lnet/klnds/socklnd/socklnd_cb.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Portals, http://www.sf.net/projects/sandiaportals/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Portals; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include "socklnd.h"
  28
  29 ksock_tx_t *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         ksock_tx_t *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT (size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         cfs_list_del(&tx->tx_list);
  45                 }
  46
  47                 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         cfs_atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_desc_size  = size;
  61
  62         cfs_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         cfs_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                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
 101
 102                 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 103
 104                 cfs_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 = libcfs_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                         cfs_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                 cfs_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         cfs_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         cfs_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         ENTRY;
 347
 348         if (ksocknal_data.ksnd_stall_rx != 0) {
 349                 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
 350         }
 351
 352         rc = ksocknal_connsock_addref(conn);
 353         if (rc != 0) {
 354                 LASSERT (conn->ksnc_closing);
 355                 return (-ESHUTDOWN);
 356         }
 357
 358         for (;;) {
 359                 if (conn->ksnc_rx_niov != 0)
 360                         rc = ksocknal_recv_iov (conn);
 361                 else
 362                         rc = ksocknal_recv_kiov (conn);
 363
 364                 if (rc <= 0) {
 365                         /* error/EOF or partial receive */
 366                         if (rc == -EAGAIN) {
 367                                 rc = 1;
 368                         } else if (rc == 0 && conn->ksnc_rx_started) {
 369                                 /* EOF in the middle of a message */
 370                                 rc = -EPROTO;
 371                         }
 372                         break;
 373                 }
 374
 375                 /* Completed a fragment */
 376
 377                 if (conn->ksnc_rx_nob_wanted == 0) {
 378                         rc = 1;
 379                         break;
 380                 }
 381         }
 382
 383         ksocknal_connsock_decref(conn);
 384         RETURN (rc);
 385 }
 386
 387 void
 388 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
 389 {
 390         lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
 391         int          rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
 392         ENTRY;
 393
 394         LASSERT(ni != NULL || tx->tx_conn != NULL);
 395
 396         if (tx->tx_conn != NULL)
 397                 ksocknal_conn_decref(tx->tx_conn);
 398
 399         if (ni == NULL && tx->tx_conn != NULL)
 400                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
 401
 402         ksocknal_free_tx (tx);
 403         if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
 404                 lnet_finalize (ni, lnetmsg, rc);
 405
 406         EXIT;
 407 }
 408
 409 void
 410 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
 411 {
 412         ksock_tx_t *tx;
 413
 414         while (!cfs_list_empty (txlist)) {
 415                 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
 416
 417                 if (error && tx->tx_lnetmsg != NULL) {
 418                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 419                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
 420                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
 421                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 422                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 423                 } else if (error) {
 424                         CNETERR("Deleting noop packet\n");
 425                 }
 426
 427                 cfs_list_del (&tx->tx_list);
 428
 429                 LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1);
 430                 ksocknal_tx_done (ni, tx);
 431         }
 432 }
 433
 434 static void
 435 ksocknal_check_zc_req(ksock_tx_t *tx)
 436 {
 437         ksock_conn_t   *conn = tx->tx_conn;
 438         ksock_peer_t   *peer = conn->ksnc_peer;
 439
 440         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 441          * to ksnp_zc_req_list if some fragment of this message should be sent
 442          * zero-copy.  Our peer will send an ACK containing this cookie when
 443          * she has received this message to tell us we can signal completion.
 444          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 445          * ksnp_zc_req_list. */
 446         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 447         LASSERT (tx->tx_zc_capable);
 448
 449         tx->tx_zc_checked = 1;
 450
 451         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 452             !conn->ksnc_zc_capable)
 453                 return;
 454
 455         /* assign cookie and queue tx to pending list, it will be released when
 456          * a matching ack is received. See ksocknal_handle_zcack() */
 457
 458         ksocknal_tx_addref(tx);
 459
 460         cfs_spin_lock(&peer->ksnp_lock);
 461
 462         /* ZC_REQ is going to be pinned to the peer */
 463         tx->tx_deadline =
 464                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 465
 466         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 467
 468         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 469
 470         if (peer->ksnp_zc_next_cookie == 0)
 471                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 472
 473         cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 474
 475         cfs_spin_unlock(&peer->ksnp_lock);
 476 }
 477
 478 static void
 479 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 480 {
 481         ksock_peer_t   *peer = tx->tx_conn->ksnc_peer;
 482
 483         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 484         LASSERT (tx->tx_zc_capable);
 485
 486         tx->tx_zc_checked = 0;
 487
 488         cfs_spin_lock(&peer->ksnp_lock);
 489
 490         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 491                 /* Not waiting for an ACK */
 492                 cfs_spin_unlock(&peer->ksnp_lock);
 493                 return;
 494         }
 495
 496         tx->tx_msg.ksm_zc_cookies[0] = 0;
 497         cfs_list_del(&tx->tx_zc_list);
 498
 499         cfs_spin_unlock(&peer->ksnp_lock);
 500
 501         ksocknal_tx_decref(tx);
 502 }
 503
 504 int
 505 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 506 {
 507         int            rc;
 508
 509         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 510                 ksocknal_check_zc_req(tx);
 511
 512         rc = ksocknal_transmit (conn, tx);
 513
 514         CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 515
 516         if (tx->tx_resid == 0) {
 517                 /* Sent everything OK */
 518                 LASSERT (rc == 0);
 519
 520                 return (0);
 521         }
 522
 523         if (rc == -EAGAIN)
 524                 return (rc);
 525
 526         if (rc == -ENOMEM) {
 527                 static int counter;
 528
 529                 counter++;   /* exponential backoff warnings */
 530                 if ((counter & (-counter)) == counter)
 531                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 532                               counter, conn, cfs_atomic_read(&libcfs_kmemory));
 533
 534                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 535                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
 536
 537                 /* enomem list takes over scheduler's ref... */
 538                 LASSERT (conn->ksnc_tx_scheduled);
 539                 cfs_list_add_tail(&conn->ksnc_tx_list,
 540                                   &ksocknal_data.ksnd_enomem_conns);
 541                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 542                                                    SOCKNAL_ENOMEM_RETRY),
 543                                    ksocknal_data.ksnd_reaper_waketime))
 544                         cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
 545
 546                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
 547                 return (rc);
 548         }
 549
 550         /* Actual error */
 551         LASSERT (rc < 0);
 552
 553         if (!conn->ksnc_closing) {
 554                 switch (rc) {
 555                 case -ECONNRESET:
 556                         LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
 557                                       "while we were sending data; it may have "
 558                                       "rebooted.\n",
 559                                       HIPQUAD(conn->ksnc_ipaddr));
 560                         break;
 561                 default:
 562                         LCONSOLE_WARN("There was an unexpected network error "
 563                                       "while writing to %u.%u.%u.%u: %d.\n",
 564                                       HIPQUAD(conn->ksnc_ipaddr), rc);
 565                         break;
 566                 }
 567                 CDEBUG(D_NET, "[%p] Error %d on write to %s"
 568                        " ip %d.%d.%d.%d:%d\n", conn, rc,
 569                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 570                        HIPQUAD(conn->ksnc_ipaddr),
 571                        conn->ksnc_port);
 572         }
 573
 574         if (tx->tx_zc_checked)
 575                 ksocknal_uncheck_zc_req(tx);
 576
 577         /* it's not an error if conn is being closed */
 578         ksocknal_close_conn_and_siblings (conn,
 579                                           (conn->ksnc_closing) ? 0 : rc);
 580
 581         return (rc);
 582 }
 583
 584 void
 585 ksocknal_launch_connection_locked (ksock_route_t *route)
 586 {
 587
 588         /* called holding write lock on ksnd_global_lock */
 589
 590         LASSERT (!route->ksnr_scheduled);
 591         LASSERT (!route->ksnr_connecting);
 592         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 593
 594         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 595         ksocknal_route_addref(route);           /* extra ref for connd */
 596
 597         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
 598
 599         cfs_list_add_tail (&route->ksnr_connd_list,
 600                            &ksocknal_data.ksnd_connd_routes);
 601         cfs_waitq_signal (&ksocknal_data.ksnd_connd_waitq);
 602
 603         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
 604 }
 605
 606 void
 607 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 608 {
 609         ksock_route_t *route;
 610
 611         /* called holding write lock on ksnd_global_lock */
 612         for (;;) {
 613                 /* launch any/all connections that need it */
 614                 route = ksocknal_find_connectable_route_locked(peer);
 615                 if (route == NULL)
 616                         return;
 617
 618                 ksocknal_launch_connection_locked(route);
 619         }
 620 }
 621
 622 ksock_conn_t *
 623 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 624 {
 625         cfs_list_t       *tmp;
 626         ksock_conn_t     *conn;
 627         ksock_conn_t     *typed = NULL;
 628         ksock_conn_t     *fallback = NULL;
 629         int               tnob     = 0;
 630         int               fnob     = 0;
 631
 632         cfs_list_for_each (tmp, &peer->ksnp_conns) {
 633                 ksock_conn_t *c  = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
 634                 int           nob = cfs_atomic_read(&c->ksnc_tx_nob) +
 635                                     libcfs_sock_wmem_queued(c->ksnc_sock);
 636                 int           rc;
 637
 638                 LASSERT (!c->ksnc_closing);
 639                 LASSERT (c->ksnc_proto != NULL &&
 640                          c->ksnc_proto->pro_match_tx != NULL);
 641
 642                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 643
 644                 switch (rc) {
 645                 default:
 646                         LBUG();
 647                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 648                         continue;
 649
 650                 case SOCKNAL_MATCH_YES: /* typed connection */
 651                         if (typed == NULL || tnob > nob ||
 652                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 653                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 654                                 typed = c;
 655                                 tnob  = nob;
 656                         }
 657                         break;
 658
 659                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 660                         if (fallback == NULL || fnob > nob ||
 661                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 662                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 663                                 fallback = c;
 664                                 fnob     = nob;
 665                         }
 666                         break;
 667                 }
 668         }
 669
 670         /* prefer the typed selection */
 671         conn = (typed != NULL) ? typed : fallback;
 672
 673         if (conn != NULL)
 674                 conn->ksnc_tx_last_post = cfs_time_current();
 675
 676         return conn;
 677 }
 678
 679 void
 680 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 681 {
 682         conn->ksnc_proto->pro_pack(tx);
 683
 684         cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 685         ksocknal_conn_addref(conn); /* +1 ref for tx */
 686         tx->tx_conn = conn;
 687 }
 688
 689 void
 690 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
 691 {
 692         ksock_sched_t *sched = conn->ksnc_scheduler;
 693         ksock_msg_t   *msg = &tx->tx_msg;
 694         ksock_tx_t    *ztx = NULL;
 695         int            bufnob = 0;
 696
 697         /* called holding global lock (read or irq-write) and caller may
 698          * not have dropped this lock between finding conn and calling me,
 699          * so we don't need the {get,put}connsock dance to deref
 700          * ksnc_sock... */
 701         LASSERT(!conn->ksnc_closing);
 702
 703         CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n",
 704                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
 705                 HIPQUAD(conn->ksnc_ipaddr),
 706                 conn->ksnc_port);
 707
 708         ksocknal_tx_prep(conn, tx);
 709
 710         /* Ensure the frags we've been given EXACTLY match the number of
 711          * bytes we want to send.  Many TCP/IP stacks disregard any total
 712          * size parameters passed to them and just look at the frags.
 713          *
 714          * We always expect at least 1 mapped fragment containing the
 715          * complete ksocknal message header. */
 716         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 717                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 718                  (unsigned int)tx->tx_nob);
 719         LASSERT (tx->tx_niov >= 1);
 720         LASSERT (tx->tx_resid == tx->tx_nob);
 721
 722         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 723                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 724                                                KSOCK_MSG_NOOP,
 725                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 726
 727         /*
 728          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 729          * but they're used inside spinlocks a lot.
 730          */
 731         bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 732         cfs_spin_lock_bh (&sched->kss_lock);
 733
 734         if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 735                 /* First packet starts the timeout */
 736                 conn->ksnc_tx_deadline =
 737                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 738                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 739                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 740                 conn->ksnc_tx_bufnob = 0;
 741                 cfs_mb(); /* order with adding to tx_queue */
 742         }
 743
 744         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 745                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 746                  * on a normal packet so I don't need to send it */
 747                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 748                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 749
 750                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 751                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 752
 753         } else {
 754                 /* It's a normal packet - can it piggback a noop zc-ack that
 755                  * has been queued already? */
 756                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 757                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 758
 759                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 760                 /* ztx will be released later */
 761         }
 762
 763         if (ztx != NULL) {
 764                 cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 765                 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 766         }
 767
 768         if (conn->ksnc_tx_ready &&      /* able to send */
 769             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 770                 /* +1 ref for scheduler */
 771                 ksocknal_conn_addref(conn);
 772                 cfs_list_add_tail (&conn->ksnc_tx_list,
 773                                    &sched->kss_tx_conns);
 774                 conn->ksnc_tx_scheduled = 1;
 775                 cfs_waitq_signal (&sched->kss_waitq);
 776         }
 777
 778         cfs_spin_unlock_bh (&sched->kss_lock);
 779 }
 780
 781
 782 ksock_route_t *
 783 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 784 {
 785         cfs_time_t     now = cfs_time_current();
 786         cfs_list_t    *tmp;
 787         ksock_route_t *route;
 788
 789         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 790                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 791
 792                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 793
 794                 if (route->ksnr_scheduled)      /* connections being established */
 795                         continue;
 796
 797                 /* all route types connected ? */
 798                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 799                         continue;
 800
 801                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 802                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 803                         CDEBUG(D_NET,
 804                                "Too soon to retry route %u.%u.%u.%u "
 805                                "(cnted %d, interval %ld, %ld secs later)\n",
 806                                HIPQUAD(route->ksnr_ipaddr),
 807                                route->ksnr_connected,
 808                                route->ksnr_retry_interval,
 809                                cfs_duration_sec(route->ksnr_timeout - now));
 810                         continue;
 811                 }
 812
 813                 return (route);
 814         }
 815
 816         return (NULL);
 817 }
 818
 819 ksock_route_t *
 820 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 821 {
 822         cfs_list_t        *tmp;
 823         ksock_route_t     *route;
 824
 825         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 826                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 827
 828                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 829
 830                 if (route->ksnr_scheduled)
 831                         return (route);
 832         }
 833
 834         return (NULL);
 835 }
 836
 837 int
 838 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 839 {
 840         ksock_peer_t     *peer;
 841         ksock_conn_t     *conn;
 842         cfs_rwlock_t     *g_lock;
 843         int               retry;
 844         int               rc;
 845
 846         LASSERT (tx->tx_conn == NULL);
 847
 848         g_lock = &ksocknal_data.ksnd_global_lock;
 849
 850         for (retry = 0;; retry = 1) {
 851                 cfs_read_lock (g_lock);
 852                 peer = ksocknal_find_peer_locked(ni, id);
 853                 if (peer != NULL) {
 854                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 855                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 856                                 if (conn != NULL) {
 857                                         /* I've got no routes that need to be
 858                                          * connecting and I do have an actual
 859                                          * connection... */
 860                                         ksocknal_queue_tx_locked (tx, conn);
 861                                         cfs_read_unlock (g_lock);
 862                                         return (0);
 863                                 }
 864                         }
 865                 }
 866
 867                 /* I'll need a write lock... */
 868                 cfs_read_unlock (g_lock);
 869
 870                 cfs_write_lock_bh (g_lock);
 871
 872                 peer = ksocknal_find_peer_locked(ni, id);
 873                 if (peer != NULL)
 874                         break;
 875
 876                 cfs_write_unlock_bh (g_lock);
 877
 878                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 879                         CERROR("Refusing to create a connection to "
 880                                "userspace process %s\n", libcfs_id2str(id));
 881                         return -EHOSTUNREACH;
 882                 }
 883
 884                 if (retry) {
 885                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 886                         return -EHOSTUNREACH;
 887                 }
 888
 889                 rc = ksocknal_add_peer(ni, id,
 890                                        LNET_NIDADDR(id.nid),
 891                                        lnet_acceptor_port());
 892                 if (rc != 0) {
 893                         CERROR("Can't add peer %s: %d\n",
 894                                libcfs_id2str(id), rc);
 895                         return rc;
 896                 }
 897         }
 898
 899         ksocknal_launch_all_connections_locked(peer);
 900
 901         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 902         if (conn != NULL) {
 903                 /* Connection exists; queue message on it */
 904                 ksocknal_queue_tx_locked (tx, conn);
 905                 cfs_write_unlock_bh (g_lock);
 906                 return (0);
 907         }
 908
 909         if (peer->ksnp_accepting > 0 ||
 910             ksocknal_find_connecting_route_locked (peer) != NULL) {
 911                 /* the message is going to be pinned to the peer */
 912                 tx->tx_deadline =
 913                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 914
 915                 /* Queue the message until a connection is established */
 916                 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
 917                 cfs_write_unlock_bh (g_lock);
 918                 return 0;
 919         }
 920
 921         cfs_write_unlock_bh (g_lock);
 922
 923         /* NB Routes may be ignored if connections to them failed recently */
 924         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 925         return (-EHOSTUNREACH);
 926 }
 927
 928 int
 929 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 930 {
 931         int               mpflag = 0;
 932         int               type = lntmsg->msg_type;
 933         lnet_process_id_t target = lntmsg->msg_target;
 934         unsigned int      payload_niov = lntmsg->msg_niov;
 935         struct iovec     *payload_iov = lntmsg->msg_iov;
 936         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 937         unsigned int      payload_offset = lntmsg->msg_offset;
 938         unsigned int      payload_nob = lntmsg->msg_len;
 939         ksock_tx_t       *tx;
 940         int               desc_size;
 941         int               rc;
 942
 943         /* NB 'private' is different depending on what we're sending.
 944          * Just ignore it... */
 945
 946         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 947                payload_nob, payload_niov, libcfs_id2str(target));
 948
 949         LASSERT (payload_nob == 0 || payload_niov > 0);
 950         LASSERT (payload_niov <= LNET_MAX_IOV);
 951         /* payload is either all vaddrs or all pages */
 952         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 953         LASSERT (!cfs_in_interrupt ());
 954
 955         if (payload_iov != NULL)
 956                 desc_size = offsetof(ksock_tx_t,
 957                                      tx_frags.virt.iov[1 + payload_niov]);
 958         else
 959                 desc_size = offsetof(ksock_tx_t,
 960                                      tx_frags.paged.kiov[payload_niov]);
 961
 962         if (lntmsg->msg_vmflush)
 963                 mpflag = cfs_memory_pressure_get_and_set();
 964         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 965         if (tx == NULL) {
 966                 CERROR("Can't allocate tx desc type %d size %d\n",
 967                        type, desc_size);
 968                 if (lntmsg->msg_vmflush)
 969                         cfs_memory_pressure_restore(mpflag);
 970                 return (-ENOMEM);
 971         }
 972
 973         tx->tx_conn = NULL;                     /* set when assigned a conn */
 974         tx->tx_lnetmsg = lntmsg;
 975
 976         if (payload_iov != NULL) {
 977                 tx->tx_kiov = NULL;
 978                 tx->tx_nkiov = 0;
 979                 tx->tx_iov = tx->tx_frags.virt.iov;
 980                 tx->tx_niov = 1 +
 981                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 982                                                payload_niov, payload_iov,
 983                                                payload_offset, payload_nob);
 984         } else {
 985                 tx->tx_niov = 1;
 986                 tx->tx_iov = &tx->tx_frags.paged.iov;
 987                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 988                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 989                                                  payload_niov, payload_kiov,
 990                                                  payload_offset, payload_nob);
 991
 992                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 993                         tx->tx_zc_capable = 1;
 994         }
 995
 996         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 997
 998         /* The first fragment will be set later in pro_pack */
 999         rc = ksocknal_launch_packet(ni, tx, target);
1000         if (lntmsg->msg_vmflush)
1001                 cfs_memory_pressure_restore(mpflag);
1002         if (rc == 0)
1003                 return (0);
1004
1005         ksocknal_free_tx(tx);
1006         return (-EIO);
1007 }
1008
1009 int
1010 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1011 {
1012         long          pid = cfs_kernel_thread (fn, arg, 0);
1013
1014         if (pid < 0)
1015                 return ((int)pid);
1016
1017         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1018         ksocknal_data.ksnd_nthreads++;
1019         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1020         return (0);
1021 }
1022
1023 void
1024 ksocknal_thread_fini (void)
1025 {
1026         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1027         ksocknal_data.ksnd_nthreads--;
1028         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1029 }
1030
1031 int
1032 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1033 {
1034         static char ksocknal_slop_buffer[4096];
1035
1036         int            nob;
1037         unsigned int   niov;
1038         int            skipped;
1039
1040         LASSERT(conn->ksnc_proto != NULL);
1041
1042         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1043                 /* Remind the socket to ack eagerly... */
1044                 ksocknal_lib_eager_ack(conn);
1045         }
1046
1047         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1048                 conn->ksnc_rx_started = 0;
1049                 cfs_mb();                       /* racing with timeout thread */
1050
1051                 switch (conn->ksnc_proto->pro_version) {
1052                 case  KSOCK_PROTO_V2:
1053                 case  KSOCK_PROTO_V3:
1054                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1055                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1056                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1057
1058                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1059                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1060                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1061                         break;
1062
1063                 case KSOCK_PROTO_V1:
1064                         /* Receiving bare lnet_hdr_t */
1065                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1066                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1067                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1068
1069                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1070                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1071                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1072                         break;
1073
1074                 default:
1075                         LBUG ();
1076                 }
1077                 conn->ksnc_rx_niov = 1;
1078
1079                 conn->ksnc_rx_kiov = NULL;
1080                 conn->ksnc_rx_nkiov = 0;
1081                 conn->ksnc_rx_csum = ~0;
1082                 return (1);
1083         }
1084
1085         /* Set up to skip as much as possible now.  If there's more left
1086          * (ran out of iov entries) we'll get called again */
1087
1088         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1089         conn->ksnc_rx_nob_left = nob_to_skip;
1090         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1091         skipped = 0;
1092         niov = 0;
1093
1094         do {
1095                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1096
1097                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1098                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1099                 niov++;
1100                 skipped += nob;
1101                 nob_to_skip -=nob;
1102
1103         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1104                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1105
1106         conn->ksnc_rx_niov = niov;
1107         conn->ksnc_rx_kiov = NULL;
1108         conn->ksnc_rx_nkiov = 0;
1109         conn->ksnc_rx_nob_wanted = skipped;
1110         return (0);
1111 }
1112
1113 int
1114 ksocknal_process_receive (ksock_conn_t *conn)
1115 {
1116         lnet_hdr_t        *lhdr;
1117         lnet_process_id_t *id;
1118         int                rc;
1119
1120         LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0);
1121
1122         /* NB: sched lock NOT held */
1123         /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1124         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1125                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1126                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1127                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1128  again:
1129         if (conn->ksnc_rx_nob_wanted != 0) {
1130                 rc = ksocknal_receive(conn);
1131
1132                 if (rc <= 0) {
1133                         LASSERT (rc != -EAGAIN);
1134
1135                         if (rc == 0)
1136                                 CDEBUG (D_NET, "[%p] EOF from %s"
1137                                         " ip %d.%d.%d.%d:%d\n", conn,
1138                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1139                                         HIPQUAD(conn->ksnc_ipaddr),
1140                                         conn->ksnc_port);
1141                         else if (!conn->ksnc_closing)
1142                                 CERROR ("[%p] Error %d on read from %s"
1143                                         " ip %d.%d.%d.%d:%d\n",
1144                                         conn, rc,
1145                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1146                                         HIPQUAD(conn->ksnc_ipaddr),
1147                                         conn->ksnc_port);
1148
1149                         /* it's not an error if conn is being closed */
1150                         ksocknal_close_conn_and_siblings (conn,
1151                                                           (conn->ksnc_closing) ? 0 : rc);
1152                         return (rc == 0 ? -ESHUTDOWN : rc);
1153                 }
1154
1155                 if (conn->ksnc_rx_nob_wanted != 0) {
1156                         /* short read */
1157                         return (-EAGAIN);
1158                 }
1159         }
1160         switch (conn->ksnc_rx_state) {
1161         case SOCKNAL_RX_KSM_HEADER:
1162                 if (conn->ksnc_flip) {
1163                         __swab32s(&conn->ksnc_msg.ksm_type);
1164                         __swab32s(&conn->ksnc_msg.ksm_csum);
1165                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1166                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1167                 }
1168
1169                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1170                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1171                         CERROR("%s: Unknown message type: %x\n",
1172                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1173                                conn->ksnc_msg.ksm_type);
1174                         ksocknal_new_packet(conn, 0);
1175                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1176                         return (-EPROTO);
1177                 }
1178
1179                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1180                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1181                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1182                         /* NOOP Checksum error */
1183                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1184                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1185                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1186                         ksocknal_new_packet(conn, 0);
1187                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1188                         return (-EIO);
1189                 }
1190
1191                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1192                         __u64 cookie = 0;
1193
1194                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1195
1196                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1197                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1198
1199                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1200                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1201
1202                         if (rc != 0) {
1203                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1204                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1205                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1206                                 ksocknal_new_packet(conn, 0);
1207                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1208                                 return (rc);
1209                         }
1210                 }
1211
1212                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1213                         ksocknal_new_packet (conn, 0);
1214                         return 0;       /* NOOP is done and just return */
1215                 }
1216
1217                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1218                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1219                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1220
1221                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1222                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1223                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1224
1225                 conn->ksnc_rx_niov = 1;
1226                 conn->ksnc_rx_kiov = NULL;
1227                 conn->ksnc_rx_nkiov = 0;
1228
1229                 goto again;     /* read lnet header now */
1230
1231         case SOCKNAL_RX_LNET_HEADER:
1232                 /* unpack message header */
1233                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1234
1235                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1236                         /* Userspace peer */
1237                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1238                         id   = &conn->ksnc_peer->ksnp_id;
1239
1240                         /* Substitute process ID assigned at connection time */
1241                         lhdr->src_pid = cpu_to_le32(id->pid);
1242                         lhdr->src_nid = cpu_to_le64(id->nid);
1243                 }
1244
1245                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1246                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1247
1248                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1249                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1250                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1251                 if (rc < 0) {
1252                         /* I just received garbage: give up on this conn */
1253                         ksocknal_new_packet(conn, 0);
1254                         ksocknal_close_conn_and_siblings (conn, rc);
1255                         ksocknal_conn_decref(conn);
1256                         return (-EPROTO);
1257                 }
1258
1259                 /* I'm racing with ksocknal_recv() */
1260                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1261                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1262
1263                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1264                         return 0;
1265
1266                 /* ksocknal_recv() got called */
1267                 goto again;
1268
1269         case SOCKNAL_RX_LNET_PAYLOAD:
1270                 /* payload all received */
1271                 rc = 0;
1272
1273                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1274                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1275                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1276                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1277                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1278                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1279                         rc = -EIO;
1280                 }
1281
1282                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1283                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1284
1285                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1286                         id   = &conn->ksnc_peer->ksnp_id;
1287
1288                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1289                                         conn->ksnc_msg.ksm_zc_cookies[0],
1290                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1291                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1292                 }
1293
1294                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1295
1296                 if (rc != 0) {
1297                         ksocknal_new_packet(conn, 0);
1298                         ksocknal_close_conn_and_siblings (conn, rc);
1299                         return (-EPROTO);
1300                 }
1301                 /* Fall through */
1302
1303         case SOCKNAL_RX_SLOP:
1304                 /* starting new packet? */
1305                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1306                         return 0;       /* come back later */
1307                 goto again;             /* try to finish reading slop now */
1308
1309         default:
1310                 break;
1311         }
1312
1313         /* Not Reached */
1314         LBUG ();
1315         return (-EINVAL);                       /* keep gcc happy */
1316 }
1317
1318 int
1319 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1320                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1321                unsigned int offset, unsigned int mlen, unsigned int rlen)
1322 {
1323         ksock_conn_t  *conn = (ksock_conn_t *)private;
1324         ksock_sched_t *sched = conn->ksnc_scheduler;
1325
1326         LASSERT (mlen <= rlen);
1327         LASSERT (niov <= LNET_MAX_IOV);
1328
1329         conn->ksnc_cookie = msg;
1330         conn->ksnc_rx_nob_wanted = mlen;
1331         conn->ksnc_rx_nob_left   = rlen;
1332
1333         if (mlen == 0 || iov != NULL) {
1334                 conn->ksnc_rx_nkiov = 0;
1335                 conn->ksnc_rx_kiov = NULL;
1336                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1337                 conn->ksnc_rx_niov =
1338                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1339                                          niov, iov, offset, mlen);
1340         } else {
1341                 conn->ksnc_rx_niov = 0;
1342                 conn->ksnc_rx_iov  = NULL;
1343                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1344                 conn->ksnc_rx_nkiov =
1345                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1346                                           niov, kiov, offset, mlen);
1347         }
1348
1349         LASSERT (mlen ==
1350                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1351                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1352
1353         LASSERT (conn->ksnc_rx_scheduled);
1354
1355         cfs_spin_lock_bh (&sched->kss_lock);
1356
1357         switch (conn->ksnc_rx_state) {
1358         case SOCKNAL_RX_PARSE_WAIT:
1359                 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1360                 cfs_waitq_signal (&sched->kss_waitq);
1361                 LASSERT (conn->ksnc_rx_ready);
1362                 break;
1363
1364         case SOCKNAL_RX_PARSE:
1365                 /* scheduler hasn't noticed I'm parsing yet */
1366                 break;
1367         }
1368
1369         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1370
1371         cfs_spin_unlock_bh (&sched->kss_lock);
1372         ksocknal_conn_decref(conn);
1373         return (0);
1374 }
1375
1376 static inline int
1377 ksocknal_sched_cansleep(ksock_sched_t *sched)
1378 {
1379         int           rc;
1380
1381         cfs_spin_lock_bh (&sched->kss_lock);
1382
1383         rc = (!ksocknal_data.ksnd_shuttingdown &&
1384               cfs_list_empty(&sched->kss_rx_conns) &&
1385               cfs_list_empty(&sched->kss_tx_conns));
1386
1387         cfs_spin_unlock_bh (&sched->kss_lock);
1388         return (rc);
1389 }
1390
1391 int ksocknal_scheduler (void *arg)
1392 {
1393         ksock_sched_t     *sched = (ksock_sched_t *)arg;
1394         ksock_conn_t      *conn;
1395         ksock_tx_t        *tx;
1396         int                rc;
1397         int                nloops = 0;
1398         int                id = (int)(sched - ksocknal_data.ksnd_schedulers);
1399         char               name[16];
1400
1401         snprintf (name, sizeof (name),"socknal_sd%02d", id);
1402         cfs_daemonize (name);
1403         cfs_block_allsigs ();
1404
1405         if (ksocknal_lib_bind_thread_to_cpu(id))
1406                 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1407
1408         cfs_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 (!cfs_list_empty (&sched->kss_rx_conns)) {
1416                         conn = cfs_list_entry(sched->kss_rx_conns.next,
1417                                               ksock_conn_t, ksnc_rx_list);
1418                         cfs_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                         cfs_spin_unlock_bh (&sched->kss_lock);
1429
1430                         rc = ksocknal_process_receive(conn);
1431
1432                         cfs_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                                 cfs_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 (!cfs_list_empty (&sched->kss_tx_conns)) {
1460                         CFS_LIST_HEAD    (zlist);
1461
1462                         if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1463                                 cfs_list_add(&zlist,
1464                                              &sched->kss_zombie_noop_txs);
1465                                 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1466                         }
1467
1468                         conn = cfs_list_entry(sched->kss_tx_conns.next,
1469                                               ksock_conn_t, ksnc_tx_list);
1470                         cfs_list_del (&conn->ksnc_tx_list);
1471
1472                         LASSERT(conn->ksnc_tx_scheduled);
1473                         LASSERT(conn->ksnc_tx_ready);
1474                         LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1475
1476                         tx = cfs_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                         cfs_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                         cfs_spin_unlock_bh (&sched->kss_lock);
1491
1492                         if (!cfs_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                                 cfs_spin_lock_bh (&sched->kss_lock);
1503                                 cfs_list_add (&tx->tx_list,
1504                                               &conn->ksnc_tx_queue);
1505                         } else {
1506                                 /* Complete send; tx -ref */
1507                                 ksocknal_tx_decref (tx);
1508
1509                                 cfs_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                                    !cfs_list_empty (&conn->ksnc_tx_queue)) {
1519                                 /* reschedule for tx */
1520                                 cfs_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                         cfs_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                                 cfs_cond_resched();
1543                         }
1544
1545                         cfs_spin_lock_bh (&sched->kss_lock);
1546                 }
1547         }
1548
1549         cfs_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         ENTRY;
1562
1563         sched = conn->ksnc_scheduler;
1564
1565         cfs_spin_lock_bh (&sched->kss_lock);
1566
1567         conn->ksnc_rx_ready = 1;
1568
1569         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1570                 cfs_list_add_tail(&conn->ksnc_rx_list,
1571                                   &sched->kss_rx_conns);
1572                 conn->ksnc_rx_scheduled = 1;
1573                 /* extra ref for scheduler */
1574                 ksocknal_conn_addref(conn);
1575
1576                 cfs_waitq_signal (&sched->kss_waitq);
1577         }
1578         cfs_spin_unlock_bh (&sched->kss_lock);
1579
1580         EXIT;
1581 }
1582
1583 /*
1584  * Add connection to kss_tx_conns of scheduler
1585  * and wakeup the scheduler.
1586  */
1587 void ksocknal_write_callback (ksock_conn_t *conn)
1588 {
1589         ksock_sched_t *sched;
1590         ENTRY;
1591
1592         sched = conn->ksnc_scheduler;
1593
1594         cfs_spin_lock_bh (&sched->kss_lock);
1595
1596         conn->ksnc_tx_ready = 1;
1597
1598         if (!conn->ksnc_tx_scheduled && // not being progressed
1599             !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1600                 cfs_list_add_tail (&conn->ksnc_tx_list,
1601                                    &sched->kss_tx_conns);
1602                 conn->ksnc_tx_scheduled = 1;
1603                 /* extra ref for scheduler */
1604                 ksocknal_conn_addref(conn);
1605
1606                 cfs_waitq_signal (&sched->kss_waitq);
1607         }
1608
1609         cfs_spin_unlock_bh (&sched->kss_lock);
1610
1611         EXIT;
1612 }
1613
1614 ksock_proto_t *
1615 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1616 {
1617         __u32   version = 0;
1618
1619         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1620                 version = hello->kshm_version;
1621         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1622                 version = __swab32(hello->kshm_version);
1623
1624         if (version != 0) {
1625 #if SOCKNAL_VERSION_DEBUG
1626                 if (*ksocknal_tunables.ksnd_protocol == 1)
1627                         return NULL;
1628
1629                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1630                     version == KSOCK_PROTO_V3)
1631                         return NULL;
1632 #endif
1633                 if (version == KSOCK_PROTO_V2)
1634                         return &ksocknal_protocol_v2x;
1635
1636                 if (version == KSOCK_PROTO_V3)
1637                         return &ksocknal_protocol_v3x;
1638
1639                 return NULL;
1640         }
1641
1642         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1643                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1644
1645                 CLASSERT (sizeof (lnet_magicversion_t) ==
1646                           offsetof (ksock_hello_msg_t, kshm_src_nid));
1647
1648                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1649                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1650                         return &ksocknal_protocol_v1x;
1651         }
1652
1653         return NULL;
1654 }
1655
1656 int
1657 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1658                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1659 {
1660         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1661         ksock_net_t         *net = (ksock_net_t *)ni->ni_data;
1662
1663         LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1664
1665         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1666         LASSERT (conn->ksnc_proto != NULL);
1667
1668         hello->kshm_src_nid         = ni->ni_nid;
1669         hello->kshm_dst_nid         = peer_nid;
1670         hello->kshm_src_pid         = the_lnet.ln_pid;
1671
1672         hello->kshm_src_incarnation = net->ksnn_incarnation;
1673         hello->kshm_ctype           = conn->ksnc_type;
1674
1675         return conn->ksnc_proto->pro_send_hello(conn, hello);
1676 }
1677
1678 int
1679 ksocknal_invert_type(int type)
1680 {
1681         switch (type)
1682         {
1683         case SOCKLND_CONN_ANY:
1684         case SOCKLND_CONN_CONTROL:
1685                 return (type);
1686         case SOCKLND_CONN_BULK_IN:
1687                 return SOCKLND_CONN_BULK_OUT;
1688         case SOCKLND_CONN_BULK_OUT:
1689                 return SOCKLND_CONN_BULK_IN;
1690         default:
1691                 return (SOCKLND_CONN_NONE);
1692         }
1693 }
1694
1695 int
1696 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1697                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1698                      __u64 *incarnation)
1699 {
1700         /* Return < 0        fatal error
1701          *        0          success
1702          *        EALREADY   lost connection race
1703          *        EPROTO     protocol version mismatch
1704          */
1705         cfs_socket_t        *sock = conn->ksnc_sock;
1706         int                  active = (conn->ksnc_proto != NULL);
1707         int                  timeout;
1708         int                  proto_match;
1709         int                  rc;
1710         ksock_proto_t       *proto;
1711         lnet_process_id_t    recv_id;
1712
1713         /* socket type set on active connections - not set on passive */
1714         LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1715
1716         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1717                             lnet_acceptor_timeout();
1718
1719         rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1720         if (rc != 0) {
1721                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1722                         rc, HIPQUAD(conn->ksnc_ipaddr));
1723                 LASSERT (rc < 0);
1724                 return rc;
1725         }
1726
1727         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1728             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1729             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1730                 /* Unexpected magic! */
1731                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1732                         "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1733                         LNET_PROTO_TCP_MAGIC,
1734                         HIPQUAD(conn->ksnc_ipaddr));
1735                 return -EPROTO;
1736         }
1737
1738         rc = libcfs_sock_read(sock, &hello->kshm_version,
1739                               sizeof(hello->kshm_version), timeout);
1740         if (rc != 0) {
1741                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1742                         rc, HIPQUAD(conn->ksnc_ipaddr));
1743                 LASSERT (rc < 0);
1744                 return rc;
1745         }
1746
1747         proto = ksocknal_parse_proto_version(hello);
1748         if (proto == NULL) {
1749                 if (!active) {
1750                         /* unknown protocol from peer, tell peer my protocol */
1751                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1752 #if SOCKNAL_VERSION_DEBUG
1753                         if (*ksocknal_tunables.ksnd_protocol == 2)
1754                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1755                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1756                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1757 #endif
1758                         hello->kshm_nips = 0;
1759                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1760                 }
1761
1762                 CERROR ("Unknown protocol version (%d.x expected)"
1763                         " from %u.%u.%u.%u\n",
1764                         conn->ksnc_proto->pro_version,
1765                         HIPQUAD(conn->ksnc_ipaddr));
1766
1767                 return -EPROTO;
1768         }
1769
1770         proto_match = (conn->ksnc_proto == proto);
1771         conn->ksnc_proto = proto;
1772
1773         /* receive the rest of hello message anyway */
1774         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1775         if (rc != 0) {
1776                 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1777                        rc, HIPQUAD(conn->ksnc_ipaddr));
1778                 LASSERT (rc < 0);
1779                 return rc;
1780         }
1781
1782         *incarnation = hello->kshm_src_incarnation;
1783
1784         if (hello->kshm_src_nid == LNET_NID_ANY) {
1785                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1786                        "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1787                 return -EPROTO;
1788         }
1789
1790         if (!active &&
1791             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1792                 /* Userspace NAL assigns peer process ID from socket */
1793                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1794                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1795         } else {
1796                 recv_id.nid = hello->kshm_src_nid;
1797                 recv_id.pid = hello->kshm_src_pid;
1798         }
1799
1800         if (!active) {
1801                 *peerid = recv_id;
1802
1803                 /* peer determines type */
1804                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1805                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1806                         CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1807                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1808                                 HIPQUAD(conn->ksnc_ipaddr));
1809                         return -EPROTO;
1810                 }
1811
1812                 return 0;
1813         }
1814
1815         if (peerid->pid != recv_id.pid ||
1816             peerid->nid != recv_id.nid) {
1817                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1818                                    " %u.%u.%u.%u, but they claimed they were "
1819                                    "%s; please check your Lustre "
1820                                    "configuration.\n",
1821                                    libcfs_id2str(*peerid),
1822                                    HIPQUAD(conn->ksnc_ipaddr),
1823                                    libcfs_id2str(recv_id));
1824                 return -EPROTO;
1825         }
1826
1827         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1828                 /* Possible protocol mismatch or I lost the connection race */
1829                 return proto_match ? EALREADY : EPROTO;
1830         }
1831
1832         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1833                 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1834                         conn->ksnc_type, libcfs_id2str(*peerid),
1835                         HIPQUAD(conn->ksnc_ipaddr),
1836                         hello->kshm_ctype);
1837                 return -EPROTO;
1838         }
1839
1840         return 0;
1841 }
1842
1843 int
1844 ksocknal_connect (ksock_route_t *route)
1845 {
1846         CFS_LIST_HEAD    (zombies);
1847         ksock_peer_t     *peer = route->ksnr_peer;
1848         int               type;
1849         int               wanted;
1850         cfs_socket_t     *sock;
1851         cfs_time_t        deadline;
1852         int               retry_later = 0;
1853         int               rc = 0;
1854
1855         deadline = cfs_time_add(cfs_time_current(),
1856                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1857
1858         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1859
1860         LASSERT (route->ksnr_scheduled);
1861         LASSERT (!route->ksnr_connecting);
1862
1863         route->ksnr_connecting = 1;
1864
1865         for (;;) {
1866                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1867
1868                 /* stop connecting if peer/route got closed under me, or
1869                  * route got connected while queued */
1870                 if (peer->ksnp_closing || route->ksnr_deleted ||
1871                     wanted == 0) {
1872                         retry_later = 0;
1873                         break;
1874                 }
1875
1876                 /* reschedule if peer is connecting to me */
1877                 if (peer->ksnp_accepting > 0) {
1878                         CDEBUG(D_NET,
1879                                "peer %s(%d) already connecting to me, retry later.\n",
1880                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1881                         retry_later = 1;
1882                 }
1883
1884                 if (retry_later) /* needs reschedule */
1885                         break;
1886
1887                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1888                         type = SOCKLND_CONN_ANY;
1889                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1890                         type = SOCKLND_CONN_CONTROL;
1891                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1892                         type = SOCKLND_CONN_BULK_IN;
1893                 } else {
1894                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1895                         type = SOCKLND_CONN_BULK_OUT;
1896                 }
1897
1898                 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1899
1900                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1901                         rc = -ETIMEDOUT;
1902                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1903                                                    route->ksnr_ipaddr,
1904                                                    route->ksnr_port);
1905                         goto failed;
1906                 }
1907
1908                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1909                                   route->ksnr_myipaddr,
1910                                   route->ksnr_ipaddr, route->ksnr_port);
1911                 if (rc != 0)
1912                         goto failed;
1913
1914                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1915                 if (rc < 0) {
1916                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1917                                                    route->ksnr_ipaddr,
1918                                                    route->ksnr_port);
1919                         goto failed;
1920                 }
1921
1922                 /* A +ve RC means I have to retry because I lost the connection
1923                  * race or I have to renegotiate protocol version */
1924                 retry_later = (rc != 0);
1925                 if (retry_later)
1926                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1927                                libcfs_nid2str(peer->ksnp_id.nid));
1928
1929                 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1930         }
1931
1932         route->ksnr_scheduled = 0;
1933         route->ksnr_connecting = 0;
1934
1935         if (retry_later) {
1936                 /* re-queue for attention; this frees me up to handle
1937                  * the peer's incoming connection request */
1938
1939                 if (rc == EALREADY ||
1940                     (rc == 0 && peer->ksnp_accepting > 0)) {
1941                         /* We want to introduce a delay before next
1942                          * attempt to connect if we lost conn race,
1943                          * but the race is resolved quickly usually,
1944                          * so min_reconnectms should be good heuristic */
1945                         route->ksnr_retry_interval =
1946                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1947                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1948                                                            route->ksnr_retry_interval);
1949                 }
1950
1951                 ksocknal_launch_connection_locked(route);
1952         }
1953
1954         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1955         return retry_later;
1956
1957  failed:
1958         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1959
1960         route->ksnr_scheduled = 0;
1961         route->ksnr_connecting = 0;
1962
1963         /* This is a retry rather than a new connection */
1964         route->ksnr_retry_interval *= 2;
1965         route->ksnr_retry_interval =
1966                 MAX(route->ksnr_retry_interval,
1967                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1968         route->ksnr_retry_interval =
1969                 MIN(route->ksnr_retry_interval,
1970                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1971
1972         LASSERT (route->ksnr_retry_interval != 0);
1973         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1974                                            route->ksnr_retry_interval);
1975
1976         if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1977             peer->ksnp_accepting == 0 &&
1978             ksocknal_find_connecting_route_locked(peer) == NULL) {
1979                 ksock_conn_t *conn;
1980
1981                 /* ksnp_tx_queue is queued on a conn on successful
1982                  * connection for V1.x and V2.x */
1983                 if (!cfs_list_empty (&peer->ksnp_conns)) {
1984                         conn = cfs_list_entry(peer->ksnp_conns.next,
1985                                               ksock_conn_t, ksnc_list);
1986                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1987                 }
1988
1989                 /* take all the blocked packets while I've got the lock and
1990                  * complete below... */
1991                 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
1992         }
1993
1994 #if 0           /* irrelevent with only eager routes */
1995         if (!route->ksnr_deleted) {
1996                 /* make this route least-favourite for re-selection */
1997                 cfs_list_del(&route->ksnr_list);
1998                 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1999         }
2000 #endif
2001         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2002
2003         ksocknal_peer_failed(peer);
2004         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2005         return 0;
2006 }
2007
2008 /* Go through connd_routes queue looking for a route that
2009    we can process right now */
2010 static ksock_route_t *
2011 ksocknal_connd_get_route_locked(signed long *timeout_p)
2012 {
2013         ksock_route_t *route;
2014         cfs_time_t     now;
2015
2016         /* Only handle an outgoing connection request if there
2017          * is someone left to handle incoming connections */
2018         if ((ksocknal_data.ksnd_connd_connecting + 1) >=
2019             *ksocknal_tunables.ksnd_nconnds)
2020                 return NULL;
2021
2022         now = cfs_time_current();
2023
2024         /* connd_routes can contain both pending and ordinary routes */
2025         cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2026                                  ksnr_connd_list) {
2027
2028                 if (route->ksnr_retry_interval == 0 ||
2029                     cfs_time_aftereq(now, route->ksnr_timeout))
2030                         return route;
2031
2032                 if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT ||
2033                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2034                         *timeout_p = (int)(route->ksnr_timeout - now);
2035         }
2036
2037         return NULL;
2038 }
2039
2040 int
2041 ksocknal_connd (void *arg)
2042 {
2043         long               id = (long)(long_ptr_t)arg;
2044         char               name[16];
2045         ksock_connreq_t   *cr;
2046         ksock_route_t     *route;
2047         cfs_waitlink_t     wait;
2048         signed long        timeout;
2049         int                nloops = 0;
2050         int                cons_retry = 0;
2051         int                dropped_lock;
2052
2053         snprintf (name, sizeof (name), "socknal_cd%02ld", id);
2054         cfs_daemonize (name);
2055         cfs_block_allsigs ();
2056
2057         cfs_waitlink_init (&wait);
2058
2059         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2060
2061         while (!ksocknal_data.ksnd_shuttingdown) {
2062
2063                 dropped_lock = 0;
2064
2065                 if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2066                         /* Connection accepted by the listener */
2067                         cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \
2068                                             next, ksock_connreq_t, ksncr_list);
2069
2070                         cfs_list_del(&cr->ksncr_list);
2071                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2072                         dropped_lock = 1;
2073
2074                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2075                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2076                         lnet_ni_decref(cr->ksncr_ni);
2077                         LIBCFS_FREE(cr, sizeof(*cr));
2078
2079                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2080                 }
2081
2082                 /* Sleep till explicit wake_up if no pending routes present */
2083                 timeout = CFS_MAX_SCHEDULE_TIMEOUT;
2084
2085                 /* Connection request */
2086                 route = ksocknal_connd_get_route_locked(&timeout);
2087
2088                 if (route != NULL) {
2089                         cfs_list_del (&route->ksnr_connd_list);
2090                         ksocknal_data.ksnd_connd_connecting++;
2091                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2092                         dropped_lock = 1;
2093
2094                         if (ksocknal_connect(route)) {
2095                                 /* consecutive retry */
2096                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2097                                         CWARN("massive consecutive "
2098                                               "re-connecting to %u.%u.%u.%u\n",
2099                                               HIPQUAD(route->ksnr_ipaddr));
2100                                         cons_retry = 0;
2101                                 }
2102                         } else {
2103                                 cons_retry = 0;
2104                         }
2105
2106                         ksocknal_route_decref(route);
2107
2108                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2109                         ksocknal_data.ksnd_connd_connecting--;
2110                 }
2111
2112                 if (dropped_lock) {
2113                         if (++nloops < SOCKNAL_RESCHED)
2114                                 continue;
2115                         cfs_spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2116                         nloops = 0;
2117                         cfs_cond_resched();
2118                         cfs_spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2119                         continue;
2120                 }
2121
2122                 /* Nothing to do for 'timeout'  */
2123                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2124                 cfs_waitq_add_exclusive (&ksocknal_data.ksnd_connd_waitq,
2125                                          &wait);
2126                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2127
2128                 nloops = 0;
2129                 cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, timeout);
2130
2131                 cfs_set_current_state (CFS_TASK_RUNNING);
2132                 cfs_waitq_del (&ksocknal_data.ksnd_connd_waitq, &wait);
2133                 cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2134         }
2135
2136         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2137
2138         ksocknal_thread_fini ();
2139         return (0);
2140 }
2141
2142 ksock_conn_t *
2143 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2144 {
2145         /* We're called with a shared lock on ksnd_global_lock */
2146         ksock_conn_t      *conn;
2147         cfs_list_t        *ctmp;
2148
2149         cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2150                 int     error;
2151                 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2152
2153                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2154                 LASSERT (!conn->ksnc_closing);
2155
2156                 /* SOCK_ERROR will reset error code of socket in
2157                  * some platform (like Darwin8.x) */
2158                 error = libcfs_sock_error(conn->ksnc_sock);
2159                 if (error != 0) {
2160                         ksocknal_conn_addref(conn);
2161
2162                         switch (error) {
2163                         case ECONNRESET:
2164                                 CNETERR("A connection with %s "
2165                                         "(%u.%u.%u.%u:%d) was reset; "
2166                                         "it may have rebooted.\n",
2167                                         libcfs_id2str(peer->ksnp_id),
2168                                         HIPQUAD(conn->ksnc_ipaddr),
2169                                         conn->ksnc_port);
2170                                 break;
2171                         case ETIMEDOUT:
2172                                 CNETERR("A connection with %s "
2173                                         "(%u.%u.%u.%u:%d) timed out; the "
2174                                         "network or node may be down.\n",
2175                                         libcfs_id2str(peer->ksnp_id),
2176                                         HIPQUAD(conn->ksnc_ipaddr),
2177                                         conn->ksnc_port);
2178                                 break;
2179                         default:
2180                                 CNETERR("An unexpected network error %d "
2181                                         "occurred with %s "
2182                                         "(%u.%u.%u.%u:%d\n", error,
2183                                         libcfs_id2str(peer->ksnp_id),
2184                                         HIPQUAD(conn->ksnc_ipaddr),
2185                                         conn->ksnc_port);
2186                                 break;
2187                         }
2188
2189                         return (conn);
2190                 }
2191
2192                 if (conn->ksnc_rx_started &&
2193                     cfs_time_aftereq(cfs_time_current(),
2194                                      conn->ksnc_rx_deadline)) {
2195                         /* Timed out incomplete incoming message */
2196                         ksocknal_conn_addref(conn);
2197                         CNETERR("Timeout receiving from %s (%u.%u.%u.%u:%d), "
2198                                 "state %d wanted %d left %d\n",
2199                                 libcfs_id2str(peer->ksnp_id),
2200                                 HIPQUAD(conn->ksnc_ipaddr),
2201                                 conn->ksnc_port,
2202                                 conn->ksnc_rx_state,
2203                                 conn->ksnc_rx_nob_wanted,
2204                                 conn->ksnc_rx_nob_left);
2205                         return (conn);
2206                 }
2207
2208                 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2209                      libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2210                     cfs_time_aftereq(cfs_time_current(),
2211                                      conn->ksnc_tx_deadline)) {
2212                         /* Timed out messages queued for sending or
2213                          * buffered in the socket's send buffer */
2214                         ksocknal_conn_addref(conn);
2215                         CNETERR("Timeout sending data to %s (%u.%u.%u.%u:%d) "
2216                                 "the network or that node may be down.\n",
2217                                 libcfs_id2str(peer->ksnp_id),
2218                                 HIPQUAD(conn->ksnc_ipaddr),
2219                                 conn->ksnc_port);
2220                         return (conn);
2221                 }
2222         }
2223
2224         return (NULL);
2225 }
2226
2227 static inline void
2228 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2229 {
2230         ksock_tx_t        *tx;
2231         CFS_LIST_HEAD      (stale_txs);
2232
2233         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
2234
2235         while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2236                 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2237                                      ksock_tx_t, tx_list);
2238
2239                 if (!cfs_time_aftereq(cfs_time_current(),
2240                                       tx->tx_deadline))
2241                         break;
2242
2243                 cfs_list_del (&tx->tx_list);
2244                 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2245         }
2246
2247         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2248
2249         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2250 }
2251
2252 int
2253 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2254 {
2255         ksock_sched_t  *sched;
2256         ksock_conn_t   *conn;
2257         ksock_tx_t     *tx;
2258
2259         if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2260                 return 0;
2261
2262         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2263                 return 0;
2264
2265         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2266             cfs_time_before(cfs_time_current(),
2267                             cfs_time_add(peer->ksnp_last_alive,
2268                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2269                 return 0;
2270
2271         if (cfs_time_before(cfs_time_current(),
2272                             peer->ksnp_send_keepalive))
2273                 return 0;
2274
2275         /* retry 10 secs later, so we wouldn't put pressure
2276          * on this peer if we failed to send keepalive this time */
2277         peer->ksnp_send_keepalive = cfs_time_shift(10);
2278
2279         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2280         if (conn != NULL) {
2281                 sched = conn->ksnc_scheduler;
2282
2283                 cfs_spin_lock_bh (&sched->kss_lock);
2284                 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2285                         cfs_spin_unlock_bh(&sched->kss_lock);
2286                         /* there is an queued ACK, don't need keepalive */
2287                         return 0;
2288                 }
2289
2290                 cfs_spin_unlock_bh(&sched->kss_lock);
2291         }
2292
2293         cfs_read_unlock(&ksocknal_data.ksnd_global_lock);
2294
2295         /* cookie = 1 is reserved for keepalive PING */
2296         tx = ksocknal_alloc_tx_noop(1, 1);
2297         if (tx == NULL) {
2298                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2299                 return -ENOMEM;
2300         }
2301
2302         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2303                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2304                 return 1;
2305         }
2306
2307         ksocknal_free_tx(tx);
2308         cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2309
2310         return -EIO;
2311 }
2312
2313
2314 void
2315 ksocknal_check_peer_timeouts (int idx)
2316 {
2317         cfs_list_t       *peers = &ksocknal_data.ksnd_peers[idx];
2318         ksock_peer_t     *peer;
2319         ksock_conn_t     *conn;
2320         ksock_tx_t       *tx;
2321
2322  again:
2323         /* NB. We expect to have a look at all the peers and not find any
2324          * connections to time out, so we just use a shared lock while we
2325          * take a look... */
2326         cfs_read_lock (&ksocknal_data.ksnd_global_lock);
2327
2328         cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2329                 cfs_time_t  deadline = 0;
2330                 int         resid = 0;
2331                 int         n     = 0;
2332
2333                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2334                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2335                         goto again;
2336                 }
2337
2338                 conn = ksocknal_find_timed_out_conn (peer);
2339
2340                 if (conn != NULL) {
2341                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2342
2343                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2344
2345                         /* NB we won't find this one again, but we can't
2346                          * just proceed with the next peer, since we dropped
2347                          * ksnd_global_lock and it might be dead already! */
2348                         ksocknal_conn_decref(conn);
2349                         goto again;
2350                 }
2351
2352                 /* we can't process stale txs right here because we're
2353                  * holding only shared lock */
2354                 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2355                         ksock_tx_t *tx =
2356                                 cfs_list_entry (peer->ksnp_tx_queue.next,
2357                                                 ksock_tx_t, tx_list);
2358
2359                         if (cfs_time_aftereq(cfs_time_current(),
2360                                              tx->tx_deadline)) {
2361
2362                                 ksocknal_peer_addref(peer);
2363                                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2364
2365                                 ksocknal_flush_stale_txs(peer);
2366
2367                                 ksocknal_peer_decref(peer);
2368                                 goto again;
2369                         }
2370                 }
2371
2372                 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2373                         continue;
2374
2375                 cfs_spin_lock(&peer->ksnp_lock);
2376                 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2377                                               ksock_tx_t, tx_zc_list) {
2378                         if (!cfs_time_aftereq(cfs_time_current(),
2379                                               tx->tx_deadline))
2380                                 break;
2381                         /* ignore the TX if connection is being closed */
2382                         if (tx->tx_conn->ksnc_closing)
2383                                 continue;
2384                         n++;
2385                 }
2386
2387                 if (n == 0) {
2388                         cfs_spin_unlock(&peer->ksnp_lock);
2389                         continue;
2390                 }
2391
2392                 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2393                                     ksock_tx_t, tx_zc_list);
2394                 deadline = tx->tx_deadline;
2395                 resid    = tx->tx_resid;
2396                 conn     = tx->tx_conn;
2397                 ksocknal_conn_addref(conn);
2398
2399                 cfs_spin_unlock(&peer->ksnp_lock);
2400                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2401
2402                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2403                        "oldest(%p) timed out %ld secs ago, "
2404                        "resid: %d, wmem: %d\n",
2405                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2406                        cfs_duration_sec(cfs_time_current() - deadline),
2407                        resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2408
2409                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2410                 ksocknal_conn_decref(conn);
2411                 goto again;
2412         }
2413
2414         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2415 }
2416
2417 int
2418 ksocknal_reaper (void *arg)
2419 {
2420         cfs_waitlink_t     wait;
2421         ksock_conn_t      *conn;
2422         ksock_sched_t     *sched;
2423         cfs_list_t         enomem_conns;
2424         int                nenomem_conns;
2425         cfs_duration_t     timeout;
2426         int                i;
2427         int                peer_index = 0;
2428         cfs_time_t         deadline = cfs_time_current();
2429
2430         cfs_daemonize ("socknal_reaper");
2431         cfs_block_allsigs ();
2432
2433         CFS_INIT_LIST_HEAD(&enomem_conns);
2434         cfs_waitlink_init (&wait);
2435
2436         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2437
2438         while (!ksocknal_data.ksnd_shuttingdown) {
2439
2440                 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2441                         conn = cfs_list_entry (ksocknal_data. \
2442                                                ksnd_deathrow_conns.next,
2443                                                ksock_conn_t, ksnc_list);
2444                         cfs_list_del (&conn->ksnc_list);
2445
2446                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2447
2448                         ksocknal_terminate_conn (conn);
2449                         ksocknal_conn_decref(conn);
2450
2451                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2452                         continue;
2453                 }
2454
2455                 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2456                         conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2457                                                next, ksock_conn_t, ksnc_list);
2458                         cfs_list_del (&conn->ksnc_list);
2459
2460                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2461
2462                         ksocknal_destroy_conn (conn);
2463
2464                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2465                         continue;
2466                 }
2467
2468                 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2469                         cfs_list_add(&enomem_conns,
2470                                      &ksocknal_data.ksnd_enomem_conns);
2471                         cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2472                 }
2473
2474                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2475
2476                 /* reschedule all the connections that stalled with ENOMEM... */
2477                 nenomem_conns = 0;
2478                 while (!cfs_list_empty (&enomem_conns)) {
2479                         conn = cfs_list_entry (enomem_conns.next,
2480                                                ksock_conn_t, ksnc_tx_list);
2481                         cfs_list_del (&conn->ksnc_tx_list);
2482
2483                         sched = conn->ksnc_scheduler;
2484
2485                         cfs_spin_lock_bh (&sched->kss_lock);
2486
2487                         LASSERT (conn->ksnc_tx_scheduled);
2488                         conn->ksnc_tx_ready = 1;
2489                         cfs_list_add_tail(&conn->ksnc_tx_list,
2490                                           &sched->kss_tx_conns);
2491                         cfs_waitq_signal (&sched->kss_waitq);
2492
2493                         cfs_spin_unlock_bh (&sched->kss_lock);
2494                         nenomem_conns++;
2495                 }
2496
2497                 /* careful with the jiffy wrap... */
2498                 while ((timeout = cfs_time_sub(deadline,
2499                                                cfs_time_current())) <= 0) {
2500                         const int n = 4;
2501                         const int p = 1;
2502                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2503
2504                         /* Time to check for timeouts on a few more peers: I do
2505                          * checks every 'p' seconds on a proportion of the peer
2506                          * table and I need to check every connection 'n' times
2507                          * within a timeout interval, to ensure I detect a
2508                          * timeout on any connection within (n+1)/n times the
2509                          * timeout interval. */
2510
2511                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2512                                 chunk = (chunk * n * p) /
2513                                         *ksocknal_tunables.ksnd_timeout;
2514                         if (chunk == 0)
2515                                 chunk = 1;
2516
2517                         for (i = 0; i < chunk; i++) {
2518                                 ksocknal_check_peer_timeouts (peer_index);
2519                                 peer_index = (peer_index + 1) %
2520                                              ksocknal_data.ksnd_peer_hash_size;
2521                         }
2522
2523                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2524                 }
2525
2526                 if (nenomem_conns != 0) {
2527                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2528                          * This also prevents me getting woken immediately
2529                          * if any go back on my enomem list. */
2530                         timeout = SOCKNAL_ENOMEM_RETRY;
2531                 }
2532                 ksocknal_data.ksnd_reaper_waketime =
2533                         cfs_time_add(cfs_time_current(), timeout);
2534
2535                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2536                 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2537
2538                 if (!ksocknal_data.ksnd_shuttingdown &&
2539                     cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2540                     cfs_list_empty (&ksocknal_data.ksnd_zombie_conns))
2541                         cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE,
2542                                              timeout);
2543
2544                 cfs_set_current_state (CFS_TASK_RUNNING);
2545                 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2546
2547                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2548         }
2549
2550         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2551
2552         ksocknal_thread_fini ();
2553         return (0);
2554 }