lnet/klnds/socklnd/socklnd_cb.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   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  *   You should have received a copy of the GNU General Public License
  23  *   along with Portals; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include "socklnd.h"
  28
  29 ksock_tx_t *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         ksock_tx_t *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT (size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         cfs_list_del(&tx->tx_list);
  45                 }
  46
  47                 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         cfs_atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_capable = 0;
  58         tx->tx_zc_checked = 0;
  59         tx->tx_desc_size  = size;
  60
  61         cfs_atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  62
  63         return tx;
  64 }
  65
  66 ksock_tx_t *
  67 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  68 {
  69         ksock_tx_t *tx;
  70
  71         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  72         if (tx == NULL) {
  73                 CERROR("Can't allocate noop tx desc\n");
  74                 return NULL;
  75         }
  76
  77         tx->tx_conn     = NULL;
  78         tx->tx_lnetmsg  = NULL;
  79         tx->tx_kiov     = NULL;
  80         tx->tx_nkiov    = 0;
  81         tx->tx_iov      = tx->tx_frags.virt.iov;
  82         tx->tx_niov     = 1;
  83         tx->tx_nonblk   = nonblk;
  84
  85         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  86         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  87
  88         return tx;
  89 }
  90
  91
  92 void
  93 ksocknal_free_tx (ksock_tx_t *tx)
  94 {
  95         cfs_atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  96
  97         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  98                 /* it's a noop tx */
  99                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
 100
 101                 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 102
 103                 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
 104         } else {
 105                 LIBCFS_FREE(tx, tx->tx_desc_size);
 106         }
 107 }
 108
 109 int
 110 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
 111 {
 112         struct iovec  *iov = tx->tx_iov;
 113         int    nob;
 114         int    rc;
 115
 116         LASSERT (tx->tx_niov > 0);
 117
 118         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 119         rc = ksocknal_lib_send_iov(conn, tx);
 120
 121         if (rc <= 0)                            /* sent nothing? */
 122                 return (rc);
 123
 124         nob = rc;
 125         LASSERT (nob <= tx->tx_resid);
 126         tx->tx_resid -= nob;
 127
 128         /* "consume" iov */
 129         do {
 130                 LASSERT (tx->tx_niov > 0);
 131
 132                 if (nob < (int) iov->iov_len) {
 133                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 134                         iov->iov_len -= nob;
 135                         return (rc);
 136                 }
 137
 138                 nob -= iov->iov_len;
 139                 tx->tx_iov = ++iov;
 140                 tx->tx_niov--;
 141         } while (nob != 0);
 142
 143         return (rc);
 144 }
 145
 146 int
 147 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
 148 {
 149         lnet_kiov_t    *kiov = tx->tx_kiov;
 150         int     nob;
 151         int     rc;
 152
 153         LASSERT (tx->tx_niov == 0);
 154         LASSERT (tx->tx_nkiov > 0);
 155
 156         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 157         rc = ksocknal_lib_send_kiov(conn, tx);
 158
 159         if (rc <= 0)                            /* sent nothing? */
 160                 return (rc);
 161
 162         nob = rc;
 163         LASSERT (nob <= tx->tx_resid);
 164         tx->tx_resid -= nob;
 165
 166         /* "consume" kiov */
 167         do {
 168                 LASSERT(tx->tx_nkiov > 0);
 169
 170                 if (nob < (int)kiov->kiov_len) {
 171                         kiov->kiov_offset += nob;
 172                         kiov->kiov_len -= nob;
 173                         return rc;
 174                 }
 175
 176                 nob -= (int)kiov->kiov_len;
 177                 tx->tx_kiov = ++kiov;
 178                 tx->tx_nkiov--;
 179         } while (nob != 0);
 180
 181         return (rc);
 182 }
 183
 184 int
 185 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 186 {
 187         int      rc;
 188         int      bufnob;
 189
 190         if (ksocknal_data.ksnd_stall_tx != 0) {
 191                 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 192         }
 193
 194         LASSERT (tx->tx_resid != 0);
 195
 196         rc = ksocknal_connsock_addref(conn);
 197         if (rc != 0) {
 198                 LASSERT (conn->ksnc_closing);
 199                 return (-ESHUTDOWN);
 200         }
 201
 202         do {
 203                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 204                         /* testing... */
 205                         ksocknal_data.ksnd_enomem_tx--;
 206                         rc = -EAGAIN;
 207                 } else if (tx->tx_niov != 0) {
 208                         rc = ksocknal_send_iov (conn, tx);
 209                 } else {
 210                         rc = ksocknal_send_kiov (conn, tx);
 211                 }
 212
 213                 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 214                 if (rc > 0)                     /* sent something? */
 215                         conn->ksnc_tx_bufnob += rc; /* account it */
 216
 217                 if (bufnob < conn->ksnc_tx_bufnob) {
 218                         /* allocated send buffer bytes < computed; infer
 219                          * something got ACKed */
 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                         cfs_mb();
 225                 }
 226
 227                 if (rc <= 0) { /* Didn't write anything? */
 228
 229                         if (rc == 0) /* 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                 cfs_atomic_sub (rc, &conn->ksnc_tx_nob);
 241                 rc = 0;
 242
 243         } while (tx->tx_resid != 0);
 244
 245         ksocknal_connsock_decref(conn);
 246         return (rc);
 247 }
 248
 249 int
 250 ksocknal_recv_iov (ksock_conn_t *conn)
 251 {
 252         struct iovec *iov = conn->ksnc_rx_iov;
 253         int     nob;
 254         int     rc;
 255
 256         LASSERT (conn->ksnc_rx_niov > 0);
 257
 258         /* Never touch conn->ksnc_rx_iov or change connection
 259          * status inside ksocknal_lib_recv_iov */
 260         rc = ksocknal_lib_recv_iov(conn);
 261
 262         if (rc <= 0)
 263                 return (rc);
 264
 265         /* received something... */
 266         nob = rc;
 267
 268         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 269         conn->ksnc_rx_deadline =
 270                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 271         cfs_mb();                       /* order with setting rx_started */
 272         conn->ksnc_rx_started = 1;
 273
 274         conn->ksnc_rx_nob_wanted -= nob;
 275         conn->ksnc_rx_nob_left -= nob;
 276
 277         do {
 278                 LASSERT (conn->ksnc_rx_niov > 0);
 279
 280                 if (nob < (int)iov->iov_len) {
 281                         iov->iov_len -= nob;
 282                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 283                         return (-EAGAIN);
 284                 }
 285
 286                 nob -= iov->iov_len;
 287                 conn->ksnc_rx_iov = ++iov;
 288                 conn->ksnc_rx_niov--;
 289         } while (nob != 0);
 290
 291         return (rc);
 292 }
 293
 294 int
 295 ksocknal_recv_kiov (ksock_conn_t *conn)
 296 {
 297         lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
 298         int     nob;
 299         int     rc;
 300         LASSERT (conn->ksnc_rx_nkiov > 0);
 301
 302         /* Never touch conn->ksnc_rx_kiov or change connection
 303          * status inside ksocknal_lib_recv_iov */
 304         rc = ksocknal_lib_recv_kiov(conn);
 305
 306         if (rc <= 0)
 307                 return (rc);
 308
 309         /* received something... */
 310         nob = rc;
 311
 312         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 313         conn->ksnc_rx_deadline =
 314                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 315         cfs_mb();                       /* order with setting rx_started */
 316         conn->ksnc_rx_started = 1;
 317
 318         conn->ksnc_rx_nob_wanted -= nob;
 319         conn->ksnc_rx_nob_left -= nob;
 320
 321         do {
 322                 LASSERT (conn->ksnc_rx_nkiov > 0);
 323
 324                 if (nob < (int) kiov->kiov_len) {
 325                         kiov->kiov_offset += nob;
 326                         kiov->kiov_len -= nob;
 327                         return -EAGAIN;
 328                 }
 329
 330                 nob -= kiov->kiov_len;
 331                 conn->ksnc_rx_kiov = ++kiov;
 332                 conn->ksnc_rx_nkiov--;
 333         } while (nob != 0);
 334
 335         return 1;
 336 }
 337
 338 int
 339 ksocknal_receive (ksock_conn_t *conn)
 340 {
 341         /* Return 1 on success, 0 on EOF, < 0 on error.
 342          * Caller checks ksnc_rx_nob_wanted to determine
 343          * progress/completion. */
 344         int     rc;
 345         ENTRY;
 346
 347         if (ksocknal_data.ksnd_stall_rx != 0) {
 348                 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
 349         }
 350
 351         rc = ksocknal_connsock_addref(conn);
 352         if (rc != 0) {
 353                 LASSERT (conn->ksnc_closing);
 354                 return (-ESHUTDOWN);
 355         }
 356
 357         for (;;) {
 358                 if (conn->ksnc_rx_niov != 0)
 359                         rc = ksocknal_recv_iov (conn);
 360                 else
 361                         rc = ksocknal_recv_kiov (conn);
 362
 363                 if (rc <= 0) {
 364                         /* error/EOF or partial receive */
 365                         if (rc == -EAGAIN) {
 366                                 rc = 1;
 367                         } else if (rc == 0 && conn->ksnc_rx_started) {
 368                                 /* EOF in the middle of a message */
 369                                 rc = -EPROTO;
 370                         }
 371                         break;
 372                 }
 373
 374                 /* Completed a fragment */
 375
 376                 if (conn->ksnc_rx_nob_wanted == 0) {
 377                         rc = 1;
 378                         break;
 379                 }
 380         }
 381
 382         ksocknal_connsock_decref(conn);
 383         RETURN (rc);
 384 }
 385
 386 void
 387 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
 388 {
 389         lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
 390         int          rc = (tx->tx_resid == 0) ? 0 : -EIO;
 391         ENTRY;
 392
 393         LASSERT(ni != NULL || tx->tx_conn != NULL);
 394
 395         if (tx->tx_conn != NULL)
 396                 ksocknal_conn_decref(tx->tx_conn);
 397
 398         if (ni == NULL && tx->tx_conn != NULL)
 399                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
 400
 401         ksocknal_free_tx (tx);
 402         if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
 403                 lnet_finalize (ni, lnetmsg, rc);
 404
 405         EXIT;
 406 }
 407
 408 void
 409 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
 410 {
 411         ksock_tx_t *tx;
 412
 413         while (!cfs_list_empty (txlist)) {
 414                 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
 415
 416                 if (error && tx->tx_lnetmsg != NULL) {
 417                         CDEBUG (D_NETERROR, "Deleting packet type %d len %d %s->%s\n",
 418                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
 419                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
 420                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 421                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 422                 } else if (error) {
 423                         CDEBUG (D_NETERROR, "Deleting noop packet\n");
 424                 }
 425
 426                 cfs_list_del (&tx->tx_list);
 427
 428                 LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1);
 429                 ksocknal_tx_done (ni, tx);
 430         }
 431 }
 432
 433 static void
 434 ksocknal_check_zc_req(ksock_tx_t *tx)
 435 {
 436         ksock_conn_t   *conn = tx->tx_conn;
 437         ksock_peer_t   *peer = conn->ksnc_peer;
 438
 439         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 440          * to ksnp_zc_req_list if some fragment of this message should be sent
 441          * zero-copy.  Our peer will send an ACK containing this cookie when
 442          * she has received this message to tell us we can signal completion.
 443          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 444          * ksnp_zc_req_list. */
 445         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 446         LASSERT (tx->tx_zc_capable);
 447
 448         tx->tx_zc_checked = 1;
 449
 450         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 451             !conn->ksnc_zc_capable)
 452                 return;
 453
 454         /* assign cookie and queue tx to pending list, it will be released when
 455          * a matching ack is received. See ksocknal_handle_zcack() */
 456
 457         ksocknal_tx_addref(tx);
 458
 459         cfs_spin_lock(&peer->ksnp_lock);
 460
 461         /* ZC_REQ is going to be pinned to the peer */
 462         tx->tx_deadline =
 463                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 464
 465         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 466
 467         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 468
 469         if (peer->ksnp_zc_next_cookie == 0)
 470                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 471
 472         cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 473
 474         cfs_spin_unlock(&peer->ksnp_lock);
 475 }
 476
 477 static void
 478 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 479 {
 480         ksock_peer_t   *peer = tx->tx_conn->ksnc_peer;
 481
 482         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 483         LASSERT (tx->tx_zc_capable);
 484
 485         tx->tx_zc_checked = 0;
 486
 487         cfs_spin_lock(&peer->ksnp_lock);
 488
 489         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 490                 /* Not waiting for an ACK */
 491                 cfs_spin_unlock(&peer->ksnp_lock);
 492                 return;
 493         }
 494
 495         tx->tx_msg.ksm_zc_cookies[0] = 0;
 496         cfs_list_del(&tx->tx_zc_list);
 497
 498         cfs_spin_unlock(&peer->ksnp_lock);
 499
 500         ksocknal_tx_decref(tx);
 501 }
 502
 503 int
 504 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 505 {
 506         int            rc;
 507
 508         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 509                 ksocknal_check_zc_req(tx);
 510
 511         rc = ksocknal_transmit (conn, tx);
 512
 513         CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 514
 515         if (tx->tx_resid == 0) {
 516                 /* Sent everything OK */
 517                 LASSERT (rc == 0);
 518
 519                 return (0);
 520         }
 521
 522         if (rc == -EAGAIN)
 523                 return (rc);
 524
 525         if (rc == -ENOMEM) {
 526                 static int counter;
 527
 528                 counter++;   /* exponential backoff warnings */
 529                 if ((counter & (-counter)) == counter)
 530                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 531                               counter, conn, cfs_atomic_read(&libcfs_kmemory));
 532
 533                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 534                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
 535
 536                 /* enomem list takes over scheduler's ref... */
 537                 LASSERT (conn->ksnc_tx_scheduled);
 538                 cfs_list_add_tail(&conn->ksnc_tx_list,
 539                                   &ksocknal_data.ksnd_enomem_conns);
 540                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 541                                                    SOCKNAL_ENOMEM_RETRY),
 542                                    ksocknal_data.ksnd_reaper_waketime))
 543                         cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
 544
 545                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
 546                 return (rc);
 547         }
 548
 549         /* Actual error */
 550         LASSERT (rc < 0);
 551
 552         if (!conn->ksnc_closing) {
 553                 switch (rc) {
 554                 case -ECONNRESET:
 555                         LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
 556                                       "while we were sending data; it may have "
 557                                       "rebooted.\n",
 558                                       HIPQUAD(conn->ksnc_ipaddr));
 559                         break;
 560                 default:
 561                         LCONSOLE_WARN("There was an unexpected network error "
 562                                       "while writing to %u.%u.%u.%u: %d.\n",
 563                                       HIPQUAD(conn->ksnc_ipaddr), rc);
 564                         break;
 565                 }
 566                 CDEBUG(D_NET, "[%p] Error %d on write to %s"
 567                        " ip %d.%d.%d.%d:%d\n", conn, rc,
 568                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 569                        HIPQUAD(conn->ksnc_ipaddr),
 570                        conn->ksnc_port);
 571         }
 572
 573         if (tx->tx_zc_checked)
 574                 ksocknal_uncheck_zc_req(tx);
 575
 576         /* it's not an error if conn is being closed */
 577         ksocknal_close_conn_and_siblings (conn,
 578                                           (conn->ksnc_closing) ? 0 : rc);
 579
 580         return (rc);
 581 }
 582
 583 void
 584 ksocknal_launch_connection_locked (ksock_route_t *route)
 585 {
 586
 587         /* called holding write lock on ksnd_global_lock */
 588
 589         LASSERT (!route->ksnr_scheduled);
 590         LASSERT (!route->ksnr_connecting);
 591         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 592
 593         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 594         ksocknal_route_addref(route);           /* extra ref for connd */
 595
 596         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
 597
 598         cfs_list_add_tail (&route->ksnr_connd_list,
 599                            &ksocknal_data.ksnd_connd_routes);
 600         cfs_waitq_signal (&ksocknal_data.ksnd_connd_waitq);
 601
 602         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
 603 }
 604
 605 void
 606 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 607 {
 608         ksock_route_t *route;
 609
 610         /* called holding write lock on ksnd_global_lock */
 611         for (;;) {
 612                 /* launch any/all connections that need it */
 613                 route = ksocknal_find_connectable_route_locked(peer);
 614                 if (route == NULL)
 615                         return;
 616
 617                 ksocknal_launch_connection_locked(route);
 618         }
 619 }
 620
 621 ksock_conn_t *
 622 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 623 {
 624         cfs_list_t       *tmp;
 625         ksock_conn_t     *conn;
 626         ksock_conn_t     *typed = NULL;
 627         ksock_conn_t     *fallback = NULL;
 628         int               tnob     = 0;
 629         int               fnob     = 0;
 630
 631         cfs_list_for_each (tmp, &peer->ksnp_conns) {
 632                 ksock_conn_t *c  = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
 633                 int           nob = cfs_atomic_read(&c->ksnc_tx_nob) +
 634                                     libcfs_sock_wmem_queued(c->ksnc_sock);
 635                 int           rc;
 636
 637                 LASSERT (!c->ksnc_closing);
 638                 LASSERT (c->ksnc_proto != NULL &&
 639                          c->ksnc_proto->pro_match_tx != NULL);
 640
 641                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 642
 643                 switch (rc) {
 644                 default:
 645                         LBUG();
 646                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 647                         continue;
 648
 649                 case SOCKNAL_MATCH_YES: /* typed connection */
 650                         if (typed == NULL || tnob > nob ||
 651                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 652                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 653                                 typed = c;
 654                                 tnob  = nob;
 655                         }
 656                         break;
 657
 658                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 659                         if (fallback == NULL || fnob > nob ||
 660                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 661                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 662                                 fallback = c;
 663                                 fnob     = nob;
 664                         }
 665                         break;
 666                 }
 667         }
 668
 669         /* prefer the typed selection */
 670         conn = (typed != NULL) ? typed : fallback;
 671
 672         if (conn != NULL)
 673                 conn->ksnc_tx_last_post = cfs_time_current();
 674
 675         return conn;
 676 }
 677
 678 void
 679 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 680 {
 681         conn->ksnc_proto->pro_pack(tx);
 682
 683         cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 684         ksocknal_conn_addref(conn); /* +1 ref for tx */
 685         tx->tx_conn = conn;
 686 }
 687
 688 void
 689 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
 690 {
 691         ksock_sched_t *sched = conn->ksnc_scheduler;
 692         ksock_msg_t   *msg = &tx->tx_msg;
 693         ksock_tx_t    *ztx = NULL;
 694         int            bufnob = 0;
 695
 696         /* called holding global lock (read or irq-write) and caller may
 697          * not have dropped this lock between finding conn and calling me,
 698          * so we don't need the {get,put}connsock dance to deref
 699          * ksnc_sock... */
 700         LASSERT(!conn->ksnc_closing);
 701
 702         CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n",
 703                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
 704                 HIPQUAD(conn->ksnc_ipaddr),
 705                 conn->ksnc_port);
 706
 707         ksocknal_tx_prep(conn, tx);
 708
 709         /* Ensure the frags we've been given EXACTLY match the number of
 710          * bytes we want to send.  Many TCP/IP stacks disregard any total
 711          * size parameters passed to them and just look at the frags.
 712          *
 713          * We always expect at least 1 mapped fragment containing the
 714          * complete ksocknal message header. */
 715         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 716                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 717                  (unsigned int)tx->tx_nob);
 718         LASSERT (tx->tx_niov >= 1);
 719         LASSERT (tx->tx_resid == tx->tx_nob);
 720
 721         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 722                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 723                                                KSOCK_MSG_NOOP,
 724                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 725
 726         /*
 727          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 728          * but they're used inside spinlocks a lot.
 729          */
 730         bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 731         cfs_spin_lock_bh (&sched->kss_lock);
 732
 733         if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 734                 /* First packet starts the timeout */
 735                 conn->ksnc_tx_deadline =
 736                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 737                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 738                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 739                 conn->ksnc_tx_bufnob = 0;
 740                 cfs_mb(); /* order with adding to tx_queue */
 741         }
 742
 743         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 744                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 745                  * on a normal packet so I don't need to send it */
 746                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 747                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 748
 749                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 750                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 751
 752         } else {
 753                 /* It's a normal packet - can it piggback a noop zc-ack that
 754                  * has been queued already? */
 755                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 756                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 757
 758                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 759                 /* ztx will be released later */
 760         }
 761
 762         if (ztx != NULL) {
 763                 cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 764                 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 765         }
 766
 767         if (conn->ksnc_tx_ready &&      /* able to send */
 768             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 769                 /* +1 ref for scheduler */
 770                 ksocknal_conn_addref(conn);
 771                 cfs_list_add_tail (&conn->ksnc_tx_list,
 772                                    &sched->kss_tx_conns);
 773                 conn->ksnc_tx_scheduled = 1;
 774                 cfs_waitq_signal (&sched->kss_waitq);
 775         }
 776
 777         cfs_spin_unlock_bh (&sched->kss_lock);
 778 }
 779
 780
 781 ksock_route_t *
 782 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 783 {
 784         cfs_time_t     now = cfs_time_current();
 785         cfs_list_t    *tmp;
 786         ksock_route_t *route;
 787
 788         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 789                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 790
 791                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 792
 793                 if (route->ksnr_scheduled)      /* connections being established */
 794                         continue;
 795
 796                 /* all route types connected ? */
 797                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 798                         continue;
 799
 800                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 801                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 802                         CDEBUG(D_NET,
 803                                "Too soon to retry route %u.%u.%u.%u "
 804                                "(cnted %d, interval %ld, %ld secs later)\n",
 805                                HIPQUAD(route->ksnr_ipaddr),
 806                                route->ksnr_connected,
 807                                route->ksnr_retry_interval,
 808                                cfs_duration_sec(route->ksnr_timeout - now));
 809                         continue;
 810                 }
 811
 812                 return (route);
 813         }
 814
 815         return (NULL);
 816 }
 817
 818 ksock_route_t *
 819 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 820 {
 821         cfs_list_t        *tmp;
 822         ksock_route_t     *route;
 823
 824         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 825                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 826
 827                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 828
 829                 if (route->ksnr_scheduled)
 830                         return (route);
 831         }
 832
 833         return (NULL);
 834 }
 835
 836 int
 837 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 838 {
 839         ksock_peer_t     *peer;
 840         ksock_conn_t     *conn;
 841         cfs_rwlock_t     *g_lock;
 842         int               retry;
 843         int               rc;
 844
 845         LASSERT (tx->tx_conn == NULL);
 846
 847         g_lock = &ksocknal_data.ksnd_global_lock;
 848
 849         for (retry = 0;; retry = 1) {
 850                 cfs_read_lock (g_lock);
 851                 peer = ksocknal_find_peer_locked(ni, id);
 852                 if (peer != NULL) {
 853                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 854                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 855                                 if (conn != NULL) {
 856                                         /* I've got no routes that need to be
 857                                          * connecting and I do have an actual
 858                                          * connection... */
 859                                         ksocknal_queue_tx_locked (tx, conn);
 860                                         cfs_read_unlock (g_lock);
 861                                         return (0);
 862                                 }
 863                         }
 864                 }
 865
 866                 /* I'll need a write lock... */
 867                 cfs_read_unlock (g_lock);
 868
 869                 cfs_write_lock_bh (g_lock);
 870
 871                 peer = ksocknal_find_peer_locked(ni, id);
 872                 if (peer != NULL)
 873                         break;
 874
 875                 cfs_write_unlock_bh (g_lock);
 876
 877                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 878                         CERROR("Refusing to create a connection to "
 879                                "userspace process %s\n", libcfs_id2str(id));
 880                         return -EHOSTUNREACH;
 881                 }
 882
 883                 if (retry) {
 884                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 885                         return -EHOSTUNREACH;
 886                 }
 887
 888                 rc = ksocknal_add_peer(ni, id,
 889                                        LNET_NIDADDR(id.nid),
 890                                        lnet_acceptor_port());
 891                 if (rc != 0) {
 892                         CERROR("Can't add peer %s: %d\n",
 893                                libcfs_id2str(id), rc);
 894                         return rc;
 895                 }
 896         }
 897
 898         ksocknal_launch_all_connections_locked(peer);
 899
 900         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 901         if (conn != NULL) {
 902                 /* Connection exists; queue message on it */
 903                 ksocknal_queue_tx_locked (tx, conn);
 904                 cfs_write_unlock_bh (g_lock);
 905                 return (0);
 906         }
 907
 908         if (peer->ksnp_accepting > 0 ||
 909             ksocknal_find_connecting_route_locked (peer) != NULL) {
 910                 /* the message is going to be pinned to the peer */
 911                 tx->tx_deadline =
 912                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 913
 914                 /* Queue the message until a connection is established */
 915                 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
 916                 cfs_write_unlock_bh (g_lock);
 917                 return 0;
 918         }
 919
 920         cfs_write_unlock_bh (g_lock);
 921
 922         /* NB Routes may be ignored if connections to them failed recently */
 923         CDEBUG(D_NETERROR, "No usable routes to %s\n", libcfs_id2str(id));
 924         return (-EHOSTUNREACH);
 925 }
 926
 927 int
 928 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 929 {
 930         int               mpflag = 0;
 931         int               type = lntmsg->msg_type;
 932         lnet_process_id_t target = lntmsg->msg_target;
 933         unsigned int      payload_niov = lntmsg->msg_niov;
 934         struct iovec     *payload_iov = lntmsg->msg_iov;
 935         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 936         unsigned int      payload_offset = lntmsg->msg_offset;
 937         unsigned int      payload_nob = lntmsg->msg_len;
 938         ksock_tx_t       *tx;
 939         int               desc_size;
 940         int               rc;
 941
 942         /* NB 'private' is different depending on what we're sending.
 943          * Just ignore it... */
 944
 945         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 946                payload_nob, payload_niov, libcfs_id2str(target));
 947
 948         LASSERT (payload_nob == 0 || payload_niov > 0);
 949         LASSERT (payload_niov <= LNET_MAX_IOV);
 950         /* payload is either all vaddrs or all pages */
 951         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 952         LASSERT (!cfs_in_interrupt ());
 953
 954         if (payload_iov != NULL)
 955                 desc_size = offsetof(ksock_tx_t,
 956                                      tx_frags.virt.iov[1 + payload_niov]);
 957         else
 958                 desc_size = offsetof(ksock_tx_t,
 959                                      tx_frags.paged.kiov[payload_niov]);
 960
 961         if (lntmsg->msg_vmflush)
 962                 mpflag = cfs_memory_pressure_get_and_set();
 963         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 964         if (tx == NULL) {
 965                 CERROR("Can't allocate tx desc type %d size %d\n",
 966                        type, desc_size);
 967                 if (lntmsg->msg_vmflush)
 968                         cfs_memory_pressure_restore(mpflag);
 969                 return (-ENOMEM);
 970         }
 971
 972         tx->tx_conn = NULL;                     /* set when assigned a conn */
 973         tx->tx_lnetmsg = lntmsg;
 974
 975         if (payload_iov != NULL) {
 976                 tx->tx_kiov = NULL;
 977                 tx->tx_nkiov = 0;
 978                 tx->tx_iov = tx->tx_frags.virt.iov;
 979                 tx->tx_niov = 1 +
 980                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 981                                                payload_niov, payload_iov,
 982                                                payload_offset, payload_nob);
 983         } else {
 984                 tx->tx_niov = 1;
 985                 tx->tx_iov = &tx->tx_frags.paged.iov;
 986                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 987                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 988                                                  payload_niov, payload_kiov,
 989                                                  payload_offset, payload_nob);
 990
 991                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 992                         tx->tx_zc_capable = 1;
 993         }
 994
 995         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 996
 997         /* The first fragment will be set later in pro_pack */
 998         rc = ksocknal_launch_packet(ni, tx, target);
 999         if (lntmsg->msg_vmflush)
1000                 cfs_memory_pressure_restore(mpflag);
1001         if (rc == 0)
1002                 return (0);
1003
1004         ksocknal_free_tx(tx);
1005         return (-EIO);
1006 }
1007
1008 int
1009 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1010 {
1011         long          pid = cfs_kernel_thread (fn, arg, 0);
1012
1013         if (pid < 0)
1014                 return ((int)pid);
1015
1016         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1017         ksocknal_data.ksnd_nthreads++;
1018         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1019         return (0);
1020 }
1021
1022 void
1023 ksocknal_thread_fini (void)
1024 {
1025         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1026         ksocknal_data.ksnd_nthreads--;
1027         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1028 }
1029
1030 int
1031 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1032 {
1033         static char ksocknal_slop_buffer[4096];
1034
1035         int            nob;
1036         unsigned int   niov;
1037         int            skipped;
1038
1039         LASSERT(conn->ksnc_proto != NULL);
1040
1041         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1042                 /* Remind the socket to ack eagerly... */
1043                 ksocknal_lib_eager_ack(conn);
1044         }
1045
1046         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1047                 conn->ksnc_rx_started = 0;
1048                 cfs_mb();                       /* racing with timeout thread */
1049
1050                 switch (conn->ksnc_proto->pro_version) {
1051                 case  KSOCK_PROTO_V2:
1052                 case  KSOCK_PROTO_V3:
1053                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1054                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1055                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1056
1057                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1058                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1059                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1060                         break;
1061
1062                 case KSOCK_PROTO_V1:
1063                         /* Receiving bare lnet_hdr_t */
1064                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1065                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1066                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1067
1068                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1069                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1070                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1071                         break;
1072
1073                 default:
1074                         LBUG ();
1075                 }
1076                 conn->ksnc_rx_niov = 1;
1077
1078                 conn->ksnc_rx_kiov = NULL;
1079                 conn->ksnc_rx_nkiov = 0;
1080                 conn->ksnc_rx_csum = ~0;
1081                 return (1);
1082         }
1083
1084         /* Set up to skip as much as possible now.  If there's more left
1085          * (ran out of iov entries) we'll get called again */
1086
1087         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1088         conn->ksnc_rx_nob_left = nob_to_skip;
1089         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1090         skipped = 0;
1091         niov = 0;
1092
1093         do {
1094                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1095
1096                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1097                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1098                 niov++;
1099                 skipped += nob;
1100                 nob_to_skip -=nob;
1101
1102         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1103                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1104
1105         conn->ksnc_rx_niov = niov;
1106         conn->ksnc_rx_kiov = NULL;
1107         conn->ksnc_rx_nkiov = 0;
1108         conn->ksnc_rx_nob_wanted = skipped;
1109         return (0);
1110 }
1111
1112 int
1113 ksocknal_process_receive (ksock_conn_t *conn)
1114 {
1115         lnet_hdr_t        *lhdr;
1116         lnet_process_id_t *id;
1117         int                rc;
1118
1119         LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0);
1120
1121         /* NB: sched lock NOT held */
1122         /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1123         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1124                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1125                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1126                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1127  again:
1128         if (conn->ksnc_rx_nob_wanted != 0) {
1129                 rc = ksocknal_receive(conn);
1130
1131                 if (rc <= 0) {
1132                         LASSERT (rc != -EAGAIN);
1133
1134                         if (rc == 0)
1135                                 CDEBUG (D_NET, "[%p] EOF from %s"
1136                                         " ip %d.%d.%d.%d:%d\n", conn,
1137                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1138                                         HIPQUAD(conn->ksnc_ipaddr),
1139                                         conn->ksnc_port);
1140                         else if (!conn->ksnc_closing)
1141                                 CERROR ("[%p] Error %d on read from %s"
1142                                         " ip %d.%d.%d.%d:%d\n",
1143                                         conn, rc,
1144                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1145                                         HIPQUAD(conn->ksnc_ipaddr),
1146                                         conn->ksnc_port);
1147
1148                         /* it's not an error if conn is being closed */
1149                         ksocknal_close_conn_and_siblings (conn,
1150                                                           (conn->ksnc_closing) ? 0 : rc);
1151                         return (rc == 0 ? -ESHUTDOWN : rc);
1152                 }
1153
1154                 if (conn->ksnc_rx_nob_wanted != 0) {
1155                         /* short read */
1156                         return (-EAGAIN);
1157                 }
1158         }
1159         switch (conn->ksnc_rx_state) {
1160         case SOCKNAL_RX_KSM_HEADER:
1161                 if (conn->ksnc_flip) {
1162                         __swab32s(&conn->ksnc_msg.ksm_type);
1163                         __swab32s(&conn->ksnc_msg.ksm_csum);
1164                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1165                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1166                 }
1167
1168                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1169                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1170                         CERROR("%s: Unknown message type: %x\n",
1171                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1172                                conn->ksnc_msg.ksm_type);
1173                         ksocknal_new_packet(conn, 0);
1174                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1175                         return (-EPROTO);
1176                 }
1177
1178                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1179                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1180                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1181                         /* NOOP Checksum error */
1182                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1183                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1184                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1185                         ksocknal_new_packet(conn, 0);
1186                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1187                         return (-EIO);
1188                 }
1189
1190                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1191                         __u64 cookie = 0;
1192
1193                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1194
1195                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1196                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1197
1198                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1199                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1200
1201                         if (rc != 0) {
1202                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1203                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1204                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1205                                 ksocknal_new_packet(conn, 0);
1206                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1207                                 return (rc);
1208                         }
1209                 }
1210
1211                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1212                         ksocknal_new_packet (conn, 0);
1213                         return 0;       /* NOOP is done and just return */
1214                 }
1215
1216                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1217                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1218                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1219
1220                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1221                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1222                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1223
1224                 conn->ksnc_rx_niov = 1;
1225                 conn->ksnc_rx_kiov = NULL;
1226                 conn->ksnc_rx_nkiov = 0;
1227
1228                 goto again;     /* read lnet header now */
1229
1230         case SOCKNAL_RX_LNET_HEADER:
1231                 /* unpack message header */
1232                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1233
1234                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1235                         /* Userspace peer */
1236                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1237                         id   = &conn->ksnc_peer->ksnp_id;
1238
1239                         /* Substitute process ID assigned at connection time */
1240                         lhdr->src_pid = cpu_to_le32(id->pid);
1241                         lhdr->src_nid = cpu_to_le64(id->nid);
1242                 }
1243
1244                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1245                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1246
1247                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1248                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1249                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1250                 if (rc < 0) {
1251                         /* I just received garbage: give up on this conn */
1252                         ksocknal_new_packet(conn, 0);
1253                         ksocknal_close_conn_and_siblings (conn, rc);
1254                         ksocknal_conn_decref(conn);
1255                         return (-EPROTO);
1256                 }
1257
1258                 /* I'm racing with ksocknal_recv() */
1259                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1260                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1261
1262                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1263                         return 0;
1264
1265                 /* ksocknal_recv() got called */
1266                 goto again;
1267
1268         case SOCKNAL_RX_LNET_PAYLOAD:
1269                 /* payload all received */
1270                 rc = 0;
1271
1272                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1273                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1274                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1275                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1276                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1277                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1278                         rc = -EIO;
1279                 }
1280
1281                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1282                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1283
1284                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1285                         id   = &conn->ksnc_peer->ksnp_id;
1286
1287                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1288                                         conn->ksnc_msg.ksm_zc_cookies[0],
1289                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1290                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1291                 }
1292
1293                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1294
1295                 if (rc != 0) {
1296                         ksocknal_new_packet(conn, 0);
1297                         ksocknal_close_conn_and_siblings (conn, rc);
1298                         return (-EPROTO);
1299                 }
1300                 /* Fall through */
1301
1302         case SOCKNAL_RX_SLOP:
1303                 /* starting new packet? */
1304                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1305                         return 0;       /* come back later */
1306                 goto again;             /* try to finish reading slop now */
1307
1308         default:
1309                 break;
1310         }
1311
1312         /* Not Reached */
1313         LBUG ();
1314         return (-EINVAL);                       /* keep gcc happy */
1315 }
1316
1317 int
1318 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1319                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1320                unsigned int offset, unsigned int mlen, unsigned int rlen)
1321 {
1322         ksock_conn_t  *conn = (ksock_conn_t *)private;
1323         ksock_sched_t *sched = conn->ksnc_scheduler;
1324
1325         LASSERT (mlen <= rlen);
1326         LASSERT (niov <= LNET_MAX_IOV);
1327
1328         conn->ksnc_cookie = msg;
1329         conn->ksnc_rx_nob_wanted = mlen;
1330         conn->ksnc_rx_nob_left   = rlen;
1331
1332         if (mlen == 0 || iov != NULL) {
1333                 conn->ksnc_rx_nkiov = 0;
1334                 conn->ksnc_rx_kiov = NULL;
1335                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1336                 conn->ksnc_rx_niov =
1337                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1338                                          niov, iov, offset, mlen);
1339         } else {
1340                 conn->ksnc_rx_niov = 0;
1341                 conn->ksnc_rx_iov  = NULL;
1342                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1343                 conn->ksnc_rx_nkiov =
1344                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1345                                           niov, kiov, offset, mlen);
1346         }
1347
1348         LASSERT (mlen ==
1349                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1350                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1351
1352         LASSERT (conn->ksnc_rx_scheduled);
1353
1354         cfs_spin_lock_bh (&sched->kss_lock);
1355
1356         switch (conn->ksnc_rx_state) {
1357         case SOCKNAL_RX_PARSE_WAIT:
1358                 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1359                 cfs_waitq_signal (&sched->kss_waitq);
1360                 LASSERT (conn->ksnc_rx_ready);
1361                 break;
1362
1363         case SOCKNAL_RX_PARSE:
1364                 /* scheduler hasn't noticed I'm parsing yet */
1365                 break;
1366         }
1367
1368         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1369
1370         cfs_spin_unlock_bh (&sched->kss_lock);
1371         ksocknal_conn_decref(conn);
1372         return (0);
1373 }
1374
1375 static inline int
1376 ksocknal_sched_cansleep(ksock_sched_t *sched)
1377 {
1378         int           rc;
1379
1380         cfs_spin_lock_bh (&sched->kss_lock);
1381
1382         rc = (!ksocknal_data.ksnd_shuttingdown &&
1383               cfs_list_empty(&sched->kss_rx_conns) &&
1384               cfs_list_empty(&sched->kss_tx_conns));
1385
1386         cfs_spin_unlock_bh (&sched->kss_lock);
1387         return (rc);
1388 }
1389
1390 int ksocknal_scheduler (void *arg)
1391 {
1392         ksock_sched_t     *sched = (ksock_sched_t *)arg;
1393         ksock_conn_t      *conn;
1394         ksock_tx_t        *tx;
1395         int                rc;
1396         int                nloops = 0;
1397         int                id = (int)(sched - ksocknal_data.ksnd_schedulers);
1398         char               name[16];
1399
1400         snprintf (name, sizeof (name),"socknal_sd%02d", id);
1401         cfs_daemonize (name);
1402         cfs_block_allsigs ();
1403
1404         if (ksocknal_lib_bind_thread_to_cpu(id))
1405                 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1406
1407         cfs_spin_lock_bh (&sched->kss_lock);
1408
1409         while (!ksocknal_data.ksnd_shuttingdown) {
1410                 int did_something = 0;
1411
1412                 /* Ensure I progress everything semi-fairly */
1413
1414                 if (!cfs_list_empty (&sched->kss_rx_conns)) {
1415                         conn = cfs_list_entry(sched->kss_rx_conns.next,
1416                                               ksock_conn_t, ksnc_rx_list);
1417                         cfs_list_del(&conn->ksnc_rx_list);
1418
1419                         LASSERT(conn->ksnc_rx_scheduled);
1420                         LASSERT(conn->ksnc_rx_ready);
1421
1422                         /* clear rx_ready in case receive isn't complete.
1423                          * Do it BEFORE we call process_recv, since
1424                          * data_ready can set it any time after we release
1425                          * kss_lock. */
1426                         conn->ksnc_rx_ready = 0;
1427                         cfs_spin_unlock_bh (&sched->kss_lock);
1428
1429                         rc = ksocknal_process_receive(conn);
1430
1431                         cfs_spin_lock_bh (&sched->kss_lock);
1432
1433                         /* I'm the only one that can clear this flag */
1434                         LASSERT(conn->ksnc_rx_scheduled);
1435
1436                         /* Did process_receive get everything it wanted? */
1437                         if (rc == 0)
1438                                 conn->ksnc_rx_ready = 1;
1439
1440                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1441                                 /* Conn blocked waiting for ksocknal_recv()
1442                                  * I change its state (under lock) to signal
1443                                  * it can be rescheduled */
1444                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1445                         } else if (conn->ksnc_rx_ready) {
1446                                 /* reschedule for rx */
1447                                 cfs_list_add_tail (&conn->ksnc_rx_list,
1448                                                    &sched->kss_rx_conns);
1449                         } else {
1450                                 conn->ksnc_rx_scheduled = 0;
1451                                 /* drop my ref */
1452                                 ksocknal_conn_decref(conn);
1453                         }
1454
1455                         did_something = 1;
1456                 }
1457
1458                 if (!cfs_list_empty (&sched->kss_tx_conns)) {
1459                         CFS_LIST_HEAD    (zlist);
1460
1461                         if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1462                                 cfs_list_add(&zlist,
1463                                              &sched->kss_zombie_noop_txs);
1464                                 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1465                         }
1466
1467                         conn = cfs_list_entry(sched->kss_tx_conns.next,
1468                                               ksock_conn_t, ksnc_tx_list);
1469                         cfs_list_del (&conn->ksnc_tx_list);
1470
1471                         LASSERT(conn->ksnc_tx_scheduled);
1472                         LASSERT(conn->ksnc_tx_ready);
1473                         LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1474
1475                         tx = cfs_list_entry(conn->ksnc_tx_queue.next,
1476                                             ksock_tx_t, tx_list);
1477
1478                         if (conn->ksnc_tx_carrier == tx)
1479                                 ksocknal_next_tx_carrier(conn);
1480
1481                         /* dequeue now so empty list => more to send */
1482                         cfs_list_del(&tx->tx_list);
1483
1484                         /* Clear tx_ready in case send isn't complete.  Do
1485                          * it BEFORE we call process_transmit, since
1486                          * write_space can set it any time after we release
1487                          * kss_lock. */
1488                         conn->ksnc_tx_ready = 0;
1489                         cfs_spin_unlock_bh (&sched->kss_lock);
1490
1491                         if (!cfs_list_empty(&zlist)) {
1492                                 /* free zombie noop txs, it's fast because
1493                                  * noop txs are just put in freelist */
1494                                 ksocknal_txlist_done(NULL, &zlist, 0);
1495                         }
1496
1497                         rc = ksocknal_process_transmit(conn, tx);
1498
1499                         if (rc == -ENOMEM || rc == -EAGAIN) {
1500                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1501                                 cfs_spin_lock_bh (&sched->kss_lock);
1502                                 cfs_list_add (&tx->tx_list,
1503                                               &conn->ksnc_tx_queue);
1504                         } else {
1505                                 /* Complete send; tx -ref */
1506                                 ksocknal_tx_decref (tx);
1507
1508                                 cfs_spin_lock_bh (&sched->kss_lock);
1509                                 /* assume space for more */
1510                                 conn->ksnc_tx_ready = 1;
1511                         }
1512
1513                         if (rc == -ENOMEM) {
1514                                 /* Do nothing; after a short timeout, this
1515                                  * conn will be reposted on kss_tx_conns. */
1516                         } else if (conn->ksnc_tx_ready &&
1517                                    !cfs_list_empty (&conn->ksnc_tx_queue)) {
1518                                 /* reschedule for tx */
1519                                 cfs_list_add_tail (&conn->ksnc_tx_list,
1520                                                    &sched->kss_tx_conns);
1521                         } else {
1522                                 conn->ksnc_tx_scheduled = 0;
1523                                 /* drop my ref */
1524                                 ksocknal_conn_decref(conn);
1525                         }
1526
1527                         did_something = 1;
1528                 }
1529                 if (!did_something ||           /* nothing to do */
1530                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1531                         cfs_spin_unlock_bh (&sched->kss_lock);
1532
1533                         nloops = 0;
1534
1535                         if (!did_something) {   /* wait for something to do */
1536                                 cfs_wait_event_interruptible_exclusive(
1537                                         sched->kss_waitq,
1538                                         !ksocknal_sched_cansleep(sched), rc);
1539                                 LASSERT (rc == 0);
1540                         } else {
1541                                 cfs_cond_resched();
1542                         }
1543
1544                         cfs_spin_lock_bh (&sched->kss_lock);
1545                 }
1546         }
1547
1548         cfs_spin_unlock_bh (&sched->kss_lock);
1549         ksocknal_thread_fini ();
1550         return (0);
1551 }
1552
1553 /*
1554  * Add connection to kss_rx_conns of scheduler
1555  * and wakeup the scheduler.
1556  */
1557 void ksocknal_read_callback (ksock_conn_t *conn)
1558 {
1559         ksock_sched_t *sched;
1560         ENTRY;
1561
1562         sched = conn->ksnc_scheduler;
1563
1564         cfs_spin_lock_bh (&sched->kss_lock);
1565
1566         conn->ksnc_rx_ready = 1;
1567
1568         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1569                 cfs_list_add_tail(&conn->ksnc_rx_list,
1570                                   &sched->kss_rx_conns);
1571                 conn->ksnc_rx_scheduled = 1;
1572                 /* extra ref for scheduler */
1573                 ksocknal_conn_addref(conn);
1574
1575                 cfs_waitq_signal (&sched->kss_waitq);
1576         }
1577         cfs_spin_unlock_bh (&sched->kss_lock);
1578
1579         EXIT;
1580 }
1581
1582 /*
1583  * Add connection to kss_tx_conns of scheduler
1584  * and wakeup the scheduler.
1585  */
1586 void ksocknal_write_callback (ksock_conn_t *conn)
1587 {
1588         ksock_sched_t *sched;
1589         ENTRY;
1590
1591         sched = conn->ksnc_scheduler;
1592
1593         cfs_spin_lock_bh (&sched->kss_lock);
1594
1595         conn->ksnc_tx_ready = 1;
1596
1597         if (!conn->ksnc_tx_scheduled && // not being progressed
1598             !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1599                 cfs_list_add_tail (&conn->ksnc_tx_list,
1600                                    &sched->kss_tx_conns);
1601                 conn->ksnc_tx_scheduled = 1;
1602                 /* extra ref for scheduler */
1603                 ksocknal_conn_addref(conn);
1604
1605                 cfs_waitq_signal (&sched->kss_waitq);
1606         }
1607
1608         cfs_spin_unlock_bh (&sched->kss_lock);
1609
1610         EXIT;
1611 }
1612
1613 ksock_proto_t *
1614 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1615 {
1616         __u32   version = 0;
1617
1618         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1619                 version = hello->kshm_version;
1620         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1621                 version = __swab32(hello->kshm_version);
1622
1623         if (version != 0) {
1624 #if SOCKNAL_VERSION_DEBUG
1625                 if (*ksocknal_tunables.ksnd_protocol == 1)
1626                         return NULL;
1627
1628                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1629                     version == KSOCK_PROTO_V3)
1630                         return NULL;
1631 #endif
1632                 if (version == KSOCK_PROTO_V2)
1633                         return &ksocknal_protocol_v2x;
1634
1635                 if (version == KSOCK_PROTO_V3)
1636                         return &ksocknal_protocol_v3x;
1637
1638                 return NULL;
1639         }
1640
1641         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1642                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1643
1644                 CLASSERT (sizeof (lnet_magicversion_t) ==
1645                           offsetof (ksock_hello_msg_t, kshm_src_nid));
1646
1647                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1648                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1649                         return &ksocknal_protocol_v1x;
1650         }
1651
1652         return NULL;
1653 }
1654
1655 int
1656 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1657                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1658 {
1659         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1660         ksock_net_t         *net = (ksock_net_t *)ni->ni_data;
1661
1662         LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1663
1664         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1665         LASSERT (conn->ksnc_proto != NULL);
1666
1667         hello->kshm_src_nid         = ni->ni_nid;
1668         hello->kshm_dst_nid         = peer_nid;
1669         hello->kshm_src_pid         = the_lnet.ln_pid;
1670
1671         hello->kshm_src_incarnation = net->ksnn_incarnation;
1672         hello->kshm_ctype           = conn->ksnc_type;
1673
1674         return conn->ksnc_proto->pro_send_hello(conn, hello);
1675 }
1676
1677 int
1678 ksocknal_invert_type(int type)
1679 {
1680         switch (type)
1681         {
1682         case SOCKLND_CONN_ANY:
1683         case SOCKLND_CONN_CONTROL:
1684                 return (type);
1685         case SOCKLND_CONN_BULK_IN:
1686                 return SOCKLND_CONN_BULK_OUT;
1687         case SOCKLND_CONN_BULK_OUT:
1688                 return SOCKLND_CONN_BULK_IN;
1689         default:
1690                 return (SOCKLND_CONN_NONE);
1691         }
1692 }
1693
1694 int
1695 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1696                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1697                      __u64 *incarnation)
1698 {
1699         /* Return < 0        fatal error
1700          *        0          success
1701          *        EALREADY   lost connection race
1702          *        EPROTO     protocol version mismatch
1703          */
1704         cfs_socket_t        *sock = conn->ksnc_sock;
1705         int                  active = (conn->ksnc_proto != NULL);
1706         int                  timeout;
1707         int                  proto_match;
1708         int                  rc;
1709         ksock_proto_t       *proto;
1710         lnet_process_id_t    recv_id;
1711
1712         /* socket type set on active connections - not set on passive */
1713         LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1714
1715         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1716                             lnet_acceptor_timeout();
1717
1718         rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1719         if (rc != 0) {
1720                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1721                         rc, HIPQUAD(conn->ksnc_ipaddr));
1722                 LASSERT (rc < 0);
1723                 return rc;
1724         }
1725
1726         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1727             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1728             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1729                 /* Unexpected magic! */
1730                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1731                         "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1732                         LNET_PROTO_TCP_MAGIC,
1733                         HIPQUAD(conn->ksnc_ipaddr));
1734                 return -EPROTO;
1735         }
1736
1737         rc = libcfs_sock_read(sock, &hello->kshm_version,
1738                               sizeof(hello->kshm_version), timeout);
1739         if (rc != 0) {
1740                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1741                         rc, HIPQUAD(conn->ksnc_ipaddr));
1742                 LASSERT (rc < 0);
1743                 return rc;
1744         }
1745
1746         proto = ksocknal_parse_proto_version(hello);
1747         if (proto == NULL) {
1748                 if (!active) {
1749                         /* unknown protocol from peer, tell peer my protocol */
1750                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1751 #if SOCKNAL_VERSION_DEBUG
1752                         if (*ksocknal_tunables.ksnd_protocol == 2)
1753                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1754                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1755                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1756 #endif
1757                         hello->kshm_nips = 0;
1758                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1759                 }
1760
1761                 CERROR ("Unknown protocol version (%d.x expected)"
1762                         " from %u.%u.%u.%u\n",
1763                         conn->ksnc_proto->pro_version,
1764                         HIPQUAD(conn->ksnc_ipaddr));
1765
1766                 return -EPROTO;
1767         }
1768
1769         proto_match = (conn->ksnc_proto == proto);
1770         conn->ksnc_proto = proto;
1771
1772         /* receive the rest of hello message anyway */
1773         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1774         if (rc != 0) {
1775                 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1776                        rc, HIPQUAD(conn->ksnc_ipaddr));
1777                 LASSERT (rc < 0);
1778                 return rc;
1779         }
1780
1781         *incarnation = hello->kshm_src_incarnation;
1782
1783         if (hello->kshm_src_nid == LNET_NID_ANY) {
1784                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1785                        "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1786                 return -EPROTO;
1787         }
1788
1789         if (!active &&
1790             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1791                 /* Userspace NAL assigns peer process ID from socket */
1792                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1793                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1794         } else {
1795                 recv_id.nid = hello->kshm_src_nid;
1796                 recv_id.pid = hello->kshm_src_pid;
1797         }
1798
1799         if (!active) {
1800                 *peerid = recv_id;
1801
1802                 /* peer determines type */
1803                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1804                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1805                         CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1806                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1807                                 HIPQUAD(conn->ksnc_ipaddr));
1808                         return -EPROTO;
1809                 }
1810
1811                 return 0;
1812         }
1813
1814         if (peerid->pid != recv_id.pid ||
1815             peerid->nid != recv_id.nid) {
1816                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1817                                    " %u.%u.%u.%u, but they claimed they were "
1818                                    "%s; please check your Lustre "
1819                                    "configuration.\n",
1820                                    libcfs_id2str(*peerid),
1821                                    HIPQUAD(conn->ksnc_ipaddr),
1822                                    libcfs_id2str(recv_id));
1823                 return -EPROTO;
1824         }
1825
1826         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1827                 /* Possible protocol mismatch or I lost the connection race */
1828                 return proto_match ? EALREADY : EPROTO;
1829         }
1830
1831         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1832                 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1833                         conn->ksnc_type, libcfs_id2str(*peerid),
1834                         HIPQUAD(conn->ksnc_ipaddr),
1835                         hello->kshm_ctype);
1836                 return -EPROTO;
1837         }
1838
1839         return 0;
1840 }
1841
1842 int
1843 ksocknal_connect (ksock_route_t *route)
1844 {
1845         CFS_LIST_HEAD    (zombies);
1846         ksock_peer_t     *peer = route->ksnr_peer;
1847         int               type;
1848         int               wanted;
1849         cfs_socket_t     *sock;
1850         cfs_time_t        deadline;
1851         int               retry_later = 0;
1852         int               rc = 0;
1853
1854         deadline = cfs_time_add(cfs_time_current(),
1855                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1856
1857         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1858
1859         LASSERT (route->ksnr_scheduled);
1860         LASSERT (!route->ksnr_connecting);
1861
1862         route->ksnr_connecting = 1;
1863
1864         for (;;) {
1865                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1866
1867                 /* stop connecting if peer/route got closed under me, or
1868                  * route got connected while queued */
1869                 if (peer->ksnp_closing || route->ksnr_deleted ||
1870                     wanted == 0) {
1871                         retry_later = 0;
1872                         break;
1873                 }
1874
1875                 /* reschedule if peer is connecting to me */
1876                 if (peer->ksnp_accepting > 0) {
1877                         CDEBUG(D_NET,
1878                                "peer %s(%d) already connecting to me, retry later.\n",
1879                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1880                         retry_later = 1;
1881                 }
1882
1883                 if (retry_later) /* needs reschedule */
1884                         break;
1885
1886                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1887                         type = SOCKLND_CONN_ANY;
1888                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1889                         type = SOCKLND_CONN_CONTROL;
1890                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1891                         type = SOCKLND_CONN_BULK_IN;
1892                 } else {
1893                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1894                         type = SOCKLND_CONN_BULK_OUT;
1895                 }
1896
1897                 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1898
1899                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1900                         rc = -ETIMEDOUT;
1901                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1902                                                    route->ksnr_ipaddr,
1903                                                    route->ksnr_port);
1904                         goto failed;
1905                 }
1906
1907                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1908                                   route->ksnr_myipaddr,
1909                                   route->ksnr_ipaddr, route->ksnr_port);
1910                 if (rc != 0)
1911                         goto failed;
1912
1913                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1914                 if (rc < 0) {
1915                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1916                                                    route->ksnr_ipaddr,
1917                                                    route->ksnr_port);
1918                         goto failed;
1919                 }
1920
1921                 /* A +ve RC means I have to retry because I lost the connection
1922                  * race or I have to renegotiate protocol version */
1923                 retry_later = (rc != 0);
1924                 if (retry_later)
1925                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1926                                libcfs_nid2str(peer->ksnp_id.nid));
1927
1928                 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1929         }
1930
1931         route->ksnr_scheduled = 0;
1932         route->ksnr_connecting = 0;
1933
1934         if (retry_later) {
1935                 /* re-queue for attention; this frees me up to handle
1936                  * the peer's incoming connection request */
1937
1938                 if (rc == EALREADY ||
1939                     (rc == 0 && peer->ksnp_accepting > 0)) {
1940                         /* We want to introduce a delay before next
1941                          * attempt to connect if we lost conn race,
1942                          * but the race is resolved quickly usually,
1943                          * so min_reconnectms should be good heuristic */
1944                         route->ksnr_retry_interval =
1945                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1946                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1947                                                            route->ksnr_retry_interval);
1948                 }
1949
1950                 ksocknal_launch_connection_locked(route);
1951         }
1952
1953         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1954         return retry_later;
1955
1956  failed:
1957         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1958
1959         route->ksnr_scheduled = 0;
1960         route->ksnr_connecting = 0;
1961
1962         /* This is a retry rather than a new connection */
1963         route->ksnr_retry_interval *= 2;
1964         route->ksnr_retry_interval =
1965                 MAX(route->ksnr_retry_interval,
1966                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1967         route->ksnr_retry_interval =
1968                 MIN(route->ksnr_retry_interval,
1969                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1970
1971         LASSERT (route->ksnr_retry_interval != 0);
1972         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1973                                            route->ksnr_retry_interval);
1974
1975         if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1976             peer->ksnp_accepting == 0 &&
1977             ksocknal_find_connecting_route_locked(peer) == NULL) {
1978                 ksock_conn_t *conn;
1979
1980                 /* ksnp_tx_queue is queued on a conn on successful
1981                  * connection for V1.x and V2.x */
1982                 if (!cfs_list_empty (&peer->ksnp_conns)) {
1983                         conn = cfs_list_entry(peer->ksnp_conns.next,
1984                                               ksock_conn_t, ksnc_list);
1985                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1986                 }
1987
1988                 /* take all the blocked packets while I've got the lock and
1989                  * complete below... */
1990                 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
1991         }
1992
1993 #if 0           /* irrelevent with only eager routes */
1994         if (!route->ksnr_deleted) {
1995                 /* make this route least-favourite for re-selection */
1996                 cfs_list_del(&route->ksnr_list);
1997                 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1998         }
1999 #endif
2000         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2001
2002         ksocknal_peer_failed(peer);
2003         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2004         return 0;
2005 }
2006
2007 /* Go through connd_routes queue looking for a route that
2008    we can process right now */
2009 static ksock_route_t *
2010 ksocknal_connd_get_route_locked(signed long *timeout_p)
2011 {
2012         ksock_route_t *route;
2013         cfs_time_t     now;
2014
2015         /* Only handle an outgoing connection request if there
2016          * is someone left to handle incoming connections */
2017         if ((ksocknal_data.ksnd_connd_connecting + 1) >=
2018             *ksocknal_tunables.ksnd_nconnds)
2019                 return NULL;
2020
2021         now = cfs_time_current();
2022
2023         /* connd_routes can contain both pending and ordinary routes */
2024         cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2025                                  ksnr_connd_list) {
2026
2027                 if (route->ksnr_retry_interval == 0 ||
2028                     cfs_time_aftereq(now, route->ksnr_timeout))
2029                         return route;
2030
2031                 if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT ||
2032                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2033                         *timeout_p = (int)(route->ksnr_timeout - now);
2034         }
2035
2036         return NULL;
2037 }
2038
2039 int
2040 ksocknal_connd (void *arg)
2041 {
2042         long               id = (long)(long_ptr_t)arg;
2043         char               name[16];
2044         ksock_connreq_t   *cr;
2045         ksock_route_t     *route;
2046         cfs_waitlink_t     wait;
2047         signed long        timeout;
2048         int                nloops = 0;
2049         int                cons_retry = 0;
2050         int                dropped_lock;
2051
2052         snprintf (name, sizeof (name), "socknal_cd%02ld", id);
2053         cfs_daemonize (name);
2054         cfs_block_allsigs ();
2055
2056         cfs_waitlink_init (&wait);
2057
2058         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2059
2060         while (!ksocknal_data.ksnd_shuttingdown) {
2061
2062                 dropped_lock = 0;
2063
2064                 if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2065                         /* Connection accepted by the listener */
2066                         cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \
2067                                             next, ksock_connreq_t, ksncr_list);
2068
2069                         cfs_list_del(&cr->ksncr_list);
2070                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2071                         dropped_lock = 1;
2072
2073                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2074                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2075                         lnet_ni_decref(cr->ksncr_ni);
2076                         LIBCFS_FREE(cr, sizeof(*cr));
2077
2078                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2079                 }
2080
2081                 /* Sleep till explicit wake_up if no pending routes present */
2082                 timeout = CFS_MAX_SCHEDULE_TIMEOUT;
2083
2084                 /* Connection request */
2085                 route = ksocknal_connd_get_route_locked(&timeout);
2086
2087                 if (route != NULL) {
2088                         cfs_list_del (&route->ksnr_connd_list);
2089                         ksocknal_data.ksnd_connd_connecting++;
2090                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2091                         dropped_lock = 1;
2092
2093                         if (ksocknal_connect(route)) {
2094                                 /* consecutive retry */
2095                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2096                                         CWARN("massive consecutive "
2097                                               "re-connecting to %u.%u.%u.%u\n",
2098                                               HIPQUAD(route->ksnr_ipaddr));
2099                                         cons_retry = 0;
2100                                 }
2101                         } else {
2102                                 cons_retry = 0;
2103                         }
2104
2105                         ksocknal_route_decref(route);
2106
2107                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2108                         ksocknal_data.ksnd_connd_connecting--;
2109                 }
2110
2111                 if (dropped_lock) {
2112                         if (++nloops < SOCKNAL_RESCHED)
2113                                 continue;
2114                         cfs_spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2115                         nloops = 0;
2116                         cfs_cond_resched();
2117                         cfs_spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2118                         continue;
2119                 }
2120
2121                 /* Nothing to do for 'timeout'  */
2122                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2123                 cfs_waitq_add_exclusive (&ksocknal_data.ksnd_connd_waitq,
2124                                          &wait);
2125                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2126
2127                 nloops = 0;
2128                 cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, timeout);
2129
2130                 cfs_set_current_state (CFS_TASK_RUNNING);
2131                 cfs_waitq_del (&ksocknal_data.ksnd_connd_waitq, &wait);
2132                 cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2133         }
2134
2135         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2136
2137         ksocknal_thread_fini ();
2138         return (0);
2139 }
2140
2141 ksock_conn_t *
2142 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2143 {
2144         /* We're called with a shared lock on ksnd_global_lock */
2145         ksock_conn_t      *conn;
2146         cfs_list_t        *ctmp;
2147
2148         cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2149                 int     error;
2150                 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2151
2152                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2153                 LASSERT (!conn->ksnc_closing);
2154
2155                 /* SOCK_ERROR will reset error code of socket in
2156                  * some platform (like Darwin8.x) */
2157                 error = libcfs_sock_error(conn->ksnc_sock);
2158                 if (error != 0) {
2159                         ksocknal_conn_addref(conn);
2160
2161                         switch (error) {
2162                         case ECONNRESET:
2163                                 CDEBUG(D_NETERROR, "A connection with %s "
2164                                        "(%u.%u.%u.%u:%d) was reset; "
2165                                        "it may have rebooted.\n",
2166                                        libcfs_id2str(peer->ksnp_id),
2167                                        HIPQUAD(conn->ksnc_ipaddr),
2168                                        conn->ksnc_port);
2169                                 break;
2170                         case ETIMEDOUT:
2171                                 CDEBUG(D_NETERROR, "A connection with %s "
2172                                        "(%u.%u.%u.%u:%d) timed out; the "
2173                                        "network or node may be down.\n",
2174                                        libcfs_id2str(peer->ksnp_id),
2175                                        HIPQUAD(conn->ksnc_ipaddr),
2176                                        conn->ksnc_port);
2177                                 break;
2178                         default:
2179                                 CDEBUG(D_NETERROR, "An unexpected network error %d "
2180                                        "occurred with %s "
2181                                        "(%u.%u.%u.%u:%d\n", error,
2182                                        libcfs_id2str(peer->ksnp_id),
2183                                        HIPQUAD(conn->ksnc_ipaddr),
2184                                        conn->ksnc_port);
2185                                 break;
2186                         }
2187
2188                         return (conn);
2189                 }
2190
2191                 if (conn->ksnc_rx_started &&
2192                     cfs_time_aftereq(cfs_time_current(),
2193                                      conn->ksnc_rx_deadline)) {
2194                         /* Timed out incomplete incoming message */
2195                         ksocknal_conn_addref(conn);
2196                         CDEBUG(D_NETERROR, "Timeout receiving from %s "
2197                                "(%u.%u.%u.%u:%d), state %d wanted %d left %d\n",
2198                                libcfs_id2str(peer->ksnp_id),
2199                                HIPQUAD(conn->ksnc_ipaddr),
2200                                conn->ksnc_port,
2201                                conn->ksnc_rx_state,
2202                                conn->ksnc_rx_nob_wanted,
2203                                conn->ksnc_rx_nob_left);
2204                         return (conn);
2205                 }
2206
2207                 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2208                      libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2209                     cfs_time_aftereq(cfs_time_current(),
2210                                      conn->ksnc_tx_deadline)) {
2211                         /* Timed out messages queued for sending or
2212                          * buffered in the socket's send buffer */
2213                         ksocknal_conn_addref(conn);
2214                         CDEBUG(D_NETERROR, "Timeout sending data to %s "
2215                                "(%u.%u.%u.%u:%d) the network or that "
2216                                "node may be down.\n",
2217                                libcfs_id2str(peer->ksnp_id),
2218                                HIPQUAD(conn->ksnc_ipaddr),
2219                                conn->ksnc_port);
2220                         return (conn);
2221                 }
2222         }
2223
2224         return (NULL);
2225 }
2226
2227 static inline void
2228 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2229 {
2230         ksock_tx_t        *tx;
2231         CFS_LIST_HEAD      (stale_txs);
2232
2233         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
2234
2235         while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2236                 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2237                                      ksock_tx_t, tx_list);
2238
2239                 if (!cfs_time_aftereq(cfs_time_current(),
2240                                       tx->tx_deadline))
2241                         break;
2242
2243                 cfs_list_del (&tx->tx_list);
2244                 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2245         }
2246
2247         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2248
2249         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2250 }
2251
2252 int
2253 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2254 {
2255         ksock_sched_t  *sched;
2256         ksock_conn_t   *conn;
2257         ksock_tx_t     *tx;
2258
2259         if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2260                 return 0;
2261
2262         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2263                 return 0;
2264
2265         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2266             cfs_time_before(cfs_time_current(),
2267                             cfs_time_add(peer->ksnp_last_alive,
2268                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2269                 return 0;
2270
2271         if (cfs_time_before(cfs_time_current(),
2272                             peer->ksnp_send_keepalive))
2273                 return 0;
2274
2275         /* retry 10 secs later, so we wouldn't put pressure
2276          * on this peer if we failed to send keepalive this time */
2277         peer->ksnp_send_keepalive = cfs_time_shift(10);
2278
2279         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2280         if (conn != NULL) {
2281                 sched = conn->ksnc_scheduler;
2282
2283                 cfs_spin_lock_bh (&sched->kss_lock);
2284                 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2285                         cfs_spin_unlock_bh(&sched->kss_lock);
2286                         /* there is an queued ACK, don't need keepalive */
2287                         return 0;
2288                 }
2289
2290                 cfs_spin_unlock_bh(&sched->kss_lock);
2291         }
2292
2293         cfs_read_unlock(&ksocknal_data.ksnd_global_lock);
2294
2295         /* cookie = 1 is reserved for keepalive PING */
2296         tx = ksocknal_alloc_tx_noop(1, 1);
2297         if (tx == NULL) {
2298                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2299                 return -ENOMEM;
2300         }
2301
2302         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2303                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2304                 return 1;
2305         }
2306
2307         ksocknal_free_tx(tx);
2308         cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2309
2310         return -EIO;
2311 }
2312
2313
2314 void
2315 ksocknal_check_peer_timeouts (int idx)
2316 {
2317         cfs_list_t       *peers = &ksocknal_data.ksnd_peers[idx];
2318         ksock_peer_t     *peer;
2319         ksock_conn_t     *conn;
2320         ksock_tx_t       *tx;
2321
2322  again:
2323         /* NB. We expect to have a look at all the peers and not find any
2324          * connections to time out, so we just use a shared lock while we
2325          * take a look... */
2326         cfs_read_lock (&ksocknal_data.ksnd_global_lock);
2327
2328         cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2329                 cfs_time_t  deadline = 0;
2330                 int         resid = 0;
2331                 int         n     = 0;
2332
2333                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2334                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2335                         goto again;
2336                 }
2337
2338                 conn = ksocknal_find_timed_out_conn (peer);
2339
2340                 if (conn != NULL) {
2341                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2342
2343                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2344
2345                         /* NB we won't find this one again, but we can't
2346                          * just proceed with the next peer, since we dropped
2347                          * ksnd_global_lock and it might be dead already! */
2348                         ksocknal_conn_decref(conn);
2349                         goto again;
2350                 }
2351
2352                 /* we can't process stale txs right here because we're
2353                  * holding only shared lock */
2354                 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2355                         ksock_tx_t *tx =
2356                                 cfs_list_entry (peer->ksnp_tx_queue.next,
2357                                                 ksock_tx_t, tx_list);
2358
2359                         if (cfs_time_aftereq(cfs_time_current(),
2360                                              tx->tx_deadline)) {
2361
2362                                 ksocknal_peer_addref(peer);
2363                                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2364
2365                                 ksocknal_flush_stale_txs(peer);
2366
2367                                 ksocknal_peer_decref(peer);
2368                                 goto again;
2369                         }
2370                 }
2371
2372                 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2373                         continue;
2374
2375                 cfs_spin_lock(&peer->ksnp_lock);
2376                 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2377                                               ksock_tx_t, tx_zc_list) {
2378                         if (!cfs_time_aftereq(cfs_time_current(),
2379                                               tx->tx_deadline))
2380                                 break;
2381                         /* ignore the TX if connection is being closed */
2382                         if (tx->tx_conn->ksnc_closing)
2383                                 continue;
2384                         n++;
2385                 }
2386
2387                 if (n == 0) {
2388                         cfs_spin_unlock(&peer->ksnp_lock);
2389                         continue;
2390                 }
2391
2392                 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2393                                     ksock_tx_t, tx_zc_list);
2394                 deadline = tx->tx_deadline;
2395                 resid    = tx->tx_resid;
2396                 conn     = tx->tx_conn;
2397                 ksocknal_conn_addref(conn);
2398
2399                 cfs_spin_unlock(&peer->ksnp_lock);
2400                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2401
2402                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2403                        "oldest(%p) timed out %ld secs ago, "
2404                        "resid: %d, wmem: %d\n",
2405                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2406                        cfs_duration_sec(cfs_time_current() - deadline),
2407                        resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2408
2409                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2410                 ksocknal_conn_decref(conn);
2411                 goto again;
2412         }
2413
2414         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2415 }
2416
2417 int
2418 ksocknal_reaper (void *arg)
2419 {
2420         cfs_waitlink_t     wait;
2421         ksock_conn_t      *conn;
2422         ksock_sched_t     *sched;
2423         cfs_list_t         enomem_conns;
2424         int                nenomem_conns;
2425         cfs_duration_t     timeout;
2426         int                i;
2427         int                peer_index = 0;
2428         cfs_time_t         deadline = cfs_time_current();
2429
2430         cfs_daemonize ("socknal_reaper");
2431         cfs_block_allsigs ();
2432
2433         CFS_INIT_LIST_HEAD(&enomem_conns);
2434         cfs_waitlink_init (&wait);
2435
2436         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2437
2438         while (!ksocknal_data.ksnd_shuttingdown) {
2439
2440                 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2441                         conn = cfs_list_entry (ksocknal_data. \
2442                                                ksnd_deathrow_conns.next,
2443                                                ksock_conn_t, ksnc_list);
2444                         cfs_list_del (&conn->ksnc_list);
2445
2446                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2447
2448                         ksocknal_terminate_conn (conn);
2449                         ksocknal_conn_decref(conn);
2450
2451                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2452                         continue;
2453                 }
2454
2455                 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2456                         conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2457                                                next, ksock_conn_t, ksnc_list);
2458                         cfs_list_del (&conn->ksnc_list);
2459
2460                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2461
2462                         ksocknal_destroy_conn (conn);
2463
2464                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2465                         continue;
2466                 }
2467
2468                 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2469                         cfs_list_add(&enomem_conns,
2470                                      &ksocknal_data.ksnd_enomem_conns);
2471                         cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2472                 }
2473
2474                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2475
2476                 /* reschedule all the connections that stalled with ENOMEM... */
2477                 nenomem_conns = 0;
2478                 while (!cfs_list_empty (&enomem_conns)) {
2479                         conn = cfs_list_entry (enomem_conns.next,
2480                                                ksock_conn_t, ksnc_tx_list);
2481                         cfs_list_del (&conn->ksnc_tx_list);
2482
2483                         sched = conn->ksnc_scheduler;
2484
2485                         cfs_spin_lock_bh (&sched->kss_lock);
2486
2487                         LASSERT (conn->ksnc_tx_scheduled);
2488                         conn->ksnc_tx_ready = 1;
2489                         cfs_list_add_tail(&conn->ksnc_tx_list,
2490                                           &sched->kss_tx_conns);
2491                         cfs_waitq_signal (&sched->kss_waitq);
2492
2493                         cfs_spin_unlock_bh (&sched->kss_lock);
2494                         nenomem_conns++;
2495                 }
2496
2497                 /* careful with the jiffy wrap... */
2498                 while ((timeout = cfs_time_sub(deadline,
2499                                                cfs_time_current())) <= 0) {
2500                         const int n = 4;
2501                         const int p = 1;
2502                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2503
2504                         /* Time to check for timeouts on a few more peers: I do
2505                          * checks every 'p' seconds on a proportion of the peer
2506                          * table and I need to check every connection 'n' times
2507                          * within a timeout interval, to ensure I detect a
2508                          * timeout on any connection within (n+1)/n times the
2509                          * timeout interval. */
2510
2511                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2512                                 chunk = (chunk * n * p) /
2513                                         *ksocknal_tunables.ksnd_timeout;
2514                         if (chunk == 0)
2515                                 chunk = 1;
2516
2517                         for (i = 0; i < chunk; i++) {
2518                                 ksocknal_check_peer_timeouts (peer_index);
2519                                 peer_index = (peer_index + 1) %
2520                                              ksocknal_data.ksnd_peer_hash_size;
2521                         }
2522
2523                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2524                 }
2525
2526                 if (nenomem_conns != 0) {
2527                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2528                          * This also prevents me getting woken immediately
2529                          * if any go back on my enomem list. */
2530                         timeout = SOCKNAL_ENOMEM_RETRY;
2531                 }
2532                 ksocknal_data.ksnd_reaper_waketime =
2533                         cfs_time_add(cfs_time_current(), timeout);
2534
2535                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2536                 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2537
2538                 if (!ksocknal_data.ksnd_shuttingdown &&
2539                     cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2540                     cfs_list_empty (&ksocknal_data.ksnd_zombie_conns))
2541                         cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE,
2542                                              timeout);
2543
2544                 cfs_set_current_state (CFS_TASK_RUNNING);
2545                 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2546
2547                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2548         }
2549
2550         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2551
2552         ksocknal_thread_fini ();
2553         return (0);
2554 }