queue_t wait;
3642         unsigned long flags;
3643         int rc;
3644
3645         LASSERT(*kiblnd_tunables.kib_dev_failover);
3646
3647         cfs_block_allsigs();
3648
3649         init_waitqueue_entry(&wait, current);
3650         write_lock_irqsave(glock, flags);
3651
3652         while (!kiblnd_data.kib_shutdown) {
3653                 int do_failover = 0;
3654                 int long_sleep;
3655
3656                 list_for_each_entry(dev, &kiblnd_data.kib_failed_devs,
3657                                     ibd_fail_list) {
3658                         if (time_before(cfs_time_current(),
3659                                         dev->ibd_next_failover))
3660                                 continue;
3661                         do_failover = 1;
3662                         break;
3663                 }
3664
3665                 if (do_failover) {
3666                         list_del_init(&dev->ibd_fail_list);
3667                         dev->ibd_failover = 1;
3668                         write_unlock_irqrestore(glock, flags);
3669
3670                         rc = kiblnd_dev_failover(dev);
3671
3672                         write_lock_irqsave(glock, flags);
3673
3674                         LASSERT(dev->ibd_failover);
3675                         dev->ibd_failover = 0;
3676                         if (rc >= 0) { /* Device is OK or failover succeed */
3677                                 dev->ibd_next_failover = cfs_time_shift(3);
3678                                 continue;
3679                         }
3680
3681                         /* failed to failover, retry later */
3682                         dev->ibd_next_failover =
3683                                 cfs_time_shift(min(dev->ibd_failed_failover, 10));
3684                         if (kiblnd_dev_can_failover(dev)) {
3685                                 list_add_tail(&dev->ibd_fail_list,
3686                                               &kiblnd_data.kib_failed_devs);
3687                         }
3688
3689                         continue;
3690                 }
3691
3692                 /* long sleep if no more pending failover */
3693                 long_sleep = list_empty(&kiblnd_data.kib_failed_devs);
3694
3695                 set_current_state(TASK_INTERRUPTIBLE);
3696                 add_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3697                 write_unlock_irqrestore(glock, flags);
3698
3699                 rc = schedule_timeout(long_sleep ? cfs_time_seconds(10) :
3700                                                    cfs_time_seconds(1));
3701                 remove_wait_queue(&kiblnd_data.kib_failover_waitq, &wait);
3702                 write_lock_irqsave(glock, flags);
3703
3704                 if (!long_sleep || rc)
3705                         continue;
3706
3707                 /*
3708                  * have a long sleep, routine check all active devices,
3709                  * we need checking like this because if there is not active
3710                  * connection on the dev and no SEND from local, we may listen
3711                  * on wrong HCA for ever while there is a bonding failover
3712                  */
3713                 list_for_each_entry(dev, &kiblnd_data.kib_devs, ibd_list) {
3714                         if (kiblnd_dev_can_failover(dev)) {
3715                                 list_add_tail(&dev->ibd_fail_list,
3716                                               &kiblnd_data.kib_failed_devs);
3717                         }
3718                 }
3719         }
3720
3721         write_unlock_irqrestore(glock, flags);
3722
3723         kiblnd_thread_fini();
3724         return 0;
3725 }