lnet/klnds/socklnd/socklnd_cb.c

   1 /*
   2  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2012, Intel Corporation.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Portals, http://www.sf.net/projects/sandiaportals/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Portals; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include "socklnd.h"
  28
  29 ksock_tx_t *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         ksock_tx_t *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         list_del(&tx->tx_list);
  45                 }
  46
  47                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_desc_size  = size;
  61
  62         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  63
  64         return tx;
  65 }
  66
  67 ksock_tx_t *
  68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  69 {
  70         ksock_tx_t *tx;
  71
  72         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  73         if (tx == NULL) {
  74                 CERROR("Can't allocate noop tx desc\n");
  75                 return NULL;
  76         }
  77
  78         tx->tx_conn     = NULL;
  79         tx->tx_lnetmsg  = NULL;
  80         tx->tx_kiov     = NULL;
  81         tx->tx_nkiov    = 0;
  82         tx->tx_iov      = tx->tx_frags.virt.iov;
  83         tx->tx_niov     = 1;
  84         tx->tx_nonblk   = nonblk;
  85
  86         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  87         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  88
  89         return tx;
  90 }
  91
  92
  93 void
  94 ksocknal_free_tx (ksock_tx_t *tx)
  95 {
  96         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  97
  98         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  99                 /* it's a noop tx */
 100                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 101
 102                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 103
 104                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 105         } else {
 106                 LIBCFS_FREE(tx, tx->tx_desc_size);
 107         }
 108 }
 109
 110 int
 111 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
 112 {
 113         struct iovec  *iov = tx->tx_iov;
 114         int    nob;
 115         int    rc;
 116
 117         LASSERT (tx->tx_niov > 0);
 118
 119         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 120         rc = ksocknal_lib_send_iov(conn, tx);
 121
 122         if (rc <= 0)                            /* sent nothing? */
 123                 return (rc);
 124
 125         nob = rc;
 126         LASSERT (nob <= tx->tx_resid);
 127         tx->tx_resid -= nob;
 128
 129         /* "consume" iov */
 130         do {
 131                 LASSERT (tx->tx_niov > 0);
 132
 133                 if (nob < (int) iov->iov_len) {
 134                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 135                         iov->iov_len -= nob;
 136                         return (rc);
 137                 }
 138
 139                 nob -= iov->iov_len;
 140                 tx->tx_iov = ++iov;
 141                 tx->tx_niov--;
 142         } while (nob != 0);
 143
 144         return (rc);
 145 }
 146
 147 int
 148 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
 149 {
 150         lnet_kiov_t    *kiov = tx->tx_kiov;
 151         int     nob;
 152         int     rc;
 153
 154         LASSERT (tx->tx_niov == 0);
 155         LASSERT (tx->tx_nkiov > 0);
 156
 157         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 158         rc = ksocknal_lib_send_kiov(conn, tx);
 159
 160         if (rc <= 0)                            /* sent nothing? */
 161                 return (rc);
 162
 163         nob = rc;
 164         LASSERT (nob <= tx->tx_resid);
 165         tx->tx_resid -= nob;
 166
 167         /* "consume" kiov */
 168         do {
 169                 LASSERT(tx->tx_nkiov > 0);
 170
 171                 if (nob < (int)kiov->kiov_len) {
 172                         kiov->kiov_offset += nob;
 173                         kiov->kiov_len -= nob;
 174                         return rc;
 175                 }
 176
 177                 nob -= (int)kiov->kiov_len;
 178                 tx->tx_kiov = ++kiov;
 179                 tx->tx_nkiov--;
 180         } while (nob != 0);
 181
 182         return (rc);
 183 }
 184
 185 int
 186 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 187 {
 188         int      rc;
 189         int      bufnob;
 190
 191         if (ksocknal_data.ksnd_stall_tx != 0) {
 192                 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 193         }
 194
 195         LASSERT (tx->tx_resid != 0);
 196
 197         rc = ksocknal_connsock_addref(conn);
 198         if (rc != 0) {
 199                 LASSERT (conn->ksnc_closing);
 200                 return (-ESHUTDOWN);
 201         }
 202
 203         do {
 204                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 205                         /* testing... */
 206                         ksocknal_data.ksnd_enomem_tx--;
 207                         rc = -EAGAIN;
 208                 } else if (tx->tx_niov != 0) {
 209                         rc = ksocknal_send_iov (conn, tx);
 210                 } else {
 211                         rc = ksocknal_send_kiov (conn, tx);
 212                 }
 213
 214                 bufnob = 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                         smp_mb();
 226                 }
 227
 228                 if (rc <= 0) { /* Didn't write anything? */
 229
 230                         if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
 231                                 rc = -EAGAIN;
 232
 233                         /* Check if EAGAIN is due to memory pressure */
 234                         if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 235                                 rc = -ENOMEM;
 236
 237                         break;
 238                 }
 239
 240                 /* socket's wmem_queued now includes 'rc' bytes */
 241                 atomic_sub (rc, &conn->ksnc_tx_nob);
 242                 rc = 0;
 243
 244         } while (tx->tx_resid != 0);
 245
 246         ksocknal_connsock_decref(conn);
 247         return (rc);
 248 }
 249
 250 int
 251 ksocknal_recv_iov (ksock_conn_t *conn)
 252 {
 253         struct iovec *iov = conn->ksnc_rx_iov;
 254         int     nob;
 255         int     rc;
 256
 257         LASSERT (conn->ksnc_rx_niov > 0);
 258
 259         /* Never touch conn->ksnc_rx_iov or change connection
 260          * status inside ksocknal_lib_recv_iov */
 261         rc = ksocknal_lib_recv_iov(conn);
 262
 263         if (rc <= 0)
 264                 return (rc);
 265
 266         /* received something... */
 267         nob = rc;
 268
 269         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 270         conn->ksnc_rx_deadline =
 271                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 272         smp_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         smp_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, struct list_head *txlist, int error)
 411 {
 412         ksock_tx_t *tx;
 413
 414         while (!list_empty(txlist)) {
 415                 tx = 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                 list_del(&tx->tx_list);
 428
 429                 LASSERT (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         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         list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 474
 475         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         spin_lock(&peer->ksnp_lock);
 489
 490         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 491                 /* Not waiting for an ACK */
 492                 spin_unlock(&peer->ksnp_lock);
 493                 return;
 494         }
 495
 496         tx->tx_msg.ksm_zc_cookies[0] = 0;
 497         list_del(&tx->tx_zc_list);
 498
 499         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, atomic_read(&libcfs_kmemory));
 533
 534                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 535                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 536
 537                 /* enomem list takes over scheduler's ref... */
 538                 LASSERT (conn->ksnc_tx_scheduled);
 539                 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                         wake_up(&ksocknal_data.ksnd_reaper_waitq);
 545
 546                 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         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 598
 599         list_add_tail(&route->ksnr_connd_list,
 600                           &ksocknal_data.ksnd_connd_routes);
 601         wake_up(&ksocknal_data.ksnd_connd_waitq);
 602
 603         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 604 }
 605
 606 void
 607 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 608 {
 609         ksock_route_t *route;
 610
 611         /* called holding write lock on ksnd_global_lock */
 612         for (;;) {
 613                 /* launch any/all connections that need it */
 614                 route = ksocknal_find_connectable_route_locked(peer);
 615                 if (route == 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         struct list_head *tmp;
 626         ksock_conn_t     *conn;
 627         ksock_conn_t     *typed = NULL;
 628         ksock_conn_t     *fallback = NULL;
 629         int               tnob     = 0;
 630         int               fnob     = 0;
 631
 632         list_for_each(tmp, &peer->ksnp_conns) {
 633                 ksock_conn_t *c  = list_entry(tmp, ksock_conn_t, ksnc_list);
 634                 int           nob = atomic_read(&c->ksnc_tx_nob) +
 635                                     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         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         bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 728         spin_lock_bh(&sched->kss_lock);
 729
 730         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 731                 /* First packet starts the timeout */
 732                 conn->ksnc_tx_deadline =
 733                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 734                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 735                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 736                 conn->ksnc_tx_bufnob = 0;
 737                 smp_mb(); /* order with adding to tx_queue */
 738         }
 739
 740         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 741                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 742                  * on a normal packet so I don't need to send it */
 743                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 744                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 745
 746                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 747                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 748
 749         } else {
 750                 /* It's a normal packet - can it piggback a noop zc-ack that
 751                  * has been queued already? */
 752                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 753                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 754
 755                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 756                 /* ztx will be released later */
 757         }
 758
 759         if (ztx != NULL) {
 760                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 761                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 762         }
 763
 764         if (conn->ksnc_tx_ready &&      /* able to send */
 765             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 766                 /* +1 ref for scheduler */
 767                 ksocknal_conn_addref(conn);
 768                 list_add_tail(&conn->ksnc_tx_list,
 769                                    &sched->kss_tx_conns);
 770                 conn->ksnc_tx_scheduled = 1;
 771                 wake_up(&sched->kss_waitq);
 772         }
 773
 774         spin_unlock_bh(&sched->kss_lock);
 775 }
 776
 777
 778 ksock_route_t *
 779 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 780 {
 781         cfs_time_t     now = cfs_time_current();
 782         struct list_head    *tmp;
 783         ksock_route_t *route;
 784
 785         list_for_each(tmp, &peer->ksnp_routes) {
 786                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 787
 788                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 789
 790                 if (route->ksnr_scheduled)      /* connections being established */
 791                         continue;
 792
 793                 /* all route types connected ? */
 794                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 795                         continue;
 796
 797                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 798                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 799                         CDEBUG(D_NET,
 800                                "Too soon to retry route %u.%u.%u.%u "
 801                                "(cnted %d, interval %ld, %ld secs later)\n",
 802                                HIPQUAD(route->ksnr_ipaddr),
 803                                route->ksnr_connected,
 804                                route->ksnr_retry_interval,
 805                                cfs_duration_sec(route->ksnr_timeout - now));
 806                         continue;
 807                 }
 808
 809                 return (route);
 810         }
 811
 812         return (NULL);
 813 }
 814
 815 ksock_route_t *
 816 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 817 {
 818         struct list_head        *tmp;
 819         ksock_route_t     *route;
 820
 821         list_for_each(tmp, &peer->ksnp_routes) {
 822                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 823
 824                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 825
 826                 if (route->ksnr_scheduled)
 827                         return (route);
 828         }
 829
 830         return (NULL);
 831 }
 832
 833 int
 834 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 835 {
 836         ksock_peer_t     *peer;
 837         ksock_conn_t     *conn;
 838         rwlock_t     *g_lock;
 839         int               retry;
 840         int               rc;
 841
 842         LASSERT (tx->tx_conn == NULL);
 843
 844         g_lock = &ksocknal_data.ksnd_global_lock;
 845
 846         for (retry = 0;; retry = 1) {
 847                 read_lock(g_lock);
 848                 peer = ksocknal_find_peer_locked(ni, id);
 849                 if (peer != NULL) {
 850                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 851                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 852                                 if (conn != NULL) {
 853                                         /* I've got no routes that need to be
 854                                          * connecting and I do have an actual
 855                                          * connection... */
 856                                         ksocknal_queue_tx_locked (tx, conn);
 857                                         read_unlock(g_lock);
 858                                         return (0);
 859                                 }
 860                         }
 861                 }
 862
 863                 /* I'll need a write lock... */
 864                 read_unlock(g_lock);
 865
 866                 write_lock_bh(g_lock);
 867
 868                 peer = ksocknal_find_peer_locked(ni, id);
 869                 if (peer != NULL)
 870                         break;
 871
 872                 write_unlock_bh(g_lock);
 873
 874                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 875                         CERROR("Refusing to create a connection to "
 876                                "userspace process %s\n", libcfs_id2str(id));
 877                         return -EHOSTUNREACH;
 878                 }
 879
 880                 if (retry) {
 881                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 882                         return -EHOSTUNREACH;
 883                 }
 884
 885                 rc = ksocknal_add_peer(ni, id,
 886                                        LNET_NIDADDR(id.nid),
 887                                        lnet_acceptor_port());
 888                 if (rc != 0) {
 889                         CERROR("Can't add peer %s: %d\n",
 890                                libcfs_id2str(id), rc);
 891                         return rc;
 892                 }
 893         }
 894
 895         ksocknal_launch_all_connections_locked(peer);
 896
 897         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 898         if (conn != NULL) {
 899                 /* Connection exists; queue message on it */
 900                 ksocknal_queue_tx_locked (tx, conn);
 901                 write_unlock_bh(g_lock);
 902                 return (0);
 903         }
 904
 905         if (peer->ksnp_accepting > 0 ||
 906             ksocknal_find_connecting_route_locked (peer) != NULL) {
 907                 /* the message is going to be pinned to the peer */
 908                 tx->tx_deadline =
 909                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 910
 911                 /* Queue the message until a connection is established */
 912                 list_add_tail(&tx->tx_list, &peer->ksnp_tx_queue);
 913                 write_unlock_bh(g_lock);
 914                 return 0;
 915         }
 916
 917         write_unlock_bh(g_lock);
 918
 919         /* NB Routes may be ignored if connections to them failed recently */
 920         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 921         return (-EHOSTUNREACH);
 922 }
 923
 924 int
 925 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 926 {
 927         int               mpflag = 1;
 928         int               type = lntmsg->msg_type;
 929         lnet_process_id_t target = lntmsg->msg_target;
 930         unsigned int      payload_niov = lntmsg->msg_niov;
 931         struct iovec     *payload_iov = lntmsg->msg_iov;
 932         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 933         unsigned int      payload_offset = lntmsg->msg_offset;
 934         unsigned int      payload_nob = lntmsg->msg_len;
 935         ksock_tx_t       *tx;
 936         int               desc_size;
 937         int               rc;
 938
 939         /* NB 'private' is different depending on what we're sending.
 940          * Just ignore it... */
 941
 942         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 943                payload_nob, payload_niov, libcfs_id2str(target));
 944
 945         LASSERT (payload_nob == 0 || payload_niov > 0);
 946         LASSERT (payload_niov <= LNET_MAX_IOV);
 947         /* payload is either all vaddrs or all pages */
 948         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 949         LASSERT (!in_interrupt ());
 950
 951         if (payload_iov != NULL)
 952                 desc_size = offsetof(ksock_tx_t,
 953                                      tx_frags.virt.iov[1 + payload_niov]);
 954         else
 955                 desc_size = offsetof(ksock_tx_t,
 956                                      tx_frags.paged.kiov[payload_niov]);
 957
 958         if (lntmsg->msg_vmflush)
 959                 mpflag = cfs_memory_pressure_get_and_set();
 960         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 961         if (tx == NULL) {
 962                 CERROR("Can't allocate tx desc type %d size %d\n",
 963                        type, desc_size);
 964                 if (lntmsg->msg_vmflush)
 965                         cfs_memory_pressure_restore(mpflag);
 966                 return (-ENOMEM);
 967         }
 968
 969         tx->tx_conn = NULL;                     /* set when assigned a conn */
 970         tx->tx_lnetmsg = lntmsg;
 971
 972         if (payload_iov != NULL) {
 973                 tx->tx_kiov = NULL;
 974                 tx->tx_nkiov = 0;
 975                 tx->tx_iov = tx->tx_frags.virt.iov;
 976                 tx->tx_niov = 1 +
 977                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 978                                                payload_niov, payload_iov,
 979                                                payload_offset, payload_nob);
 980         } else {
 981                 tx->tx_niov = 1;
 982                 tx->tx_iov = &tx->tx_frags.paged.iov;
 983                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 984                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 985                                                  payload_niov, payload_kiov,
 986                                                  payload_offset, payload_nob);
 987
 988                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 989                         tx->tx_zc_capable = 1;
 990         }
 991
 992         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 993
 994         /* The first fragment will be set later in pro_pack */
 995         rc = ksocknal_launch_packet(ni, tx, target);
 996         if (!mpflag)
 997                 cfs_memory_pressure_restore(mpflag);
 998
 999         if (rc == 0)
1000                 return (0);
1001
1002         ksocknal_free_tx(tx);
1003         return (-EIO);
1004 }
1005
1006 int
1007 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1008 {
1009         struct task_struct *task = kthread_run(fn, arg, name);
1010
1011         if (IS_ERR(task))
1012                 return PTR_ERR(task);
1013
1014         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1015         ksocknal_data.ksnd_nthreads++;
1016         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1017         return 0;
1018 }
1019
1020 void
1021 ksocknal_thread_fini (void)
1022 {
1023         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1024         ksocknal_data.ksnd_nthreads--;
1025         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1026 }
1027
1028 int
1029 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1030 {
1031         static char ksocknal_slop_buffer[4096];
1032
1033         int            nob;
1034         unsigned int   niov;
1035         int            skipped;
1036
1037         LASSERT(conn->ksnc_proto != NULL);
1038
1039         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1040                 /* Remind the socket to ack eagerly... */
1041                 ksocknal_lib_eager_ack(conn);
1042         }
1043
1044         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1045                 conn->ksnc_rx_started = 0;
1046                 smp_mb();                       /* racing with timeout thread */
1047
1048                 switch (conn->ksnc_proto->pro_version) {
1049                 case  KSOCK_PROTO_V2:
1050                 case  KSOCK_PROTO_V3:
1051                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1052                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1053                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1054
1055                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1056                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1057                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1058                         break;
1059
1060                 case KSOCK_PROTO_V1:
1061                         /* Receiving bare lnet_hdr_t */
1062                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1063                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1064                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1065
1066                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1067                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1068                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1069                         break;
1070
1071                 default:
1072                         LBUG ();
1073                 }
1074                 conn->ksnc_rx_niov = 1;
1075
1076                 conn->ksnc_rx_kiov = NULL;
1077                 conn->ksnc_rx_nkiov = 0;
1078                 conn->ksnc_rx_csum = ~0;
1079                 return (1);
1080         }
1081
1082         /* Set up to skip as much as possible now.  If there's more left
1083          * (ran out of iov entries) we'll get called again */
1084
1085         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1086         conn->ksnc_rx_nob_left = nob_to_skip;
1087         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1088         skipped = 0;
1089         niov = 0;
1090
1091         do {
1092                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1093
1094                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1095                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1096                 niov++;
1097                 skipped += nob;
1098                 nob_to_skip -=nob;
1099
1100         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1101                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1102
1103         conn->ksnc_rx_niov = niov;
1104         conn->ksnc_rx_kiov = NULL;
1105         conn->ksnc_rx_nkiov = 0;
1106         conn->ksnc_rx_nob_wanted = skipped;
1107         return (0);
1108 }
1109
1110 int
1111 ksocknal_process_receive (ksock_conn_t *conn)
1112 {
1113         lnet_hdr_t        *lhdr;
1114         lnet_process_id_t *id;
1115         int                rc;
1116
1117         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1118
1119         /* NB: sched lock NOT held */
1120         /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1121         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1122                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1123                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1124                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1125  again:
1126         if (conn->ksnc_rx_nob_wanted != 0) {
1127                 rc = ksocknal_receive(conn);
1128
1129                 if (rc <= 0) {
1130                         LASSERT (rc != -EAGAIN);
1131
1132                         if (rc == 0)
1133                                 CDEBUG (D_NET, "[%p] EOF from %s"
1134                                         " ip %d.%d.%d.%d:%d\n", conn,
1135                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1136                                         HIPQUAD(conn->ksnc_ipaddr),
1137                                         conn->ksnc_port);
1138                         else if (!conn->ksnc_closing)
1139                                 CERROR ("[%p] Error %d on read from %s"
1140                                         " ip %d.%d.%d.%d:%d\n",
1141                                         conn, rc,
1142                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1143                                         HIPQUAD(conn->ksnc_ipaddr),
1144                                         conn->ksnc_port);
1145
1146                         /* it's not an error if conn is being closed */
1147                         ksocknal_close_conn_and_siblings (conn,
1148                                                           (conn->ksnc_closing) ? 0 : rc);
1149                         return (rc == 0 ? -ESHUTDOWN : rc);
1150                 }
1151
1152                 if (conn->ksnc_rx_nob_wanted != 0) {
1153                         /* short read */
1154                         return (-EAGAIN);
1155                 }
1156         }
1157         switch (conn->ksnc_rx_state) {
1158         case SOCKNAL_RX_KSM_HEADER:
1159                 if (conn->ksnc_flip) {
1160                         __swab32s(&conn->ksnc_msg.ksm_type);
1161                         __swab32s(&conn->ksnc_msg.ksm_csum);
1162                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1163                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1164                 }
1165
1166                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1167                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1168                         CERROR("%s: Unknown message type: %x\n",
1169                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1170                                conn->ksnc_msg.ksm_type);
1171                         ksocknal_new_packet(conn, 0);
1172                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1173                         return (-EPROTO);
1174                 }
1175
1176                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1177                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1178                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1179                         /* NOOP Checksum error */
1180                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1181                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1182                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1183                         ksocknal_new_packet(conn, 0);
1184                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1185                         return (-EIO);
1186                 }
1187
1188                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1189                         __u64 cookie = 0;
1190
1191                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1192
1193                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1194                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1195
1196                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1197                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1198
1199                         if (rc != 0) {
1200                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1201                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1202                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1203                                 ksocknal_new_packet(conn, 0);
1204                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1205                                 return (rc);
1206                         }
1207                 }
1208
1209                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1210                         ksocknal_new_packet (conn, 0);
1211                         return 0;       /* NOOP is done and just return */
1212                 }
1213
1214                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1215                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1216                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1217
1218                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1219                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1220                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1221
1222                 conn->ksnc_rx_niov = 1;
1223                 conn->ksnc_rx_kiov = NULL;
1224                 conn->ksnc_rx_nkiov = 0;
1225
1226                 goto again;     /* read lnet header now */
1227
1228         case SOCKNAL_RX_LNET_HEADER:
1229                 /* unpack message header */
1230                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1231
1232                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1233                         /* Userspace peer */
1234                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1235                         id   = &conn->ksnc_peer->ksnp_id;
1236
1237                         /* Substitute process ID assigned at connection time */
1238                         lhdr->src_pid = cpu_to_le32(id->pid);
1239                         lhdr->src_nid = cpu_to_le64(id->nid);
1240                 }
1241
1242                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1243                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1244
1245                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1246                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1247                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1248                 if (rc < 0) {
1249                         /* I just received garbage: give up on this conn */
1250                         ksocknal_new_packet(conn, 0);
1251                         ksocknal_close_conn_and_siblings (conn, rc);
1252                         ksocknal_conn_decref(conn);
1253                         return (-EPROTO);
1254                 }
1255
1256                 /* I'm racing with ksocknal_recv() */
1257                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1258                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1259
1260                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1261                         return 0;
1262
1263                 /* ksocknal_recv() got called */
1264                 goto again;
1265
1266         case SOCKNAL_RX_LNET_PAYLOAD:
1267                 /* payload all received */
1268                 rc = 0;
1269
1270                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1271                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1272                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1273                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1274                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1275                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1276                         rc = -EIO;
1277                 }
1278
1279                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1280                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1281
1282                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1283                         id   = &conn->ksnc_peer->ksnp_id;
1284
1285                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1286                                         conn->ksnc_msg.ksm_zc_cookies[0],
1287                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1288                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1289                 }
1290
1291                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1292
1293                 if (rc != 0) {
1294                         ksocknal_new_packet(conn, 0);
1295                         ksocknal_close_conn_and_siblings (conn, rc);
1296                         return (-EPROTO);
1297                 }
1298                 /* Fall through */
1299
1300         case SOCKNAL_RX_SLOP:
1301                 /* starting new packet? */
1302                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1303                         return 0;       /* come back later */
1304                 goto again;             /* try to finish reading slop now */
1305
1306         default:
1307                 break;
1308         }
1309
1310         /* Not Reached */
1311         LBUG ();
1312         return (-EINVAL);                       /* keep gcc happy */
1313 }
1314
1315 int
1316 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1317                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1318                unsigned int offset, unsigned int mlen, unsigned int rlen)
1319 {
1320         ksock_conn_t  *conn = (ksock_conn_t *)private;
1321         ksock_sched_t *sched = conn->ksnc_scheduler;
1322
1323         LASSERT (mlen <= rlen);
1324         LASSERT (niov <= LNET_MAX_IOV);
1325
1326         conn->ksnc_cookie = msg;
1327         conn->ksnc_rx_nob_wanted = mlen;
1328         conn->ksnc_rx_nob_left   = rlen;
1329
1330         if (mlen == 0 || iov != NULL) {
1331                 conn->ksnc_rx_nkiov = 0;
1332                 conn->ksnc_rx_kiov = NULL;
1333                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1334                 conn->ksnc_rx_niov =
1335                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1336                                          niov, iov, offset, mlen);
1337         } else {
1338                 conn->ksnc_rx_niov = 0;
1339                 conn->ksnc_rx_iov  = NULL;
1340                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1341                 conn->ksnc_rx_nkiov =
1342                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1343                                           niov, kiov, offset, mlen);
1344         }
1345
1346         LASSERT (mlen ==
1347                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1348                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1349
1350         LASSERT (conn->ksnc_rx_scheduled);
1351
1352         spin_lock_bh(&sched->kss_lock);
1353
1354         switch (conn->ksnc_rx_state) {
1355         case SOCKNAL_RX_PARSE_WAIT:
1356                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1357                 wake_up(&sched->kss_waitq);
1358                 LASSERT(conn->ksnc_rx_ready);
1359                 break;
1360
1361         case SOCKNAL_RX_PARSE:
1362                 /* scheduler hasn't noticed I'm parsing yet */
1363                 break;
1364         }
1365
1366         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1367
1368         spin_unlock_bh(&sched->kss_lock);
1369         ksocknal_conn_decref(conn);
1370         return 0;
1371 }
1372
1373 static inline int
1374 ksocknal_sched_cansleep(ksock_sched_t *sched)
1375 {
1376         int           rc;
1377
1378         spin_lock_bh(&sched->kss_lock);
1379
1380         rc = (!ksocknal_data.ksnd_shuttingdown &&
1381               list_empty(&sched->kss_rx_conns) &&
1382               list_empty(&sched->kss_tx_conns));
1383
1384         spin_unlock_bh(&sched->kss_lock);
1385         return rc;
1386 }
1387
1388 int ksocknal_scheduler(void *arg)
1389 {
1390         struct ksock_sched_info *info;
1391         ksock_sched_t           *sched;
1392         ksock_conn_t            *conn;
1393         ksock_tx_t              *tx;
1394         int                     rc;
1395         int                     nloops = 0;
1396         long                    id = (long)arg;
1397
1398         info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1399         sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1400
1401         cfs_block_allsigs();
1402
1403         rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1404         if (rc != 0) {
1405                 CERROR("Can't set CPT affinity to %d: %d\n",
1406                        info->ksi_cpt, rc);
1407         }
1408
1409         spin_lock_bh(&sched->kss_lock);
1410
1411         while (!ksocknal_data.ksnd_shuttingdown) {
1412                 int did_something = 0;
1413
1414                 /* Ensure I progress everything semi-fairly */
1415
1416                 if (!list_empty(&sched->kss_rx_conns)) {
1417                         conn = list_entry(sched->kss_rx_conns.next,
1418                                               ksock_conn_t, ksnc_rx_list);
1419                         list_del(&conn->ksnc_rx_list);
1420
1421                         LASSERT(conn->ksnc_rx_scheduled);
1422                         LASSERT(conn->ksnc_rx_ready);
1423
1424                         /* clear rx_ready in case receive isn't complete.
1425                          * Do it BEFORE we call process_recv, since
1426                          * data_ready can set it any time after we release
1427                          * kss_lock. */
1428                         conn->ksnc_rx_ready = 0;
1429                         spin_unlock_bh(&sched->kss_lock);
1430
1431                         rc = ksocknal_process_receive(conn);
1432
1433                         spin_lock_bh(&sched->kss_lock);
1434
1435                         /* I'm the only one that can clear this flag */
1436                         LASSERT(conn->ksnc_rx_scheduled);
1437
1438                         /* Did process_receive get everything it wanted? */
1439                         if (rc == 0)
1440                                 conn->ksnc_rx_ready = 1;
1441
1442                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1443                                 /* Conn blocked waiting for ksocknal_recv()
1444                                  * I change its state (under lock) to signal
1445                                  * it can be rescheduled */
1446                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1447                         } else if (conn->ksnc_rx_ready) {
1448                                 /* reschedule for rx */
1449                                 list_add_tail(&conn->ksnc_rx_list,
1450                                                    &sched->kss_rx_conns);
1451                         } else {
1452                                 conn->ksnc_rx_scheduled = 0;
1453                                 /* drop my ref */
1454                                 ksocknal_conn_decref(conn);
1455                         }
1456
1457                         did_something = 1;
1458                 }
1459
1460                 if (!list_empty(&sched->kss_tx_conns)) {
1461                         struct list_head zlist = LIST_HEAD_INIT(zlist);
1462
1463                         if (!list_empty(&sched->kss_zombie_noop_txs)) {
1464                                 list_add(&zlist,
1465                                              &sched->kss_zombie_noop_txs);
1466                                 list_del_init(&sched->kss_zombie_noop_txs);
1467                         }
1468
1469                         conn = list_entry(sched->kss_tx_conns.next,
1470                                               ksock_conn_t, ksnc_tx_list);
1471                         list_del(&conn->ksnc_tx_list);
1472
1473                         LASSERT(conn->ksnc_tx_scheduled);
1474                         LASSERT(conn->ksnc_tx_ready);
1475                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1476
1477                         tx = list_entry(conn->ksnc_tx_queue.next,
1478                                             ksock_tx_t, tx_list);
1479
1480                         if (conn->ksnc_tx_carrier == tx)
1481                                 ksocknal_next_tx_carrier(conn);
1482
1483                         /* dequeue now so empty list => more to send */
1484                         list_del(&tx->tx_list);
1485
1486                         /* Clear tx_ready in case send isn't complete.  Do
1487                          * it BEFORE we call process_transmit, since
1488                          * write_space can set it any time after we release
1489                          * kss_lock. */
1490                         conn->ksnc_tx_ready = 0;
1491                         spin_unlock_bh(&sched->kss_lock);
1492
1493                         if (!list_empty(&zlist)) {
1494                                 /* free zombie noop txs, it's fast because
1495                                  * noop txs are just put in freelist */
1496                                 ksocknal_txlist_done(NULL, &zlist, 0);
1497                         }
1498
1499                         rc = ksocknal_process_transmit(conn, tx);
1500
1501                         if (rc == -ENOMEM || rc == -EAGAIN) {
1502                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1503                                 spin_lock_bh(&sched->kss_lock);
1504                                 list_add(&tx->tx_list,
1505                                              &conn->ksnc_tx_queue);
1506                         } else {
1507                                 /* Complete send; tx -ref */
1508                                 ksocknal_tx_decref(tx);
1509
1510                                 spin_lock_bh(&sched->kss_lock);
1511                                 /* assume space for more */
1512                                 conn->ksnc_tx_ready = 1;
1513                         }
1514
1515                         if (rc == -ENOMEM) {
1516                                 /* Do nothing; after a short timeout, this
1517                                  * conn will be reposted on kss_tx_conns. */
1518                         } else if (conn->ksnc_tx_ready &&
1519                                    !list_empty(&conn->ksnc_tx_queue)) {
1520                                 /* reschedule for tx */
1521                                 list_add_tail(&conn->ksnc_tx_list,
1522                                                    &sched->kss_tx_conns);
1523                         } else {
1524                                 conn->ksnc_tx_scheduled = 0;
1525                                 /* drop my ref */
1526                                 ksocknal_conn_decref(conn);
1527                         }
1528
1529                         did_something = 1;
1530                 }
1531                 if (!did_something ||           /* nothing to do */
1532                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1533                         spin_unlock_bh(&sched->kss_lock);
1534
1535                         nloops = 0;
1536
1537                         if (!did_something) {   /* wait for something to do */
1538                                 rc = wait_event_interruptible_exclusive(
1539                                         sched->kss_waitq,
1540                                         !ksocknal_sched_cansleep(sched));
1541                                 LASSERT (rc == 0);
1542                         } else {
1543                                 cond_resched();
1544                         }
1545
1546                         spin_lock_bh(&sched->kss_lock);
1547                 }
1548         }
1549
1550         spin_unlock_bh(&sched->kss_lock);
1551         ksocknal_thread_fini();
1552         return 0;
1553 }
1554
1555 /*
1556  * Add connection to kss_rx_conns of scheduler
1557  * and wakeup the scheduler.
1558  */
1559 void ksocknal_read_callback (ksock_conn_t *conn)
1560 {
1561         ksock_sched_t *sched;
1562         ENTRY;
1563
1564         sched = conn->ksnc_scheduler;
1565
1566         spin_lock_bh(&sched->kss_lock);
1567
1568         conn->ksnc_rx_ready = 1;
1569
1570         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1571                 list_add_tail(&conn->ksnc_rx_list,
1572                                   &sched->kss_rx_conns);
1573                 conn->ksnc_rx_scheduled = 1;
1574                 /* extra ref for scheduler */
1575                 ksocknal_conn_addref(conn);
1576
1577                 wake_up (&sched->kss_waitq);
1578         }
1579         spin_unlock_bh(&sched->kss_lock);
1580
1581         EXIT;
1582 }
1583
1584 /*
1585  * Add connection to kss_tx_conns of scheduler
1586  * and wakeup the scheduler.
1587  */
1588 void ksocknal_write_callback(ksock_conn_t *conn)
1589 {
1590         ksock_sched_t *sched;
1591         ENTRY;
1592
1593         sched = conn->ksnc_scheduler;
1594
1595         spin_lock_bh(&sched->kss_lock);
1596
1597         conn->ksnc_tx_ready = 1;
1598
1599         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1600             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1601                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1602                 conn->ksnc_tx_scheduled = 1;
1603                 /* extra ref for scheduler */
1604                 ksocknal_conn_addref(conn);
1605
1606                 wake_up(&sched->kss_waitq);
1607         }
1608
1609         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         struct list_head        zombies = LIST_HEAD_INIT(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         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                 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                 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         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1955         return retry_later;
1956
1957  failed:
1958         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 (!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 (!list_empty(&peer->ksnp_conns)) {
1984                         conn = 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                 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                 list_del(&route->ksnr_list);
1998                 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1999         }
2000 #endif
2001         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 /*
2009  * check whether we need to create more connds.
2010  * It will try to create new thread if it's necessary, @timeout can
2011  * be updated if failed to create, so caller wouldn't keep try while
2012  * running out of resource.
2013  */
2014 static int
2015 ksocknal_connd_check_start(long sec, long *timeout)
2016 {
2017         char name[16];
2018         int rc;
2019         int total = ksocknal_data.ksnd_connd_starting +
2020                     ksocknal_data.ksnd_connd_running;
2021
2022         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2023                 /* still in initializing */
2024                 return 0;
2025         }
2026
2027         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2028             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2029                 /* can't create more connd, or still have enough
2030                  * threads to handle more connecting */
2031                 return 0;
2032         }
2033
2034         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2035                 /* no pending connecting request */
2036                 return 0;
2037         }
2038
2039         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2040                 /* may run out of resource, retry later */
2041                 *timeout = cfs_time_seconds(1);
2042                 return 0;
2043         }
2044
2045         if (ksocknal_data.ksnd_connd_starting > 0) {
2046                 /* serialize starting to avoid flood */
2047                 return 0;
2048         }
2049
2050         ksocknal_data.ksnd_connd_starting_stamp = sec;
2051         ksocknal_data.ksnd_connd_starting++;
2052         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2053
2054         /* NB: total is the next id */
2055         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2056         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2057
2058         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2059         if (rc == 0)
2060                 return 1;
2061
2062         /* we tried ... */
2063         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2064         ksocknal_data.ksnd_connd_starting--;
2065         ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2066
2067         return 1;
2068 }
2069
2070 /*
2071  * check whether current thread can exit, it will return 1 if there are too
2072  * many threads and no creating in past 120 seconds.
2073  * Also, this function may update @timeout to make caller come back
2074  * again to recheck these conditions.
2075  */
2076 static int
2077 ksocknal_connd_check_stop(long sec, long *timeout)
2078 {
2079         int val;
2080
2081         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2082                 /* still in initializing */
2083                 return 0;
2084         }
2085
2086         if (ksocknal_data.ksnd_connd_starting > 0) {
2087                 /* in progress of starting new thread */
2088                 return 0;
2089         }
2090
2091         if (ksocknal_data.ksnd_connd_running <=
2092             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2093                 return 0;
2094         }
2095
2096         /* created thread in past 120 seconds? */
2097         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2098                     SOCKNAL_CONND_TIMEOUT - sec);
2099
2100         *timeout = (val > 0) ? cfs_time_seconds(val) :
2101                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2102         if (val > 0)
2103                 return 0;
2104
2105         /* no creating in past 120 seconds */
2106
2107         return ksocknal_data.ksnd_connd_running >
2108                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2109 }
2110
2111 /* Go through connd_routes queue looking for a route that we can process
2112  * right now, @timeout_p can be updated if we need to come back later */
2113 static ksock_route_t *
2114 ksocknal_connd_get_route_locked(signed long *timeout_p)
2115 {
2116         ksock_route_t *route;
2117         cfs_time_t     now;
2118
2119         now = cfs_time_current();
2120
2121         /* connd_routes can contain both pending and ordinary routes */
2122         list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2123                                  ksnr_connd_list) {
2124
2125                 if (route->ksnr_retry_interval == 0 ||
2126                     cfs_time_aftereq(now, route->ksnr_timeout))
2127                         return route;
2128
2129                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2130                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2131                         *timeout_p = (int)(route->ksnr_timeout - now);
2132         }
2133
2134         return NULL;
2135 }
2136
2137 int
2138 ksocknal_connd (void *arg)
2139 {
2140         spinlock_t    *connd_lock = &ksocknal_data.ksnd_connd_lock;
2141         ksock_connreq_t   *cr;
2142         wait_queue_t     wait;
2143         int                nloops = 0;
2144         int                cons_retry = 0;
2145
2146         cfs_block_allsigs ();
2147
2148         init_waitqueue_entry_current(&wait);
2149
2150         spin_lock_bh(connd_lock);
2151
2152         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2153         ksocknal_data.ksnd_connd_starting--;
2154         ksocknal_data.ksnd_connd_running++;
2155
2156         while (!ksocknal_data.ksnd_shuttingdown) {
2157                 ksock_route_t *route = NULL;
2158                 long sec = cfs_time_current_sec();
2159                 long timeout = MAX_SCHEDULE_TIMEOUT;
2160                 int  dropped_lock = 0;
2161
2162                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2163                         /* wakeup another one to check stop */
2164                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2165                         break;
2166                 }
2167
2168                 if (ksocknal_connd_check_start(sec, &timeout)) {
2169                         /* created new thread */
2170                         dropped_lock = 1;
2171                 }
2172
2173                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2174                         /* Connection accepted by the listener */
2175                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2176                                             next, ksock_connreq_t, ksncr_list);
2177
2178                         list_del(&cr->ksncr_list);
2179                         spin_unlock_bh(connd_lock);
2180                         dropped_lock = 1;
2181
2182                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2183                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2184                         lnet_ni_decref(cr->ksncr_ni);
2185                         LIBCFS_FREE(cr, sizeof(*cr));
2186
2187                         spin_lock_bh(connd_lock);
2188                 }
2189
2190                 /* Only handle an outgoing connection request if there
2191                  * is a thread left to handle incoming connections and
2192                  * create new connd */
2193                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2194                     ksocknal_data.ksnd_connd_running) {
2195                         route = ksocknal_connd_get_route_locked(&timeout);
2196                 }
2197                 if (route != NULL) {
2198                         list_del(&route->ksnr_connd_list);
2199                         ksocknal_data.ksnd_connd_connecting++;
2200                         spin_unlock_bh(connd_lock);
2201                         dropped_lock = 1;
2202
2203                         if (ksocknal_connect(route)) {
2204                                 /* consecutive retry */
2205                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2206                                         CWARN("massive consecutive "
2207                                               "re-connecting to %u.%u.%u.%u\n",
2208                                               HIPQUAD(route->ksnr_ipaddr));
2209                                         cons_retry = 0;
2210                                 }
2211                         } else {
2212                                 cons_retry = 0;
2213                         }
2214
2215                         ksocknal_route_decref(route);
2216
2217                         spin_lock_bh(connd_lock);
2218                         ksocknal_data.ksnd_connd_connecting--;
2219                 }
2220
2221                 if (dropped_lock) {
2222                         if (++nloops < SOCKNAL_RESCHED)
2223                                 continue;
2224                         spin_unlock_bh(connd_lock);
2225                         nloops = 0;
2226                         cond_resched();
2227                         spin_lock_bh(connd_lock);
2228                         continue;
2229                 }
2230
2231                 /* Nothing to do for 'timeout'  */
2232                 set_current_state(TASK_INTERRUPTIBLE);
2233                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2234                 spin_unlock_bh(connd_lock);
2235
2236                 nloops = 0;
2237                 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2238
2239                 set_current_state(TASK_RUNNING);
2240                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2241                 spin_lock_bh(connd_lock);
2242         }
2243         ksocknal_data.ksnd_connd_running--;
2244         spin_unlock_bh(connd_lock);
2245
2246         ksocknal_thread_fini();
2247         return 0;
2248 }
2249
2250 ksock_conn_t *
2251 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2252 {
2253         /* We're called with a shared lock on ksnd_global_lock */
2254         ksock_conn_t      *conn;
2255         struct list_head        *ctmp;
2256
2257         list_for_each(ctmp, &peer->ksnp_conns) {
2258                 int     error;
2259                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
2260
2261                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2262                 LASSERT (!conn->ksnc_closing);
2263
2264                 error = libcfs_sock_error(conn->ksnc_sock);
2265                 if (error != 0) {
2266                         ksocknal_conn_addref(conn);
2267
2268                         switch (error) {
2269                         case ECONNRESET:
2270                                 CNETERR("A connection with %s "
2271                                         "(%u.%u.%u.%u:%d) was reset; "
2272                                         "it may have rebooted.\n",
2273                                         libcfs_id2str(peer->ksnp_id),
2274                                         HIPQUAD(conn->ksnc_ipaddr),
2275                                         conn->ksnc_port);
2276                                 break;
2277                         case ETIMEDOUT:
2278                                 CNETERR("A connection with %s "
2279                                         "(%u.%u.%u.%u:%d) timed out; the "
2280                                         "network or node may be down.\n",
2281                                         libcfs_id2str(peer->ksnp_id),
2282                                         HIPQUAD(conn->ksnc_ipaddr),
2283                                         conn->ksnc_port);
2284                                 break;
2285                         default:
2286                                 CNETERR("An unexpected network error %d "
2287                                         "occurred with %s "
2288                                         "(%u.%u.%u.%u:%d\n", error,
2289                                         libcfs_id2str(peer->ksnp_id),
2290                                         HIPQUAD(conn->ksnc_ipaddr),
2291                                         conn->ksnc_port);
2292                                 break;
2293                         }
2294
2295                         return (conn);
2296                 }
2297
2298                 if (conn->ksnc_rx_started &&
2299                     cfs_time_aftereq(cfs_time_current(),
2300                                      conn->ksnc_rx_deadline)) {
2301                         /* Timed out incomplete incoming message */
2302                         ksocknal_conn_addref(conn);
2303                         CNETERR("Timeout receiving from %s (%u.%u.%u.%u:%d), "
2304                                 "state %d wanted %d left %d\n",
2305                                 libcfs_id2str(peer->ksnp_id),
2306                                 HIPQUAD(conn->ksnc_ipaddr),
2307                                 conn->ksnc_port,
2308                                 conn->ksnc_rx_state,
2309                                 conn->ksnc_rx_nob_wanted,
2310                                 conn->ksnc_rx_nob_left);
2311                         return (conn);
2312                 }
2313
2314                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2315                      libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2316                     cfs_time_aftereq(cfs_time_current(),
2317                                      conn->ksnc_tx_deadline)) {
2318                         /* Timed out messages queued for sending or
2319                          * buffered in the socket's send buffer */
2320                         ksocknal_conn_addref(conn);
2321                         CNETERR("Timeout sending data to %s (%u.%u.%u.%u:%d) "
2322                                 "the network or that node may be down.\n",
2323                                 libcfs_id2str(peer->ksnp_id),
2324                                 HIPQUAD(conn->ksnc_ipaddr),
2325                                 conn->ksnc_port);
2326                         return (conn);
2327                 }
2328         }
2329
2330         return (NULL);
2331 }
2332
2333 static inline void
2334 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2335 {
2336         ksock_tx_t        *tx;
2337         struct list_head        stale_txs = LIST_HEAD_INIT(stale_txs);
2338
2339         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2340
2341         while (!list_empty(&peer->ksnp_tx_queue)) {
2342                 tx = list_entry(peer->ksnp_tx_queue.next,
2343                                      ksock_tx_t, tx_list);
2344
2345                 if (!cfs_time_aftereq(cfs_time_current(),
2346                                       tx->tx_deadline))
2347                         break;
2348
2349                 list_del(&tx->tx_list);
2350                 list_add_tail(&tx->tx_list, &stale_txs);
2351         }
2352
2353         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2354
2355         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2356 }
2357
2358 int
2359 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2360 __must_hold(&ksocknal_data.ksnd_global_lock)
2361 {
2362         ksock_sched_t  *sched;
2363         ksock_conn_t   *conn;
2364         ksock_tx_t     *tx;
2365
2366         /* last_alive will be updated by create_conn */
2367         if (list_empty(&peer->ksnp_conns))
2368                 return 0;
2369
2370         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2371                 return 0;
2372
2373         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2374             cfs_time_before(cfs_time_current(),
2375                             cfs_time_add(peer->ksnp_last_alive,
2376                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2377                 return 0;
2378
2379         if (cfs_time_before(cfs_time_current(),
2380                             peer->ksnp_send_keepalive))
2381                 return 0;
2382
2383         /* retry 10 secs later, so we wouldn't put pressure
2384          * on this peer if we failed to send keepalive this time */
2385         peer->ksnp_send_keepalive = cfs_time_shift(10);
2386
2387         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2388         if (conn != NULL) {
2389                 sched = conn->ksnc_scheduler;
2390
2391                 spin_lock_bh(&sched->kss_lock);
2392                 if (!list_empty(&conn->ksnc_tx_queue)) {
2393                         spin_unlock_bh(&sched->kss_lock);
2394                         /* there is an queued ACK, don't need keepalive */
2395                         return 0;
2396                 }
2397
2398                 spin_unlock_bh(&sched->kss_lock);
2399         }
2400
2401         read_unlock(&ksocknal_data.ksnd_global_lock);
2402
2403         /* cookie = 1 is reserved for keepalive PING */
2404         tx = ksocknal_alloc_tx_noop(1, 1);
2405         if (tx == NULL) {
2406                 read_lock(&ksocknal_data.ksnd_global_lock);
2407                 return -ENOMEM;
2408         }
2409
2410         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2411                 read_lock(&ksocknal_data.ksnd_global_lock);
2412                 return 1;
2413         }
2414
2415         ksocknal_free_tx(tx);
2416         read_lock(&ksocknal_data.ksnd_global_lock);
2417
2418         return -EIO;
2419 }
2420
2421
2422 void
2423 ksocknal_check_peer_timeouts (int idx)
2424 {
2425         struct list_head       *peers = &ksocknal_data.ksnd_peers[idx];
2426         ksock_peer_t     *peer;
2427         ksock_conn_t     *conn;
2428         ksock_tx_t       *tx;
2429
2430  again:
2431         /* NB. We expect to have a look at all the peers and not find any
2432          * connections to time out, so we just use a shared lock while we
2433          * take a look... */
2434         read_lock(&ksocknal_data.ksnd_global_lock);
2435
2436         list_for_each_entry(peer, peers, ksnp_list) {
2437                 cfs_time_t  deadline = 0;
2438                 int         resid = 0;
2439                 int         n     = 0;
2440
2441                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2442                         read_unlock(&ksocknal_data.ksnd_global_lock);
2443                         goto again;
2444                 }
2445
2446                 conn = ksocknal_find_timed_out_conn (peer);
2447
2448                 if (conn != NULL) {
2449                         read_unlock(&ksocknal_data.ksnd_global_lock);
2450
2451                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2452
2453                         /* NB we won't find this one again, but we can't
2454                          * just proceed with the next peer, since we dropped
2455                          * ksnd_global_lock and it might be dead already! */
2456                         ksocknal_conn_decref(conn);
2457                         goto again;
2458                 }
2459
2460                 /* we can't process stale txs right here because we're
2461                  * holding only shared lock */
2462                 if (!list_empty(&peer->ksnp_tx_queue)) {
2463                         ksock_tx_t *tx =
2464                                 list_entry(peer->ksnp_tx_queue.next,
2465                                                 ksock_tx_t, tx_list);
2466
2467                         if (cfs_time_aftereq(cfs_time_current(),
2468                                              tx->tx_deadline)) {
2469
2470                                 ksocknal_peer_addref(peer);
2471                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2472
2473                                 ksocknal_flush_stale_txs(peer);
2474
2475                                 ksocknal_peer_decref(peer);
2476                                 goto again;
2477                         }
2478                 }
2479
2480                 if (list_empty(&peer->ksnp_zc_req_list))
2481                         continue;
2482
2483                 spin_lock(&peer->ksnp_lock);
2484                 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2485                         if (!cfs_time_aftereq(cfs_time_current(),
2486                                               tx->tx_deadline))
2487                                 break;
2488                         /* ignore the TX if connection is being closed */
2489                         if (tx->tx_conn->ksnc_closing)
2490                                 continue;
2491                         n++;
2492                 }
2493
2494                 if (n == 0) {
2495                         spin_unlock(&peer->ksnp_lock);
2496                         continue;
2497                 }
2498
2499                 tx = list_entry(peer->ksnp_zc_req_list.next,
2500                                     ksock_tx_t, tx_zc_list);
2501                 deadline = tx->tx_deadline;
2502                 resid    = tx->tx_resid;
2503                 conn     = tx->tx_conn;
2504                 ksocknal_conn_addref(conn);
2505
2506                 spin_unlock(&peer->ksnp_lock);
2507                 read_unlock(&ksocknal_data.ksnd_global_lock);
2508
2509                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2510                        "oldest(%p) timed out %ld secs ago, "
2511                        "resid: %d, wmem: %d\n",
2512                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2513                        cfs_duration_sec(cfs_time_current() - deadline),
2514                        resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2515
2516                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2517                 ksocknal_conn_decref(conn);
2518                 goto again;
2519         }
2520
2521         read_unlock(&ksocknal_data.ksnd_global_lock);
2522 }
2523
2524 int ksocknal_reaper(void *arg)
2525 {
2526         wait_queue_t     wait;
2527         ksock_conn_t      *conn;
2528         ksock_sched_t     *sched;
2529         struct list_head         enomem_conns;
2530         int                nenomem_conns;
2531         cfs_duration_t     timeout;
2532         int                i;
2533         int                peer_index = 0;
2534         cfs_time_t         deadline = cfs_time_current();
2535
2536         cfs_block_allsigs ();
2537
2538         INIT_LIST_HEAD(&enomem_conns);
2539         init_waitqueue_entry_current(&wait);
2540
2541         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2542
2543         while (!ksocknal_data.ksnd_shuttingdown) {
2544
2545                 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2546                         conn = list_entry(ksocknal_data. \
2547                                                ksnd_deathrow_conns.next,
2548                                                ksock_conn_t, ksnc_list);
2549                         list_del(&conn->ksnc_list);
2550
2551                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2552
2553                         ksocknal_terminate_conn(conn);
2554                         ksocknal_conn_decref(conn);
2555
2556                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2557                         continue;
2558                 }
2559
2560                 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2561                         conn = list_entry(ksocknal_data.ksnd_zombie_conns.\
2562                                                next, ksock_conn_t, ksnc_list);
2563                         list_del(&conn->ksnc_list);
2564
2565                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2566
2567                         ksocknal_destroy_conn(conn);
2568
2569                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2570                         continue;
2571                 }
2572
2573                 if (!list_empty(&ksocknal_data.ksnd_enomem_conns)) {
2574                         list_add(&enomem_conns,
2575                                      &ksocknal_data.ksnd_enomem_conns);
2576                         list_del_init(&ksocknal_data.ksnd_enomem_conns);
2577                 }
2578
2579                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2580
2581                 /* reschedule all the connections that stalled with ENOMEM... */
2582                 nenomem_conns = 0;
2583                 while (!list_empty(&enomem_conns)) {
2584                         conn = list_entry(enomem_conns.next,
2585                                                ksock_conn_t, ksnc_tx_list);
2586                         list_del(&conn->ksnc_tx_list);
2587
2588                         sched = conn->ksnc_scheduler;
2589
2590                         spin_lock_bh(&sched->kss_lock);
2591
2592                         LASSERT(conn->ksnc_tx_scheduled);
2593                         conn->ksnc_tx_ready = 1;
2594                         list_add_tail(&conn->ksnc_tx_list,
2595                                           &sched->kss_tx_conns);
2596                         wake_up(&sched->kss_waitq);
2597
2598                         spin_unlock_bh(&sched->kss_lock);
2599                         nenomem_conns++;
2600                 }
2601
2602                 /* careful with the jiffy wrap... */
2603                 while ((timeout = cfs_time_sub(deadline,
2604                                                cfs_time_current())) <= 0) {
2605                         const int n = 4;
2606                         const int p = 1;
2607                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2608
2609                         /* Time to check for timeouts on a few more peers: I do
2610                          * checks every 'p' seconds on a proportion of the peer
2611                          * table and I need to check every connection 'n' times
2612                          * within a timeout interval, to ensure I detect a
2613                          * timeout on any connection within (n+1)/n times the
2614                          * timeout interval. */
2615
2616                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2617                                 chunk = (chunk * n * p) /
2618                                         *ksocknal_tunables.ksnd_timeout;
2619                         if (chunk == 0)
2620                                 chunk = 1;
2621
2622                         for (i = 0; i < chunk; i++) {
2623                                 ksocknal_check_peer_timeouts (peer_index);
2624                                 peer_index = (peer_index + 1) %
2625                                              ksocknal_data.ksnd_peer_hash_size;
2626                         }
2627
2628                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2629                 }
2630
2631                 if (nenomem_conns != 0) {
2632                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2633                          * This also prevents me getting woken immediately
2634                          * if any go back on my enomem list. */
2635                         timeout = SOCKNAL_ENOMEM_RETRY;
2636                 }
2637                 ksocknal_data.ksnd_reaper_waketime =
2638                         cfs_time_add(cfs_time_current(), timeout);
2639
2640                         set_current_state(TASK_INTERRUPTIBLE);
2641                 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2642
2643                 if (!ksocknal_data.ksnd_shuttingdown &&
2644                     list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2645                     list_empty(&ksocknal_data.ksnd_zombie_conns))
2646                         waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2647
2648                 set_current_state(TASK_RUNNING);
2649                 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2650
2651                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2652         }
2653
2654         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2655
2656         ksocknal_thread_fini();
2657         return 0;
2658 }