drivers/staging/lustre/lnet/klnds/socklnd/socklnd_cb.c

   1 /*
   2  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2012, Intel Corporation.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Portals, http://www.sf.net/projects/sandiaportals/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  */
  23
  24 #include "socklnd.h"
  25
  26 ksock_tx_t *
  27 ksocknal_alloc_tx(int type, int size)
  28 {
  29         ksock_tx_t *tx = NULL;
  30
  31         if (type == KSOCK_MSG_NOOP) {
  32                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  33
  34                 /* searching for a noop tx in free list */
  35                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  36
  37                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  38                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  39                                             next, ksock_tx_t, tx_list);
  40                         LASSERT(tx->tx_desc_size == size);
  41                         list_del(&tx->tx_list);
  42                 }
  43
  44                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  45         }
  46
  47         if (!tx)
  48                 LIBCFS_ALLOC(tx, size);
  49
  50         if (!tx)
  51                 return NULL;
  52
  53         atomic_set(&tx->tx_refcount, 1);
  54         tx->tx_zc_aborted = 0;
  55         tx->tx_zc_capable = 0;
  56         tx->tx_zc_checked = 0;
  57         tx->tx_desc_size  = size;
  58
  59         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  60
  61         return tx;
  62 }
  63
  64 ksock_tx_t *
  65 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  66 {
  67         ksock_tx_t *tx;
  68
  69         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  70         if (!tx) {
  71                 CERROR("Can't allocate noop tx desc\n");
  72                 return NULL;
  73         }
  74
  75         tx->tx_conn    = NULL;
  76         tx->tx_lnetmsg = NULL;
  77         tx->tx_kiov    = NULL;
  78         tx->tx_nkiov   = 0;
  79         tx->tx_iov     = tx->tx_frags.virt.iov;
  80         tx->tx_niov    = 1;
  81         tx->tx_nonblk  = nonblk;
  82
  83         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  84         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  85
  86         return tx;
  87 }
  88
  89 void
  90 ksocknal_free_tx(ksock_tx_t *tx)
  91 {
  92         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  93
  94         if (!tx->tx_lnetmsg && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  95                 /* it's a noop tx */
  96                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  97
  98                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
  99
 100                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 101         } else {
 102                 LIBCFS_FREE(tx, tx->tx_desc_size);
 103         }
 104 }
 105
 106 static int
 107 ksocknal_send_iov(ksock_conn_t *conn, ksock_tx_t *tx)
 108 {
 109         struct kvec *iov = tx->tx_iov;
 110         int nob;
 111         int rc;
 112
 113         LASSERT(tx->tx_niov > 0);
 114
 115         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 116         rc = ksocknal_lib_send_iov(conn, tx);
 117
 118         if (rc <= 0)                        /* sent nothing? */
 119                 return rc;
 120
 121         nob = rc;
 122         LASSERT(nob <= tx->tx_resid);
 123         tx->tx_resid -= nob;
 124
 125         /* "consume" iov */
 126         do {
 127                 LASSERT(tx->tx_niov > 0);
 128
 129                 if (nob < (int) iov->iov_len) {
 130                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 131                         iov->iov_len -= nob;
 132                         return rc;
 133                 }
 134
 135                 nob -= iov->iov_len;
 136                 tx->tx_iov = ++iov;
 137                 tx->tx_niov--;
 138         } while (nob);
 139
 140         return rc;
 141 }
 142
 143 static int
 144 ksocknal_send_kiov(ksock_conn_t *conn, ksock_tx_t *tx)
 145 {
 146         lnet_kiov_t *kiov = tx->tx_kiov;
 147         int nob;
 148         int rc;
 149
 150         LASSERT(!tx->tx_niov);
 151         LASSERT(tx->tx_nkiov > 0);
 152
 153         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 154         rc = ksocknal_lib_send_kiov(conn, tx);
 155
 156         if (rc <= 0)                        /* sent nothing? */
 157                 return rc;
 158
 159         nob = rc;
 160         LASSERT(nob <= tx->tx_resid);
 161         tx->tx_resid -= nob;
 162
 163         /* "consume" kiov */
 164         do {
 165                 LASSERT(tx->tx_nkiov > 0);
 166
 167                 if (nob < (int)kiov->kiov_len) {
 168                         kiov->kiov_offset += nob;
 169                         kiov->kiov_len -= nob;
 170                         return rc;
 171                 }
 172
 173                 nob -= (int)kiov->kiov_len;
 174                 tx->tx_kiov = ++kiov;
 175                 tx->tx_nkiov--;
 176         } while (nob);
 177
 178         return rc;
 179 }
 180
 181 static int
 182 ksocknal_transmit(ksock_conn_t *conn, ksock_tx_t *tx)
 183 {
 184         int rc;
 185         int bufnob;
 186
 187         if (ksocknal_data.ksnd_stall_tx) {
 188                 set_current_state(TASK_UNINTERRUPTIBLE);
 189                 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 190         }
 191
 192         LASSERT(tx->tx_resid);
 193
 194         rc = ksocknal_connsock_addref(conn);
 195         if (rc) {
 196                 LASSERT(conn->ksnc_closing);
 197                 return -ESHUTDOWN;
 198         }
 199
 200         do {
 201                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 202                         /* testing... */
 203                         ksocknal_data.ksnd_enomem_tx--;
 204                         rc = -EAGAIN;
 205                 } else if (tx->tx_niov) {
 206                         rc = ksocknal_send_iov(conn, tx);
 207                 } else {
 208                         rc = ksocknal_send_kiov(conn, tx);
 209                 }
 210
 211                 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 212                 if (rc > 0)                  /* sent something? */
 213                         conn->ksnc_tx_bufnob += rc; /* account it */
 214
 215                 if (bufnob < conn->ksnc_tx_bufnob) {
 216                         /*
 217                          * allocated send buffer bytes < computed; infer
 218                          * something got ACKed
 219                          */
 220                         conn->ksnc_tx_deadline =
 221                                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 222                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 223                         conn->ksnc_tx_bufnob = bufnob;
 224                         mb();
 225                 }
 226
 227                 if (rc <= 0) { /* Didn't write anything? */
 228
 229                         if (!rc) /* some stacks return 0 instead of -EAGAIN */
 230                                 rc = -EAGAIN;
 231
 232                         /* Check if EAGAIN is due to memory pressure */
 233                         if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 234                                 rc = -ENOMEM;
 235
 236                         break;
 237                 }
 238
 239                 /* socket's wmem_queued now includes 'rc' bytes */
 240                 atomic_sub(rc, &conn->ksnc_tx_nob);
 241                 rc = 0;
 242
 243         } while (tx->tx_resid);
 244
 245         ksocknal_connsock_decref(conn);
 246         return rc;
 247 }
 248
 249 static int
 250 ksocknal_recv_iov(ksock_conn_t *conn)
 251 {
 252         struct kvec *iov = conn->ksnc_rx_iov;
 253         int nob;
 254         int rc;
 255
 256         LASSERT(conn->ksnc_rx_niov > 0);
 257
 258         /*
 259          * Never touch conn->ksnc_rx_iov or change connection
 260          * status inside ksocknal_lib_recv_iov
 261          */
 262         rc = ksocknal_lib_recv_iov(conn);
 263
 264         if (rc <= 0)
 265                 return rc;
 266
 267         /* received something... */
 268         nob = rc;
 269
 270         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 271         conn->ksnc_rx_deadline =
 272                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 273         mb();                  /* order with setting rx_started */
 274         conn->ksnc_rx_started = 1;
 275
 276         conn->ksnc_rx_nob_wanted -= nob;
 277         conn->ksnc_rx_nob_left -= nob;
 278
 279         do {
 280                 LASSERT(conn->ksnc_rx_niov > 0);
 281
 282                 if (nob < (int)iov->iov_len) {
 283                         iov->iov_len -= nob;
 284                         iov->iov_base += nob;
 285                         return -EAGAIN;
 286                 }
 287
 288                 nob -= iov->iov_len;
 289                 conn->ksnc_rx_iov = ++iov;
 290                 conn->ksnc_rx_niov--;
 291         } while (nob);
 292
 293         return rc;
 294 }
 295
 296 static int
 297 ksocknal_recv_kiov(ksock_conn_t *conn)
 298 {
 299         lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
 300         int nob;
 301         int rc;
 302
 303         LASSERT(conn->ksnc_rx_nkiov > 0);
 304
 305         /*
 306          * Never touch conn->ksnc_rx_kiov or change connection
 307          * status inside ksocknal_lib_recv_iov
 308          */
 309         rc = ksocknal_lib_recv_kiov(conn);
 310
 311         if (rc <= 0)
 312                 return rc;
 313
 314         /* received something... */
 315         nob = rc;
 316
 317         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 318         conn->ksnc_rx_deadline =
 319                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 320         mb();                  /* order with setting rx_started */
 321         conn->ksnc_rx_started = 1;
 322
 323         conn->ksnc_rx_nob_wanted -= nob;
 324         conn->ksnc_rx_nob_left -= nob;
 325
 326         do {
 327                 LASSERT(conn->ksnc_rx_nkiov > 0);
 328
 329                 if (nob < (int) kiov->kiov_len) {
 330                         kiov->kiov_offset += nob;
 331                         kiov->kiov_len -= nob;
 332                         return -EAGAIN;
 333                 }
 334
 335                 nob -= kiov->kiov_len;
 336                 conn->ksnc_rx_kiov = ++kiov;
 337                 conn->ksnc_rx_nkiov--;
 338         } while (nob);
 339
 340         return 1;
 341 }
 342
 343 static int
 344 ksocknal_receive(ksock_conn_t *conn)
 345 {
 346         /*
 347          * Return 1 on success, 0 on EOF, < 0 on error.
 348          * Caller checks ksnc_rx_nob_wanted to determine
 349          * progress/completion.
 350          */
 351         int rc;
 352
 353         if (ksocknal_data.ksnd_stall_rx) {
 354                 set_current_state(TASK_UNINTERRUPTIBLE);
 355                 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
 356         }
 357
 358         rc = ksocknal_connsock_addref(conn);
 359         if (rc) {
 360                 LASSERT(conn->ksnc_closing);
 361                 return -ESHUTDOWN;
 362         }
 363
 364         for (;;) {
 365                 if (conn->ksnc_rx_niov)
 366                         rc = ksocknal_recv_iov(conn);
 367                 else
 368                         rc = ksocknal_recv_kiov(conn);
 369
 370                 if (rc <= 0) {
 371                         /* error/EOF or partial receive */
 372                         if (rc == -EAGAIN) {
 373                                 rc = 1;
 374                         } else if (!rc && conn->ksnc_rx_started) {
 375                                 /* EOF in the middle of a message */
 376                                 rc = -EPROTO;
 377                         }
 378                         break;
 379                 }
 380
 381                 /* Completed a fragment */
 382
 383                 if (!conn->ksnc_rx_nob_wanted) {
 384                         rc = 1;
 385                         break;
 386                 }
 387         }
 388
 389         ksocknal_connsock_decref(conn);
 390         return rc;
 391 }
 392
 393 void
 394 ksocknal_tx_done(lnet_ni_t *ni, ksock_tx_t *tx)
 395 {
 396         lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
 397         int rc = (!tx->tx_resid && !tx->tx_zc_aborted) ? 0 : -EIO;
 398
 399         LASSERT(ni || tx->tx_conn);
 400
 401         if (tx->tx_conn)
 402                 ksocknal_conn_decref(tx->tx_conn);
 403
 404         if (!ni && tx->tx_conn)
 405                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
 406
 407         ksocknal_free_tx(tx);
 408         if (lnetmsg) /* KSOCK_MSG_NOOP go without lnetmsg */
 409                 lnet_finalize(ni, lnetmsg, rc);
 410 }
 411
 412 void
 413 ksocknal_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int error)
 414 {
 415         ksock_tx_t *tx;
 416
 417         while (!list_empty(txlist)) {
 418                 tx = list_entry(txlist->next, ksock_tx_t, tx_list);
 419
 420                 if (error && tx->tx_lnetmsg) {
 421                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 422                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
 423                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
 424                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 425                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 426                 } else if (error) {
 427                         CNETERR("Deleting noop packet\n");
 428                 }
 429
 430                 list_del(&tx->tx_list);
 431
 432                 LASSERT(atomic_read(&tx->tx_refcount) == 1);
 433                 ksocknal_tx_done(ni, tx);
 434         }
 435 }
 436
 437 static void
 438 ksocknal_check_zc_req(ksock_tx_t *tx)
 439 {
 440         ksock_conn_t *conn = tx->tx_conn;
 441         ksock_peer_t *peer = conn->ksnc_peer;
 442
 443         /*
 444          * Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 445          * to ksnp_zc_req_list if some fragment of this message should be sent
 446          * zero-copy.  Our peer will send an ACK containing this cookie when
 447          * she has received this message to tell us we can signal completion.
 448          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 449          * ksnp_zc_req_list.
 450          */
 451         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 452         LASSERT(tx->tx_zc_capable);
 453
 454         tx->tx_zc_checked = 1;
 455
 456         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 457             !conn->ksnc_zc_capable)
 458                 return;
 459
 460         /*
 461          * assign cookie and queue tx to pending list, it will be released when
 462          * a matching ack is received. See ksocknal_handle_zcack()
 463          */
 464         ksocknal_tx_addref(tx);
 465
 466         spin_lock(&peer->ksnp_lock);
 467
 468         /* ZC_REQ is going to be pinned to the peer */
 469         tx->tx_deadline =
 470                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 471
 472         LASSERT(!tx->tx_msg.ksm_zc_cookies[0]);
 473
 474         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 475
 476         if (!peer->ksnp_zc_next_cookie)
 477                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 478
 479         list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 480
 481         spin_unlock(&peer->ksnp_lock);
 482 }
 483
 484 static void
 485 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 486 {
 487         ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
 488
 489         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 490         LASSERT(tx->tx_zc_capable);
 491
 492         tx->tx_zc_checked = 0;
 493
 494         spin_lock(&peer->ksnp_lock);
 495
 496         if (!tx->tx_msg.ksm_zc_cookies[0]) {
 497                 /* Not waiting for an ACK */
 498                 spin_unlock(&peer->ksnp_lock);
 499                 return;
 500         }
 501
 502         tx->tx_msg.ksm_zc_cookies[0] = 0;
 503         list_del(&tx->tx_zc_list);
 504
 505         spin_unlock(&peer->ksnp_lock);
 506
 507         ksocknal_tx_decref(tx);
 508 }
 509
 510 static int
 511 ksocknal_process_transmit(ksock_conn_t *conn, ksock_tx_t *tx)
 512 {
 513         int rc;
 514
 515         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 516                 ksocknal_check_zc_req(tx);
 517
 518         rc = ksocknal_transmit(conn, tx);
 519
 520         CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 521
 522         if (!tx->tx_resid) {
 523                 /* Sent everything OK */
 524                 LASSERT(!rc);
 525
 526                 return 0;
 527         }
 528
 529         if (rc == -EAGAIN)
 530                 return rc;
 531
 532         if (rc == -ENOMEM) {
 533                 static int counter;
 534
 535                 counter++;   /* exponential backoff warnings */
 536                 if ((counter & (-counter)) == counter)
 537                         CWARN("%u ENOMEM tx %p\n", counter, conn);
 538
 539                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 540                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 541
 542                 /* enomem list takes over scheduler's ref... */
 543                 LASSERT(conn->ksnc_tx_scheduled);
 544                 list_add_tail(&conn->ksnc_tx_list,
 545                               &ksocknal_data.ksnd_enomem_conns);
 546                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 547                                                    SOCKNAL_ENOMEM_RETRY),
 548                                    ksocknal_data.ksnd_reaper_waketime))
 549                         wake_up(&ksocknal_data.ksnd_reaper_waitq);
 550
 551                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 552                 return rc;
 553         }
 554
 555         /* Actual error */
 556         LASSERT(rc < 0);
 557
 558         if (!conn->ksnc_closing) {
 559                 switch (rc) {
 560                 case -ECONNRESET:
 561                         LCONSOLE_WARN("Host %pI4h reset our connection while we were sending data; it may have rebooted.\n",
 562                                       &conn->ksnc_ipaddr);
 563                         break;
 564                 default:
 565                         LCONSOLE_WARN("There was an unexpected network error while writing to %pI4h: %d.\n",
 566                                       &conn->ksnc_ipaddr, rc);
 567                         break;
 568                 }
 569                 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
 570                        conn, rc,
 571                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 572                        &conn->ksnc_ipaddr,
 573                        conn->ksnc_port);
 574         }
 575
 576         if (tx->tx_zc_checked)
 577                 ksocknal_uncheck_zc_req(tx);
 578
 579         /* it's not an error if conn is being closed */
 580         ksocknal_close_conn_and_siblings(conn, (conn->ksnc_closing) ? 0 : rc);
 581
 582         return rc;
 583 }
 584
 585 static void
 586 ksocknal_launch_connection_locked(ksock_route_t *route)
 587 {
 588         /* called holding write lock on ksnd_global_lock */
 589
 590         LASSERT(!route->ksnr_scheduled);
 591         LASSERT(!route->ksnr_connecting);
 592         LASSERT(ksocknal_route_mask() & ~route->ksnr_connected);
 593
 594         route->ksnr_scheduled = 1;            /* scheduling conn for connd */
 595         ksocknal_route_addref(route);      /* extra ref for connd */
 596
 597         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 598
 599         list_add_tail(&route->ksnr_connd_list,
 600                       &ksocknal_data.ksnd_connd_routes);
 601         wake_up(&ksocknal_data.ksnd_connd_waitq);
 602
 603         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 604 }
 605
 606 void
 607 ksocknal_launch_all_connections_locked(ksock_peer_t *peer)
 608 {
 609         ksock_route_t *route;
 610
 611         /* called holding write lock on ksnd_global_lock */
 612         for (;;) {
 613                 /* launch any/all connections that need it */
 614                 route = ksocknal_find_connectable_route_locked(peer);
 615                 if (!route)
 616                         return;
 617
 618                 ksocknal_launch_connection_locked(route);
 619         }
 620 }
 621
 622 ksock_conn_t *
 623 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 624 {
 625         struct list_head *tmp;
 626         ksock_conn_t *conn;
 627         ksock_conn_t *typed = NULL;
 628         ksock_conn_t *fallback = NULL;
 629         int tnob = 0;
 630         int fnob = 0;
 631
 632         list_for_each(tmp, &peer->ksnp_conns) {
 633                 ksock_conn_t *c  = list_entry(tmp, ksock_conn_t, ksnc_list);
 634                 int nob = atomic_read(&c->ksnc_tx_nob) +
 635                         c->ksnc_sock->sk->sk_wmem_queued;
 636                 int rc;
 637
 638                 LASSERT(!c->ksnc_closing);
 639                 LASSERT(c->ksnc_proto &&
 640                         c->ksnc_proto->pro_match_tx);
 641
 642                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 643
 644                 switch (rc) {
 645                 default:
 646                         LBUG();
 647                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 648                         continue;
 649
 650                 case SOCKNAL_MATCH_YES: /* typed connection */
 651                         if (!typed || tnob > nob ||
 652                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 653                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 654                                 typed = c;
 655                                 tnob  = nob;
 656                         }
 657                         break;
 658
 659                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 660                         if (!fallback || fnob > nob ||
 661                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 662                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 663                                 fallback = c;
 664                                 fnob = nob;
 665                         }
 666                         break;
 667                 }
 668         }
 669
 670         /* prefer the typed selection */
 671         conn = (typed) ? typed : fallback;
 672
 673         if (conn)
 674                 conn->ksnc_tx_last_post = cfs_time_current();
 675
 676         return conn;
 677 }
 678
 679 void
 680 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 681 {
 682         conn->ksnc_proto->pro_pack(tx);
 683
 684         atomic_add(tx->tx_nob, &conn->ksnc_tx_nob);
 685         ksocknal_conn_addref(conn); /* +1 ref for tx */
 686         tx->tx_conn = conn;
 687 }
 688
 689 void
 690 ksocknal_queue_tx_locked(ksock_tx_t *tx, ksock_conn_t *conn)
 691 {
 692         ksock_sched_t *sched = conn->ksnc_scheduler;
 693         ksock_msg_t *msg = &tx->tx_msg;
 694         ksock_tx_t *ztx = NULL;
 695         int bufnob = 0;
 696
 697         /*
 698          * called holding global lock (read or irq-write) and caller may
 699          * not have dropped this lock between finding conn and calling me,
 700          * so we don't need the {get,put}connsock dance to deref
 701          * ksnc_sock...
 702          */
 703         LASSERT(!conn->ksnc_closing);
 704
 705         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 706                libcfs_id2str(conn->ksnc_peer->ksnp_id),
 707                &conn->ksnc_ipaddr, conn->ksnc_port);
 708
 709         ksocknal_tx_prep(conn, tx);
 710
 711         /*
 712          * Ensure the frags we've been given EXACTLY match the number of
 713          * bytes we want to send.  Many TCP/IP stacks disregard any total
 714          * size parameters passed to them and just look at the frags.
 715          *
 716          * We always expect at least 1 mapped fragment containing the
 717          * complete ksocknal message header.
 718          */
 719         LASSERT(lnet_iov_nob(tx->tx_niov, tx->tx_iov) +
 720                 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 721                 (unsigned int)tx->tx_nob);
 722         LASSERT(tx->tx_niov >= 1);
 723         LASSERT(tx->tx_resid == tx->tx_nob);
 724
 725         CDEBUG(D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 726                tx, (tx->tx_lnetmsg) ? tx->tx_lnetmsg->msg_hdr.type :
 727                                               KSOCK_MSG_NOOP,
 728                tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 729
 730         /*
 731          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 732          * but they're used inside spinlocks a lot.
 733          */
 734         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 735         spin_lock_bh(&sched->kss_lock);
 736
 737         if (list_empty(&conn->ksnc_tx_queue) && !bufnob) {
 738                 /* First packet starts the timeout */
 739                 conn->ksnc_tx_deadline =
 740                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 741                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 742                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 743                 conn->ksnc_tx_bufnob = 0;
 744                 mb(); /* order with adding to tx_queue */
 745         }
 746
 747         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 748                 /*
 749                  * The packet is noop ZC ACK, try to piggyback the ack_cookie
 750                  * on a normal packet so I don't need to send it
 751                  */
 752                 LASSERT(msg->ksm_zc_cookies[1]);
 753                 LASSERT(conn->ksnc_proto->pro_queue_tx_zcack);
 754
 755                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 756                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 757
 758         } else {
 759                 /*
 760                  * It's a normal packet - can it piggback a noop zc-ack that
 761                  * has been queued already?
 762                  */
 763                 LASSERT(!msg->ksm_zc_cookies[1]);
 764                 LASSERT(conn->ksnc_proto->pro_queue_tx_msg);
 765
 766                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 767                 /* ztx will be released later */
 768         }
 769
 770         if (ztx) {
 771                 atomic_sub(ztx->tx_nob, &conn->ksnc_tx_nob);
 772                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 773         }
 774
 775         if (conn->ksnc_tx_ready &&      /* able to send */
 776             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 777                 /* +1 ref for scheduler */
 778                 ksocknal_conn_addref(conn);
 779                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
 780                 conn->ksnc_tx_scheduled = 1;
 781                 wake_up(&sched->kss_waitq);
 782         }
 783
 784         spin_unlock_bh(&sched->kss_lock);
 785 }
 786
 787 ksock_route_t *
 788 ksocknal_find_connectable_route_locked(ksock_peer_t *peer)
 789 {
 790         unsigned long now = cfs_time_current();
 791         struct list_head *tmp;
 792         ksock_route_t *route;
 793
 794         list_for_each(tmp, &peer->ksnp_routes) {
 795                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 796
 797                 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
 798
 799                 if (route->ksnr_scheduled)      /* connections being established */
 800                         continue;
 801
 802                 /* all route types connected ? */
 803                 if (!(ksocknal_route_mask() & ~route->ksnr_connected))
 804                         continue;
 805
 806                 if (!(!route->ksnr_retry_interval || /* first attempt */
 807                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 808                         CDEBUG(D_NET,
 809                                "Too soon to retry route %pI4h (cnted %d, interval %ld, %ld secs later)\n",
 810                                &route->ksnr_ipaddr,
 811                                route->ksnr_connected,
 812                                route->ksnr_retry_interval,
 813                                cfs_duration_sec(route->ksnr_timeout - now));
 814                         continue;
 815                 }
 816
 817                 return route;
 818         }
 819
 820         return NULL;
 821 }
 822
 823 ksock_route_t *
 824 ksocknal_find_connecting_route_locked(ksock_peer_t *peer)
 825 {
 826         struct list_head *tmp;
 827         ksock_route_t *route;
 828
 829         list_for_each(tmp, &peer->ksnp_routes) {
 830                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 831
 832                 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
 833
 834                 if (route->ksnr_scheduled)
 835                         return route;
 836         }
 837
 838         return NULL;
 839 }
 840
 841 int
 842 ksocknal_launch_packet(lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 843 {
 844         ksock_peer_t *peer;
 845         ksock_conn_t *conn;
 846         rwlock_t *g_lock;
 847         int retry;
 848         int rc;
 849
 850         LASSERT(!tx->tx_conn);
 851
 852         g_lock = &ksocknal_data.ksnd_global_lock;
 853
 854         for (retry = 0;; retry = 1) {
 855                 read_lock(g_lock);
 856                 peer = ksocknal_find_peer_locked(ni, id);
 857                 if (peer) {
 858                         if (!ksocknal_find_connectable_route_locked(peer)) {
 859                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 860                                 if (conn) {
 861                                         /*
 862                                          * I've got no routes that need to be
 863                                          * connecting and I do have an actual
 864                                          * connection...
 865                                          */
 866                                         ksocknal_queue_tx_locked(tx, conn);
 867                                         read_unlock(g_lock);
 868                                         return 0;
 869                                 }
 870                         }
 871                 }
 872
 873                 /* I'll need a write lock... */
 874                 read_unlock(g_lock);
 875
 876                 write_lock_bh(g_lock);
 877
 878                 peer = ksocknal_find_peer_locked(ni, id);
 879                 if (peer)
 880                         break;
 881
 882                 write_unlock_bh(g_lock);
 883
 884                 if (id.pid & LNET_PID_USERFLAG) {
 885                         CERROR("Refusing to create a connection to userspace process %s\n",
 886                                libcfs_id2str(id));
 887                         return -EHOSTUNREACH;
 888                 }
 889
 890                 if (retry) {
 891                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 892                         return -EHOSTUNREACH;
 893                 }
 894
 895                 rc = ksocknal_add_peer(ni, id,
 896                                        LNET_NIDADDR(id.nid),
 897                                        lnet_acceptor_port());
 898                 if (rc) {
 899                         CERROR("Can't add peer %s: %d\n",
 900                                libcfs_id2str(id), rc);
 901                         return rc;
 902                 }
 903         }
 904
 905         ksocknal_launch_all_connections_locked(peer);
 906
 907         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 908         if (conn) {
 909                 /* Connection exists; queue message on it */
 910                 ksocknal_queue_tx_locked(tx, conn);
 911                 write_unlock_bh(g_lock);
 912                 return 0;
 913         }
 914
 915         if (peer->ksnp_accepting > 0 ||
 916             ksocknal_find_connecting_route_locked(peer)) {
 917                 /* the message is going to be pinned to the peer */
 918                 tx->tx_deadline =
 919                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 920
 921                 /* Queue the message until a connection is established */
 922                 list_add_tail(&tx->tx_list, &peer->ksnp_tx_queue);
 923                 write_unlock_bh(g_lock);
 924                 return 0;
 925         }
 926
 927         write_unlock_bh(g_lock);
 928
 929         /* NB Routes may be ignored if connections to them failed recently */
 930         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 931         return -EHOSTUNREACH;
 932 }
 933
 934 int
 935 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 936 {
 937         int mpflag = 1;
 938         int type = lntmsg->msg_type;
 939         lnet_process_id_t target = lntmsg->msg_target;
 940         unsigned int payload_niov = lntmsg->msg_niov;
 941         struct kvec *payload_iov = lntmsg->msg_iov;
 942         lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
 943         unsigned int payload_offset = lntmsg->msg_offset;
 944         unsigned int payload_nob = lntmsg->msg_len;
 945         ksock_tx_t *tx;
 946         int desc_size;
 947         int rc;
 948
 949         /*
 950          * NB 'private' is different depending on what we're sending.
 951          * Just ignore it...
 952          */
 953         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 954                payload_nob, payload_niov, libcfs_id2str(target));
 955
 956         LASSERT(!payload_nob || payload_niov > 0);
 957         LASSERT(payload_niov <= LNET_MAX_IOV);
 958         /* payload is either all vaddrs or all pages */
 959         LASSERT(!(payload_kiov && payload_iov));
 960         LASSERT(!in_interrupt());
 961
 962         if (payload_iov)
 963                 desc_size = offsetof(ksock_tx_t,
 964                                      tx_frags.virt.iov[1 + payload_niov]);
 965         else
 966                 desc_size = offsetof(ksock_tx_t,
 967                                      tx_frags.paged.kiov[payload_niov]);
 968
 969         if (lntmsg->msg_vmflush)
 970                 mpflag = cfs_memory_pressure_get_and_set();
 971         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 972         if (!tx) {
 973                 CERROR("Can't allocate tx desc type %d size %d\n",
 974                        type, desc_size);
 975                 if (lntmsg->msg_vmflush)
 976                         cfs_memory_pressure_restore(mpflag);
 977                 return -ENOMEM;
 978         }
 979
 980         tx->tx_conn = NULL;                  /* set when assigned a conn */
 981         tx->tx_lnetmsg = lntmsg;
 982
 983         if (payload_iov) {
 984                 tx->tx_kiov = NULL;
 985                 tx->tx_nkiov = 0;
 986                 tx->tx_iov = tx->tx_frags.virt.iov;
 987                 tx->tx_niov = 1 +
 988                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 989                                                payload_niov, payload_iov,
 990                                                payload_offset, payload_nob);
 991         } else {
 992                 tx->tx_niov = 1;
 993                 tx->tx_iov = &tx->tx_frags.paged.iov;
 994                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 995                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 996                                                  payload_niov, payload_kiov,
 997                                                  payload_offset, payload_nob);
 998
 999                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1000                         tx->tx_zc_capable = 1;
1001         }
1002
1003         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
1004
1005         /* The first fragment will be set later in pro_pack */
1006         rc = ksocknal_launch_packet(ni, tx, target);
1007         if (!mpflag)
1008                 cfs_memory_pressure_restore(mpflag);
1009
1010         if (!rc)
1011                 return 0;
1012
1013         ksocknal_free_tx(tx);
1014         return -EIO;
1015 }
1016
1017 int
1018 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1019 {
1020         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1021
1022         if (IS_ERR(task))
1023                 return PTR_ERR(task);
1024
1025         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1026         ksocknal_data.ksnd_nthreads++;
1027         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1028         return 0;
1029 }
1030
1031 void
1032 ksocknal_thread_fini(void)
1033 {
1034         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1035         ksocknal_data.ksnd_nthreads--;
1036         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1037 }
1038
1039 int
1040 ksocknal_new_packet(ksock_conn_t *conn, int nob_to_skip)
1041 {
1042         static char ksocknal_slop_buffer[4096];
1043
1044         int nob;
1045         unsigned int niov;
1046         int skipped;
1047
1048         LASSERT(conn->ksnc_proto);
1049
1050         if (*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) {
1051                 /* Remind the socket to ack eagerly... */
1052                 ksocknal_lib_eager_ack(conn);
1053         }
1054
1055         if (!nob_to_skip) {      /* right at next packet boundary now */
1056                 conn->ksnc_rx_started = 0;
1057                 mb();                  /* racing with timeout thread */
1058
1059                 switch (conn->ksnc_proto->pro_version) {
1060                 case  KSOCK_PROTO_V2:
1061                 case  KSOCK_PROTO_V3:
1062                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1063                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1064                         conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg;
1065
1066                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1067                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1068                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1069                         break;
1070
1071                 case KSOCK_PROTO_V1:
1072                         /* Receiving bare lnet_hdr_t */
1073                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1074                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1075                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1076
1077                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1078                         conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1079                         conn->ksnc_rx_iov[0].iov_len = sizeof(lnet_hdr_t);
1080                         break;
1081
1082                 default:
1083                         LBUG();
1084                 }
1085                 conn->ksnc_rx_niov = 1;
1086
1087                 conn->ksnc_rx_kiov = NULL;
1088                 conn->ksnc_rx_nkiov = 0;
1089                 conn->ksnc_rx_csum = ~0;
1090                 return 1;
1091         }
1092
1093         /*
1094          * Set up to skip as much as possible now.  If there's more left
1095          * (ran out of iov entries) we'll get called again
1096          */
1097         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1098         conn->ksnc_rx_nob_left = nob_to_skip;
1099         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1100         skipped = 0;
1101         niov = 0;
1102
1103         do {
1104                 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1105
1106                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1107                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1108                 niov++;
1109                 skipped += nob;
1110                 nob_to_skip -= nob;
1111
1112         } while (nob_to_skip &&    /* mustn't overflow conn's rx iov */
1113                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct iovec));
1114
1115         conn->ksnc_rx_niov = niov;
1116         conn->ksnc_rx_kiov = NULL;
1117         conn->ksnc_rx_nkiov = 0;
1118         conn->ksnc_rx_nob_wanted = skipped;
1119         return 0;
1120 }
1121
1122 static int
1123 ksocknal_process_receive(ksock_conn_t *conn)
1124 {
1125         lnet_hdr_t *lhdr;
1126         lnet_process_id_t *id;
1127         int rc;
1128
1129         LASSERT(atomic_read(&conn->ksnc_conn_refcount) > 0);
1130
1131         /* NB: sched lock NOT held */
1132         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1133         LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1134                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1135                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1136                 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1137  again:
1138         if (conn->ksnc_rx_nob_wanted) {
1139                 rc = ksocknal_receive(conn);
1140
1141                 if (rc <= 0) {
1142                         LASSERT(rc != -EAGAIN);
1143
1144                         if (!rc)
1145                                 CDEBUG(D_NET, "[%p] EOF from %s ip %pI4h:%d\n",
1146                                        conn,
1147                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1148                                        &conn->ksnc_ipaddr,
1149                                        conn->ksnc_port);
1150                         else if (!conn->ksnc_closing)
1151                                 CERROR("[%p] Error %d on read from %s ip %pI4h:%d\n",
1152                                        conn, rc,
1153                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1154                                        &conn->ksnc_ipaddr,
1155                                        conn->ksnc_port);
1156
1157                         /* it's not an error if conn is being closed */
1158                         ksocknal_close_conn_and_siblings(conn,
1159                                                          (conn->ksnc_closing) ? 0 : rc);
1160                         return (!rc ? -ESHUTDOWN : rc);
1161                 }
1162
1163                 if (conn->ksnc_rx_nob_wanted) {
1164                         /* short read */
1165                         return -EAGAIN;
1166                 }
1167         }
1168         switch (conn->ksnc_rx_state) {
1169         case SOCKNAL_RX_KSM_HEADER:
1170                 if (conn->ksnc_flip) {
1171                         __swab32s(&conn->ksnc_msg.ksm_type);
1172                         __swab32s(&conn->ksnc_msg.ksm_csum);
1173                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1174                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1175                 }
1176
1177                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1178                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1179                         CERROR("%s: Unknown message type: %x\n",
1180                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1181                                conn->ksnc_msg.ksm_type);
1182                         ksocknal_new_packet(conn, 0);
1183                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1184                         return -EPROTO;
1185                 }
1186
1187                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1188                     conn->ksnc_msg.ksm_csum &&     /* has checksum */
1189                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1190                         /* NOOP Checksum error */
1191                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1192                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1193                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1194                         ksocknal_new_packet(conn, 0);
1195                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1196                         return -EIO;
1197                 }
1198
1199                 if (conn->ksnc_msg.ksm_zc_cookies[1]) {
1200                         __u64 cookie = 0;
1201
1202                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1203
1204                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1205                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1206
1207                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1208                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1209
1210                         if (rc) {
1211                                 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1212                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1213                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1214                                 ksocknal_new_packet(conn, 0);
1215                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1216                                 return rc;
1217                         }
1218                 }
1219
1220                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1221                         ksocknal_new_packet(conn, 0);
1222                         return 0;       /* NOOP is done and just return */
1223                 }
1224
1225                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1226                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1227                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1228
1229                 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1230                 conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1231                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1232
1233                 conn->ksnc_rx_niov = 1;
1234                 conn->ksnc_rx_kiov = NULL;
1235                 conn->ksnc_rx_nkiov = 0;
1236
1237                 goto again;     /* read lnet header now */
1238
1239         case SOCKNAL_RX_LNET_HEADER:
1240                 /* unpack message header */
1241                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1242
1243                 if (conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) {
1244                         /* Userspace peer */
1245                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1246                         id = &conn->ksnc_peer->ksnp_id;
1247
1248                         /* Substitute process ID assigned at connection time */
1249                         lhdr->src_pid = cpu_to_le32(id->pid);
1250                         lhdr->src_nid = cpu_to_le64(id->nid);
1251                 }
1252
1253                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1254                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1255
1256                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1257                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1258                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1259                 if (rc < 0) {
1260                         /* I just received garbage: give up on this conn */
1261                         ksocknal_new_packet(conn, 0);
1262                         ksocknal_close_conn_and_siblings(conn, rc);
1263                         ksocknal_conn_decref(conn);
1264                         return -EPROTO;
1265                 }
1266
1267                 /* I'm racing with ksocknal_recv() */
1268                 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1269                         conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1270
1271                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1272                         return 0;
1273
1274                 /* ksocknal_recv() got called */
1275                 goto again;
1276
1277         case SOCKNAL_RX_LNET_PAYLOAD:
1278                 /* payload all received */
1279                 rc = 0;
1280
1281                 if (!conn->ksnc_rx_nob_left &&   /* not truncating */
1282                     conn->ksnc_msg.ksm_csum &&  /* has checksum */
1283                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1284                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1285                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1286                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1287                         rc = -EIO;
1288                 }
1289
1290                 if (!rc && conn->ksnc_msg.ksm_zc_cookies[0]) {
1291                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1292
1293                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1294                         id = &conn->ksnc_peer->ksnp_id;
1295
1296                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1297                                         conn->ksnc_msg.ksm_zc_cookies[0],
1298                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1299                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1300                 }
1301
1302                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1303
1304                 if (rc) {
1305                         ksocknal_new_packet(conn, 0);
1306                         ksocknal_close_conn_and_siblings(conn, rc);
1307                         return -EPROTO;
1308                 }
1309                 /* Fall through */
1310
1311         case SOCKNAL_RX_SLOP:
1312                 /* starting new packet? */
1313                 if (ksocknal_new_packet(conn, conn->ksnc_rx_nob_left))
1314                         return 0;       /* come back later */
1315                 goto again;          /* try to finish reading slop now */
1316
1317         default:
1318                 break;
1319         }
1320
1321         /* Not Reached */
1322         LBUG();
1323         return -EINVAL;                /* keep gcc happy */
1324 }
1325
1326 int
1327 ksocknal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1328               unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1329               unsigned int offset, unsigned int mlen, unsigned int rlen)
1330 {
1331         ksock_conn_t *conn = private;
1332         ksock_sched_t *sched = conn->ksnc_scheduler;
1333
1334         LASSERT(mlen <= rlen);
1335         LASSERT(niov <= LNET_MAX_IOV);
1336
1337         conn->ksnc_cookie = msg;
1338         conn->ksnc_rx_nob_wanted = mlen;
1339         conn->ksnc_rx_nob_left = rlen;
1340
1341         if (!mlen || iov) {
1342                 conn->ksnc_rx_nkiov = 0;
1343                 conn->ksnc_rx_kiov = NULL;
1344                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1345                 conn->ksnc_rx_niov =
1346                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1347                                          niov, iov, offset, mlen);
1348         } else {
1349                 conn->ksnc_rx_niov = 0;
1350                 conn->ksnc_rx_iov = NULL;
1351                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1352                 conn->ksnc_rx_nkiov =
1353                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1354                                           niov, kiov, offset, mlen);
1355         }
1356
1357         LASSERT(mlen ==
1358                 lnet_iov_nob(conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1359                 lnet_kiov_nob(conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1360
1361         LASSERT(conn->ksnc_rx_scheduled);
1362
1363         spin_lock_bh(&sched->kss_lock);
1364
1365         switch (conn->ksnc_rx_state) {
1366         case SOCKNAL_RX_PARSE_WAIT:
1367                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1368                 wake_up(&sched->kss_waitq);
1369                 LASSERT(conn->ksnc_rx_ready);
1370                 break;
1371
1372         case SOCKNAL_RX_PARSE:
1373                 /* scheduler hasn't noticed I'm parsing yet */
1374                 break;
1375         }
1376
1377         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1378
1379         spin_unlock_bh(&sched->kss_lock);
1380         ksocknal_conn_decref(conn);
1381         return 0;
1382 }
1383
1384 static inline int
1385 ksocknal_sched_cansleep(ksock_sched_t *sched)
1386 {
1387         int rc;
1388
1389         spin_lock_bh(&sched->kss_lock);
1390
1391         rc = !ksocknal_data.ksnd_shuttingdown &&
1392               list_empty(&sched->kss_rx_conns) &&
1393               list_empty(&sched->kss_tx_conns);
1394
1395         spin_unlock_bh(&sched->kss_lock);
1396         return rc;
1397 }
1398
1399 int ksocknal_scheduler(void *arg)
1400 {
1401         struct ksock_sched_info *info;
1402         ksock_sched_t *sched;
1403         ksock_conn_t *conn;
1404         ksock_tx_t *tx;
1405         int rc;
1406         int nloops = 0;
1407         long id = (long)arg;
1408
1409         info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1410         sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1411
1412         cfs_block_allsigs();
1413
1414         rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1415         if (rc) {
1416                 CERROR("Can't set CPT affinity to %d: %d\n",
1417                        info->ksi_cpt, rc);
1418         }
1419
1420         spin_lock_bh(&sched->kss_lock);
1421
1422         while (!ksocknal_data.ksnd_shuttingdown) {
1423                 int did_something = 0;
1424
1425                 /* Ensure I progress everything semi-fairly */
1426
1427                 if (!list_empty(&sched->kss_rx_conns)) {
1428                         conn = list_entry(sched->kss_rx_conns.next,
1429                                           ksock_conn_t, ksnc_rx_list);
1430                         list_del(&conn->ksnc_rx_list);
1431
1432                         LASSERT(conn->ksnc_rx_scheduled);
1433                         LASSERT(conn->ksnc_rx_ready);
1434
1435                         /*
1436                          * clear rx_ready in case receive isn't complete.
1437                          * Do it BEFORE we call process_recv, since
1438                          * data_ready can set it any time after we release
1439                          * kss_lock.
1440                          */
1441                         conn->ksnc_rx_ready = 0;
1442                         spin_unlock_bh(&sched->kss_lock);
1443
1444                         rc = ksocknal_process_receive(conn);
1445
1446                         spin_lock_bh(&sched->kss_lock);
1447
1448                         /* I'm the only one that can clear this flag */
1449                         LASSERT(conn->ksnc_rx_scheduled);
1450
1451                         /* Did process_receive get everything it wanted? */
1452                         if (!rc)
1453                                 conn->ksnc_rx_ready = 1;
1454
1455                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1456                                 /*
1457                                  * Conn blocked waiting for ksocknal_recv()
1458                                  * I change its state (under lock) to signal
1459                                  * it can be rescheduled
1460                                  */
1461                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1462                         } else if (conn->ksnc_rx_ready) {
1463                                 /* reschedule for rx */
1464                                 list_add_tail(&conn->ksnc_rx_list,
1465                                               &sched->kss_rx_conns);
1466                         } else {
1467                                 conn->ksnc_rx_scheduled = 0;
1468                                 /* drop my ref */
1469                                 ksocknal_conn_decref(conn);
1470                         }
1471
1472                         did_something = 1;
1473                 }
1474
1475                 if (!list_empty(&sched->kss_tx_conns)) {
1476                         LIST_HEAD(zlist);
1477
1478                         if (!list_empty(&sched->kss_zombie_noop_txs)) {
1479                                 list_add(&zlist, &sched->kss_zombie_noop_txs);
1480                                 list_del_init(&sched->kss_zombie_noop_txs);
1481                         }
1482
1483                         conn = list_entry(sched->kss_tx_conns.next,
1484                                           ksock_conn_t, ksnc_tx_list);
1485                         list_del(&conn->ksnc_tx_list);
1486
1487                         LASSERT(conn->ksnc_tx_scheduled);
1488                         LASSERT(conn->ksnc_tx_ready);
1489                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1490
1491                         tx = list_entry(conn->ksnc_tx_queue.next,
1492                                         ksock_tx_t, tx_list);
1493
1494                         if (conn->ksnc_tx_carrier == tx)
1495                                 ksocknal_next_tx_carrier(conn);
1496
1497                         /* dequeue now so empty list => more to send */
1498                         list_del(&tx->tx_list);
1499
1500                         /*
1501                          * Clear tx_ready in case send isn't complete.  Do
1502                          * it BEFORE we call process_transmit, since
1503                          * write_space can set it any time after we release
1504                          * kss_lock.
1505                          */
1506                         conn->ksnc_tx_ready = 0;
1507                         spin_unlock_bh(&sched->kss_lock);
1508
1509                         if (!list_empty(&zlist)) {
1510                                 /*
1511                                  * free zombie noop txs, it's fast because
1512                                  * noop txs are just put in freelist
1513                                  */
1514                                 ksocknal_txlist_done(NULL, &zlist, 0);
1515                         }
1516
1517                         rc = ksocknal_process_transmit(conn, tx);
1518
1519                         if (rc == -ENOMEM || rc == -EAGAIN) {
1520                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1521                                 spin_lock_bh(&sched->kss_lock);
1522                                 list_add(&tx->tx_list, &conn->ksnc_tx_queue);
1523                         } else {
1524                                 /* Complete send; tx -ref */
1525                                 ksocknal_tx_decref(tx);
1526
1527                                 spin_lock_bh(&sched->kss_lock);
1528                                 /* assume space for more */
1529                                 conn->ksnc_tx_ready = 1;
1530                         }
1531
1532                         if (rc == -ENOMEM) {
1533                                 /*
1534                                  * Do nothing; after a short timeout, this
1535                                  * conn will be reposted on kss_tx_conns.
1536                                  */
1537                         } else if (conn->ksnc_tx_ready &&
1538                                    !list_empty(&conn->ksnc_tx_queue)) {
1539                                 /* reschedule for tx */
1540                                 list_add_tail(&conn->ksnc_tx_list,
1541                                               &sched->kss_tx_conns);
1542                         } else {
1543                                 conn->ksnc_tx_scheduled = 0;
1544                                 /* drop my ref */
1545                                 ksocknal_conn_decref(conn);
1546                         }
1547
1548                         did_something = 1;
1549                 }
1550                 if (!did_something ||      /* nothing to do */
1551                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1552                         spin_unlock_bh(&sched->kss_lock);
1553
1554                         nloops = 0;
1555
1556                         if (!did_something) {   /* wait for something to do */
1557                                 rc = wait_event_interruptible_exclusive(
1558                                         sched->kss_waitq,
1559                                         !ksocknal_sched_cansleep(sched));
1560                                 LASSERT(!rc);
1561                         } else {
1562                                 cond_resched();
1563                         }
1564
1565                         spin_lock_bh(&sched->kss_lock);
1566                 }
1567         }
1568
1569         spin_unlock_bh(&sched->kss_lock);
1570         ksocknal_thread_fini();
1571         return 0;
1572 }
1573
1574 /*
1575  * Add connection to kss_rx_conns of scheduler
1576  * and wakeup the scheduler.
1577  */
1578 void ksocknal_read_callback(ksock_conn_t *conn)
1579 {
1580         ksock_sched_t *sched;
1581
1582         sched = conn->ksnc_scheduler;
1583
1584         spin_lock_bh(&sched->kss_lock);
1585
1586         conn->ksnc_rx_ready = 1;
1587
1588         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1589                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1590                 conn->ksnc_rx_scheduled = 1;
1591                 /* extra ref for scheduler */
1592                 ksocknal_conn_addref(conn);
1593
1594                 wake_up(&sched->kss_waitq);
1595         }
1596         spin_unlock_bh(&sched->kss_lock);
1597 }
1598
1599 /*
1600  * Add connection to kss_tx_conns of scheduler
1601  * and wakeup the scheduler.
1602  */
1603 void ksocknal_write_callback(ksock_conn_t *conn)
1604 {
1605         ksock_sched_t *sched;
1606
1607         sched = conn->ksnc_scheduler;
1608
1609         spin_lock_bh(&sched->kss_lock);
1610
1611         conn->ksnc_tx_ready = 1;
1612
1613         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1614             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1615                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1616                 conn->ksnc_tx_scheduled = 1;
1617                 /* extra ref for scheduler */
1618                 ksocknal_conn_addref(conn);
1619
1620                 wake_up(&sched->kss_waitq);
1621         }
1622
1623         spin_unlock_bh(&sched->kss_lock);
1624 }
1625
1626 static ksock_proto_t *
1627 ksocknal_parse_proto_version(ksock_hello_msg_t *hello)
1628 {
1629         __u32 version = 0;
1630
1631         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1632                 version = hello->kshm_version;
1633         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1634                 version = __swab32(hello->kshm_version);
1635
1636         if (version) {
1637 #if SOCKNAL_VERSION_DEBUG
1638                 if (*ksocknal_tunables.ksnd_protocol == 1)
1639                         return NULL;
1640
1641                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1642                     version == KSOCK_PROTO_V3)
1643                         return NULL;
1644 #endif
1645                 if (version == KSOCK_PROTO_V2)
1646                         return &ksocknal_protocol_v2x;
1647
1648                 if (version == KSOCK_PROTO_V3)
1649                         return &ksocknal_protocol_v3x;
1650
1651                 return NULL;
1652         }
1653
1654         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1655                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1656
1657                 CLASSERT(sizeof(lnet_magicversion_t) ==
1658                          offsetof(ksock_hello_msg_t, kshm_src_nid));
1659
1660                 if (hmv->version_major == cpu_to_le16(KSOCK_PROTO_V1_MAJOR) &&
1661                     hmv->version_minor == cpu_to_le16(KSOCK_PROTO_V1_MINOR))
1662                         return &ksocknal_protocol_v1x;
1663         }
1664
1665         return NULL;
1666 }
1667
1668 int
1669 ksocknal_send_hello(lnet_ni_t *ni, ksock_conn_t *conn,
1670                     lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1671 {
1672         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1673         ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1674
1675         LASSERT(hello->kshm_nips <= LNET_MAX_INTERFACES);
1676
1677         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1678         LASSERT(conn->ksnc_proto);
1679
1680         hello->kshm_src_nid = ni->ni_nid;
1681         hello->kshm_dst_nid = peer_nid;
1682         hello->kshm_src_pid = the_lnet.ln_pid;
1683
1684         hello->kshm_src_incarnation = net->ksnn_incarnation;
1685         hello->kshm_ctype = conn->ksnc_type;
1686
1687         return conn->ksnc_proto->pro_send_hello(conn, hello);
1688 }
1689
1690 static int
1691 ksocknal_invert_type(int type)
1692 {
1693         switch (type) {
1694         case SOCKLND_CONN_ANY:
1695         case SOCKLND_CONN_CONTROL:
1696                 return type;
1697         case SOCKLND_CONN_BULK_IN:
1698                 return SOCKLND_CONN_BULK_OUT;
1699         case SOCKLND_CONN_BULK_OUT:
1700                 return SOCKLND_CONN_BULK_IN;
1701         default:
1702                 return SOCKLND_CONN_NONE;
1703         }
1704 }
1705
1706 int
1707 ksocknal_recv_hello(lnet_ni_t *ni, ksock_conn_t *conn,
1708                     ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1709                     __u64 *incarnation)
1710 {
1711         /* Return < 0   fatal error
1712          *      0         success
1713          *      EALREADY   lost connection race
1714          *      EPROTO     protocol version mismatch
1715          */
1716         struct socket *sock = conn->ksnc_sock;
1717         int active = !!conn->ksnc_proto;
1718         int timeout;
1719         int proto_match;
1720         int rc;
1721         ksock_proto_t *proto;
1722         lnet_process_id_t recv_id;
1723
1724         /* socket type set on active connections - not set on passive */
1725         LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1726
1727         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1728                             lnet_acceptor_timeout();
1729
1730         rc = lnet_sock_read(sock, &hello->kshm_magic, sizeof(hello->kshm_magic), timeout);
1731         if (rc) {
1732                 CERROR("Error %d reading HELLO from %pI4h\n",
1733                        rc, &conn->ksnc_ipaddr);
1734                 LASSERT(rc < 0);
1735                 return rc;
1736         }
1737
1738         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1739             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1740             hello->kshm_magic != le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1741                 /* Unexpected magic! */
1742                 CERROR("Bad magic(1) %#08x (%#08x expected) from %pI4h\n",
1743                        __cpu_to_le32(hello->kshm_magic),
1744                        LNET_PROTO_TCP_MAGIC,
1745                        &conn->ksnc_ipaddr);
1746                 return -EPROTO;
1747         }
1748
1749         rc = lnet_sock_read(sock, &hello->kshm_version,
1750                             sizeof(hello->kshm_version), timeout);
1751         if (rc) {
1752                 CERROR("Error %d reading HELLO from %pI4h\n",
1753                        rc, &conn->ksnc_ipaddr);
1754                 LASSERT(rc < 0);
1755                 return rc;
1756         }
1757
1758         proto = ksocknal_parse_proto_version(hello);
1759         if (!proto) {
1760                 if (!active) {
1761                         /* unknown protocol from peer, tell peer my protocol */
1762                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1763 #if SOCKNAL_VERSION_DEBUG
1764                         if (*ksocknal_tunables.ksnd_protocol == 2)
1765                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1766                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1767                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1768 #endif
1769                         hello->kshm_nips = 0;
1770                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1771                 }
1772
1773                 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1774                        conn->ksnc_proto->pro_version,
1775                        &conn->ksnc_ipaddr);
1776
1777                 return -EPROTO;
1778         }
1779
1780         proto_match = (conn->ksnc_proto == proto);
1781         conn->ksnc_proto = proto;
1782
1783         /* receive the rest of hello message anyway */
1784         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1785         if (rc) {
1786                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1787                        rc, &conn->ksnc_ipaddr);
1788                 LASSERT(rc < 0);
1789                 return rc;
1790         }
1791
1792         *incarnation = hello->kshm_src_incarnation;
1793
1794         if (hello->kshm_src_nid == LNET_NID_ANY) {
1795                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1796                        &conn->ksnc_ipaddr);
1797                 return -EPROTO;
1798         }
1799
1800         if (!active &&
1801             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1802                 /* Userspace NAL assigns peer process ID from socket */
1803                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1804                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1805         } else {
1806                 recv_id.nid = hello->kshm_src_nid;
1807                 recv_id.pid = hello->kshm_src_pid;
1808         }
1809
1810         if (!active) {
1811                 *peerid = recv_id;
1812
1813                 /* peer determines type */
1814                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1815                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1816                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1817                                hello->kshm_ctype, libcfs_id2str(*peerid),
1818                                &conn->ksnc_ipaddr);
1819                         return -EPROTO;
1820                 }
1821
1822                 return 0;
1823         }
1824
1825         if (peerid->pid != recv_id.pid ||
1826             peerid->nid != recv_id.nid) {
1827                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host %pI4h, but they claimed they were %s; please check your Lustre configuration.\n",
1828                                    libcfs_id2str(*peerid),
1829                                    &conn->ksnc_ipaddr,
1830                                    libcfs_id2str(recv_id));
1831                 return -EPROTO;
1832         }
1833
1834         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1835                 /* Possible protocol mismatch or I lost the connection race */
1836                 return proto_match ? EALREADY : EPROTO;
1837         }
1838
1839         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1840                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1841                        conn->ksnc_type, libcfs_id2str(*peerid),
1842                        &conn->ksnc_ipaddr, hello->kshm_ctype);
1843                 return -EPROTO;
1844         }
1845
1846         return 0;
1847 }
1848
1849 static int
1850 ksocknal_connect(ksock_route_t *route)
1851 {
1852         LIST_HEAD(zombies);
1853         ksock_peer_t *peer = route->ksnr_peer;
1854         int type;
1855         int wanted;
1856         struct socket *sock;
1857         unsigned long deadline;
1858         int retry_later = 0;
1859         int rc = 0;
1860
1861         deadline = cfs_time_add(cfs_time_current(),
1862                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1863
1864         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1865
1866         LASSERT(route->ksnr_scheduled);
1867         LASSERT(!route->ksnr_connecting);
1868
1869         route->ksnr_connecting = 1;
1870
1871         for (;;) {
1872                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1873
1874                 /*
1875                  * stop connecting if peer/route got closed under me, or
1876                  * route got connected while queued
1877                  */
1878                 if (peer->ksnp_closing || route->ksnr_deleted ||
1879                     !wanted) {
1880                         retry_later = 0;
1881                         break;
1882                 }
1883
1884                 /* reschedule if peer is connecting to me */
1885                 if (peer->ksnp_accepting > 0) {
1886                         CDEBUG(D_NET,
1887                                "peer %s(%d) already connecting to me, retry later.\n",
1888                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1889                         retry_later = 1;
1890                 }
1891
1892                 if (retry_later) /* needs reschedule */
1893                         break;
1894
1895                 if (wanted & (1 << SOCKLND_CONN_ANY)) {
1896                         type = SOCKLND_CONN_ANY;
1897                 } else if (wanted & (1 << SOCKLND_CONN_CONTROL)) {
1898                         type = SOCKLND_CONN_CONTROL;
1899                 } else if (wanted & (1 << SOCKLND_CONN_BULK_IN)) {
1900                         type = SOCKLND_CONN_BULK_IN;
1901                 } else {
1902                         LASSERT(wanted & (1 << SOCKLND_CONN_BULK_OUT));
1903                         type = SOCKLND_CONN_BULK_OUT;
1904                 }
1905
1906                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1907
1908                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1909                         rc = -ETIMEDOUT;
1910                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1911                                                    route->ksnr_ipaddr,
1912                                                    route->ksnr_port);
1913                         goto failed;
1914                 }
1915
1916                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1917                                   route->ksnr_myipaddr,
1918                                   route->ksnr_ipaddr, route->ksnr_port);
1919                 if (rc)
1920                         goto failed;
1921
1922                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1923                 if (rc < 0) {
1924                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1925                                                    route->ksnr_ipaddr,
1926                                                    route->ksnr_port);
1927                         goto failed;
1928                 }
1929
1930                 /*
1931                  * A +ve RC means I have to retry because I lost the connection
1932                  * race or I have to renegotiate protocol version
1933                  */
1934                 retry_later = (rc);
1935                 if (retry_later)
1936                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1937                                libcfs_nid2str(peer->ksnp_id.nid));
1938
1939                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1940         }
1941
1942         route->ksnr_scheduled = 0;
1943         route->ksnr_connecting = 0;
1944
1945         if (retry_later) {
1946                 /*
1947                  * re-queue for attention; this frees me up to handle
1948                  * the peer's incoming connection request
1949                  */
1950                 if (rc == EALREADY ||
1951                     (!rc && peer->ksnp_accepting > 0)) {
1952                         /*
1953                          * We want to introduce a delay before next
1954                          * attempt to connect if we lost conn race,
1955                          * but the race is resolved quickly usually,
1956                          * so min_reconnectms should be good heuristic
1957                          */
1958                         route->ksnr_retry_interval =
1959                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms) / 1000;
1960                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1961                                                            route->ksnr_retry_interval);
1962                 }
1963
1964                 ksocknal_launch_connection_locked(route);
1965         }
1966
1967         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1968         return retry_later;
1969
1970  failed:
1971         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1972
1973         route->ksnr_scheduled = 0;
1974         route->ksnr_connecting = 0;
1975
1976         /* This is a retry rather than a new connection */
1977         route->ksnr_retry_interval *= 2;
1978         route->ksnr_retry_interval =
1979                 max(route->ksnr_retry_interval,
1980                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms) / 1000);
1981         route->ksnr_retry_interval =
1982                 min(route->ksnr_retry_interval,
1983                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms) / 1000);
1984
1985         LASSERT(route->ksnr_retry_interval);
1986         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1987                                            route->ksnr_retry_interval);
1988
1989         if (!list_empty(&peer->ksnp_tx_queue) &&
1990             !peer->ksnp_accepting &&
1991             !ksocknal_find_connecting_route_locked(peer)) {
1992                 ksock_conn_t *conn;
1993
1994                 /*
1995                  * ksnp_tx_queue is queued on a conn on successful
1996                  * connection for V1.x and V2.x
1997                  */
1998                 if (!list_empty(&peer->ksnp_conns)) {
1999                         conn = list_entry(peer->ksnp_conns.next,
2000                                           ksock_conn_t, ksnc_list);
2001                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
2002                 }
2003
2004                 /*
2005                  * take all the blocked packets while I've got the lock and
2006                  * complete below...
2007                  */
2008                 list_splice_init(&peer->ksnp_tx_queue, &zombies);
2009         }
2010
2011 #if 0      /* irrelevant with only eager routes */
2012         if (!route->ksnr_deleted) {
2013                 /* make this route least-favourite for re-selection */
2014                 list_del(&route->ksnr_list);
2015                 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2016         }
2017 #endif
2018         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2019
2020         ksocknal_peer_failed(peer);
2021         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2022         return 0;
2023 }
2024
2025 /*
2026  * check whether we need to create more connds.
2027  * It will try to create new thread if it's necessary, @timeout can
2028  * be updated if failed to create, so caller wouldn't keep try while
2029  * running out of resource.
2030  */
2031 static int
2032 ksocknal_connd_check_start(time64_t sec, long *timeout)
2033 {
2034         char name[16];
2035         int rc;
2036         int total = ksocknal_data.ksnd_connd_starting +
2037                     ksocknal_data.ksnd_connd_running;
2038
2039         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2040                 /* still in initializing */
2041                 return 0;
2042         }
2043
2044         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2045             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2046                 /*
2047                  * can't create more connd, or still have enough
2048                  * threads to handle more connecting
2049                  */
2050                 return 0;
2051         }
2052
2053         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2054                 /* no pending connecting request */
2055                 return 0;
2056         }
2057
2058         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2059                 /* may run out of resource, retry later */
2060                 *timeout = cfs_time_seconds(1);
2061                 return 0;
2062         }
2063
2064         if (ksocknal_data.ksnd_connd_starting > 0) {
2065                 /* serialize starting to avoid flood */
2066                 return 0;
2067         }
2068
2069         ksocknal_data.ksnd_connd_starting_stamp = sec;
2070         ksocknal_data.ksnd_connd_starting++;
2071         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2072
2073         /* NB: total is the next id */
2074         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2075         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2076
2077         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2078         if (!rc)
2079                 return 1;
2080
2081         /* we tried ... */
2082         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2083         ksocknal_data.ksnd_connd_starting--;
2084         ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2085
2086         return 1;
2087 }
2088
2089 /*
2090  * check whether current thread can exit, it will return 1 if there are too
2091  * many threads and no creating in past 120 seconds.
2092  * Also, this function may update @timeout to make caller come back
2093  * again to recheck these conditions.
2094  */
2095 static int
2096 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2097 {
2098         int val;
2099
2100         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2101                 /* still in initializing */
2102                 return 0;
2103         }
2104
2105         if (ksocknal_data.ksnd_connd_starting > 0) {
2106                 /* in progress of starting new thread */
2107                 return 0;
2108         }
2109
2110         if (ksocknal_data.ksnd_connd_running <=
2111             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2112                 return 0;
2113         }
2114
2115         /* created thread in past 120 seconds? */
2116         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2117                     SOCKNAL_CONND_TIMEOUT - sec);
2118
2119         *timeout = (val > 0) ? cfs_time_seconds(val) :
2120                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2121         if (val > 0)
2122                 return 0;
2123
2124         /* no creating in past 120 seconds */
2125
2126         return ksocknal_data.ksnd_connd_running >
2127                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2128 }
2129
2130 /*
2131  * Go through connd_routes queue looking for a route that we can process
2132  * right now, @timeout_p can be updated if we need to come back later
2133  */
2134 static ksock_route_t *
2135 ksocknal_connd_get_route_locked(signed long *timeout_p)
2136 {
2137         ksock_route_t *route;
2138         unsigned long now;
2139
2140         now = cfs_time_current();
2141
2142         /* connd_routes can contain both pending and ordinary routes */
2143         list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2144                             ksnr_connd_list) {
2145                 if (!route->ksnr_retry_interval ||
2146                     cfs_time_aftereq(now, route->ksnr_timeout))
2147                         return route;
2148
2149                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2150                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2151                         *timeout_p = (int)(route->ksnr_timeout - now);
2152         }
2153
2154         return NULL;
2155 }
2156
2157 int
2158 ksocknal_connd(void *arg)
2159 {
2160         spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2161         ksock_connreq_t *cr;
2162         wait_queue_t wait;
2163         int nloops = 0;
2164         int cons_retry = 0;
2165
2166         cfs_block_allsigs();
2167
2168         init_waitqueue_entry(&wait, current);
2169
2170         spin_lock_bh(connd_lock);
2171
2172         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2173         ksocknal_data.ksnd_connd_starting--;
2174         ksocknal_data.ksnd_connd_running++;
2175
2176         while (!ksocknal_data.ksnd_shuttingdown) {
2177                 ksock_route_t *route = NULL;
2178                 time64_t sec = ktime_get_real_seconds();
2179                 long timeout = MAX_SCHEDULE_TIMEOUT;
2180                 int dropped_lock = 0;
2181
2182                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2183                         /* wakeup another one to check stop */
2184                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2185                         break;
2186                 }
2187
2188                 if (ksocknal_connd_check_start(sec, &timeout)) {
2189                         /* created new thread */
2190                         dropped_lock = 1;
2191                 }
2192
2193                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2194                         /* Connection accepted by the listener */
2195                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2196                                             next, ksock_connreq_t, ksncr_list);
2197
2198                         list_del(&cr->ksncr_list);
2199                         spin_unlock_bh(connd_lock);
2200                         dropped_lock = 1;
2201
2202                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2203                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2204                         lnet_ni_decref(cr->ksncr_ni);
2205                         LIBCFS_FREE(cr, sizeof(*cr));
2206
2207                         spin_lock_bh(connd_lock);
2208                 }
2209
2210                 /*
2211                  * Only handle an outgoing connection request if there
2212                  * is a thread left to handle incoming connections and
2213                  * create new connd
2214                  */
2215                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2216                     ksocknal_data.ksnd_connd_running) {
2217                         route = ksocknal_connd_get_route_locked(&timeout);
2218                 }
2219                 if (route) {
2220                         list_del(&route->ksnr_connd_list);
2221                         ksocknal_data.ksnd_connd_connecting++;
2222                         spin_unlock_bh(connd_lock);
2223                         dropped_lock = 1;
2224
2225                         if (ksocknal_connect(route)) {
2226                                 /* consecutive retry */
2227                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2228                                         CWARN("massive consecutive re-connecting to %pI4h\n",
2229                                               &route->ksnr_ipaddr);
2230                                         cons_retry = 0;
2231                                 }
2232                         } else {
2233                                 cons_retry = 0;
2234                         }
2235
2236                         ksocknal_route_decref(route);
2237
2238                         spin_lock_bh(connd_lock);
2239                         ksocknal_data.ksnd_connd_connecting--;
2240                 }
2241
2242                 if (dropped_lock) {
2243                         if (++nloops < SOCKNAL_RESCHED)
2244                                 continue;
2245                         spin_unlock_bh(connd_lock);
2246                         nloops = 0;
2247                         cond_resched();
2248                         spin_lock_bh(connd_lock);
2249                         continue;
2250                 }
2251
2252                 /* Nothing to do for 'timeout'  */
2253                 set_current_state(TASK_INTERRUPTIBLE);
2254                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2255                 spin_unlock_bh(connd_lock);
2256
2257                 nloops = 0;
2258                 schedule_timeout(timeout);
2259
2260                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2261                 spin_lock_bh(connd_lock);
2262         }
2263         ksocknal_data.ksnd_connd_running--;
2264         spin_unlock_bh(connd_lock);
2265
2266         ksocknal_thread_fini();
2267         return 0;
2268 }
2269
2270 static ksock_conn_t *
2271 ksocknal_find_timed_out_conn(ksock_peer_t *peer)
2272 {
2273         /* We're called with a shared lock on ksnd_global_lock */
2274         ksock_conn_t *conn;
2275         struct list_head *ctmp;
2276
2277         list_for_each(ctmp, &peer->ksnp_conns) {
2278                 int error;
2279
2280                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
2281
2282                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2283                 LASSERT(!conn->ksnc_closing);
2284
2285                 /*
2286                  * SOCK_ERROR will reset error code of socket in
2287                  * some platform (like Darwin8.x)
2288                  */
2289                 error = conn->ksnc_sock->sk->sk_err;
2290                 if (error) {
2291                         ksocknal_conn_addref(conn);
2292
2293                         switch (error) {
2294                         case ECONNRESET:
2295                                 CNETERR("A connection with %s (%pI4h:%d) was reset; it may have rebooted.\n",
2296                                         libcfs_id2str(peer->ksnp_id),
2297                                         &conn->ksnc_ipaddr,
2298                                         conn->ksnc_port);
2299                                 break;
2300                         case ETIMEDOUT:
2301                                 CNETERR("A connection with %s (%pI4h:%d) timed out; the network or node may be down.\n",
2302                                         libcfs_id2str(peer->ksnp_id),
2303                                         &conn->ksnc_ipaddr,
2304                                         conn->ksnc_port);
2305                                 break;
2306                         default:
2307                                 CNETERR("An unexpected network error %d occurred with %s (%pI4h:%d\n",
2308                                         error,
2309                                         libcfs_id2str(peer->ksnp_id),
2310                                         &conn->ksnc_ipaddr,
2311                                         conn->ksnc_port);
2312                                 break;
2313                         }
2314
2315                         return conn;
2316                 }
2317
2318                 if (conn->ksnc_rx_started &&
2319                     cfs_time_aftereq(cfs_time_current(),
2320                                      conn->ksnc_rx_deadline)) {
2321                         /* Timed out incomplete incoming message */
2322                         ksocknal_conn_addref(conn);
2323                         CNETERR("Timeout receiving from %s (%pI4h:%d), state %d wanted %d left %d\n",
2324                                 libcfs_id2str(peer->ksnp_id),
2325                                 &conn->ksnc_ipaddr,
2326                                 conn->ksnc_port,
2327                                 conn->ksnc_rx_state,
2328                                 conn->ksnc_rx_nob_wanted,
2329                                 conn->ksnc_rx_nob_left);
2330                         return conn;
2331                 }
2332
2333                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2334                      conn->ksnc_sock->sk->sk_wmem_queued) &&
2335                     cfs_time_aftereq(cfs_time_current(),
2336                                      conn->ksnc_tx_deadline)) {
2337                         /*
2338                          * Timed out messages queued for sending or
2339                          * buffered in the socket's send buffer
2340                          */
2341                         ksocknal_conn_addref(conn);
2342                         CNETERR("Timeout sending data to %s (%pI4h:%d) the network or that node may be down.\n",
2343                                 libcfs_id2str(peer->ksnp_id),
2344                                 &conn->ksnc_ipaddr,
2345                                 conn->ksnc_port);
2346                         return conn;
2347                 }
2348         }
2349
2350         return NULL;
2351 }
2352
2353 static inline void
2354 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2355 {
2356         ksock_tx_t *tx;
2357         LIST_HEAD(stale_txs);
2358
2359         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2360
2361         while (!list_empty(&peer->ksnp_tx_queue)) {
2362                 tx = list_entry(peer->ksnp_tx_queue.next, ksock_tx_t, tx_list);
2363
2364                 if (!cfs_time_aftereq(cfs_time_current(),
2365                                       tx->tx_deadline))
2366                         break;
2367
2368                 list_del(&tx->tx_list);
2369                 list_add_tail(&tx->tx_list, &stale_txs);
2370         }
2371
2372         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2373
2374         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2375 }
2376
2377 static int
2378 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2379         __must_hold(&ksocknal_data.ksnd_global_lock)
2380 {
2381         ksock_sched_t *sched;
2382         ksock_conn_t *conn;
2383         ksock_tx_t *tx;
2384
2385         if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2386                 return 0;
2387
2388         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2389                 return 0;
2390
2391         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2392             time_before(cfs_time_current(),
2393                         cfs_time_add(peer->ksnp_last_alive,
2394                                      cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2395                 return 0;
2396
2397         if (time_before(cfs_time_current(), peer->ksnp_send_keepalive))
2398                 return 0;
2399
2400         /*
2401          * retry 10 secs later, so we wouldn't put pressure
2402          * on this peer if we failed to send keepalive this time
2403          */
2404         peer->ksnp_send_keepalive = cfs_time_shift(10);
2405
2406         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2407         if (conn) {
2408                 sched = conn->ksnc_scheduler;
2409
2410                 spin_lock_bh(&sched->kss_lock);
2411                 if (!list_empty(&conn->ksnc_tx_queue)) {
2412                         spin_unlock_bh(&sched->kss_lock);
2413                         /* there is an queued ACK, don't need keepalive */
2414                         return 0;
2415                 }
2416
2417                 spin_unlock_bh(&sched->kss_lock);
2418         }
2419
2420         read_unlock(&ksocknal_data.ksnd_global_lock);
2421
2422         /* cookie = 1 is reserved for keepalive PING */
2423         tx = ksocknal_alloc_tx_noop(1, 1);
2424         if (!tx) {
2425                 read_lock(&ksocknal_data.ksnd_global_lock);
2426                 return -ENOMEM;
2427         }
2428
2429         if (!ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id)) {
2430                 read_lock(&ksocknal_data.ksnd_global_lock);
2431                 return 1;
2432         }
2433
2434         ksocknal_free_tx(tx);
2435         read_lock(&ksocknal_data.ksnd_global_lock);
2436
2437         return -EIO;
2438 }
2439
2440 static void
2441 ksocknal_check_peer_timeouts(int idx)
2442 {
2443         struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2444         ksock_peer_t *peer;
2445         ksock_conn_t *conn;
2446         ksock_tx_t *tx;
2447
2448  again:
2449         /*
2450          * NB. We expect to have a look at all the peers and not find any
2451          * connections to time out, so we just use a shared lock while we
2452          * take a look...
2453          */
2454         read_lock(&ksocknal_data.ksnd_global_lock);
2455
2456         list_for_each_entry(peer, peers, ksnp_list) {
2457                 unsigned long deadline = 0;
2458                 int resid = 0;
2459                 int n = 0;
2460
2461                 if (ksocknal_send_keepalive_locked(peer)) {
2462                         read_unlock(&ksocknal_data.ksnd_global_lock);
2463                         goto again;
2464                 }
2465
2466                 conn = ksocknal_find_timed_out_conn(peer);
2467
2468                 if (conn) {
2469                         read_unlock(&ksocknal_data.ksnd_global_lock);
2470
2471                         ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2472
2473                         /*
2474                          * NB we won't find this one again, but we can't
2475                          * just proceed with the next peer, since we dropped
2476                          * ksnd_global_lock and it might be dead already!
2477                          */
2478                         ksocknal_conn_decref(conn);
2479                         goto again;
2480                 }
2481
2482                 /*
2483                  * we can't process stale txs right here because we're
2484                  * holding only shared lock
2485                  */
2486                 if (!list_empty(&peer->ksnp_tx_queue)) {
2487                         ksock_tx_t *tx = list_entry(peer->ksnp_tx_queue.next,
2488                                                     ksock_tx_t, tx_list);
2489
2490                         if (cfs_time_aftereq(cfs_time_current(),
2491                                              tx->tx_deadline)) {
2492                                 ksocknal_peer_addref(peer);
2493                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2494
2495                                 ksocknal_flush_stale_txs(peer);
2496
2497                                 ksocknal_peer_decref(peer);
2498                                 goto again;
2499                         }
2500                 }
2501
2502                 if (list_empty(&peer->ksnp_zc_req_list))
2503                         continue;
2504
2505                 spin_lock(&peer->ksnp_lock);
2506                 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2507                         if (!cfs_time_aftereq(cfs_time_current(),
2508                                               tx->tx_deadline))
2509                                 break;
2510                         /* ignore the TX if connection is being closed */
2511                         if (tx->tx_conn->ksnc_closing)
2512                                 continue;
2513                         n++;
2514                 }
2515
2516                 if (!n) {
2517                         spin_unlock(&peer->ksnp_lock);
2518                         continue;
2519                 }
2520
2521                 tx = list_entry(peer->ksnp_zc_req_list.next,
2522                                 ksock_tx_t, tx_zc_list);
2523                 deadline = tx->tx_deadline;
2524                 resid = tx->tx_resid;
2525                 conn = tx->tx_conn;
2526                 ksocknal_conn_addref(conn);
2527
2528                 spin_unlock(&peer->ksnp_lock);
2529                 read_unlock(&ksocknal_data.ksnd_global_lock);
2530
2531                 CERROR("Total %d stale ZC_REQs for peer %s detected; the oldest(%p) timed out %ld secs ago, resid: %d, wmem: %d\n",
2532                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2533                        cfs_duration_sec(cfs_time_current() - deadline),
2534                        resid, conn->ksnc_sock->sk->sk_wmem_queued);
2535
2536                 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2537                 ksocknal_conn_decref(conn);
2538                 goto again;
2539         }
2540
2541         read_unlock(&ksocknal_data.ksnd_global_lock);
2542 }
2543
2544 int
2545 ksocknal_reaper(void *arg)
2546 {
2547         wait_queue_t wait;
2548         ksock_conn_t *conn;
2549         ksock_sched_t *sched;
2550         struct list_head enomem_conns;
2551         int nenomem_conns;
2552         long timeout;
2553         int i;
2554         int peer_index = 0;
2555         unsigned long deadline = cfs_time_current();
2556
2557         cfs_block_allsigs();
2558
2559         INIT_LIST_HEAD(&enomem_conns);
2560         init_waitqueue_entry(&wait, current);
2561
2562         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2563
2564         while (!ksocknal_data.ksnd_shuttingdown) {
2565                 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2566                         conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2567                                           ksock_conn_t, ksnc_list);
2568                         list_del(&conn->ksnc_list);
2569
2570                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2571
2572                         ksocknal_terminate_conn(conn);
2573                         ksocknal_conn_decref(conn);
2574
2575                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2576                         continue;
2577                 }
2578
2579                 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2580                         conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2581                                           ksock_conn_t, ksnc_list);
2582                         list_del(&conn->ksnc_list);
2583
2584                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2585
2586                         ksocknal_destroy_conn(conn);
2587
2588                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2589                         continue;
2590                 }
2591
2592                 if (!list_empty(&ksocknal_data.ksnd_enomem_conns)) {
2593                         list_add(&enomem_conns,
2594                                  &ksocknal_data.ksnd_enomem_conns);
2595                         list_del_init(&ksocknal_data.ksnd_enomem_conns);
2596                 }
2597
2598                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2599
2600                 /* reschedule all the connections that stalled with ENOMEM... */
2601                 nenomem_conns = 0;
2602                 while (!list_empty(&enomem_conns)) {
2603                         conn = list_entry(enomem_conns.next, ksock_conn_t,
2604                                           ksnc_tx_list);
2605                         list_del(&conn->ksnc_tx_list);
2606
2607                         sched = conn->ksnc_scheduler;
2608
2609                         spin_lock_bh(&sched->kss_lock);
2610
2611                         LASSERT(conn->ksnc_tx_scheduled);
2612                         conn->ksnc_tx_ready = 1;
2613                         list_add_tail(&conn->ksnc_tx_list,
2614                                       &sched->kss_tx_conns);
2615                         wake_up(&sched->kss_waitq);
2616
2617                         spin_unlock_bh(&sched->kss_lock);
2618                         nenomem_conns++;
2619                 }
2620
2621                 /* careful with the jiffy wrap... */
2622                 while ((timeout = cfs_time_sub(deadline,
2623                                                cfs_time_current())) <= 0) {
2624                         const int n = 4;
2625                         const int p = 1;
2626                         int chunk = ksocknal_data.ksnd_peer_hash_size;
2627
2628                         /*
2629                          * Time to check for timeouts on a few more peers: I do
2630                          * checks every 'p' seconds on a proportion of the peer
2631                          * table and I need to check every connection 'n' times
2632                          * within a timeout interval, to ensure I detect a
2633                          * timeout on any connection within (n+1)/n times the
2634                          * timeout interval.
2635                          */
2636                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2637                                 chunk = (chunk * n * p) /
2638                                         *ksocknal_tunables.ksnd_timeout;
2639                         if (!chunk)
2640                                 chunk = 1;
2641
2642                         for (i = 0; i < chunk; i++) {
2643                                 ksocknal_check_peer_timeouts(peer_index);
2644                                 peer_index = (peer_index + 1) %
2645                                              ksocknal_data.ksnd_peer_hash_size;
2646                         }
2647
2648                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2649                 }
2650
2651                 if (nenomem_conns) {
2652                         /*
2653                          * Reduce my timeout if I rescheduled ENOMEM conns.
2654                          * This also prevents me getting woken immediately
2655                          * if any go back on my enomem list.
2656                          */
2657                         timeout = SOCKNAL_ENOMEM_RETRY;
2658                 }
2659                 ksocknal_data.ksnd_reaper_waketime =
2660                         cfs_time_add(cfs_time_current(), timeout);
2661
2662                 set_current_state(TASK_INTERRUPTIBLE);
2663                 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2664
2665                 if (!ksocknal_data.ksnd_shuttingdown &&
2666                     list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2667                     list_empty(&ksocknal_data.ksnd_zombie_conns))
2668                         schedule_timeout(timeout);
2669
2670                 set_current_state(TASK_RUNNING);
2671                 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2672
2673                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2674         }
2675
2676         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2677
2678         ksocknal_thread_fini();
2679         return 0;
2680 }