Pileus Git - ~andy/linux/blob - drivers/staging/lustre/lnet/klnds/socklnd/socklnd_cb.c

   1 /*
   2  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2012, Intel Corporation.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Portals, http://www.sf.net/projects/sandiaportals/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   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                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         list_del(&tx->tx_list);
  45                 }
  46
  47                 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         atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_desc_size  = size;
  61
  62         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  63
  64         return tx;
  65 }
  66
  67 ksock_tx_t *
  68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  69 {
  70         ksock_tx_t *tx;
  71
  72         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  73         if (tx == NULL) {
  74                 CERROR("Can't allocate noop tx desc\n");
  75                 return NULL;
  76         }
  77
  78         tx->tx_conn     = NULL;
  79         tx->tx_lnetmsg  = NULL;
  80         tx->tx_kiov     = NULL;
  81         tx->tx_nkiov    = 0;
  82         tx->tx_iov      = tx->tx_frags.virt.iov;
  83         tx->tx_niov     = 1;
  84         tx->tx_nonblk   = nonblk;
  85
  86         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  87         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  88
  89         return tx;
  90 }
  91
  92
  93 void
  94 ksocknal_free_tx (ksock_tx_t *tx)
  95 {
  96         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  97
  98         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  99                 /* it's a noop tx */
 100                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 101
 102                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 103
 104                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 105         } else {
 106                 LIBCFS_FREE(tx, tx->tx_desc_size);
 107         }
 108 }
 109
 110 int
 111 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
 112 {
 113         struct iovec  *iov = tx->tx_iov;
 114         int    nob;
 115         int    rc;
 116
 117         LASSERT (tx->tx_niov > 0);
 118
 119         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 120         rc = ksocknal_lib_send_iov(conn, tx);
 121
 122         if (rc <= 0)                        /* sent nothing? */
 123                 return (rc);
 124
 125         nob = rc;
 126         LASSERT (nob <= tx->tx_resid);
 127         tx->tx_resid -= nob;
 128
 129         /* "consume" iov */
 130         do {
 131                 LASSERT (tx->tx_niov > 0);
 132
 133                 if (nob < (int) iov->iov_len) {
 134                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 135                         iov->iov_len -= nob;
 136                         return (rc);
 137                 }
 138
 139                 nob -= iov->iov_len;
 140                 tx->tx_iov = ++iov;
 141                 tx->tx_niov--;
 142         } while (nob != 0);
 143
 144         return (rc);
 145 }
 146
 147 int
 148 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
 149 {
 150         lnet_kiov_t    *kiov = tx->tx_kiov;
 151         int     nob;
 152         int     rc;
 153
 154         LASSERT (tx->tx_niov == 0);
 155         LASSERT (tx->tx_nkiov > 0);
 156
 157         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 158         rc = ksocknal_lib_send_kiov(conn, tx);
 159
 160         if (rc <= 0)                        /* sent nothing? */
 161                 return (rc);
 162
 163         nob = rc;
 164         LASSERT (nob <= tx->tx_resid);
 165         tx->tx_resid -= nob;
 166
 167         /* "consume" kiov */
 168         do {
 169                 LASSERT(tx->tx_nkiov > 0);
 170
 171                 if (nob < (int)kiov->kiov_len) {
 172                         kiov->kiov_offset += nob;
 173                         kiov->kiov_len -= nob;
 174                         return rc;
 175                 }
 176
 177                 nob -= (int)kiov->kiov_len;
 178                 tx->tx_kiov = ++kiov;
 179                 tx->tx_nkiov--;
 180         } while (nob != 0);
 181
 182         return (rc);
 183 }
 184
 185 int
 186 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 187 {
 188         int      rc;
 189         int      bufnob;
 190
 191         if (ksocknal_data.ksnd_stall_tx != 0) {
 192                 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 193         }
 194
 195         LASSERT (tx->tx_resid != 0);
 196
 197         rc = ksocknal_connsock_addref(conn);
 198         if (rc != 0) {
 199                 LASSERT (conn->ksnc_closing);
 200                 return (-ESHUTDOWN);
 201         }
 202
 203         do {
 204                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 205                         /* testing... */
 206                         ksocknal_data.ksnd_enomem_tx--;
 207                         rc = -EAGAIN;
 208                 } else if (tx->tx_niov != 0) {
 209                         rc = ksocknal_send_iov (conn, tx);
 210                 } else {
 211                         rc = ksocknal_send_kiov (conn, tx);
 212                 }
 213
 214                 bufnob = cfs_sock_wmem_queued(conn->ksnc_sock);
 215                 if (rc > 0)                  /* sent something? */
 216                         conn->ksnc_tx_bufnob += rc; /* account it */
 217
 218                 if (bufnob < conn->ksnc_tx_bufnob) {
 219                         /* allocated send buffer bytes < computed; infer
 220                          * something got ACKed */
 221                         conn->ksnc_tx_deadline =
 222                                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 223                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 224                         conn->ksnc_tx_bufnob = bufnob;
 225                         mb();
 226                 }
 227
 228                 if (rc <= 0) { /* Didn't write anything? */
 229
 230                         if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
 231                                 rc = -EAGAIN;
 232
 233                         /* Check if EAGAIN is due to memory pressure */
 234                         if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 235                                 rc = -ENOMEM;
 236
 237                         break;
 238                 }
 239
 240                 /* socket's wmem_queued now includes 'rc' bytes */
 241                 atomic_sub (rc, &conn->ksnc_tx_nob);
 242                 rc = 0;
 243
 244         } while (tx->tx_resid != 0);
 245
 246         ksocknal_connsock_decref(conn);
 247         return (rc);
 248 }
 249
 250 int
 251 ksocknal_recv_iov (ksock_conn_t *conn)
 252 {
 253         struct iovec *iov = conn->ksnc_rx_iov;
 254         int     nob;
 255         int     rc;
 256
 257         LASSERT (conn->ksnc_rx_niov > 0);
 258
 259         /* Never touch conn->ksnc_rx_iov or change connection
 260          * status inside ksocknal_lib_recv_iov */
 261         rc = ksocknal_lib_recv_iov(conn);
 262
 263         if (rc <= 0)
 264                 return (rc);
 265
 266         /* received something... */
 267         nob = rc;
 268
 269         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 270         conn->ksnc_rx_deadline =
 271                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 272         mb();                  /* order with setting rx_started */
 273         conn->ksnc_rx_started = 1;
 274
 275         conn->ksnc_rx_nob_wanted -= nob;
 276         conn->ksnc_rx_nob_left -= nob;
 277
 278         do {
 279                 LASSERT (conn->ksnc_rx_niov > 0);
 280
 281                 if (nob < (int)iov->iov_len) {
 282                         iov->iov_len -= nob;
 283                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 284                         return (-EAGAIN);
 285                 }
 286
 287                 nob -= iov->iov_len;
 288                 conn->ksnc_rx_iov = ++iov;
 289                 conn->ksnc_rx_niov--;
 290         } while (nob != 0);
 291
 292         return (rc);
 293 }
 294
 295 int
 296 ksocknal_recv_kiov (ksock_conn_t *conn)
 297 {
 298         lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
 299         int     nob;
 300         int     rc;
 301         LASSERT (conn->ksnc_rx_nkiov > 0);
 302
 303         /* Never touch conn->ksnc_rx_kiov or change connection
 304          * status inside ksocknal_lib_recv_iov */
 305         rc = ksocknal_lib_recv_kiov(conn);
 306
 307         if (rc <= 0)
 308                 return (rc);
 309
 310         /* received something... */
 311         nob = rc;
 312
 313         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 314         conn->ksnc_rx_deadline =
 315                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 316         mb();                  /* order with setting rx_started */
 317         conn->ksnc_rx_started = 1;
 318
 319         conn->ksnc_rx_nob_wanted -= nob;
 320         conn->ksnc_rx_nob_left -= nob;
 321
 322         do {
 323                 LASSERT (conn->ksnc_rx_nkiov > 0);
 324
 325                 if (nob < (int) kiov->kiov_len) {
 326                         kiov->kiov_offset += nob;
 327                         kiov->kiov_len -= nob;
 328                         return -EAGAIN;
 329                 }
 330
 331                 nob -= kiov->kiov_len;
 332                 conn->ksnc_rx_kiov = ++kiov;
 333                 conn->ksnc_rx_nkiov--;
 334         } while (nob != 0);
 335
 336         return 1;
 337 }
 338
 339 int
 340 ksocknal_receive (ksock_conn_t *conn)
 341 {
 342         /* Return 1 on success, 0 on EOF, < 0 on error.
 343          * Caller checks ksnc_rx_nob_wanted to determine
 344          * progress/completion. */
 345         int     rc;
 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 && !tx->tx_zc_aborted) ? 0 : -EIO;
 391
 392         LASSERT(ni != NULL || tx->tx_conn != NULL);
 393
 394         if (tx->tx_conn != NULL)
 395                 ksocknal_conn_decref(tx->tx_conn);
 396
 397         if (ni == NULL && tx->tx_conn != NULL)
 398                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
 399
 400         ksocknal_free_tx (tx);
 401         if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
 402                 lnet_finalize (ni, lnetmsg, rc);
 403 }
 404
 405 void
 406 ksocknal_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int error)
 407 {
 408         ksock_tx_t *tx;
 409
 410         while (!list_empty (txlist)) {
 411                 tx = list_entry (txlist->next, ksock_tx_t, tx_list);
 412
 413                 if (error && tx->tx_lnetmsg != NULL) {
 414                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 415                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
 416                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
 417                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 418                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 419                 } else if (error) {
 420                         CNETERR("Deleting noop packet\n");
 421                 }
 422
 423                 list_del (&tx->tx_list);
 424
 425                 LASSERT (atomic_read(&tx->tx_refcount) == 1);
 426                 ksocknal_tx_done (ni, tx);
 427         }
 428 }
 429
 430 static void
 431 ksocknal_check_zc_req(ksock_tx_t *tx)
 432 {
 433         ksock_conn_t   *conn = tx->tx_conn;
 434         ksock_peer_t   *peer = conn->ksnc_peer;
 435
 436         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 437          * to ksnp_zc_req_list if some fragment of this message should be sent
 438          * zero-copy.  Our peer will send an ACK containing this cookie when
 439          * she has received this message to tell us we can signal completion.
 440          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 441          * ksnp_zc_req_list. */
 442         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 443         LASSERT (tx->tx_zc_capable);
 444
 445         tx->tx_zc_checked = 1;
 446
 447         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 448             !conn->ksnc_zc_capable)
 449                 return;
 450
 451         /* assign cookie and queue tx to pending list, it will be released when
 452          * a matching ack is received. See ksocknal_handle_zcack() */
 453
 454         ksocknal_tx_addref(tx);
 455
 456         spin_lock(&peer->ksnp_lock);
 457
 458         /* ZC_REQ is going to be pinned to the peer */
 459         tx->tx_deadline =
 460                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 461
 462         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 463
 464         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 465
 466         if (peer->ksnp_zc_next_cookie == 0)
 467                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 468
 469         list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 470
 471         spin_unlock(&peer->ksnp_lock);
 472 }
 473
 474 static void
 475 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 476 {
 477         ksock_peer_t   *peer = tx->tx_conn->ksnc_peer;
 478
 479         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 480         LASSERT(tx->tx_zc_capable);
 481
 482         tx->tx_zc_checked = 0;
 483
 484         spin_lock(&peer->ksnp_lock);
 485
 486         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 487                 /* Not waiting for an ACK */
 488                 spin_unlock(&peer->ksnp_lock);
 489                 return;
 490         }
 491
 492         tx->tx_msg.ksm_zc_cookies[0] = 0;
 493         list_del(&tx->tx_zc_list);
 494
 495         spin_unlock(&peer->ksnp_lock);
 496
 497         ksocknal_tx_decref(tx);
 498 }
 499
 500 int
 501 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 502 {
 503         int         rc;
 504
 505         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 506                 ksocknal_check_zc_req(tx);
 507
 508         rc = ksocknal_transmit (conn, tx);
 509
 510         CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 511
 512         if (tx->tx_resid == 0) {
 513                 /* Sent everything OK */
 514                 LASSERT (rc == 0);
 515
 516                 return (0);
 517         }
 518
 519         if (rc == -EAGAIN)
 520                 return (rc);
 521
 522         if (rc == -ENOMEM) {
 523                 static int counter;
 524
 525                 counter++;   /* exponential backoff warnings */
 526                 if ((counter & (-counter)) == counter)
 527                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 528                               counter, conn, atomic_read(&libcfs_kmemory));
 529
 530                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 531                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 532
 533                 /* enomem list takes over scheduler's ref... */
 534                 LASSERT (conn->ksnc_tx_scheduled);
 535                 list_add_tail(&conn->ksnc_tx_list,
 536                                   &ksocknal_data.ksnd_enomem_conns);
 537                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 538                                                    SOCKNAL_ENOMEM_RETRY),
 539                                    ksocknal_data.ksnd_reaper_waketime))
 540                         wake_up (&ksocknal_data.ksnd_reaper_waitq);
 541
 542                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 543                 return (rc);
 544         }
 545
 546         /* Actual error */
 547         LASSERT (rc < 0);
 548
 549         if (!conn->ksnc_closing) {
 550                 switch (rc) {
 551                 case -ECONNRESET:
 552                         LCONSOLE_WARN("Host %pI4h reset our connection "
 553                                       "while we were sending data; it may have "
 554                                       "rebooted.\n",
 555                                       &conn->ksnc_ipaddr);
 556                         break;
 557                 default:
 558                         LCONSOLE_WARN("There was an unexpected network error "
 559                                       "while writing to %pI4h: %d.\n",
 560                                       &conn->ksnc_ipaddr, rc);
 561                         break;
 562                 }
 563                 CDEBUG(D_NET, "[%p] Error %d on write to %s"
 564                        " ip %pI4h:%d\n", conn, rc,
 565                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 566                        &conn->ksnc_ipaddr,
 567                        conn->ksnc_port);
 568         }
 569
 570         if (tx->tx_zc_checked)
 571                 ksocknal_uncheck_zc_req(tx);
 572
 573         /* it's not an error if conn is being closed */
 574         ksocknal_close_conn_and_siblings (conn,
 575                                           (conn->ksnc_closing) ? 0 : rc);
 576
 577         return (rc);
 578 }
 579
 580 void
 581 ksocknal_launch_connection_locked (ksock_route_t *route)
 582 {
 583
 584         /* called holding write lock on ksnd_global_lock */
 585
 586         LASSERT (!route->ksnr_scheduled);
 587         LASSERT (!route->ksnr_connecting);
 588         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 589
 590         route->ksnr_scheduled = 1;            /* scheduling conn for connd */
 591         ksocknal_route_addref(route);      /* extra ref for connd */
 592
 593         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 594
 595         list_add_tail(&route->ksnr_connd_list,
 596                           &ksocknal_data.ksnd_connd_routes);
 597         wake_up(&ksocknal_data.ksnd_connd_waitq);
 598
 599         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 600 }
 601
 602 void
 603 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 604 {
 605         ksock_route_t *route;
 606
 607         /* called holding write lock on ksnd_global_lock */
 608         for (;;) {
 609                 /* launch any/all connections that need it */
 610                 route = ksocknal_find_connectable_route_locked(peer);
 611                 if (route == NULL)
 612                         return;
 613
 614                 ksocknal_launch_connection_locked(route);
 615         }
 616 }
 617
 618 ksock_conn_t *
 619 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 620 {
 621         struct list_head       *tmp;
 622         ksock_conn_t     *conn;
 623         ksock_conn_t     *typed = NULL;
 624         ksock_conn_t     *fallback = NULL;
 625         int            tnob     = 0;
 626         int            fnob     = 0;
 627
 628         list_for_each (tmp, &peer->ksnp_conns) {
 629                 ksock_conn_t *c  = list_entry(tmp, ksock_conn_t, ksnc_list);
 630                 int        nob = atomic_read(&c->ksnc_tx_nob) +
 631                                     cfs_sock_wmem_queued(c->ksnc_sock);
 632                 int        rc;
 633
 634                 LASSERT (!c->ksnc_closing);
 635                 LASSERT (c->ksnc_proto != NULL &&
 636                          c->ksnc_proto->pro_match_tx != NULL);
 637
 638                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 639
 640                 switch (rc) {
 641                 default:
 642                         LBUG();
 643                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 644                         continue;
 645
 646                 case SOCKNAL_MATCH_YES: /* typed connection */
 647                         if (typed == NULL || tnob > nob ||
 648                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 649                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 650                                 typed = c;
 651                                 tnob  = nob;
 652                         }
 653                         break;
 654
 655                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 656                         if (fallback == NULL || fnob > nob ||
 657                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 658                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 659                                 fallback = c;
 660                                 fnob     = nob;
 661                         }
 662                         break;
 663                 }
 664         }
 665
 666         /* prefer the typed selection */
 667         conn = (typed != NULL) ? typed : fallback;
 668
 669         if (conn != NULL)
 670                 conn->ksnc_tx_last_post = cfs_time_current();
 671
 672         return conn;
 673 }
 674
 675 void
 676 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 677 {
 678         conn->ksnc_proto->pro_pack(tx);
 679
 680         atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 681         ksocknal_conn_addref(conn); /* +1 ref for tx */
 682         tx->tx_conn = conn;
 683 }
 684
 685 void
 686 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
 687 {
 688         ksock_sched_t *sched = conn->ksnc_scheduler;
 689         ksock_msg_t   *msg = &tx->tx_msg;
 690         ksock_tx_t    *ztx = NULL;
 691         int         bufnob = 0;
 692
 693         /* called holding global lock (read or irq-write) and caller may
 694          * not have dropped this lock between finding conn and calling me,
 695          * so we don't need the {get,put}connsock dance to deref
 696          * ksnc_sock... */
 697         LASSERT(!conn->ksnc_closing);
 698
 699         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 700                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
 701                 &conn->ksnc_ipaddr,
 702                 conn->ksnc_port);
 703
 704         ksocknal_tx_prep(conn, tx);
 705
 706         /* Ensure the frags we've been given EXACTLY match the number of
 707          * bytes we want to send.  Many TCP/IP stacks disregard any total
 708          * size parameters passed to them and just look at the frags.
 709          *
 710          * We always expect at least 1 mapped fragment containing the
 711          * complete ksocknal message header. */
 712         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 713                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 714                  (unsigned int)tx->tx_nob);
 715         LASSERT (tx->tx_niov >= 1);
 716         LASSERT (tx->tx_resid == tx->tx_nob);
 717
 718         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 719                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 720                                                KSOCK_MSG_NOOP,
 721                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 722
 723         /*
 724          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 725          * but they're used inside spinlocks a lot.
 726          */
 727         bufnob = cfs_sock_wmem_queued(conn->ksnc_sock);
 728         spin_lock_bh(&sched->kss_lock);
 729
 730         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 731                 /* First packet starts the timeout */
 732                 conn->ksnc_tx_deadline =
 733                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 734                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 735                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 736                 conn->ksnc_tx_bufnob = 0;
 737                 mb(); /* order with adding to tx_queue */
 738         }
 739
 740         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 741                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 742                  * on a normal packet so I don't need to send it */
 743                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 744                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 745
 746                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 747                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 748
 749         } else {
 750                 /* It's a normal packet - can it piggback a noop zc-ack that
 751                  * has been queued already? */
 752                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 753                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 754
 755                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 756                 /* ztx will be released later */
 757         }
 758
 759         if (ztx != NULL) {
 760                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 761                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 762         }
 763
 764         if (conn->ksnc_tx_ready &&      /* able to send */
 765             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 766                 /* +1 ref for scheduler */
 767                 ksocknal_conn_addref(conn);
 768                 list_add_tail (&conn->ksnc_tx_list,
 769                                    &sched->kss_tx_conns);
 770                 conn->ksnc_tx_scheduled = 1;
 771                 wake_up (&sched->kss_waitq);
 772         }
 773
 774         spin_unlock_bh(&sched->kss_lock);
 775 }
 776
 777
 778 ksock_route_t *
 779 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 780 {
 781         cfs_time_t     now = cfs_time_current();
 782         struct list_head    *tmp;
 783         ksock_route_t *route;
 784
 785         list_for_each (tmp, &peer->ksnp_routes) {
 786                 route = list_entry (tmp, ksock_route_t, ksnr_list);
 787
 788                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 789
 790                 if (route->ksnr_scheduled)      /* connections being established */
 791                         continue;
 792
 793                 /* all route types connected ? */
 794                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 795                         continue;
 796
 797                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 798                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 799                         CDEBUG(D_NET,
 800                                "Too soon to retry route %pI4h "
 801                                "(cnted %d, interval %ld, %ld secs later)\n",
 802                                &route->ksnr_ipaddr,
 803                                route->ksnr_connected,
 804                                route->ksnr_retry_interval,
 805                                cfs_duration_sec(route->ksnr_timeout - now));
 806                         continue;
 807                 }
 808
 809                 return (route);
 810         }
 811
 812         return (NULL);
 813 }
 814
 815 ksock_route_t *
 816 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 817 {
 818         struct list_head        *tmp;
 819         ksock_route_t     *route;
 820
 821         list_for_each (tmp, &peer->ksnp_routes) {
 822                 route = list_entry (tmp, ksock_route_t, ksnr_list);
 823
 824                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 825
 826                 if (route->ksnr_scheduled)
 827                         return (route);
 828         }
 829
 830         return (NULL);
 831 }
 832
 833 int
 834 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 835 {
 836         ksock_peer_t     *peer;
 837         ksock_conn_t     *conn;
 838         rwlock_t     *g_lock;
 839         int            retry;
 840         int            rc;
 841
 842         LASSERT (tx->tx_conn == NULL);
 843
 844         g_lock = &ksocknal_data.ksnd_global_lock;
 845
 846         for (retry = 0;; retry = 1) {
 847                 read_lock(g_lock);
 848                 peer = ksocknal_find_peer_locked(ni, id);
 849                 if (peer != NULL) {
 850                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 851                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 852                                 if (conn != NULL) {
 853                                         /* I've got no routes that need to be
 854                                          * connecting and I do have an actual
 855                                          * connection... */
 856                                         ksocknal_queue_tx_locked (tx, conn);
 857                                         read_unlock(g_lock);
 858                                         return (0);
 859                                 }
 860                         }
 861                 }
 862
 863                 /* I'll need a write lock... */
 864                 read_unlock(g_lock);
 865
 866                 write_lock_bh(g_lock);
 867
 868                 peer = ksocknal_find_peer_locked(ni, id);
 869                 if (peer != NULL)
 870                         break;
 871
 872                 write_unlock_bh(g_lock);
 873
 874                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 875                         CERROR("Refusing to create a connection to "
 876                                "userspace process %s\n", libcfs_id2str(id));
 877                         return -EHOSTUNREACH;
 878                 }
 879
 880                 if (retry) {
 881                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 882                         return -EHOSTUNREACH;
 883                 }
 884
 885                 rc = ksocknal_add_peer(ni, id,
 886                                        LNET_NIDADDR(id.nid),
 887                                        lnet_acceptor_port());
 888                 if (rc != 0) {
 889                         CERROR("Can't add peer %s: %d\n",
 890                                libcfs_id2str(id), rc);
 891                         return rc;
 892                 }
 893         }
 894
 895         ksocknal_launch_all_connections_locked(peer);
 896
 897         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 898         if (conn != NULL) {
 899                 /* Connection exists; queue message on it */
 900                 ksocknal_queue_tx_locked (tx, conn);
 901                 write_unlock_bh(g_lock);
 902                 return (0);
 903         }
 904
 905         if (peer->ksnp_accepting > 0 ||
 906             ksocknal_find_connecting_route_locked (peer) != NULL) {
 907                 /* the message is going to be pinned to the peer */
 908                 tx->tx_deadline =
 909                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 910
 911                 /* Queue the message until a connection is established */
 912                 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
 913                 write_unlock_bh(g_lock);
 914                 return 0;
 915         }
 916
 917         write_unlock_bh(g_lock);
 918
 919         /* NB Routes may be ignored if connections to them failed recently */
 920         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 921         return (-EHOSTUNREACH);
 922 }
 923
 924 int
 925 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 926 {
 927         int            mpflag = 0;
 928         int            type = lntmsg->msg_type;
 929         lnet_process_id_t target = lntmsg->msg_target;
 930         unsigned int      payload_niov = lntmsg->msg_niov;
 931         struct iovec     *payload_iov = lntmsg->msg_iov;
 932         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 933         unsigned int      payload_offset = lntmsg->msg_offset;
 934         unsigned int      payload_nob = lntmsg->msg_len;
 935         ksock_tx_t       *tx;
 936         int            desc_size;
 937         int            rc;
 938
 939         /* NB 'private' is different depending on what we're sending.
 940          * Just ignore it... */
 941
 942         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 943                payload_nob, payload_niov, libcfs_id2str(target));
 944
 945         LASSERT (payload_nob == 0 || payload_niov > 0);
 946         LASSERT (payload_niov <= LNET_MAX_IOV);
 947         /* payload is either all vaddrs or all pages */
 948         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 949         LASSERT (!in_interrupt ());
 950
 951         if (payload_iov != NULL)
 952                 desc_size = offsetof(ksock_tx_t,
 953                                      tx_frags.virt.iov[1 + payload_niov]);
 954         else
 955                 desc_size = offsetof(ksock_tx_t,
 956                                      tx_frags.paged.kiov[payload_niov]);
 957
 958         if (lntmsg->msg_vmflush)
 959                 mpflag = cfs_memory_pressure_get_and_set();
 960         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 961         if (tx == NULL) {
 962                 CERROR("Can't allocate tx desc type %d size %d\n",
 963                        type, desc_size);
 964                 if (lntmsg->msg_vmflush)
 965                         cfs_memory_pressure_restore(mpflag);
 966                 return (-ENOMEM);
 967         }
 968
 969         tx->tx_conn = NULL;                  /* set when assigned a conn */
 970         tx->tx_lnetmsg = lntmsg;
 971
 972         if (payload_iov != NULL) {
 973                 tx->tx_kiov = NULL;
 974                 tx->tx_nkiov = 0;
 975                 tx->tx_iov = tx->tx_frags.virt.iov;
 976                 tx->tx_niov = 1 +
 977                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 978                                                payload_niov, payload_iov,
 979                                                payload_offset, payload_nob);
 980         } else {
 981                 tx->tx_niov = 1;
 982                 tx->tx_iov = &tx->tx_frags.paged.iov;
 983                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 984                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 985                                                  payload_niov, payload_kiov,
 986                                                  payload_offset, payload_nob);
 987
 988                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 989                         tx->tx_zc_capable = 1;
 990         }
 991
 992         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 993
 994         /* The first fragment will be set later in pro_pack */
 995         rc = ksocknal_launch_packet(ni, tx, target);
 996         if (lntmsg->msg_vmflush)
 997                 cfs_memory_pressure_restore(mpflag);
 998         if (rc == 0)
 999                 return (0);
1000
1001         ksocknal_free_tx(tx);
1002         return (-EIO);
1003 }
1004
1005 int
1006 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1007 {
1008         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1009
1010         if (IS_ERR(task))
1011                 return PTR_ERR(task);
1012
1013         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1014         ksocknal_data.ksnd_nthreads++;
1015         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1016         return 0;
1017 }
1018
1019 void
1020 ksocknal_thread_fini (void)
1021 {
1022         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1023         ksocknal_data.ksnd_nthreads--;
1024         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1025 }
1026
1027 int
1028 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1029 {
1030         static char ksocknal_slop_buffer[4096];
1031
1032         int         nob;
1033         unsigned int   niov;
1034         int         skipped;
1035
1036         LASSERT(conn->ksnc_proto != NULL);
1037
1038         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1039                 /* Remind the socket to ack eagerly... */
1040                 ksocknal_lib_eager_ack(conn);
1041         }
1042
1043         if (nob_to_skip == 0) {  /* right at next packet boundary now */
1044                 conn->ksnc_rx_started = 0;
1045                 mb();                  /* racing with timeout thread */
1046
1047                 switch (conn->ksnc_proto->pro_version) {
1048                 case  KSOCK_PROTO_V2:
1049                 case  KSOCK_PROTO_V3:
1050                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1051                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1052                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1053
1054                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1055                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1056                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1057                         break;
1058
1059                 case KSOCK_PROTO_V1:
1060                         /* Receiving bare lnet_hdr_t */
1061                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1062                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1063                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1064
1065                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1066                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1067                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1068                         break;
1069
1070                 default:
1071                         LBUG ();
1072                 }
1073                 conn->ksnc_rx_niov = 1;
1074
1075                 conn->ksnc_rx_kiov = NULL;
1076                 conn->ksnc_rx_nkiov = 0;
1077                 conn->ksnc_rx_csum = ~0;
1078                 return (1);
1079         }
1080
1081         /* Set up to skip as much as possible now.  If there's more left
1082          * (ran out of iov entries) we'll get called again */
1083
1084         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1085         conn->ksnc_rx_nob_left = nob_to_skip;
1086         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1087         skipped = 0;
1088         niov = 0;
1089
1090         do {
1091                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1092
1093                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1094                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1095                 niov++;
1096                 skipped += nob;
1097                 nob_to_skip -=nob;
1098
1099         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1100                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1101
1102         conn->ksnc_rx_niov = niov;
1103         conn->ksnc_rx_kiov = NULL;
1104         conn->ksnc_rx_nkiov = 0;
1105         conn->ksnc_rx_nob_wanted = skipped;
1106         return (0);
1107 }
1108
1109 int
1110 ksocknal_process_receive (ksock_conn_t *conn)
1111 {
1112         lnet_hdr_t      *lhdr;
1113         lnet_process_id_t *id;
1114         int             rc;
1115
1116         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1117
1118         /* NB: sched lock NOT held */
1119         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1120         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1121                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1122                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1123                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1124  again:
1125         if (conn->ksnc_rx_nob_wanted != 0) {
1126                 rc = ksocknal_receive(conn);
1127
1128                 if (rc <= 0) {
1129                         LASSERT (rc != -EAGAIN);
1130
1131                         if (rc == 0)
1132                                 CDEBUG(D_NET, "[%p] EOF from %s"
1133                                         " ip %pI4h:%d\n", conn,
1134                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1135                                         &conn->ksnc_ipaddr,
1136                                         conn->ksnc_port);
1137                         else if (!conn->ksnc_closing)
1138                                 CERROR("[%p] Error %d on read from %s"
1139                                         " ip %pI4h:%d\n",
1140                                         conn, rc,
1141                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1142                                         &conn->ksnc_ipaddr,
1143                                         conn->ksnc_port);
1144
1145                         /* it's not an error if conn is being closed */
1146                         ksocknal_close_conn_and_siblings (conn,
1147                                                           (conn->ksnc_closing) ? 0 : rc);
1148                         return (rc == 0 ? -ESHUTDOWN : rc);
1149                 }
1150
1151                 if (conn->ksnc_rx_nob_wanted != 0) {
1152                         /* short read */
1153                         return (-EAGAIN);
1154                 }
1155         }
1156         switch (conn->ksnc_rx_state) {
1157         case SOCKNAL_RX_KSM_HEADER:
1158                 if (conn->ksnc_flip) {
1159                         __swab32s(&conn->ksnc_msg.ksm_type);
1160                         __swab32s(&conn->ksnc_msg.ksm_csum);
1161                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1162                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1163                 }
1164
1165                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1166                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1167                         CERROR("%s: Unknown message type: %x\n",
1168                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1169                                conn->ksnc_msg.ksm_type);
1170                         ksocknal_new_packet(conn, 0);
1171                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1172                         return (-EPROTO);
1173                 }
1174
1175                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1176                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1177                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1178                         /* NOOP Checksum error */
1179                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1180                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1181                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1182                         ksocknal_new_packet(conn, 0);
1183                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1184                         return (-EIO);
1185                 }
1186
1187                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1188                         __u64 cookie = 0;
1189
1190                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1191
1192                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1193                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1194
1195                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1196                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1197
1198                         if (rc != 0) {
1199                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1200                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1201                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1202                                 ksocknal_new_packet(conn, 0);
1203                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1204                                 return (rc);
1205                         }
1206                 }
1207
1208                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1209                         ksocknal_new_packet (conn, 0);
1210                         return 0;       /* NOOP is done and just return */
1211                 }
1212
1213                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1214                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1215                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1216
1217                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1218                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1219                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1220
1221                 conn->ksnc_rx_niov = 1;
1222                 conn->ksnc_rx_kiov = NULL;
1223                 conn->ksnc_rx_nkiov = 0;
1224
1225                 goto again;     /* read lnet header now */
1226
1227         case SOCKNAL_RX_LNET_HEADER:
1228                 /* unpack message header */
1229                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1230
1231                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1232                         /* Userspace peer */
1233                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1234                         id   = &conn->ksnc_peer->ksnp_id;
1235
1236                         /* Substitute process ID assigned at connection time */
1237                         lhdr->src_pid = cpu_to_le32(id->pid);
1238                         lhdr->src_nid = cpu_to_le64(id->nid);
1239                 }
1240
1241                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1242                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1243
1244                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1245                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1246                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1247                 if (rc < 0) {
1248                         /* I just received garbage: give up on this conn */
1249                         ksocknal_new_packet(conn, 0);
1250                         ksocknal_close_conn_and_siblings (conn, rc);
1251                         ksocknal_conn_decref(conn);
1252                         return (-EPROTO);
1253                 }
1254
1255                 /* I'm racing with ksocknal_recv() */
1256                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1257                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1258
1259                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1260                         return 0;
1261
1262                 /* ksocknal_recv() got called */
1263                 goto again;
1264
1265         case SOCKNAL_RX_LNET_PAYLOAD:
1266                 /* payload all received */
1267                 rc = 0;
1268
1269                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1270                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1271                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1272                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1273                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1274                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1275                         rc = -EIO;
1276                 }
1277
1278                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1279                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1280
1281                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1282                         id   = &conn->ksnc_peer->ksnp_id;
1283
1284                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1285                                         conn->ksnc_msg.ksm_zc_cookies[0],
1286                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1287                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1288                 }
1289
1290                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1291
1292                 if (rc != 0) {
1293                         ksocknal_new_packet(conn, 0);
1294                         ksocknal_close_conn_and_siblings (conn, rc);
1295                         return (-EPROTO);
1296                 }
1297                 /* Fall through */
1298
1299         case SOCKNAL_RX_SLOP:
1300                 /* starting new packet? */
1301                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1302                         return 0;       /* come back later */
1303                 goto again;          /* try to finish reading slop now */
1304
1305         default:
1306                 break;
1307         }
1308
1309         /* Not Reached */
1310         LBUG ();
1311         return (-EINVAL);                      /* keep gcc happy */
1312 }
1313
1314 int
1315 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1316                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1317                unsigned int offset, unsigned int mlen, unsigned int rlen)
1318 {
1319         ksock_conn_t  *conn = (ksock_conn_t *)private;
1320         ksock_sched_t *sched = conn->ksnc_scheduler;
1321
1322         LASSERT (mlen <= rlen);
1323         LASSERT (niov <= LNET_MAX_IOV);
1324
1325         conn->ksnc_cookie = msg;
1326         conn->ksnc_rx_nob_wanted = mlen;
1327         conn->ksnc_rx_nob_left   = rlen;
1328
1329         if (mlen == 0 || iov != NULL) {
1330                 conn->ksnc_rx_nkiov = 0;
1331                 conn->ksnc_rx_kiov = NULL;
1332                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1333                 conn->ksnc_rx_niov =
1334                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1335                                          niov, iov, offset, mlen);
1336         } else {
1337                 conn->ksnc_rx_niov = 0;
1338                 conn->ksnc_rx_iov  = NULL;
1339                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1340                 conn->ksnc_rx_nkiov =
1341                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1342                                           niov, kiov, offset, mlen);
1343         }
1344
1345         LASSERT (mlen ==
1346                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1347                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1348
1349         LASSERT (conn->ksnc_rx_scheduled);
1350
1351         spin_lock_bh(&sched->kss_lock);
1352
1353         switch (conn->ksnc_rx_state) {
1354         case SOCKNAL_RX_PARSE_WAIT:
1355                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1356                 wake_up (&sched->kss_waitq);
1357                 LASSERT (conn->ksnc_rx_ready);
1358                 break;
1359
1360         case SOCKNAL_RX_PARSE:
1361                 /* scheduler hasn't noticed I'm parsing yet */
1362                 break;
1363         }
1364
1365         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1366
1367         spin_unlock_bh(&sched->kss_lock);
1368         ksocknal_conn_decref(conn);
1369         return 0;
1370 }
1371
1372 static inline int
1373 ksocknal_sched_cansleep(ksock_sched_t *sched)
1374 {
1375         int        rc;
1376
1377         spin_lock_bh(&sched->kss_lock);
1378
1379         rc = (!ksocknal_data.ksnd_shuttingdown &&
1380               list_empty(&sched->kss_rx_conns) &&
1381               list_empty(&sched->kss_tx_conns));
1382
1383         spin_unlock_bh(&sched->kss_lock);
1384         return rc;
1385 }
1386
1387 int ksocknal_scheduler(void *arg)
1388 {
1389         struct ksock_sched_info *info;
1390         ksock_sched_t           *sched;
1391         ksock_conn_t            *conn;
1392         ksock_tx_t              *tx;
1393         int                     rc;
1394         int                     nloops = 0;
1395         long                    id = (long)arg;
1396
1397         info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1398         sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1399
1400         cfs_block_allsigs();
1401
1402         rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1403         if (rc != 0) {
1404                 CERROR("Can't set CPT affinity to %d: %d\n",
1405                        info->ksi_cpt, rc);
1406         }
1407
1408         spin_lock_bh(&sched->kss_lock);
1409
1410         while (!ksocknal_data.ksnd_shuttingdown) {
1411                 int did_something = 0;
1412
1413                 /* Ensure I progress everything semi-fairly */
1414
1415                 if (!list_empty (&sched->kss_rx_conns)) {
1416                         conn = list_entry(sched->kss_rx_conns.next,
1417                                               ksock_conn_t, ksnc_rx_list);
1418                         list_del(&conn->ksnc_rx_list);
1419
1420                         LASSERT(conn->ksnc_rx_scheduled);
1421                         LASSERT(conn->ksnc_rx_ready);
1422
1423                         /* clear rx_ready in case receive isn't complete.
1424                          * Do it BEFORE we call process_recv, since
1425                          * data_ready can set it any time after we release
1426                          * kss_lock. */
1427                         conn->ksnc_rx_ready = 0;
1428                         spin_unlock_bh(&sched->kss_lock);
1429
1430                         rc = ksocknal_process_receive(conn);
1431
1432                         spin_lock_bh(&sched->kss_lock);
1433
1434                         /* I'm the only one that can clear this flag */
1435                         LASSERT(conn->ksnc_rx_scheduled);
1436
1437                         /* Did process_receive get everything it wanted? */
1438                         if (rc == 0)
1439                                 conn->ksnc_rx_ready = 1;
1440
1441                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1442                                 /* Conn blocked waiting for ksocknal_recv()
1443                                  * I change its state (under lock) to signal
1444                                  * it can be rescheduled */
1445                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1446                         } else if (conn->ksnc_rx_ready) {
1447                                 /* reschedule for rx */
1448                                 list_add_tail (&conn->ksnc_rx_list,
1449                                                    &sched->kss_rx_conns);
1450                         } else {
1451                                 conn->ksnc_rx_scheduled = 0;
1452                                 /* drop my ref */
1453                                 ksocknal_conn_decref(conn);
1454                         }
1455
1456                         did_something = 1;
1457                 }
1458
1459                 if (!list_empty (&sched->kss_tx_conns)) {
1460                         LIST_HEAD    (zlist);
1461
1462                         if (!list_empty(&sched->kss_zombie_noop_txs)) {
1463                                 list_add(&zlist,
1464                                              &sched->kss_zombie_noop_txs);
1465                                 list_del_init(&sched->kss_zombie_noop_txs);
1466                         }
1467
1468                         conn = list_entry(sched->kss_tx_conns.next,
1469                                               ksock_conn_t, ksnc_tx_list);
1470                         list_del (&conn->ksnc_tx_list);
1471
1472                         LASSERT(conn->ksnc_tx_scheduled);
1473                         LASSERT(conn->ksnc_tx_ready);
1474                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1475
1476                         tx = list_entry(conn->ksnc_tx_queue.next,
1477                                             ksock_tx_t, tx_list);
1478
1479                         if (conn->ksnc_tx_carrier == tx)
1480                                 ksocknal_next_tx_carrier(conn);
1481
1482                         /* dequeue now so empty list => more to send */
1483                         list_del(&tx->tx_list);
1484
1485                         /* Clear tx_ready in case send isn't complete.  Do
1486                          * it BEFORE we call process_transmit, since
1487                          * write_space can set it any time after we release
1488                          * kss_lock. */
1489                         conn->ksnc_tx_ready = 0;
1490                         spin_unlock_bh(&sched->kss_lock);
1491
1492                         if (!list_empty(&zlist)) {
1493                                 /* free zombie noop txs, it's fast because
1494                                  * noop txs are just put in freelist */
1495                                 ksocknal_txlist_done(NULL, &zlist, 0);
1496                         }
1497
1498                         rc = ksocknal_process_transmit(conn, tx);
1499
1500                         if (rc == -ENOMEM || rc == -EAGAIN) {
1501                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1502                                 spin_lock_bh(&sched->kss_lock);
1503                                 list_add(&tx->tx_list,
1504                                              &conn->ksnc_tx_queue);
1505                         } else {
1506                                 /* Complete send; tx -ref */
1507                                 ksocknal_tx_decref(tx);
1508
1509                                 spin_lock_bh(&sched->kss_lock);
1510                                 /* assume space for more */
1511                                 conn->ksnc_tx_ready = 1;
1512                         }
1513
1514                         if (rc == -ENOMEM) {
1515                                 /* Do nothing; after a short timeout, this
1516                                  * conn will be reposted on kss_tx_conns. */
1517                         } else if (conn->ksnc_tx_ready &&
1518                                    !list_empty (&conn->ksnc_tx_queue)) {
1519                                 /* reschedule for tx */
1520                                 list_add_tail (&conn->ksnc_tx_list,
1521                                                    &sched->kss_tx_conns);
1522                         } else {
1523                                 conn->ksnc_tx_scheduled = 0;
1524                                 /* drop my ref */
1525                                 ksocknal_conn_decref(conn);
1526                         }
1527
1528                         did_something = 1;
1529                 }
1530                 if (!did_something ||      /* nothing to do */
1531                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1532                         spin_unlock_bh(&sched->kss_lock);
1533
1534                         nloops = 0;
1535
1536                         if (!did_something) {   /* wait for something to do */
1537                                 cfs_wait_event_interruptible_exclusive(
1538                                         sched->kss_waitq,
1539                                         !ksocknal_sched_cansleep(sched), rc);
1540                                 LASSERT (rc == 0);
1541                         } else {
1542                                 cond_resched();
1543                         }
1544
1545                         spin_lock_bh(&sched->kss_lock);
1546                 }
1547         }
1548
1549         spin_unlock_bh(&sched->kss_lock);
1550         ksocknal_thread_fini();
1551         return 0;
1552 }
1553
1554 /*
1555  * Add connection to kss_rx_conns of scheduler
1556  * and wakeup the scheduler.
1557  */
1558 void ksocknal_read_callback (ksock_conn_t *conn)
1559 {
1560         ksock_sched_t *sched;
1561
1562         sched = conn->ksnc_scheduler;
1563
1564         spin_lock_bh(&sched->kss_lock);
1565
1566         conn->ksnc_rx_ready = 1;
1567
1568         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1569                 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                 wake_up (&sched->kss_waitq);
1576         }
1577         spin_unlock_bh(&sched->kss_lock);
1578 }
1579
1580 /*
1581  * Add connection to kss_tx_conns of scheduler
1582  * and wakeup the scheduler.
1583  */
1584 void ksocknal_write_callback (ksock_conn_t *conn)
1585 {
1586         ksock_sched_t *sched;
1587
1588         sched = conn->ksnc_scheduler;
1589
1590         spin_lock_bh(&sched->kss_lock);
1591
1592         conn->ksnc_tx_ready = 1;
1593
1594         if (!conn->ksnc_tx_scheduled && // not being progressed
1595             !list_empty(&conn->ksnc_tx_queue)){//packets to send
1596                 list_add_tail (&conn->ksnc_tx_list,
1597                                    &sched->kss_tx_conns);
1598                 conn->ksnc_tx_scheduled = 1;
1599                 /* extra ref for scheduler */
1600                 ksocknal_conn_addref(conn);
1601
1602                 wake_up (&sched->kss_waitq);
1603         }
1604
1605         spin_unlock_bh(&sched->kss_lock);
1606 }
1607
1608 ksock_proto_t *
1609 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1610 {
1611         __u32   version = 0;
1612
1613         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1614                 version = hello->kshm_version;
1615         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1616                 version = __swab32(hello->kshm_version);
1617
1618         if (version != 0) {
1619 #if SOCKNAL_VERSION_DEBUG
1620                 if (*ksocknal_tunables.ksnd_protocol == 1)
1621                         return NULL;
1622
1623                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1624                     version == KSOCK_PROTO_V3)
1625                         return NULL;
1626 #endif
1627                 if (version == KSOCK_PROTO_V2)
1628                         return &ksocknal_protocol_v2x;
1629
1630                 if (version == KSOCK_PROTO_V3)
1631                         return &ksocknal_protocol_v3x;
1632
1633                 return NULL;
1634         }
1635
1636         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1637                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1638
1639                 CLASSERT (sizeof (lnet_magicversion_t) ==
1640                           offsetof (ksock_hello_msg_t, kshm_src_nid));
1641
1642                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1643                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1644                         return &ksocknal_protocol_v1x;
1645         }
1646
1647         return NULL;
1648 }
1649
1650 int
1651 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1652                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1653 {
1654         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1655         ksock_net_t      *net = (ksock_net_t *)ni->ni_data;
1656
1657         LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1658
1659         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1660         LASSERT (conn->ksnc_proto != NULL);
1661
1662         hello->kshm_src_nid      = ni->ni_nid;
1663         hello->kshm_dst_nid      = peer_nid;
1664         hello->kshm_src_pid      = the_lnet.ln_pid;
1665
1666         hello->kshm_src_incarnation = net->ksnn_incarnation;
1667         hello->kshm_ctype          = conn->ksnc_type;
1668
1669         return conn->ksnc_proto->pro_send_hello(conn, hello);
1670 }
1671
1672 int
1673 ksocknal_invert_type(int type)
1674 {
1675         switch (type)
1676         {
1677         case SOCKLND_CONN_ANY:
1678         case SOCKLND_CONN_CONTROL:
1679                 return (type);
1680         case SOCKLND_CONN_BULK_IN:
1681                 return SOCKLND_CONN_BULK_OUT;
1682         case SOCKLND_CONN_BULK_OUT:
1683                 return SOCKLND_CONN_BULK_IN;
1684         default:
1685                 return (SOCKLND_CONN_NONE);
1686         }
1687 }
1688
1689 int
1690 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1691                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1692                      __u64 *incarnation)
1693 {
1694         /* Return < 0   fatal error
1695          *      0         success
1696          *      EALREADY   lost connection race
1697          *      EPROTO     protocol version mismatch
1698          */
1699         socket_t        *sock = conn->ksnc_sock;
1700         int               active = (conn->ksnc_proto != NULL);
1701         int               timeout;
1702         int               proto_match;
1703         int               rc;
1704         ksock_proto_t       *proto;
1705         lnet_process_id_t    recv_id;
1706
1707         /* socket type set on active connections - not set on passive */
1708         LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1709
1710         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1711                             lnet_acceptor_timeout();
1712
1713         rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1714         if (rc != 0) {
1715                 CERROR("Error %d reading HELLO from %pI4h\n",
1716                         rc, &conn->ksnc_ipaddr);
1717                 LASSERT (rc < 0);
1718                 return rc;
1719         }
1720
1721         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1722             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1723             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1724                 /* Unexpected magic! */
1725                 CERROR("Bad magic(1) %#08x (%#08x expected) from "
1726                         "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1727                         LNET_PROTO_TCP_MAGIC,
1728                         &conn->ksnc_ipaddr);
1729                 return -EPROTO;
1730         }
1731
1732         rc = libcfs_sock_read(sock, &hello->kshm_version,
1733                               sizeof(hello->kshm_version), timeout);
1734         if (rc != 0) {
1735                 CERROR("Error %d reading HELLO from %pI4h\n",
1736                         rc, &conn->ksnc_ipaddr);
1737                 LASSERT (rc < 0);
1738                 return rc;
1739         }
1740
1741         proto = ksocknal_parse_proto_version(hello);
1742         if (proto == NULL) {
1743                 if (!active) {
1744                         /* unknown protocol from peer, tell peer my protocol */
1745                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1746 #if SOCKNAL_VERSION_DEBUG
1747                         if (*ksocknal_tunables.ksnd_protocol == 2)
1748                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1749                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1750                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1751 #endif
1752                         hello->kshm_nips = 0;
1753                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1754                 }
1755
1756                 CERROR("Unknown protocol version (%d.x expected)"
1757                         " from %pI4h\n",
1758                         conn->ksnc_proto->pro_version,
1759                         &conn->ksnc_ipaddr);
1760
1761                 return -EPROTO;
1762         }
1763
1764         proto_match = (conn->ksnc_proto == proto);
1765         conn->ksnc_proto = proto;
1766
1767         /* receive the rest of hello message anyway */
1768         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1769         if (rc != 0) {
1770                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1771                        rc, &conn->ksnc_ipaddr);
1772                 LASSERT (rc < 0);
1773                 return rc;
1774         }
1775
1776         *incarnation = hello->kshm_src_incarnation;
1777
1778         if (hello->kshm_src_nid == LNET_NID_ANY) {
1779                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1780                        "from %pI4h\n", &conn->ksnc_ipaddr);
1781                 return -EPROTO;
1782         }
1783
1784         if (!active &&
1785             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1786                 /* Userspace NAL assigns peer process ID from socket */
1787                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1788                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1789         } else {
1790                 recv_id.nid = hello->kshm_src_nid;
1791                 recv_id.pid = hello->kshm_src_pid;
1792         }
1793
1794         if (!active) {
1795                 *peerid = recv_id;
1796
1797                 /* peer determines type */
1798                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1799                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1800                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1801                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1802                                 &conn->ksnc_ipaddr);
1803                         return -EPROTO;
1804                 }
1805
1806                 return 0;
1807         }
1808
1809         if (peerid->pid != recv_id.pid ||
1810             peerid->nid != recv_id.nid) {
1811                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1812                                    " %pI4h, but they claimed they were "
1813                                    "%s; please check your Lustre "
1814                                    "configuration.\n",
1815                                    libcfs_id2str(*peerid),
1816                                    &conn->ksnc_ipaddr,
1817                                    libcfs_id2str(recv_id));
1818                 return -EPROTO;
1819         }
1820
1821         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1822                 /* Possible protocol mismatch or I lost the connection race */
1823                 return proto_match ? EALREADY : EPROTO;
1824         }
1825
1826         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1827                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1828                         conn->ksnc_type, libcfs_id2str(*peerid),
1829                         &conn->ksnc_ipaddr,
1830                         hello->kshm_ctype);
1831                 return -EPROTO;
1832         }
1833
1834         return 0;
1835 }
1836
1837 int
1838 ksocknal_connect (ksock_route_t *route)
1839 {
1840         LIST_HEAD    (zombies);
1841         ksock_peer_t     *peer = route->ksnr_peer;
1842         int            type;
1843         int            wanted;
1844         socket_t     *sock;
1845         cfs_time_t      deadline;
1846         int            retry_later = 0;
1847         int            rc = 0;
1848
1849         deadline = cfs_time_add(cfs_time_current(),
1850                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1851
1852         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1853
1854         LASSERT (route->ksnr_scheduled);
1855         LASSERT (!route->ksnr_connecting);
1856
1857         route->ksnr_connecting = 1;
1858
1859         for (;;) {
1860                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1861
1862                 /* stop connecting if peer/route got closed under me, or
1863                  * route got connected while queued */
1864                 if (peer->ksnp_closing || route->ksnr_deleted ||
1865                     wanted == 0) {
1866                         retry_later = 0;
1867                         break;
1868                 }
1869
1870                 /* reschedule if peer is connecting to me */
1871                 if (peer->ksnp_accepting > 0) {
1872                         CDEBUG(D_NET,
1873                                "peer %s(%d) already connecting to me, retry later.\n",
1874                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1875                         retry_later = 1;
1876                 }
1877
1878                 if (retry_later) /* needs reschedule */
1879                         break;
1880
1881                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1882                         type = SOCKLND_CONN_ANY;
1883                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1884                         type = SOCKLND_CONN_CONTROL;
1885                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1886                         type = SOCKLND_CONN_BULK_IN;
1887                 } else {
1888                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1889                         type = SOCKLND_CONN_BULK_OUT;
1890                 }
1891
1892                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1893
1894                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1895                         rc = -ETIMEDOUT;
1896                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1897                                                    route->ksnr_ipaddr,
1898                                                    route->ksnr_port);
1899                         goto failed;
1900                 }
1901
1902                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1903                                   route->ksnr_myipaddr,
1904                                   route->ksnr_ipaddr, route->ksnr_port);
1905                 if (rc != 0)
1906                         goto failed;
1907
1908                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1909                 if (rc < 0) {
1910                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1911                                                    route->ksnr_ipaddr,
1912                                                    route->ksnr_port);
1913                         goto failed;
1914                 }
1915
1916                 /* A +ve RC means I have to retry because I lost the connection
1917                  * race or I have to renegotiate protocol version */
1918                 retry_later = (rc != 0);
1919                 if (retry_later)
1920                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1921                                libcfs_nid2str(peer->ksnp_id.nid));
1922
1923                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1924         }
1925
1926         route->ksnr_scheduled = 0;
1927         route->ksnr_connecting = 0;
1928
1929         if (retry_later) {
1930                 /* re-queue for attention; this frees me up to handle
1931                  * the peer's incoming connection request */
1932
1933                 if (rc == EALREADY ||
1934                     (rc == 0 && peer->ksnp_accepting > 0)) {
1935                         /* We want to introduce a delay before next
1936                          * attempt to connect if we lost conn race,
1937                          * but the race is resolved quickly usually,
1938                          * so min_reconnectms should be good heuristic */
1939                         route->ksnr_retry_interval =
1940                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1941                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1942                                                            route->ksnr_retry_interval);
1943                 }
1944
1945                 ksocknal_launch_connection_locked(route);
1946         }
1947
1948         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1949         return retry_later;
1950
1951  failed:
1952         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1953
1954         route->ksnr_scheduled = 0;
1955         route->ksnr_connecting = 0;
1956
1957         /* This is a retry rather than a new connection */
1958         route->ksnr_retry_interval *= 2;
1959         route->ksnr_retry_interval =
1960                 MAX(route->ksnr_retry_interval,
1961                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1962         route->ksnr_retry_interval =
1963                 MIN(route->ksnr_retry_interval,
1964                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1965
1966         LASSERT (route->ksnr_retry_interval != 0);
1967         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1968                                            route->ksnr_retry_interval);
1969
1970         if (!list_empty(&peer->ksnp_tx_queue) &&
1971             peer->ksnp_accepting == 0 &&
1972             ksocknal_find_connecting_route_locked(peer) == NULL) {
1973                 ksock_conn_t *conn;
1974
1975                 /* ksnp_tx_queue is queued on a conn on successful
1976                  * connection for V1.x and V2.x */
1977                 if (!list_empty (&peer->ksnp_conns)) {
1978                         conn = list_entry(peer->ksnp_conns.next,
1979                                               ksock_conn_t, ksnc_list);
1980                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1981                 }
1982
1983                 /* take all the blocked packets while I've got the lock and
1984                  * complete below... */
1985                 list_splice_init(&peer->ksnp_tx_queue, &zombies);
1986         }
1987
1988 #if 0      /* irrelevant with only eager routes */
1989         if (!route->ksnr_deleted) {
1990                 /* make this route least-favourite for re-selection */
1991                 list_del(&route->ksnr_list);
1992                 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1993         }
1994 #endif
1995         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1996
1997         ksocknal_peer_failed(peer);
1998         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
1999         return 0;
2000 }
2001
2002 /*
2003  * check whether we need to create more connds.
2004  * It will try to create new thread if it's necessary, @timeout can
2005  * be updated if failed to create, so caller wouldn't keep try while
2006  * running out of resource.
2007  */
2008 static int
2009 ksocknal_connd_check_start(long sec, long *timeout)
2010 {
2011         char name[16];
2012         int rc;
2013         int total = ksocknal_data.ksnd_connd_starting +
2014                     ksocknal_data.ksnd_connd_running;
2015
2016         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2017                 /* still in initializing */
2018                 return 0;
2019         }
2020
2021         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2022             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2023                 /* can't create more connd, or still have enough
2024                  * threads to handle more connecting */
2025                 return 0;
2026         }
2027
2028         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2029                 /* no pending connecting request */
2030                 return 0;
2031         }
2032
2033         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2034                 /* may run out of resource, retry later */
2035                 *timeout = cfs_time_seconds(1);
2036                 return 0;
2037         }
2038
2039         if (ksocknal_data.ksnd_connd_starting > 0) {
2040                 /* serialize starting to avoid flood */
2041                 return 0;
2042         }
2043
2044         ksocknal_data.ksnd_connd_starting_stamp = sec;
2045         ksocknal_data.ksnd_connd_starting++;
2046         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2047
2048         /* NB: total is the next id */
2049         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2050         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2051
2052         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2053         if (rc == 0)
2054                 return 1;
2055
2056         /* we tried ... */
2057         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2058         ksocknal_data.ksnd_connd_starting--;
2059         ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2060
2061         return 1;
2062 }
2063
2064 /*
2065  * check whether current thread can exit, it will return 1 if there are too
2066  * many threads and no creating in past 120 seconds.
2067  * Also, this function may update @timeout to make caller come back
2068  * again to recheck these conditions.
2069  */
2070 static int
2071 ksocknal_connd_check_stop(long sec, long *timeout)
2072 {
2073         int val;
2074
2075         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2076                 /* still in initializing */
2077                 return 0;
2078         }
2079
2080         if (ksocknal_data.ksnd_connd_starting > 0) {
2081                 /* in progress of starting new thread */
2082                 return 0;
2083         }
2084
2085         if (ksocknal_data.ksnd_connd_running <=
2086             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2087                 return 0;
2088         }
2089
2090         /* created thread in past 120 seconds? */
2091         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2092                     SOCKNAL_CONND_TIMEOUT - sec);
2093
2094         *timeout = (val > 0) ? cfs_time_seconds(val) :
2095                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2096         if (val > 0)
2097                 return 0;
2098
2099         /* no creating in past 120 seconds */
2100
2101         return ksocknal_data.ksnd_connd_running >
2102                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2103 }
2104
2105 /* Go through connd_routes queue looking for a route that we can process
2106  * right now, @timeout_p can be updated if we need to come back later */
2107 static ksock_route_t *
2108 ksocknal_connd_get_route_locked(signed long *timeout_p)
2109 {
2110         ksock_route_t *route;
2111         cfs_time_t     now;
2112
2113         now = cfs_time_current();
2114
2115         /* connd_routes can contain both pending and ordinary routes */
2116         list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2117                                  ksnr_connd_list) {
2118
2119                 if (route->ksnr_retry_interval == 0 ||
2120                     cfs_time_aftereq(now, route->ksnr_timeout))
2121                         return route;
2122
2123                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2124                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2125                         *timeout_p = (int)(route->ksnr_timeout - now);
2126         }
2127
2128         return NULL;
2129 }
2130
2131 int
2132 ksocknal_connd (void *arg)
2133 {
2134         spinlock_t    *connd_lock = &ksocknal_data.ksnd_connd_lock;
2135         ksock_connreq_t   *cr;
2136         wait_queue_t     wait;
2137         int             nloops = 0;
2138         int             cons_retry = 0;
2139
2140         cfs_block_allsigs ();
2141
2142         init_waitqueue_entry_current (&wait);
2143
2144         spin_lock_bh(connd_lock);
2145
2146         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2147         ksocknal_data.ksnd_connd_starting--;
2148         ksocknal_data.ksnd_connd_running++;
2149
2150         while (!ksocknal_data.ksnd_shuttingdown) {
2151                 ksock_route_t *route = NULL;
2152                 long sec = cfs_time_current_sec();
2153                 long timeout = MAX_SCHEDULE_TIMEOUT;
2154                 int  dropped_lock = 0;
2155
2156                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2157                         /* wakeup another one to check stop */
2158                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2159                         break;
2160                 }
2161
2162                 if (ksocknal_connd_check_start(sec, &timeout)) {
2163                         /* created new thread */
2164                         dropped_lock = 1;
2165                 }
2166
2167                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2168                         /* Connection accepted by the listener */
2169                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2170                                             next, ksock_connreq_t, ksncr_list);
2171
2172                         list_del(&cr->ksncr_list);
2173                         spin_unlock_bh(connd_lock);
2174                         dropped_lock = 1;
2175
2176                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2177                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2178                         lnet_ni_decref(cr->ksncr_ni);
2179                         LIBCFS_FREE(cr, sizeof(*cr));
2180
2181                         spin_lock_bh(connd_lock);
2182                 }
2183
2184                 /* Only handle an outgoing connection request if there
2185                  * is a thread left to handle incoming connections and
2186                  * create new connd */
2187                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2188                     ksocknal_data.ksnd_connd_running) {
2189                         route = ksocknal_connd_get_route_locked(&timeout);
2190                 }
2191                 if (route != NULL) {
2192                         list_del (&route->ksnr_connd_list);
2193                         ksocknal_data.ksnd_connd_connecting++;
2194                         spin_unlock_bh(connd_lock);
2195                         dropped_lock = 1;
2196
2197                         if (ksocknal_connect(route)) {
2198                                 /* consecutive retry */
2199                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2200                                         CWARN("massive consecutive "
2201                                               "re-connecting to %pI4h\n",
2202                                               &route->ksnr_ipaddr);
2203                                         cons_retry = 0;
2204                                 }
2205                         } else {
2206                                 cons_retry = 0;
2207                         }
2208
2209                         ksocknal_route_decref(route);
2210
2211                         spin_lock_bh(connd_lock);
2212                         ksocknal_data.ksnd_connd_connecting--;
2213                 }
2214
2215                 if (dropped_lock) {
2216                         if (++nloops < SOCKNAL_RESCHED)
2217                                 continue;
2218                         spin_unlock_bh(connd_lock);
2219                         nloops = 0;
2220                         cond_resched();
2221                         spin_lock_bh(connd_lock);
2222                         continue;
2223                 }
2224
2225                 /* Nothing to do for 'timeout'  */
2226                 set_current_state(TASK_INTERRUPTIBLE);
2227                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2228                 spin_unlock_bh(connd_lock);
2229
2230                 nloops = 0;
2231                 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2232
2233                 set_current_state(TASK_RUNNING);
2234                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2235                 spin_lock_bh(connd_lock);
2236         }
2237         ksocknal_data.ksnd_connd_running--;
2238         spin_unlock_bh(connd_lock);
2239
2240         ksocknal_thread_fini();
2241         return 0;
2242 }
2243
2244 ksock_conn_t *
2245 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2246 {
2247         /* We're called with a shared lock on ksnd_global_lock */
2248         ksock_conn_t      *conn;
2249         struct list_head        *ctmp;
2250
2251         list_for_each (ctmp, &peer->ksnp_conns) {
2252                 int     error;
2253                 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2254
2255                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2256                 LASSERT (!conn->ksnc_closing);
2257
2258                 /* SOCK_ERROR will reset error code of socket in
2259                  * some platform (like Darwin8.x) */
2260                 error = cfs_sock_error(conn->ksnc_sock);
2261                 if (error != 0) {
2262                         ksocknal_conn_addref(conn);
2263
2264                         switch (error) {
2265                         case ECONNRESET:
2266                                 CNETERR("A connection with %s "
2267                                         "(%pI4h:%d) was reset; "
2268                                         "it may have rebooted.\n",
2269                                         libcfs_id2str(peer->ksnp_id),
2270                                         &conn->ksnc_ipaddr,
2271                                         conn->ksnc_port);
2272                                 break;
2273                         case ETIMEDOUT:
2274                                 CNETERR("A connection with %s "
2275                                         "(%pI4h:%d) timed out; the "
2276                                         "network or node may be down.\n",
2277                                         libcfs_id2str(peer->ksnp_id),
2278                                         &conn->ksnc_ipaddr,
2279                                         conn->ksnc_port);
2280                                 break;
2281                         default:
2282                                 CNETERR("An unexpected network error %d "
2283                                         "occurred with %s "
2284                                         "(%pI4h:%d\n", error,
2285                                         libcfs_id2str(peer->ksnp_id),
2286                                         &conn->ksnc_ipaddr,
2287                                         conn->ksnc_port);
2288                                 break;
2289                         }
2290
2291                         return (conn);
2292                 }
2293
2294                 if (conn->ksnc_rx_started &&
2295                     cfs_time_aftereq(cfs_time_current(),
2296                                      conn->ksnc_rx_deadline)) {
2297                         /* Timed out incomplete incoming message */
2298                         ksocknal_conn_addref(conn);
2299                         CNETERR("Timeout receiving from %s (%pI4h:%d), "
2300                                 "state %d wanted %d left %d\n",
2301                                 libcfs_id2str(peer->ksnp_id),
2302                                 &conn->ksnc_ipaddr,
2303                                 conn->ksnc_port,
2304                                 conn->ksnc_rx_state,
2305                                 conn->ksnc_rx_nob_wanted,
2306                                 conn->ksnc_rx_nob_left);
2307                         return (conn);
2308                 }
2309
2310                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2311                      cfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2312                     cfs_time_aftereq(cfs_time_current(),
2313                                      conn->ksnc_tx_deadline)) {
2314                         /* Timed out messages queued for sending or
2315                          * buffered in the socket's send buffer */
2316                         ksocknal_conn_addref(conn);
2317                         CNETERR("Timeout sending data to %s (%pI4h:%d) "
2318                                 "the network or that node may be down.\n",
2319                                 libcfs_id2str(peer->ksnp_id),
2320                                 &conn->ksnc_ipaddr,
2321                                 conn->ksnc_port);
2322                         return (conn);
2323                 }
2324         }
2325
2326         return (NULL);
2327 }
2328
2329 static inline void
2330 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2331 {
2332         ksock_tx_t      *tx;
2333         LIST_HEAD      (stale_txs);
2334
2335         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2336
2337         while (!list_empty (&peer->ksnp_tx_queue)) {
2338                 tx = list_entry (peer->ksnp_tx_queue.next,
2339                                      ksock_tx_t, tx_list);
2340
2341                 if (!cfs_time_aftereq(cfs_time_current(),
2342                                       tx->tx_deadline))
2343                         break;
2344
2345                 list_del (&tx->tx_list);
2346                 list_add_tail (&tx->tx_list, &stale_txs);
2347         }
2348
2349         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2350
2351         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2352 }
2353
2354 int
2355 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2356 {
2357         ksock_sched_t  *sched;
2358         ksock_conn_t   *conn;
2359         ksock_tx_t     *tx;
2360
2361         if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2362                 return 0;
2363
2364         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2365                 return 0;
2366
2367         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2368             cfs_time_before(cfs_time_current(),
2369                             cfs_time_add(peer->ksnp_last_alive,
2370                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2371                 return 0;
2372
2373         if (cfs_time_before(cfs_time_current(),
2374                             peer->ksnp_send_keepalive))
2375                 return 0;
2376
2377         /* retry 10 secs later, so we wouldn't put pressure
2378          * on this peer if we failed to send keepalive this time */
2379         peer->ksnp_send_keepalive = cfs_time_shift(10);
2380
2381         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2382         if (conn != NULL) {
2383                 sched = conn->ksnc_scheduler;
2384
2385                 spin_lock_bh(&sched->kss_lock);
2386                 if (!list_empty(&conn->ksnc_tx_queue)) {
2387                         spin_unlock_bh(&sched->kss_lock);
2388                         /* there is an queued ACK, don't need keepalive */
2389                         return 0;
2390                 }
2391
2392                 spin_unlock_bh(&sched->kss_lock);
2393         }
2394
2395         read_unlock(&ksocknal_data.ksnd_global_lock);
2396
2397         /* cookie = 1 is reserved for keepalive PING */
2398         tx = ksocknal_alloc_tx_noop(1, 1);
2399         if (tx == NULL) {
2400                 read_lock(&ksocknal_data.ksnd_global_lock);
2401                 return -ENOMEM;
2402         }
2403
2404         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2405                 read_lock(&ksocknal_data.ksnd_global_lock);
2406                 return 1;
2407         }
2408
2409         ksocknal_free_tx(tx);
2410         read_lock(&ksocknal_data.ksnd_global_lock);
2411
2412         return -EIO;
2413 }
2414
2415
2416 void
2417 ksocknal_check_peer_timeouts (int idx)
2418 {
2419         struct list_head       *peers = &ksocknal_data.ksnd_peers[idx];
2420         ksock_peer_t     *peer;
2421         ksock_conn_t     *conn;
2422         ksock_tx_t       *tx;
2423
2424  again:
2425         /* NB. We expect to have a look at all the peers and not find any
2426          * connections to time out, so we just use a shared lock while we
2427          * take a look... */
2428         read_lock(&ksocknal_data.ksnd_global_lock);
2429
2430         list_for_each_entry(peer, peers, ksnp_list) {
2431                 cfs_time_t  deadline = 0;
2432                 int      resid = 0;
2433                 int      n     = 0;
2434
2435                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2436                         read_unlock(&ksocknal_data.ksnd_global_lock);
2437                         goto again;
2438                 }
2439
2440                 conn = ksocknal_find_timed_out_conn (peer);
2441
2442                 if (conn != NULL) {
2443                         read_unlock(&ksocknal_data.ksnd_global_lock);
2444
2445                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2446
2447                         /* NB we won't find this one again, but we can't
2448                          * just proceed with the next peer, since we dropped
2449                          * ksnd_global_lock and it might be dead already! */
2450                         ksocknal_conn_decref(conn);
2451                         goto again;
2452                 }
2453
2454                 /* we can't process stale txs right here because we're
2455                  * holding only shared lock */
2456                 if (!list_empty (&peer->ksnp_tx_queue)) {
2457                         ksock_tx_t *tx =
2458                                 list_entry (peer->ksnp_tx_queue.next,
2459                                                 ksock_tx_t, tx_list);
2460
2461                         if (cfs_time_aftereq(cfs_time_current(),
2462                                              tx->tx_deadline)) {
2463
2464                                 ksocknal_peer_addref(peer);
2465                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2466
2467                                 ksocknal_flush_stale_txs(peer);
2468
2469                                 ksocknal_peer_decref(peer);
2470                                 goto again;
2471                         }
2472                 }
2473
2474                 if (list_empty(&peer->ksnp_zc_req_list))
2475                         continue;
2476
2477                 spin_lock(&peer->ksnp_lock);
2478                 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2479                         if (!cfs_time_aftereq(cfs_time_current(),
2480                                               tx->tx_deadline))
2481                                 break;
2482                         /* ignore the TX if connection is being closed */
2483                         if (tx->tx_conn->ksnc_closing)
2484                                 continue;
2485                         n++;
2486                 }
2487
2488                 if (n == 0) {
2489                         spin_unlock(&peer->ksnp_lock);
2490                         continue;
2491                 }
2492
2493                 tx = list_entry(peer->ksnp_zc_req_list.next,
2494                                     ksock_tx_t, tx_zc_list);
2495                 deadline = tx->tx_deadline;
2496                 resid    = tx->tx_resid;
2497                 conn     = tx->tx_conn;
2498                 ksocknal_conn_addref(conn);
2499
2500                 spin_unlock(&peer->ksnp_lock);
2501                 read_unlock(&ksocknal_data.ksnd_global_lock);
2502
2503                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2504                        "oldest(%p) timed out %ld secs ago, "
2505                        "resid: %d, wmem: %d\n",
2506                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2507                        cfs_duration_sec(cfs_time_current() - deadline),
2508                        resid, cfs_sock_wmem_queued(conn->ksnc_sock));
2509
2510                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2511                 ksocknal_conn_decref(conn);
2512                 goto again;
2513         }
2514
2515         read_unlock(&ksocknal_data.ksnd_global_lock);
2516 }
2517
2518 int
2519 ksocknal_reaper (void *arg)
2520 {
2521         wait_queue_t     wait;
2522         ksock_conn_t      *conn;
2523         ksock_sched_t     *sched;
2524         struct list_head         enomem_conns;
2525         int             nenomem_conns;
2526         cfs_duration_t     timeout;
2527         int             i;
2528         int             peer_index = 0;
2529         cfs_time_t       deadline = cfs_time_current();
2530
2531         cfs_block_allsigs ();
2532
2533         INIT_LIST_HEAD(&enomem_conns);
2534         init_waitqueue_entry_current (&wait);
2535
2536         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2537
2538         while (!ksocknal_data.ksnd_shuttingdown) {
2539
2540                 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2541                         conn = list_entry (ksocknal_data. \
2542                                                ksnd_deathrow_conns.next,
2543                                                ksock_conn_t, ksnc_list);
2544                         list_del (&conn->ksnc_list);
2545
2546                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2547
2548                         ksocknal_terminate_conn(conn);
2549                         ksocknal_conn_decref(conn);
2550
2551                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2552                         continue;
2553                 }
2554
2555                 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2556                         conn = list_entry (ksocknal_data.ksnd_zombie_conns.\
2557                                                next, ksock_conn_t, ksnc_list);
2558                         list_del (&conn->ksnc_list);
2559
2560                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2561
2562                         ksocknal_destroy_conn(conn);
2563
2564                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2565                         continue;
2566                 }
2567
2568                 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2569                         list_add(&enomem_conns,
2570                                      &ksocknal_data.ksnd_enomem_conns);
2571                         list_del_init(&ksocknal_data.ksnd_enomem_conns);
2572                 }
2573
2574                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2575
2576                 /* reschedule all the connections that stalled with ENOMEM... */
2577                 nenomem_conns = 0;
2578                 while (!list_empty (&enomem_conns)) {
2579                         conn = list_entry (enomem_conns.next,
2580                                                ksock_conn_t, ksnc_tx_list);
2581                         list_del (&conn->ksnc_tx_list);
2582
2583                         sched = conn->ksnc_scheduler;
2584
2585                         spin_lock_bh(&sched->kss_lock);
2586
2587                         LASSERT(conn->ksnc_tx_scheduled);
2588                         conn->ksnc_tx_ready = 1;
2589                         list_add_tail(&conn->ksnc_tx_list,
2590                                           &sched->kss_tx_conns);
2591                         wake_up(&sched->kss_waitq);
2592
2593                         spin_unlock_bh(&sched->kss_lock);
2594                         nenomem_conns++;
2595                 }
2596
2597                 /* careful with the jiffy wrap... */
2598                 while ((timeout = cfs_time_sub(deadline,
2599                                                cfs_time_current())) <= 0) {
2600                         const int n = 4;
2601                         const int p = 1;
2602                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2603
2604                         /* Time to check for timeouts on a few more peers: I do
2605                          * checks every 'p' seconds on a proportion of the peer
2606                          * table and I need to check every connection 'n' times
2607                          * within a timeout interval, to ensure I detect a
2608                          * timeout on any connection within (n+1)/n times the
2609                          * timeout interval. */
2610
2611                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2612                                 chunk = (chunk * n * p) /
2613                                         *ksocknal_tunables.ksnd_timeout;
2614                         if (chunk == 0)
2615                                 chunk = 1;
2616
2617                         for (i = 0; i < chunk; i++) {
2618                                 ksocknal_check_peer_timeouts (peer_index);
2619                                 peer_index = (peer_index + 1) %
2620                                              ksocknal_data.ksnd_peer_hash_size;
2621                         }
2622
2623                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2624                 }
2625
2626                 if (nenomem_conns != 0) {
2627                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2628                          * This also prevents me getting woken immediately
2629                          * if any go back on my enomem list. */
2630                         timeout = SOCKNAL_ENOMEM_RETRY;
2631                 }
2632                 ksocknal_data.ksnd_reaper_waketime =
2633                         cfs_time_add(cfs_time_current(), timeout);
2634
2635                 set_current_state (TASK_INTERRUPTIBLE);
2636                 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2637
2638                 if (!ksocknal_data.ksnd_shuttingdown &&
2639                     list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2640                     list_empty (&ksocknal_data.ksnd_zombie_conns))
2641                         waitq_timedwait (&wait, TASK_INTERRUPTIBLE,
2642                                              timeout);
2643
2644                 set_current_state (TASK_RUNNING);
2645                 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2646
2647                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2648         }
2649
2650         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2651
2652         ksocknal_thread_fini();
2653         return 0;
2654 }