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

   1 /*
   2  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2012, Intel Corporation.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Portals, http://www.sf.net/projects/sandiaportals/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Portals; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include "socklnd.h"
  28
  29 ksock_tx_t *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         ksock_tx_t *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         list_del(&tx->tx_list);
  45                 }
  46
  47                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_desc_size  = size;
  61
  62         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  63
  64         return tx;
  65 }
  66
  67 ksock_tx_t *
  68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  69 {
  70         ksock_tx_t *tx;
  71
  72         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  73         if (tx == NULL) {
  74                 CERROR("Can't allocate noop tx desc\n");
  75                 return NULL;
  76         }
  77
  78         tx->tx_conn    = NULL;
  79         tx->tx_lnetmsg = NULL;
  80         tx->tx_kiov    = NULL;
  81         tx->tx_nkiov   = 0;
  82         tx->tx_iov     = tx->tx_frags.virt.iov;
  83         tx->tx_niov    = 1;
  84         tx->tx_nonblk  = nonblk;
  85
  86         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  87         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  88
  89         return tx;
  90 }
  91
  92 void
  93 ksocknal_free_tx (ksock_tx_t *tx)
  94 {
  95         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  96
  97         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  98                 /* it's a noop tx */
  99                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 100
 101                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 102
 103                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 104         } else {
 105                 LIBCFS_FREE(tx, tx->tx_desc_size);
 106         }
 107 }
 108
 109 static int
 110 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
 111 {
 112         struct kvec *iov = tx->tx_iov;
 113         int nob;
 114         int rc;
 115
 116         LASSERT(tx->tx_niov > 0);
 117
 118         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 119         rc = ksocknal_lib_send_iov(conn, tx);
 120
 121         if (rc <= 0)                        /* sent nothing? */
 122                 return rc;
 123
 124         nob = rc;
 125         LASSERT (nob <= tx->tx_resid);
 126         tx->tx_resid -= nob;
 127
 128         /* "consume" iov */
 129         do {
 130                 LASSERT(tx->tx_niov > 0);
 131
 132                 if (nob < (int) iov->iov_len) {
 133                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 134                         iov->iov_len -= nob;
 135                         return rc;
 136                 }
 137
 138                 nob -= iov->iov_len;
 139                 tx->tx_iov = ++iov;
 140                 tx->tx_niov--;
 141         } while (nob != 0);
 142
 143         return rc;
 144 }
 145
 146 static int
 147 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
 148 {
 149         lnet_kiov_t *kiov = tx->tx_kiov;
 150         int nob;
 151         int rc;
 152
 153         LASSERT(tx->tx_niov == 0);
 154         LASSERT(tx->tx_nkiov > 0);
 155
 156         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 157         rc = ksocknal_lib_send_kiov(conn, tx);
 158
 159         if (rc <= 0)                        /* sent nothing? */
 160                 return rc;
 161
 162         nob = rc;
 163         LASSERT (nob <= tx->tx_resid);
 164         tx->tx_resid -= nob;
 165
 166         /* "consume" kiov */
 167         do {
 168                 LASSERT(tx->tx_nkiov > 0);
 169
 170                 if (nob < (int)kiov->kiov_len) {
 171                         kiov->kiov_offset += nob;
 172                         kiov->kiov_len -= nob;
 173                         return rc;
 174                 }
 175
 176                 nob -= (int)kiov->kiov_len;
 177                 tx->tx_kiov = ++kiov;
 178                 tx->tx_nkiov--;
 179         } while (nob != 0);
 180
 181         return rc;
 182 }
 183
 184 static int
 185 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 186 {
 187         int rc;
 188         int bufnob;
 189
 190         if (ksocknal_data.ksnd_stall_tx != 0) {
 191                 set_current_state(TASK_UNINTERRUPTIBLE);
 192                 schedule_timeout(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 = conn->ksnc_sock->sk->sk_wmem_queued;
 215                 if (rc > 0)                  /* sent something? */
 216                         conn->ksnc_tx_bufnob += rc; /* account it */
 217
 218                 if (bufnob < conn->ksnc_tx_bufnob) {
 219                         /* allocated send buffer bytes < computed; infer
 220                          * something got ACKed */
 221                         conn->ksnc_tx_deadline =
 222                                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 223                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 224                         conn->ksnc_tx_bufnob = bufnob;
 225                         mb();
 226                 }
 227
 228                 if (rc <= 0) { /* Didn't write anything? */
 229
 230                         if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
 231                                 rc = -EAGAIN;
 232
 233                         /* Check if EAGAIN is due to memory pressure */
 234                         if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 235                                 rc = -ENOMEM;
 236
 237                         break;
 238                 }
 239
 240                 /* socket's wmem_queued now includes 'rc' bytes */
 241                 atomic_sub (rc, &conn->ksnc_tx_nob);
 242                 rc = 0;
 243
 244         } while (tx->tx_resid != 0);
 245
 246         ksocknal_connsock_decref(conn);
 247         return rc;
 248 }
 249
 250 static int
 251 ksocknal_recv_iov (ksock_conn_t *conn)
 252 {
 253         struct kvec *iov = conn->ksnc_rx_iov;
 254         int nob;
 255         int rc;
 256
 257         LASSERT(conn->ksnc_rx_niov > 0);
 258
 259         /* Never touch conn->ksnc_rx_iov or change connection
 260          * status inside ksocknal_lib_recv_iov */
 261         rc = ksocknal_lib_recv_iov(conn);
 262
 263         if (rc <= 0)
 264                 return rc;
 265
 266         /* received something... */
 267         nob = rc;
 268
 269         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 270         conn->ksnc_rx_deadline =
 271                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 272         mb();                  /* order with setting rx_started */
 273         conn->ksnc_rx_started = 1;
 274
 275         conn->ksnc_rx_nob_wanted -= nob;
 276         conn->ksnc_rx_nob_left -= nob;
 277
 278         do {
 279                 LASSERT(conn->ksnc_rx_niov > 0);
 280
 281                 if (nob < (int)iov->iov_len) {
 282                         iov->iov_len -= nob;
 283                         iov->iov_base += 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 static 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
 302         LASSERT(conn->ksnc_rx_nkiov > 0);
 303
 304         /* Never touch conn->ksnc_rx_kiov or change connection
 305          * status inside ksocknal_lib_recv_iov */
 306         rc = ksocknal_lib_recv_kiov(conn);
 307
 308         if (rc <= 0)
 309                 return rc;
 310
 311         /* received something... */
 312         nob = rc;
 313
 314         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 315         conn->ksnc_rx_deadline =
 316                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 317         mb();                  /* order with setting rx_started */
 318         conn->ksnc_rx_started = 1;
 319
 320         conn->ksnc_rx_nob_wanted -= nob;
 321         conn->ksnc_rx_nob_left -= nob;
 322
 323         do {
 324                 LASSERT(conn->ksnc_rx_nkiov > 0);
 325
 326                 if (nob < (int) kiov->kiov_len) {
 327                         kiov->kiov_offset += nob;
 328                         kiov->kiov_len -= nob;
 329                         return -EAGAIN;
 330                 }
 331
 332                 nob -= kiov->kiov_len;
 333                 conn->ksnc_rx_kiov = ++kiov;
 334                 conn->ksnc_rx_nkiov--;
 335         } while (nob != 0);
 336
 337         return 1;
 338 }
 339
 340 static int
 341 ksocknal_receive (ksock_conn_t *conn)
 342 {
 343         /* Return 1 on success, 0 on EOF, < 0 on error.
 344          * Caller checks ksnc_rx_nob_wanted to determine
 345          * progress/completion. */
 346         int rc;
 347
 348         if (ksocknal_data.ksnd_stall_rx != 0) {
 349                 set_current_state(TASK_UNINTERRUPTIBLE);
 350                 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
 351         }
 352
 353         rc = ksocknal_connsock_addref(conn);
 354         if (rc != 0) {
 355                 LASSERT (conn->ksnc_closing);
 356                 return -ESHUTDOWN;
 357         }
 358
 359         for (;;) {
 360                 if (conn->ksnc_rx_niov != 0)
 361                         rc = ksocknal_recv_iov (conn);
 362                 else
 363                         rc = ksocknal_recv_kiov (conn);
 364
 365                 if (rc <= 0) {
 366                         /* error/EOF or partial receive */
 367                         if (rc == -EAGAIN) {
 368                                 rc = 1;
 369                         } else if (rc == 0 && conn->ksnc_rx_started) {
 370                                 /* EOF in the middle of a message */
 371                                 rc = -EPROTO;
 372                         }
 373                         break;
 374                 }
 375
 376                 /* Completed a fragment */
 377
 378                 if (conn->ksnc_rx_nob_wanted == 0) {
 379                         rc = 1;
 380                         break;
 381                 }
 382         }
 383
 384         ksocknal_connsock_decref(conn);
 385         return rc;
 386 }
 387
 388 void
 389 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
 390 {
 391         lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
 392         int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
 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
 407 void
 408 ksocknal_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int error)
 409 {
 410         ksock_tx_t *tx;
 411
 412         while (!list_empty (txlist)) {
 413                 tx = list_entry(txlist->next, ksock_tx_t, tx_list);
 414
 415                 if (error && tx->tx_lnetmsg != NULL) {
 416                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 417                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
 418                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
 419                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 420                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 421                 } else if (error) {
 422                         CNETERR("Deleting noop packet\n");
 423                 }
 424
 425                 list_del(&tx->tx_list);
 426
 427                 LASSERT(atomic_read(&tx->tx_refcount) == 1);
 428                 ksocknal_tx_done(ni, tx);
 429         }
 430 }
 431
 432 static void
 433 ksocknal_check_zc_req(ksock_tx_t *tx)
 434 {
 435         ksock_conn_t *conn = tx->tx_conn;
 436         ksock_peer_t *peer = conn->ksnc_peer;
 437
 438         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 439          * to ksnp_zc_req_list if some fragment of this message should be sent
 440          * zero-copy.  Our peer will send an ACK containing this cookie when
 441          * she has received this message to tell us we can signal completion.
 442          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 443          * ksnp_zc_req_list. */
 444         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 445         LASSERT(tx->tx_zc_capable);
 446
 447         tx->tx_zc_checked = 1;
 448
 449         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 450             !conn->ksnc_zc_capable)
 451                 return;
 452
 453         /* assign cookie and queue tx to pending list, it will be released when
 454          * a matching ack is received. See ksocknal_handle_zcack() */
 455
 456         ksocknal_tx_addref(tx);
 457
 458         spin_lock(&peer->ksnp_lock);
 459
 460         /* ZC_REQ is going to be pinned to the peer */
 461         tx->tx_deadline =
 462                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 463
 464         LASSERT(tx->tx_msg.ksm_zc_cookies[0] == 0);
 465
 466         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 467
 468         if (peer->ksnp_zc_next_cookie == 0)
 469                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 470
 471         list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 472
 473         spin_unlock(&peer->ksnp_lock);
 474 }
 475
 476 static void
 477 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 478 {
 479         ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
 480
 481         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 482         LASSERT(tx->tx_zc_capable);
 483
 484         tx->tx_zc_checked = 0;
 485
 486         spin_lock(&peer->ksnp_lock);
 487
 488         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 489                 /* Not waiting for an ACK */
 490                 spin_unlock(&peer->ksnp_lock);
 491                 return;
 492         }
 493
 494         tx->tx_msg.ksm_zc_cookies[0] = 0;
 495         list_del(&tx->tx_zc_list);
 496
 497         spin_unlock(&peer->ksnp_lock);
 498
 499         ksocknal_tx_decref(tx);
 500 }
 501
 502 static int
 503 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 504 {
 505         int rc;
 506
 507         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 508                 ksocknal_check_zc_req(tx);
 509
 510         rc = ksocknal_transmit (conn, tx);
 511
 512         CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 513
 514         if (tx->tx_resid == 0) {
 515                 /* Sent everything OK */
 516                 LASSERT (rc == 0);
 517
 518                 return 0;
 519         }
 520
 521         if (rc == -EAGAIN)
 522                 return rc;
 523
 524         if (rc == -ENOMEM) {
 525                 static int counter;
 526
 527                 counter++;   /* exponential backoff warnings */
 528                 if ((counter & (-counter)) == counter)
 529                         CWARN("%u ENOMEM tx %p\n", counter, conn);
 530
 531                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 532                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 533
 534                 /* enomem list takes over scheduler's ref... */
 535                 LASSERT (conn->ksnc_tx_scheduled);
 536                 list_add_tail(&conn->ksnc_tx_list,
 537                                   &ksocknal_data.ksnd_enomem_conns);
 538                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 539                                                    SOCKNAL_ENOMEM_RETRY),
 540                                    ksocknal_data.ksnd_reaper_waketime))
 541                         wake_up (&ksocknal_data.ksnd_reaper_waitq);
 542
 543                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 544                 return rc;
 545         }
 546
 547         /* Actual error */
 548         LASSERT(rc < 0);
 549
 550         if (!conn->ksnc_closing) {
 551                 switch (rc) {
 552                 case -ECONNRESET:
 553                         LCONSOLE_WARN("Host %pI4h reset our connection while we were sending data; it may have rebooted.\n",
 554                                       &conn->ksnc_ipaddr);
 555                         break;
 556                 default:
 557                         LCONSOLE_WARN("There was an unexpected network error while writing to %pI4h: %d.\n",
 558                                       &conn->ksnc_ipaddr, rc);
 559                         break;
 560                 }
 561                 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
 562                        conn, rc,
 563                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 564                        &conn->ksnc_ipaddr,
 565                        conn->ksnc_port);
 566         }
 567
 568         if (tx->tx_zc_checked)
 569                 ksocknal_uncheck_zc_req(tx);
 570
 571         /* it's not an error if conn is being closed */
 572         ksocknal_close_conn_and_siblings (conn,
 573                                           (conn->ksnc_closing) ? 0 : rc);
 574
 575         return rc;
 576 }
 577
 578 static void
 579 ksocknal_launch_connection_locked (ksock_route_t *route)
 580 {
 581
 582         /* called holding write lock on ksnd_global_lock */
 583
 584         LASSERT(!route->ksnr_scheduled);
 585         LASSERT(!route->ksnr_connecting);
 586         LASSERT((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 587
 588         route->ksnr_scheduled = 1;            /* scheduling conn for connd */
 589         ksocknal_route_addref(route);      /* extra ref for connd */
 590
 591         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 592
 593         list_add_tail(&route->ksnr_connd_list,
 594                           &ksocknal_data.ksnd_connd_routes);
 595         wake_up(&ksocknal_data.ksnd_connd_waitq);
 596
 597         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 598 }
 599
 600 void
 601 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 602 {
 603         ksock_route_t *route;
 604
 605         /* called holding write lock on ksnd_global_lock */
 606         for (;;) {
 607                 /* launch any/all connections that need it */
 608                 route = ksocknal_find_connectable_route_locked(peer);
 609                 if (route == NULL)
 610                         return;
 611
 612                 ksocknal_launch_connection_locked(route);
 613         }
 614 }
 615
 616 ksock_conn_t *
 617 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 618 {
 619         struct list_head *tmp;
 620         ksock_conn_t *conn;
 621         ksock_conn_t *typed = NULL;
 622         ksock_conn_t *fallback = NULL;
 623         int tnob = 0;
 624         int fnob = 0;
 625
 626         list_for_each (tmp, &peer->ksnp_conns) {
 627                 ksock_conn_t *c  = list_entry(tmp, ksock_conn_t, ksnc_list);
 628                 int nob = atomic_read(&c->ksnc_tx_nob) +
 629                         c->ksnc_sock->sk->sk_wmem_queued;
 630                 int rc;
 631
 632                 LASSERT(!c->ksnc_closing);
 633                 LASSERT(c->ksnc_proto != NULL &&
 634                         c->ksnc_proto->pro_match_tx != NULL);
 635
 636                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 637
 638                 switch (rc) {
 639                 default:
 640                         LBUG();
 641                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 642                         continue;
 643
 644                 case SOCKNAL_MATCH_YES: /* typed connection */
 645                         if (typed == NULL || tnob > nob ||
 646                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 647                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 648                                 typed = c;
 649                                 tnob  = nob;
 650                         }
 651                         break;
 652
 653                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 654                         if (fallback == NULL || fnob > nob ||
 655                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 656                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 657                                 fallback = c;
 658                                 fnob = nob;
 659                         }
 660                         break;
 661                 }
 662         }
 663
 664         /* prefer the typed selection */
 665         conn = (typed != NULL) ? typed : fallback;
 666
 667         if (conn != NULL)
 668                 conn->ksnc_tx_last_post = cfs_time_current();
 669
 670         return conn;
 671 }
 672
 673 void
 674 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 675 {
 676         conn->ksnc_proto->pro_pack(tx);
 677
 678         atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 679         ksocknal_conn_addref(conn); /* +1 ref for tx */
 680         tx->tx_conn = conn;
 681 }
 682
 683 void
 684 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
 685 {
 686         ksock_sched_t *sched = conn->ksnc_scheduler;
 687         ksock_msg_t *msg = &tx->tx_msg;
 688         ksock_tx_t *ztx = NULL;
 689         int bufnob = 0;
 690
 691         /* called holding global lock (read or irq-write) and caller may
 692          * not have dropped this lock between finding conn and calling me,
 693          * so we don't need the {get,put}connsock dance to deref
 694          * ksnc_sock... */
 695         LASSERT(!conn->ksnc_closing);
 696
 697         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 698                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
 699                 &conn->ksnc_ipaddr,
 700                 conn->ksnc_port);
 701
 702         ksocknal_tx_prep(conn, tx);
 703
 704         /* Ensure the frags we've been given EXACTLY match the number of
 705          * bytes we want to send.  Many TCP/IP stacks disregard any total
 706          * size parameters passed to them and just look at the frags.
 707          *
 708          * We always expect at least 1 mapped fragment containing the
 709          * complete ksocknal message header. */
 710         LASSERT(lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 711                 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 712                 (unsigned int)tx->tx_nob);
 713         LASSERT(tx->tx_niov >= 1);
 714         LASSERT(tx->tx_resid == tx->tx_nob);
 715
 716         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 717                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type :
 718                                                KSOCK_MSG_NOOP,
 719                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 720
 721         /*
 722          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 723          * but they're used inside spinlocks a lot.
 724          */
 725         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 726         spin_lock_bh(&sched->kss_lock);
 727
 728         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 729                 /* First packet starts the timeout */
 730                 conn->ksnc_tx_deadline =
 731                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 732                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 733                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 734                 conn->ksnc_tx_bufnob = 0;
 735                 mb(); /* order with adding to tx_queue */
 736         }
 737
 738         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 739                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 740                  * on a normal packet so I don't need to send it */
 741                 LASSERT(msg->ksm_zc_cookies[1] != 0);
 742                 LASSERT(conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 743
 744                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 745                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 746
 747         } else {
 748                 /* It's a normal packet - can it piggback a noop zc-ack that
 749                  * has been queued already? */
 750                 LASSERT(msg->ksm_zc_cookies[1] == 0);
 751                 LASSERT(conn->ksnc_proto->pro_queue_tx_msg != NULL);
 752
 753                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 754                 /* ztx will be released later */
 755         }
 756
 757         if (ztx != NULL) {
 758                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 759                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 760         }
 761
 762         if (conn->ksnc_tx_ready &&      /* able to send */
 763             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 764                 /* +1 ref for scheduler */
 765                 ksocknal_conn_addref(conn);
 766                 list_add_tail (&conn->ksnc_tx_list,
 767                                    &sched->kss_tx_conns);
 768                 conn->ksnc_tx_scheduled = 1;
 769                 wake_up (&sched->kss_waitq);
 770         }
 771
 772         spin_unlock_bh(&sched->kss_lock);
 773 }
 774
 775 ksock_route_t *
 776 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 777 {
 778         unsigned long now = cfs_time_current();
 779         struct list_head *tmp;
 780         ksock_route_t *route;
 781
 782         list_for_each (tmp, &peer->ksnp_routes) {
 783                 route = list_entry (tmp, ksock_route_t, ksnr_list);
 784
 785                 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
 786
 787                 if (route->ksnr_scheduled)      /* connections being established */
 788                         continue;
 789
 790                 /* all route types connected ? */
 791                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 792                         continue;
 793
 794                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 795                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 796                         CDEBUG(D_NET,
 797                                "Too soon to retry route %pI4h (cnted %d, interval %ld, %ld secs later)\n",
 798                                &route->ksnr_ipaddr,
 799                                route->ksnr_connected,
 800                                route->ksnr_retry_interval,
 801                                cfs_duration_sec(route->ksnr_timeout - now));
 802                         continue;
 803                 }
 804
 805                 return route;
 806         }
 807
 808         return NULL;
 809 }
 810
 811 ksock_route_t *
 812 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 813 {
 814         struct list_head *tmp;
 815         ksock_route_t *route;
 816
 817         list_for_each (tmp, &peer->ksnp_routes) {
 818                 route = list_entry (tmp, ksock_route_t, ksnr_list);
 819
 820                 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
 821
 822                 if (route->ksnr_scheduled)
 823                         return route;
 824         }
 825
 826         return NULL;
 827 }
 828
 829 int
 830 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 831 {
 832         ksock_peer_t *peer;
 833         ksock_conn_t *conn;
 834         rwlock_t *g_lock;
 835         int retry;
 836         int rc;
 837
 838         LASSERT(tx->tx_conn == NULL);
 839
 840         g_lock = &ksocknal_data.ksnd_global_lock;
 841
 842         for (retry = 0;; retry = 1) {
 843                 read_lock(g_lock);
 844                 peer = ksocknal_find_peer_locked(ni, id);
 845                 if (peer != NULL) {
 846                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 847                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 848                                 if (conn != NULL) {
 849                                         /* I've got no routes that need to be
 850                                          * connecting and I do have an actual
 851                                          * connection... */
 852                                         ksocknal_queue_tx_locked (tx, conn);
 853                                         read_unlock(g_lock);
 854                                         return 0;
 855                                 }
 856                         }
 857                 }
 858
 859                 /* I'll need a write lock... */
 860                 read_unlock(g_lock);
 861
 862                 write_lock_bh(g_lock);
 863
 864                 peer = ksocknal_find_peer_locked(ni, id);
 865                 if (peer != NULL)
 866                         break;
 867
 868                 write_unlock_bh(g_lock);
 869
 870                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 871                         CERROR("Refusing to create a connection to userspace process %s\n",
 872                                libcfs_id2str(id));
 873                         return -EHOSTUNREACH;
 874                 }
 875
 876                 if (retry) {
 877                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 878                         return -EHOSTUNREACH;
 879                 }
 880
 881                 rc = ksocknal_add_peer(ni, id,
 882                                        LNET_NIDADDR(id.nid),
 883                                        lnet_acceptor_port());
 884                 if (rc != 0) {
 885                         CERROR("Can't add peer %s: %d\n",
 886                                libcfs_id2str(id), rc);
 887                         return rc;
 888                 }
 889         }
 890
 891         ksocknal_launch_all_connections_locked(peer);
 892
 893         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 894         if (conn != NULL) {
 895                 /* Connection exists; queue message on it */
 896                 ksocknal_queue_tx_locked (tx, conn);
 897                 write_unlock_bh(g_lock);
 898                 return 0;
 899         }
 900
 901         if (peer->ksnp_accepting > 0 ||
 902             ksocknal_find_connecting_route_locked (peer) != NULL) {
 903                 /* the message is going to be pinned to the peer */
 904                 tx->tx_deadline =
 905                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 906
 907                 /* Queue the message until a connection is established */
 908                 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
 909                 write_unlock_bh(g_lock);
 910                 return 0;
 911         }
 912
 913         write_unlock_bh(g_lock);
 914
 915         /* NB Routes may be ignored if connections to them failed recently */
 916         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 917         return -EHOSTUNREACH;
 918 }
 919
 920 int
 921 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 922 {
 923         int mpflag = 1;
 924         int type = lntmsg->msg_type;
 925         lnet_process_id_t target = lntmsg->msg_target;
 926         unsigned int payload_niov = lntmsg->msg_niov;
 927         struct kvec *payload_iov = lntmsg->msg_iov;
 928         lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
 929         unsigned int payload_offset = lntmsg->msg_offset;
 930         unsigned int payload_nob = lntmsg->msg_len;
 931         ksock_tx_t *tx;
 932         int desc_size;
 933         int rc;
 934
 935         /* NB 'private' is different depending on what we're sending.
 936          * Just ignore it... */
 937
 938         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 939                payload_nob, payload_niov, libcfs_id2str(target));
 940
 941         LASSERT(payload_nob == 0 || payload_niov > 0);
 942         LASSERT(payload_niov <= LNET_MAX_IOV);
 943         /* payload is either all vaddrs or all pages */
 944         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 945         LASSERT (!in_interrupt ());
 946
 947         if (payload_iov != NULL)
 948                 desc_size = offsetof(ksock_tx_t,
 949                                      tx_frags.virt.iov[1 + payload_niov]);
 950         else
 951                 desc_size = offsetof(ksock_tx_t,
 952                                      tx_frags.paged.kiov[payload_niov]);
 953
 954         if (lntmsg->msg_vmflush)
 955                 mpflag = cfs_memory_pressure_get_and_set();
 956         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 957         if (tx == NULL) {
 958                 CERROR("Can't allocate tx desc type %d size %d\n",
 959                        type, desc_size);
 960                 if (lntmsg->msg_vmflush)
 961                         cfs_memory_pressure_restore(mpflag);
 962                 return -ENOMEM;
 963         }
 964
 965         tx->tx_conn = NULL;                  /* set when assigned a conn */
 966         tx->tx_lnetmsg = lntmsg;
 967
 968         if (payload_iov != NULL) {
 969                 tx->tx_kiov = NULL;
 970                 tx->tx_nkiov = 0;
 971                 tx->tx_iov = tx->tx_frags.virt.iov;
 972                 tx->tx_niov = 1 +
 973                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 974                                                payload_niov, payload_iov,
 975                                                payload_offset, payload_nob);
 976         } else {
 977                 tx->tx_niov = 1;
 978                 tx->tx_iov = &tx->tx_frags.paged.iov;
 979                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 980                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 981                                                  payload_niov, payload_kiov,
 982                                                  payload_offset, payload_nob);
 983
 984                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 985                         tx->tx_zc_capable = 1;
 986         }
 987
 988         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 989
 990         /* The first fragment will be set later in pro_pack */
 991         rc = ksocknal_launch_packet(ni, tx, target);
 992         if (!mpflag)
 993                 cfs_memory_pressure_restore(mpflag);
 994
 995         if (rc == 0)
 996                 return 0;
 997
 998         ksocknal_free_tx(tx);
 999         return -EIO;
1000 }
1001
1002 int
1003 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1004 {
1005         struct task_struct *task = kthread_run(fn, arg, "%s", name);
1006
1007         if (IS_ERR(task))
1008                 return PTR_ERR(task);
1009
1010         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1011         ksocknal_data.ksnd_nthreads++;
1012         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1013         return 0;
1014 }
1015
1016 void
1017 ksocknal_thread_fini (void)
1018 {
1019         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1020         ksocknal_data.ksnd_nthreads--;
1021         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1022 }
1023
1024 int
1025 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1026 {
1027         static char ksocknal_slop_buffer[4096];
1028
1029         int nob;
1030         unsigned int niov;
1031         int skipped;
1032
1033         LASSERT(conn->ksnc_proto != NULL);
1034
1035         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1036                 /* Remind the socket to ack eagerly... */
1037                 ksocknal_lib_eager_ack(conn);
1038         }
1039
1040         if (nob_to_skip == 0) {  /* right at next packet boundary now */
1041                 conn->ksnc_rx_started = 0;
1042                 mb();                  /* racing with timeout thread */
1043
1044                 switch (conn->ksnc_proto->pro_version) {
1045                 case  KSOCK_PROTO_V2:
1046                 case  KSOCK_PROTO_V3:
1047                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1048                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1049                         conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg;
1050
1051                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1052                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1053                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1054                         break;
1055
1056                 case KSOCK_PROTO_V1:
1057                         /* Receiving bare lnet_hdr_t */
1058                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1059                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1060                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1061
1062                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1063                         conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1064                         conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t);
1065                         break;
1066
1067                 default:
1068                         LBUG ();
1069                 }
1070                 conn->ksnc_rx_niov = 1;
1071
1072                 conn->ksnc_rx_kiov = NULL;
1073                 conn->ksnc_rx_nkiov = 0;
1074                 conn->ksnc_rx_csum = ~0;
1075                 return 1;
1076         }
1077
1078         /* Set up to skip as much as possible now.  If there's more left
1079          * (ran out of iov entries) we'll get called again */
1080
1081         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1082         conn->ksnc_rx_nob_left = nob_to_skip;
1083         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1084         skipped = 0;
1085         niov = 0;
1086
1087         do {
1088                 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1089
1090                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1091                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1092                 niov++;
1093                 skipped += nob;
1094                 nob_to_skip -= nob;
1095
1096         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1097                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1098
1099         conn->ksnc_rx_niov = niov;
1100         conn->ksnc_rx_kiov = NULL;
1101         conn->ksnc_rx_nkiov = 0;
1102         conn->ksnc_rx_nob_wanted = skipped;
1103         return 0;
1104 }
1105
1106 static int
1107 ksocknal_process_receive (ksock_conn_t *conn)
1108 {
1109         lnet_hdr_t *lhdr;
1110         lnet_process_id_t *id;
1111         int rc;
1112
1113         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1114
1115         /* NB: sched lock NOT held */
1116         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1117         LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1118                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1119                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1120                 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1121  again:
1122         if (conn->ksnc_rx_nob_wanted != 0) {
1123                 rc = ksocknal_receive(conn);
1124
1125                 if (rc <= 0) {
1126                         LASSERT (rc != -EAGAIN);
1127
1128                         if (rc == 0)
1129                                 CDEBUG(D_NET, "[%p] EOF from %s ip %pI4h:%d\n",
1130                                        conn,
1131                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1132                                        &conn->ksnc_ipaddr,
1133                                        conn->ksnc_port);
1134                         else if (!conn->ksnc_closing)
1135                                 CERROR("[%p] Error %d on read from %s ip %pI4h:%d\n",
1136                                        conn, rc,
1137                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1138                                        &conn->ksnc_ipaddr,
1139                                        conn->ksnc_port);
1140
1141                         /* it's not an error if conn is being closed */
1142                         ksocknal_close_conn_and_siblings (conn,
1143                                                           (conn->ksnc_closing) ? 0 : rc);
1144                         return (rc == 0 ? -ESHUTDOWN : rc);
1145                 }
1146
1147                 if (conn->ksnc_rx_nob_wanted != 0) {
1148                         /* short read */
1149                         return -EAGAIN;
1150                 }
1151         }
1152         switch (conn->ksnc_rx_state) {
1153         case SOCKNAL_RX_KSM_HEADER:
1154                 if (conn->ksnc_flip) {
1155                         __swab32s(&conn->ksnc_msg.ksm_type);
1156                         __swab32s(&conn->ksnc_msg.ksm_csum);
1157                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1158                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1159                 }
1160
1161                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1162                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1163                         CERROR("%s: Unknown message type: %x\n",
1164                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1165                                conn->ksnc_msg.ksm_type);
1166                         ksocknal_new_packet(conn, 0);
1167                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1168                         return -EPROTO;
1169                 }
1170
1171                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1172                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1173                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1174                         /* NOOP Checksum error */
1175                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1176                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1177                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1178                         ksocknal_new_packet(conn, 0);
1179                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1180                         return -EIO;
1181                 }
1182
1183                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1184                         __u64 cookie = 0;
1185
1186                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1187
1188                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1189                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1190
1191                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1192                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1193
1194                         if (rc != 0) {
1195                                 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1196                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1197                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1198                                 ksocknal_new_packet(conn, 0);
1199                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1200                                 return rc;
1201                         }
1202                 }
1203
1204                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1205                         ksocknal_new_packet (conn, 0);
1206                         return 0;       /* NOOP is done and just return */
1207                 }
1208
1209                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1210                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1211                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1212
1213                 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1214                 conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1215                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1216
1217                 conn->ksnc_rx_niov = 1;
1218                 conn->ksnc_rx_kiov = NULL;
1219                 conn->ksnc_rx_nkiov = 0;
1220
1221                 goto again;     /* read lnet header now */
1222
1223         case SOCKNAL_RX_LNET_HEADER:
1224                 /* unpack message header */
1225                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1226
1227                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1228                         /* Userspace peer */
1229                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1230                         id = &conn->ksnc_peer->ksnp_id;
1231
1232                         /* Substitute process ID assigned at connection time */
1233                         lhdr->src_pid = cpu_to_le32(id->pid);
1234                         lhdr->src_nid = cpu_to_le64(id->nid);
1235                 }
1236
1237                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1238                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1239
1240                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1241                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1242                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1243                 if (rc < 0) {
1244                         /* I just received garbage: give up on this conn */
1245                         ksocknal_new_packet(conn, 0);
1246                         ksocknal_close_conn_and_siblings (conn, rc);
1247                         ksocknal_conn_decref(conn);
1248                         return -EPROTO;
1249                 }
1250
1251                 /* I'm racing with ksocknal_recv() */
1252                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1253                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1254
1255                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1256                         return 0;
1257
1258                 /* ksocknal_recv() got called */
1259                 goto again;
1260
1261         case SOCKNAL_RX_LNET_PAYLOAD:
1262                 /* payload all received */
1263                 rc = 0;
1264
1265                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1266                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1267                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1268                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1269                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1270                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1271                         rc = -EIO;
1272                 }
1273
1274                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1275                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1276
1277                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1278                         id = &conn->ksnc_peer->ksnp_id;
1279
1280                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1281                                         conn->ksnc_msg.ksm_zc_cookies[0],
1282                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1283                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1284                 }
1285
1286                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1287
1288                 if (rc != 0) {
1289                         ksocknal_new_packet(conn, 0);
1290                         ksocknal_close_conn_and_siblings (conn, rc);
1291                         return -EPROTO;
1292                 }
1293                 /* Fall through */
1294
1295         case SOCKNAL_RX_SLOP:
1296                 /* starting new packet? */
1297                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1298                         return 0;       /* come back later */
1299                 goto again;          /* try to finish reading slop now */
1300
1301         default:
1302                 break;
1303         }
1304
1305         /* Not Reached */
1306         LBUG();
1307         return -EINVAL;                /* keep gcc happy */
1308 }
1309
1310 int
1311 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1312                unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1313                unsigned int offset, unsigned int mlen, unsigned int rlen)
1314 {
1315         ksock_conn_t *conn = private;
1316         ksock_sched_t *sched = conn->ksnc_scheduler;
1317
1318         LASSERT(mlen <= rlen);
1319         LASSERT(niov <= LNET_MAX_IOV);
1320
1321         conn->ksnc_cookie = msg;
1322         conn->ksnc_rx_nob_wanted = mlen;
1323         conn->ksnc_rx_nob_left = rlen;
1324
1325         if (mlen == 0 || iov != NULL) {
1326                 conn->ksnc_rx_nkiov = 0;
1327                 conn->ksnc_rx_kiov = NULL;
1328                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1329                 conn->ksnc_rx_niov =
1330                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1331                                          niov, iov, offset, mlen);
1332         } else {
1333                 conn->ksnc_rx_niov = 0;
1334                 conn->ksnc_rx_iov = NULL;
1335                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1336                 conn->ksnc_rx_nkiov =
1337                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1338                                           niov, kiov, offset, mlen);
1339         }
1340
1341         LASSERT(mlen ==
1342                 lnet_iov_nob(conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1343                 lnet_kiov_nob(conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1344
1345         LASSERT(conn->ksnc_rx_scheduled);
1346
1347         spin_lock_bh(&sched->kss_lock);
1348
1349         switch (conn->ksnc_rx_state) {
1350         case SOCKNAL_RX_PARSE_WAIT:
1351                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1352                 wake_up (&sched->kss_waitq);
1353                 LASSERT (conn->ksnc_rx_ready);
1354                 break;
1355
1356         case SOCKNAL_RX_PARSE:
1357                 /* scheduler hasn't noticed I'm parsing yet */
1358                 break;
1359         }
1360
1361         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1362
1363         spin_unlock_bh(&sched->kss_lock);
1364         ksocknal_conn_decref(conn);
1365         return 0;
1366 }
1367
1368 static inline int
1369 ksocknal_sched_cansleep(ksock_sched_t *sched)
1370 {
1371         int rc;
1372
1373         spin_lock_bh(&sched->kss_lock);
1374
1375         rc = !ksocknal_data.ksnd_shuttingdown &&
1376               list_empty(&sched->kss_rx_conns) &&
1377               list_empty(&sched->kss_tx_conns);
1378
1379         spin_unlock_bh(&sched->kss_lock);
1380         return rc;
1381 }
1382
1383 int ksocknal_scheduler(void *arg)
1384 {
1385         struct ksock_sched_info *info;
1386         ksock_sched_t *sched;
1387         ksock_conn_t *conn;
1388         ksock_tx_t *tx;
1389         int rc;
1390         int nloops = 0;
1391         long id = (long)arg;
1392
1393         info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1394         sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1395
1396         cfs_block_allsigs();
1397
1398         rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1399         if (rc != 0) {
1400                 CERROR("Can't set CPT affinity to %d: %d\n",
1401                        info->ksi_cpt, rc);
1402         }
1403
1404         spin_lock_bh(&sched->kss_lock);
1405
1406         while (!ksocknal_data.ksnd_shuttingdown) {
1407                 int did_something = 0;
1408
1409                 /* Ensure I progress everything semi-fairly */
1410
1411                 if (!list_empty (&sched->kss_rx_conns)) {
1412                         conn = list_entry(sched->kss_rx_conns.next,
1413                                               ksock_conn_t, ksnc_rx_list);
1414                         list_del(&conn->ksnc_rx_list);
1415
1416                         LASSERT(conn->ksnc_rx_scheduled);
1417                         LASSERT(conn->ksnc_rx_ready);
1418
1419                         /* clear rx_ready in case receive isn't complete.
1420                          * Do it BEFORE we call process_recv, since
1421                          * data_ready can set it any time after we release
1422                          * kss_lock. */
1423                         conn->ksnc_rx_ready = 0;
1424                         spin_unlock_bh(&sched->kss_lock);
1425
1426                         rc = ksocknal_process_receive(conn);
1427
1428                         spin_lock_bh(&sched->kss_lock);
1429
1430                         /* I'm the only one that can clear this flag */
1431                         LASSERT(conn->ksnc_rx_scheduled);
1432
1433                         /* Did process_receive get everything it wanted? */
1434                         if (rc == 0)
1435                                 conn->ksnc_rx_ready = 1;
1436
1437                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1438                                 /* Conn blocked waiting for ksocknal_recv()
1439                                  * I change its state (under lock) to signal
1440                                  * it can be rescheduled */
1441                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1442                         } else if (conn->ksnc_rx_ready) {
1443                                 /* reschedule for rx */
1444                                 list_add_tail (&conn->ksnc_rx_list,
1445                                                    &sched->kss_rx_conns);
1446                         } else {
1447                                 conn->ksnc_rx_scheduled = 0;
1448                                 /* drop my ref */
1449                                 ksocknal_conn_decref(conn);
1450                         }
1451
1452                         did_something = 1;
1453                 }
1454
1455                 if (!list_empty (&sched->kss_tx_conns)) {
1456                         LIST_HEAD(zlist);
1457
1458                         if (!list_empty(&sched->kss_zombie_noop_txs)) {
1459                                 list_add(&zlist,
1460                                              &sched->kss_zombie_noop_txs);
1461                                 list_del_init(&sched->kss_zombie_noop_txs);
1462                         }
1463
1464                         conn = list_entry(sched->kss_tx_conns.next,
1465                                               ksock_conn_t, ksnc_tx_list);
1466                         list_del (&conn->ksnc_tx_list);
1467
1468                         LASSERT(conn->ksnc_tx_scheduled);
1469                         LASSERT(conn->ksnc_tx_ready);
1470                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1471
1472                         tx = list_entry(conn->ksnc_tx_queue.next,
1473                                             ksock_tx_t, tx_list);
1474
1475                         if (conn->ksnc_tx_carrier == tx)
1476                                 ksocknal_next_tx_carrier(conn);
1477
1478                         /* dequeue now so empty list => more to send */
1479                         list_del(&tx->tx_list);
1480
1481                         /* Clear tx_ready in case send isn't complete.  Do
1482                          * it BEFORE we call process_transmit, since
1483                          * write_space can set it any time after we release
1484                          * kss_lock. */
1485                         conn->ksnc_tx_ready = 0;
1486                         spin_unlock_bh(&sched->kss_lock);
1487
1488                         if (!list_empty(&zlist)) {
1489                                 /* free zombie noop txs, it's fast because
1490                                  * noop txs are just put in freelist */
1491                                 ksocknal_txlist_done(NULL, &zlist, 0);
1492                         }
1493
1494                         rc = ksocknal_process_transmit(conn, tx);
1495
1496                         if (rc == -ENOMEM || rc == -EAGAIN) {
1497                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1498                                 spin_lock_bh(&sched->kss_lock);
1499                                 list_add(&tx->tx_list,
1500                                              &conn->ksnc_tx_queue);
1501                         } else {
1502                                 /* Complete send; tx -ref */
1503                                 ksocknal_tx_decref(tx);
1504
1505                                 spin_lock_bh(&sched->kss_lock);
1506                                 /* assume space for more */
1507                                 conn->ksnc_tx_ready = 1;
1508                         }
1509
1510                         if (rc == -ENOMEM) {
1511                                 /* Do nothing; after a short timeout, this
1512                                  * conn will be reposted on kss_tx_conns. */
1513                         } else if (conn->ksnc_tx_ready &&
1514                                    !list_empty(&conn->ksnc_tx_queue)) {
1515                                 /* reschedule for tx */
1516                                 list_add_tail(&conn->ksnc_tx_list,
1517                                                    &sched->kss_tx_conns);
1518                         } else {
1519                                 conn->ksnc_tx_scheduled = 0;
1520                                 /* drop my ref */
1521                                 ksocknal_conn_decref(conn);
1522                         }
1523
1524                         did_something = 1;
1525                 }
1526                 if (!did_something ||      /* nothing to do */
1527                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1528                         spin_unlock_bh(&sched->kss_lock);
1529
1530                         nloops = 0;
1531
1532                         if (!did_something) {   /* wait for something to do */
1533                                 rc = wait_event_interruptible_exclusive(
1534                                         sched->kss_waitq,
1535                                         !ksocknal_sched_cansleep(sched));
1536                                 LASSERT (rc == 0);
1537                         } else {
1538                                 cond_resched();
1539                         }
1540
1541                         spin_lock_bh(&sched->kss_lock);
1542                 }
1543         }
1544
1545         spin_unlock_bh(&sched->kss_lock);
1546         ksocknal_thread_fini();
1547         return 0;
1548 }
1549
1550 /*
1551  * Add connection to kss_rx_conns of scheduler
1552  * and wakeup the scheduler.
1553  */
1554 void ksocknal_read_callback (ksock_conn_t *conn)
1555 {
1556         ksock_sched_t *sched;
1557
1558         sched = conn->ksnc_scheduler;
1559
1560         spin_lock_bh(&sched->kss_lock);
1561
1562         conn->ksnc_rx_ready = 1;
1563
1564         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1565                 list_add_tail(&conn->ksnc_rx_list,
1566                                   &sched->kss_rx_conns);
1567                 conn->ksnc_rx_scheduled = 1;
1568                 /* extra ref for scheduler */
1569                 ksocknal_conn_addref(conn);
1570
1571                 wake_up (&sched->kss_waitq);
1572         }
1573         spin_unlock_bh(&sched->kss_lock);
1574 }
1575
1576 /*
1577  * Add connection to kss_tx_conns of scheduler
1578  * and wakeup the scheduler.
1579  */
1580 void ksocknal_write_callback (ksock_conn_t *conn)
1581 {
1582         ksock_sched_t *sched;
1583
1584         sched = conn->ksnc_scheduler;
1585
1586         spin_lock_bh(&sched->kss_lock);
1587
1588         conn->ksnc_tx_ready = 1;
1589
1590         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1591             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1592                 list_add_tail (&conn->ksnc_tx_list,
1593                                    &sched->kss_tx_conns);
1594                 conn->ksnc_tx_scheduled = 1;
1595                 /* extra ref for scheduler */
1596                 ksocknal_conn_addref(conn);
1597
1598                 wake_up (&sched->kss_waitq);
1599         }
1600
1601         spin_unlock_bh(&sched->kss_lock);
1602 }
1603
1604 static ksock_proto_t *
1605 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1606 {
1607         __u32 version = 0;
1608
1609         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1610                 version = hello->kshm_version;
1611         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1612                 version = __swab32(hello->kshm_version);
1613
1614         if (version != 0) {
1615 #if SOCKNAL_VERSION_DEBUG
1616                 if (*ksocknal_tunables.ksnd_protocol == 1)
1617                         return NULL;
1618
1619                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1620                     version == KSOCK_PROTO_V3)
1621                         return NULL;
1622 #endif
1623                 if (version == KSOCK_PROTO_V2)
1624                         return &ksocknal_protocol_v2x;
1625
1626                 if (version == KSOCK_PROTO_V3)
1627                         return &ksocknal_protocol_v3x;
1628
1629                 return NULL;
1630         }
1631
1632         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1633                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1634
1635                 CLASSERT(sizeof (lnet_magicversion_t) ==
1636                          offsetof (ksock_hello_msg_t, kshm_src_nid));
1637
1638                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1639                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1640                         return &ksocknal_protocol_v1x;
1641         }
1642
1643         return NULL;
1644 }
1645
1646 int
1647 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1648                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1649 {
1650         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1651         ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1652
1653         LASSERT(hello->kshm_nips <= LNET_MAX_INTERFACES);
1654
1655         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1656         LASSERT(conn->ksnc_proto != NULL);
1657
1658         hello->kshm_src_nid = ni->ni_nid;
1659         hello->kshm_dst_nid = peer_nid;
1660         hello->kshm_src_pid = the_lnet.ln_pid;
1661
1662         hello->kshm_src_incarnation = net->ksnn_incarnation;
1663         hello->kshm_ctype = conn->ksnc_type;
1664
1665         return conn->ksnc_proto->pro_send_hello(conn, hello);
1666 }
1667
1668 static int
1669 ksocknal_invert_type(int type)
1670 {
1671         switch (type) {
1672         case SOCKLND_CONN_ANY:
1673         case SOCKLND_CONN_CONTROL:
1674                 return type;
1675         case SOCKLND_CONN_BULK_IN:
1676                 return SOCKLND_CONN_BULK_OUT;
1677         case SOCKLND_CONN_BULK_OUT:
1678                 return SOCKLND_CONN_BULK_IN;
1679         default:
1680                 return SOCKLND_CONN_NONE;
1681         }
1682 }
1683
1684 int
1685 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1686                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1687                      __u64 *incarnation)
1688 {
1689         /* Return < 0   fatal error
1690          *      0         success
1691          *      EALREADY   lost connection race
1692          *      EPROTO     protocol version mismatch
1693          */
1694         struct socket *sock = conn->ksnc_sock;
1695         int active = (conn->ksnc_proto != NULL);
1696         int timeout;
1697         int proto_match;
1698         int rc;
1699         ksock_proto_t *proto;
1700         lnet_process_id_t recv_id;
1701
1702         /* socket type set on active connections - not set on passive */
1703         LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1704
1705         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1706                             lnet_acceptor_timeout();
1707
1708         rc = lnet_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1709         if (rc != 0) {
1710                 CERROR("Error %d reading HELLO from %pI4h\n",
1711                         rc, &conn->ksnc_ipaddr);
1712                 LASSERT (rc < 0);
1713                 return rc;
1714         }
1715
1716         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1717             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1718             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1719                 /* Unexpected magic! */
1720                 CERROR("Bad magic(1) %#08x (%#08x expected) from %pI4h\n",
1721                        __cpu_to_le32 (hello->kshm_magic),
1722                        LNET_PROTO_TCP_MAGIC,
1723                        &conn->ksnc_ipaddr);
1724                 return -EPROTO;
1725         }
1726
1727         rc = lnet_sock_read(sock, &hello->kshm_version,
1728                             sizeof(hello->kshm_version), timeout);
1729         if (rc != 0) {
1730                 CERROR("Error %d reading HELLO from %pI4h\n",
1731                         rc, &conn->ksnc_ipaddr);
1732                 LASSERT(rc < 0);
1733                 return rc;
1734         }
1735
1736         proto = ksocknal_parse_proto_version(hello);
1737         if (proto == NULL) {
1738                 if (!active) {
1739                         /* unknown protocol from peer, tell peer my protocol */
1740                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1741 #if SOCKNAL_VERSION_DEBUG
1742                         if (*ksocknal_tunables.ksnd_protocol == 2)
1743                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1744                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1745                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1746 #endif
1747                         hello->kshm_nips = 0;
1748                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1749                 }
1750
1751                 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1752                        conn->ksnc_proto->pro_version,
1753                        &conn->ksnc_ipaddr);
1754
1755                 return -EPROTO;
1756         }
1757
1758         proto_match = (conn->ksnc_proto == proto);
1759         conn->ksnc_proto = proto;
1760
1761         /* receive the rest of hello message anyway */
1762         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1763         if (rc != 0) {
1764                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1765                        rc, &conn->ksnc_ipaddr);
1766                 LASSERT(rc < 0);
1767                 return rc;
1768         }
1769
1770         *incarnation = hello->kshm_src_incarnation;
1771
1772         if (hello->kshm_src_nid == LNET_NID_ANY) {
1773                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1774                        &conn->ksnc_ipaddr);
1775                 return -EPROTO;
1776         }
1777
1778         if (!active &&
1779             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1780                 /* Userspace NAL assigns peer process ID from socket */
1781                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1782                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1783         } else {
1784                 recv_id.nid = hello->kshm_src_nid;
1785                 recv_id.pid = hello->kshm_src_pid;
1786         }
1787
1788         if (!active) {
1789                 *peerid = recv_id;
1790
1791                 /* peer determines type */
1792                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1793                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1794                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1795                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1796                                 &conn->ksnc_ipaddr);
1797                         return -EPROTO;
1798                 }
1799
1800                 return 0;
1801         }
1802
1803         if (peerid->pid != recv_id.pid ||
1804             peerid->nid != recv_id.nid) {
1805                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host %pI4h, but they claimed they were %s; please check your Lustre configuration.\n",
1806                                    libcfs_id2str(*peerid),
1807                                    &conn->ksnc_ipaddr,
1808                                    libcfs_id2str(recv_id));
1809                 return -EPROTO;
1810         }
1811
1812         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1813                 /* Possible protocol mismatch or I lost the connection race */
1814                 return proto_match ? EALREADY : EPROTO;
1815         }
1816
1817         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1818                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1819                         conn->ksnc_type, libcfs_id2str(*peerid),
1820                         &conn->ksnc_ipaddr,
1821                         hello->kshm_ctype);
1822                 return -EPROTO;
1823         }
1824
1825         return 0;
1826 }
1827
1828 static int
1829 ksocknal_connect (ksock_route_t *route)
1830 {
1831         LIST_HEAD(zombies);
1832         ksock_peer_t *peer = route->ksnr_peer;
1833         int type;
1834         int wanted;
1835         struct socket *sock;
1836         unsigned long deadline;
1837         int retry_later = 0;
1838         int rc = 0;
1839
1840         deadline = cfs_time_add(cfs_time_current(),
1841                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1842
1843         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1844
1845         LASSERT(route->ksnr_scheduled);
1846         LASSERT(!route->ksnr_connecting);
1847
1848         route->ksnr_connecting = 1;
1849
1850         for (;;) {
1851                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1852
1853                 /* stop connecting if peer/route got closed under me, or
1854                  * route got connected while queued */
1855                 if (peer->ksnp_closing || route->ksnr_deleted ||
1856                     wanted == 0) {
1857                         retry_later = 0;
1858                         break;
1859                 }
1860
1861                 /* reschedule if peer is connecting to me */
1862                 if (peer->ksnp_accepting > 0) {
1863                         CDEBUG(D_NET,
1864                                "peer %s(%d) already connecting to me, retry later.\n",
1865                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1866                         retry_later = 1;
1867                 }
1868
1869                 if (retry_later) /* needs reschedule */
1870                         break;
1871
1872                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1873                         type = SOCKLND_CONN_ANY;
1874                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1875                         type = SOCKLND_CONN_CONTROL;
1876                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1877                         type = SOCKLND_CONN_BULK_IN;
1878                 } else {
1879                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1880                         type = SOCKLND_CONN_BULK_OUT;
1881                 }
1882
1883                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1884
1885                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1886                         rc = -ETIMEDOUT;
1887                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1888                                                    route->ksnr_ipaddr,
1889                                                    route->ksnr_port);
1890                         goto failed;
1891                 }
1892
1893                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1894                                   route->ksnr_myipaddr,
1895                                   route->ksnr_ipaddr, route->ksnr_port);
1896                 if (rc != 0)
1897                         goto failed;
1898
1899                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1900                 if (rc < 0) {
1901                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1902                                                    route->ksnr_ipaddr,
1903                                                    route->ksnr_port);
1904                         goto failed;
1905                 }
1906
1907                 /* A +ve RC means I have to retry because I lost the connection
1908                  * race or I have to renegotiate protocol version */
1909                 retry_later = (rc != 0);
1910                 if (retry_later)
1911                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1912                                libcfs_nid2str(peer->ksnp_id.nid));
1913
1914                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1915         }
1916
1917         route->ksnr_scheduled = 0;
1918         route->ksnr_connecting = 0;
1919
1920         if (retry_later) {
1921                 /* re-queue for attention; this frees me up to handle
1922                  * the peer's incoming connection request */
1923
1924                 if (rc == EALREADY ||
1925                     (rc == 0 && peer->ksnp_accepting > 0)) {
1926                         /* We want to introduce a delay before next
1927                          * attempt to connect if we lost conn race,
1928                          * but the race is resolved quickly usually,
1929                          * so min_reconnectms should be good heuristic */
1930                         route->ksnr_retry_interval =
1931                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1932                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1933                                                            route->ksnr_retry_interval);
1934                 }
1935
1936                 ksocknal_launch_connection_locked(route);
1937         }
1938
1939         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1940         return retry_later;
1941
1942  failed:
1943         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1944
1945         route->ksnr_scheduled = 0;
1946         route->ksnr_connecting = 0;
1947
1948         /* This is a retry rather than a new connection */
1949         route->ksnr_retry_interval *= 2;
1950         route->ksnr_retry_interval =
1951                 max(route->ksnr_retry_interval,
1952                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1953         route->ksnr_retry_interval =
1954                 min(route->ksnr_retry_interval,
1955                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1956
1957         LASSERT (route->ksnr_retry_interval != 0);
1958         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1959                                            route->ksnr_retry_interval);
1960
1961         if (!list_empty(&peer->ksnp_tx_queue) &&
1962             peer->ksnp_accepting == 0 &&
1963             ksocknal_find_connecting_route_locked(peer) == NULL) {
1964                 ksock_conn_t *conn;
1965
1966                 /* ksnp_tx_queue is queued on a conn on successful
1967                  * connection for V1.x and V2.x */
1968                 if (!list_empty (&peer->ksnp_conns)) {
1969                         conn = list_entry(peer->ksnp_conns.next,
1970                                               ksock_conn_t, ksnc_list);
1971                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1972                 }
1973
1974                 /* take all the blocked packets while I've got the lock and
1975                  * complete below... */
1976                 list_splice_init(&peer->ksnp_tx_queue, &zombies);
1977         }
1978
1979 #if 0      /* irrelevant with only eager routes */
1980         if (!route->ksnr_deleted) {
1981                 /* make this route least-favourite for re-selection */
1982                 list_del(&route->ksnr_list);
1983                 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1984         }
1985 #endif
1986         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1987
1988         ksocknal_peer_failed(peer);
1989         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
1990         return 0;
1991 }
1992
1993 /*
1994  * check whether we need to create more connds.
1995  * It will try to create new thread if it's necessary, @timeout can
1996  * be updated if failed to create, so caller wouldn't keep try while
1997  * running out of resource.
1998  */
1999 static int
2000 ksocknal_connd_check_start(time64_t sec, long *timeout)
2001 {
2002         char name[16];
2003         int rc;
2004         int total = ksocknal_data.ksnd_connd_starting +
2005                     ksocknal_data.ksnd_connd_running;
2006
2007         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2008                 /* still in initializing */
2009                 return 0;
2010         }
2011
2012         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2013             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2014                 /* can't create more connd, or still have enough
2015                  * threads to handle more connecting */
2016                 return 0;
2017         }
2018
2019         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2020                 /* no pending connecting request */
2021                 return 0;
2022         }
2023
2024         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2025                 /* may run out of resource, retry later */
2026                 *timeout = cfs_time_seconds(1);
2027                 return 0;
2028         }
2029
2030         if (ksocknal_data.ksnd_connd_starting > 0) {
2031                 /* serialize starting to avoid flood */
2032                 return 0;
2033         }
2034
2035         ksocknal_data.ksnd_connd_starting_stamp = sec;
2036         ksocknal_data.ksnd_connd_starting++;
2037         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2038
2039         /* NB: total is the next id */
2040         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2041         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2042
2043         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2044         if (rc == 0)
2045                 return 1;
2046
2047         /* we tried ... */
2048         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2049         ksocknal_data.ksnd_connd_starting--;
2050         ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2051
2052         return 1;
2053 }
2054
2055 /*
2056  * check whether current thread can exit, it will return 1 if there are too
2057  * many threads and no creating in past 120 seconds.
2058  * Also, this function may update @timeout to make caller come back
2059  * again to recheck these conditions.
2060  */
2061 static int
2062 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2063 {
2064         int val;
2065
2066         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2067                 /* still in initializing */
2068                 return 0;
2069         }
2070
2071         if (ksocknal_data.ksnd_connd_starting > 0) {
2072                 /* in progress of starting new thread */
2073                 return 0;
2074         }
2075
2076         if (ksocknal_data.ksnd_connd_running <=
2077             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2078                 return 0;
2079         }
2080
2081         /* created thread in past 120 seconds? */
2082         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2083                     SOCKNAL_CONND_TIMEOUT - sec);
2084
2085         *timeout = (val > 0) ? cfs_time_seconds(val) :
2086                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2087         if (val > 0)
2088                 return 0;
2089
2090         /* no creating in past 120 seconds */
2091
2092         return ksocknal_data.ksnd_connd_running >
2093                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2094 }
2095
2096 /* Go through connd_routes queue looking for a route that we can process
2097  * right now, @timeout_p can be updated if we need to come back later */
2098 static ksock_route_t *
2099 ksocknal_connd_get_route_locked(signed long *timeout_p)
2100 {
2101         ksock_route_t *route;
2102         unsigned long now;
2103
2104         now = cfs_time_current();
2105
2106         /* connd_routes can contain both pending and ordinary routes */
2107         list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2108                                  ksnr_connd_list) {
2109
2110                 if (route->ksnr_retry_interval == 0 ||
2111                     cfs_time_aftereq(now, route->ksnr_timeout))
2112                         return route;
2113
2114                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2115                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2116                         *timeout_p = (int)(route->ksnr_timeout - now);
2117         }
2118
2119         return NULL;
2120 }
2121
2122 int
2123 ksocknal_connd (void *arg)
2124 {
2125         spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2126         ksock_connreq_t *cr;
2127         wait_queue_t wait;
2128         int nloops = 0;
2129         int cons_retry = 0;
2130
2131         cfs_block_allsigs();
2132
2133         init_waitqueue_entry(&wait, current);
2134
2135         spin_lock_bh(connd_lock);
2136
2137         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2138         ksocknal_data.ksnd_connd_starting--;
2139         ksocknal_data.ksnd_connd_running++;
2140
2141         while (!ksocknal_data.ksnd_shuttingdown) {
2142                 ksock_route_t *route = NULL;
2143                 time64_t sec = ktime_get_real_seconds();
2144                 long timeout = MAX_SCHEDULE_TIMEOUT;
2145                 int dropped_lock = 0;
2146
2147                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2148                         /* wakeup another one to check stop */
2149                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2150                         break;
2151                 }
2152
2153                 if (ksocknal_connd_check_start(sec, &timeout)) {
2154                         /* created new thread */
2155                         dropped_lock = 1;
2156                 }
2157
2158                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2159                         /* Connection accepted by the listener */
2160                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2161                                             next, ksock_connreq_t, ksncr_list);
2162
2163                         list_del(&cr->ksncr_list);
2164                         spin_unlock_bh(connd_lock);
2165                         dropped_lock = 1;
2166
2167                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2168                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2169                         lnet_ni_decref(cr->ksncr_ni);
2170                         LIBCFS_FREE(cr, sizeof(*cr));
2171
2172                         spin_lock_bh(connd_lock);
2173                 }
2174
2175                 /* Only handle an outgoing connection request if there
2176                  * is a thread left to handle incoming connections and
2177                  * create new connd */
2178                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2179                     ksocknal_data.ksnd_connd_running) {
2180                         route = ksocknal_connd_get_route_locked(&timeout);
2181                 }
2182                 if (route != NULL) {
2183                         list_del (&route->ksnr_connd_list);
2184                         ksocknal_data.ksnd_connd_connecting++;
2185                         spin_unlock_bh(connd_lock);
2186                         dropped_lock = 1;
2187
2188                         if (ksocknal_connect(route)) {
2189                                 /* consecutive retry */
2190                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2191                                         CWARN("massive consecutive re-connecting to %pI4h\n",
2192                                               &route->ksnr_ipaddr);
2193                                         cons_retry = 0;
2194                                 }
2195                         } else {
2196                                 cons_retry = 0;
2197                         }
2198
2199                         ksocknal_route_decref(route);
2200
2201                         spin_lock_bh(connd_lock);
2202                         ksocknal_data.ksnd_connd_connecting--;
2203                 }
2204
2205                 if (dropped_lock) {
2206                         if (++nloops < SOCKNAL_RESCHED)
2207                                 continue;
2208                         spin_unlock_bh(connd_lock);
2209                         nloops = 0;
2210                         cond_resched();
2211                         spin_lock_bh(connd_lock);
2212                         continue;
2213                 }
2214
2215                 /* Nothing to do for 'timeout'  */
2216                 set_current_state(TASK_INTERRUPTIBLE);
2217                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2218                 spin_unlock_bh(connd_lock);
2219
2220                 nloops = 0;
2221                 schedule_timeout(timeout);
2222
2223                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2224                 spin_lock_bh(connd_lock);
2225         }
2226         ksocknal_data.ksnd_connd_running--;
2227         spin_unlock_bh(connd_lock);
2228
2229         ksocknal_thread_fini();
2230         return 0;
2231 }
2232
2233 static ksock_conn_t *
2234 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2235 {
2236         /* We're called with a shared lock on ksnd_global_lock */
2237         ksock_conn_t *conn;
2238         struct list_head *ctmp;
2239
2240         list_for_each (ctmp, &peer->ksnp_conns) {
2241                 int error;
2242
2243                 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2244
2245                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2246                 LASSERT(!conn->ksnc_closing);
2247
2248                 /* SOCK_ERROR will reset error code of socket in
2249                  * some platform (like Darwin8.x) */
2250                 error = conn->ksnc_sock->sk->sk_err;
2251                 if (error != 0) {
2252                         ksocknal_conn_addref(conn);
2253
2254                         switch (error) {
2255                         case ECONNRESET:
2256                                 CNETERR("A connection with %s (%pI4h:%d) was reset; it may have rebooted.\n",
2257                                         libcfs_id2str(peer->ksnp_id),
2258                                         &conn->ksnc_ipaddr,
2259                                         conn->ksnc_port);
2260                                 break;
2261                         case ETIMEDOUT:
2262                                 CNETERR("A connection with %s (%pI4h:%d) timed out; the network or node may be down.\n",
2263                                         libcfs_id2str(peer->ksnp_id),
2264                                         &conn->ksnc_ipaddr,
2265                                         conn->ksnc_port);
2266                                 break;
2267                         default:
2268                                 CNETERR("An unexpected network error %d occurred with %s (%pI4h:%d\n",
2269                                         error,
2270                                         libcfs_id2str(peer->ksnp_id),
2271                                         &conn->ksnc_ipaddr,
2272                                         conn->ksnc_port);
2273                                 break;
2274                         }
2275
2276                         return conn;
2277                 }
2278
2279                 if (conn->ksnc_rx_started &&
2280                     cfs_time_aftereq(cfs_time_current(),
2281                                      conn->ksnc_rx_deadline)) {
2282                         /* Timed out incomplete incoming message */
2283                         ksocknal_conn_addref(conn);
2284                         CNETERR("Timeout receiving from %s (%pI4h:%d), state %d wanted %d left %d\n",
2285                                 libcfs_id2str(peer->ksnp_id),
2286                                 &conn->ksnc_ipaddr,
2287                                 conn->ksnc_port,
2288                                 conn->ksnc_rx_state,
2289                                 conn->ksnc_rx_nob_wanted,
2290                                 conn->ksnc_rx_nob_left);
2291                         return conn;
2292                 }
2293
2294                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2295                      conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2296                     cfs_time_aftereq(cfs_time_current(),
2297                                      conn->ksnc_tx_deadline)) {
2298                         /* Timed out messages queued for sending or
2299                          * buffered in the socket's send buffer */
2300                         ksocknal_conn_addref(conn);
2301                         CNETERR("Timeout sending data to %s (%pI4h:%d) the network or that node may be down.\n",
2302                                 libcfs_id2str(peer->ksnp_id),
2303                                 &conn->ksnc_ipaddr,
2304                                 conn->ksnc_port);
2305                         return conn;
2306                 }
2307         }
2308
2309         return NULL;
2310 }
2311
2312 static inline void
2313 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2314 {
2315         ksock_tx_t *tx;
2316         LIST_HEAD(stale_txs);
2317
2318         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2319
2320         while (!list_empty (&peer->ksnp_tx_queue)) {
2321                 tx = list_entry (peer->ksnp_tx_queue.next,
2322                                      ksock_tx_t, tx_list);
2323
2324                 if (!cfs_time_aftereq(cfs_time_current(),
2325                                       tx->tx_deadline))
2326                         break;
2327
2328                 list_del (&tx->tx_list);
2329                 list_add_tail (&tx->tx_list, &stale_txs);
2330         }
2331
2332         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2333
2334         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2335 }
2336
2337 static int
2338 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2339 {
2340         ksock_sched_t *sched;
2341         ksock_conn_t *conn;
2342         ksock_tx_t *tx;
2343
2344         if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2345                 return 0;
2346
2347         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2348                 return 0;
2349
2350         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2351             time_before(cfs_time_current(),
2352                         cfs_time_add(peer->ksnp_last_alive,
2353                                      cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2354                 return 0;
2355
2356         if (time_before(cfs_time_current(), peer->ksnp_send_keepalive))
2357                 return 0;
2358
2359         /* retry 10 secs later, so we wouldn't put pressure
2360          * on this peer if we failed to send keepalive this time */
2361         peer->ksnp_send_keepalive = cfs_time_shift(10);
2362
2363         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2364         if (conn != NULL) {
2365                 sched = conn->ksnc_scheduler;
2366
2367                 spin_lock_bh(&sched->kss_lock);
2368                 if (!list_empty(&conn->ksnc_tx_queue)) {
2369                         spin_unlock_bh(&sched->kss_lock);
2370                         /* there is an queued ACK, don't need keepalive */
2371                         return 0;
2372                 }
2373
2374                 spin_unlock_bh(&sched->kss_lock);
2375         }
2376
2377         read_unlock(&ksocknal_data.ksnd_global_lock);
2378
2379         /* cookie = 1 is reserved for keepalive PING */
2380         tx = ksocknal_alloc_tx_noop(1, 1);
2381         if (tx == NULL) {
2382                 read_lock(&ksocknal_data.ksnd_global_lock);
2383                 return -ENOMEM;
2384         }
2385
2386         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2387                 read_lock(&ksocknal_data.ksnd_global_lock);
2388                 return 1;
2389         }
2390
2391         ksocknal_free_tx(tx);
2392         read_lock(&ksocknal_data.ksnd_global_lock);
2393
2394         return -EIO;
2395 }
2396
2397 static void
2398 ksocknal_check_peer_timeouts (int idx)
2399 {
2400         struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2401         ksock_peer_t *peer;
2402         ksock_conn_t *conn;
2403         ksock_tx_t *tx;
2404
2405  again:
2406         /* NB. We expect to have a look at all the peers and not find any
2407          * connections to time out, so we just use a shared lock while we
2408          * take a look... */
2409         read_lock(&ksocknal_data.ksnd_global_lock);
2410
2411         list_for_each_entry(peer, peers, ksnp_list) {
2412                 unsigned long deadline = 0;
2413                 int resid = 0;
2414                 int n = 0;
2415
2416                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2417                         read_unlock(&ksocknal_data.ksnd_global_lock);
2418                         goto again;
2419                 }
2420
2421                 conn = ksocknal_find_timed_out_conn (peer);
2422
2423                 if (conn != NULL) {
2424                         read_unlock(&ksocknal_data.ksnd_global_lock);
2425
2426                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2427
2428                         /* NB we won't find this one again, but we can't
2429                          * just proceed with the next peer, since we dropped
2430                          * ksnd_global_lock and it might be dead already! */
2431                         ksocknal_conn_decref(conn);
2432                         goto again;
2433                 }
2434
2435                 /* we can't process stale txs right here because we're
2436                  * holding only shared lock */
2437                 if (!list_empty (&peer->ksnp_tx_queue)) {
2438                         ksock_tx_t *tx =
2439                                 list_entry (peer->ksnp_tx_queue.next,
2440                                                 ksock_tx_t, tx_list);
2441
2442                         if (cfs_time_aftereq(cfs_time_current(),
2443                                              tx->tx_deadline)) {
2444
2445                                 ksocknal_peer_addref(peer);
2446                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2447
2448                                 ksocknal_flush_stale_txs(peer);
2449
2450                                 ksocknal_peer_decref(peer);
2451                                 goto again;
2452                         }
2453                 }
2454
2455                 if (list_empty(&peer->ksnp_zc_req_list))
2456                         continue;
2457
2458                 spin_lock(&peer->ksnp_lock);
2459                 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2460                         if (!cfs_time_aftereq(cfs_time_current(),
2461                                               tx->tx_deadline))
2462                                 break;
2463                         /* ignore the TX if connection is being closed */
2464                         if (tx->tx_conn->ksnc_closing)
2465                                 continue;
2466                         n++;
2467                 }
2468
2469                 if (n == 0) {
2470                         spin_unlock(&peer->ksnp_lock);
2471                         continue;
2472                 }
2473
2474                 tx = list_entry(peer->ksnp_zc_req_list.next,
2475                                     ksock_tx_t, tx_zc_list);
2476                 deadline = tx->tx_deadline;
2477                 resid = tx->tx_resid;
2478                 conn = tx->tx_conn;
2479                 ksocknal_conn_addref(conn);
2480
2481                 spin_unlock(&peer->ksnp_lock);
2482                 read_unlock(&ksocknal_data.ksnd_global_lock);
2483
2484                 CERROR("Total %d stale ZC_REQs for peer %s detected; the oldest(%p) timed out %ld secs ago, resid: %d, wmem: %d\n",
2485                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2486                        cfs_duration_sec(cfs_time_current() - deadline),
2487                        resid, conn->ksnc_sock->sk->sk_wmem_queued);
2488
2489                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2490                 ksocknal_conn_decref(conn);
2491                 goto again;
2492         }
2493
2494         read_unlock(&ksocknal_data.ksnd_global_lock);
2495 }
2496
2497 int
2498 ksocknal_reaper (void *arg)
2499 {
2500         wait_queue_t wait;
2501         ksock_conn_t *conn;
2502         ksock_sched_t *sched;
2503         struct list_head enomem_conns;
2504         int nenomem_conns;
2505         long timeout;
2506         int i;
2507         int peer_index = 0;
2508         unsigned long deadline = cfs_time_current();
2509
2510         cfs_block_allsigs();
2511
2512         INIT_LIST_HEAD(&enomem_conns);
2513         init_waitqueue_entry(&wait, current);
2514
2515         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2516
2517         while (!ksocknal_data.ksnd_shuttingdown) {
2518
2519                 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2520                         conn = list_entry (ksocknal_data. \
2521                                                ksnd_deathrow_conns.next,
2522                                                ksock_conn_t, ksnc_list);
2523                         list_del (&conn->ksnc_list);
2524
2525                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2526
2527                         ksocknal_terminate_conn(conn);
2528                         ksocknal_conn_decref(conn);
2529
2530                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2531                         continue;
2532                 }
2533
2534                 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2535                         conn = list_entry (ksocknal_data.ksnd_zombie_conns.\
2536                                                next, ksock_conn_t, ksnc_list);
2537                         list_del (&conn->ksnc_list);
2538
2539                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2540
2541                         ksocknal_destroy_conn(conn);
2542
2543                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2544                         continue;
2545                 }
2546
2547                 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2548                         list_add(&enomem_conns,
2549                                      &ksocknal_data.ksnd_enomem_conns);
2550                         list_del_init(&ksocknal_data.ksnd_enomem_conns);
2551                 }
2552
2553                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2554
2555                 /* reschedule all the connections that stalled with ENOMEM... */
2556                 nenomem_conns = 0;
2557                 while (!list_empty (&enomem_conns)) {
2558                         conn = list_entry (enomem_conns.next,
2559                                                ksock_conn_t, ksnc_tx_list);
2560                         list_del (&conn->ksnc_tx_list);
2561
2562                         sched = conn->ksnc_scheduler;
2563
2564                         spin_lock_bh(&sched->kss_lock);
2565
2566                         LASSERT(conn->ksnc_tx_scheduled);
2567                         conn->ksnc_tx_ready = 1;
2568                         list_add_tail(&conn->ksnc_tx_list,
2569                                           &sched->kss_tx_conns);
2570                         wake_up(&sched->kss_waitq);
2571
2572                         spin_unlock_bh(&sched->kss_lock);
2573                         nenomem_conns++;
2574                 }
2575
2576                 /* careful with the jiffy wrap... */
2577                 while ((timeout = cfs_time_sub(deadline,
2578                                                cfs_time_current())) <= 0) {
2579                         const int n = 4;
2580                         const int p = 1;
2581                         int chunk = ksocknal_data.ksnd_peer_hash_size;
2582
2583                         /* Time to check for timeouts on a few more peers: I do
2584                          * checks every 'p' seconds on a proportion of the peer
2585                          * table and I need to check every connection 'n' times
2586                          * within a timeout interval, to ensure I detect a
2587                          * timeout on any connection within (n+1)/n times the
2588                          * timeout interval. */
2589
2590                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2591                                 chunk = (chunk * n * p) /
2592                                         *ksocknal_tunables.ksnd_timeout;
2593                         if (chunk == 0)
2594                                 chunk = 1;
2595
2596                         for (i = 0; i < chunk; i++) {
2597                                 ksocknal_check_peer_timeouts (peer_index);
2598                                 peer_index = (peer_index + 1) %
2599                                              ksocknal_data.ksnd_peer_hash_size;
2600                         }
2601
2602                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2603                 }
2604
2605                 if (nenomem_conns != 0) {
2606                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2607                          * This also prevents me getting woken immediately
2608                          * if any go back on my enomem list. */
2609                         timeout = SOCKNAL_ENOMEM_RETRY;
2610                 }
2611                 ksocknal_data.ksnd_reaper_waketime =
2612                         cfs_time_add(cfs_time_current(), timeout);
2613
2614                 set_current_state (TASK_INTERRUPTIBLE);
2615                 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2616
2617                 if (!ksocknal_data.ksnd_shuttingdown &&
2618                     list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2619                     list_empty (&ksocknal_data.ksnd_zombie_conns))
2620                         schedule_timeout(timeout);
2621
2622                 set_current_state (TASK_RUNNING);
2623                 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2624
2625                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2626         }
2627
2628         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2629
2630         ksocknal_thread_fini();
2631         return 0;
2632 }