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 (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \
  42                                             next, ksock_tx_t, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         cfs_list_del(&tx->tx_list);
  45                 }
  46
  47                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         cfs_atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_desc_size  = size;
  61
  62         cfs_atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  63
  64         return tx;
  65 }
  66
  67 ksock_tx_t *
  68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  69 {
  70         ksock_tx_t *tx;
  71
  72         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  73         if (tx == NULL) {
  74                 CERROR("Can't allocate noop tx desc\n");
  75                 return NULL;
  76         }
  77
  78         tx->tx_conn     = NULL;
  79         tx->tx_lnetmsg  = NULL;
  80         tx->tx_kiov     = NULL;
  81         tx->tx_nkiov    = 0;
  82         tx->tx_iov      = tx->tx_frags.virt.iov;
  83         tx->tx_niov     = 1;
  84         tx->tx_nonblk   = nonblk;
  85
  86         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
  87         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  88
  89         return tx;
  90 }
  91
  92
  93 void
  94 ksocknal_free_tx (ksock_tx_t *tx)
  95 {
  96         cfs_atomic_dec(&ksocknal_data.ksnd_nactive_txs);
  97
  98         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
  99                 /* it's a noop tx */
 100                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 101
 102                 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 103
 104                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 105         } else {
 106                 LIBCFS_FREE(tx, tx->tx_desc_size);
 107         }
 108 }
 109
 110 int
 111 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
 112 {
 113         struct iovec  *iov = tx->tx_iov;
 114         int    nob;
 115         int    rc;
 116
 117         LASSERT (tx->tx_niov > 0);
 118
 119         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 120         rc = ksocknal_lib_send_iov(conn, tx);
 121
 122         if (rc <= 0)                            /* sent nothing? */
 123                 return (rc);
 124
 125         nob = rc;
 126         LASSERT (nob <= tx->tx_resid);
 127         tx->tx_resid -= nob;
 128
 129         /* "consume" iov */
 130         do {
 131                 LASSERT (tx->tx_niov > 0);
 132
 133                 if (nob < (int) iov->iov_len) {
 134                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 135                         iov->iov_len -= nob;
 136                         return (rc);
 137                 }
 138
 139                 nob -= iov->iov_len;
 140                 tx->tx_iov = ++iov;
 141                 tx->tx_niov--;
 142         } while (nob != 0);
 143
 144         return (rc);
 145 }
 146
 147 int
 148 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
 149 {
 150         lnet_kiov_t    *kiov = tx->tx_kiov;
 151         int     nob;
 152         int     rc;
 153
 154         LASSERT (tx->tx_niov == 0);
 155         LASSERT (tx->tx_nkiov > 0);
 156
 157         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 158         rc = ksocknal_lib_send_kiov(conn, tx);
 159
 160         if (rc <= 0)                            /* sent nothing? */
 161                 return (rc);
 162
 163         nob = rc;
 164         LASSERT (nob <= tx->tx_resid);
 165         tx->tx_resid -= nob;
 166
 167         /* "consume" kiov */
 168         do {
 169                 LASSERT(tx->tx_nkiov > 0);
 170
 171                 if (nob < (int)kiov->kiov_len) {
 172                         kiov->kiov_offset += nob;
 173                         kiov->kiov_len -= nob;
 174                         return rc;
 175                 }
 176
 177                 nob -= (int)kiov->kiov_len;
 178                 tx->tx_kiov = ++kiov;
 179                 tx->tx_nkiov--;
 180         } while (nob != 0);
 181
 182         return (rc);
 183 }
 184
 185 int
 186 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 187 {
 188         int      rc;
 189         int      bufnob;
 190
 191         if (ksocknal_data.ksnd_stall_tx != 0) {
 192                 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 193         }
 194
 195         LASSERT (tx->tx_resid != 0);
 196
 197         rc = ksocknal_connsock_addref(conn);
 198         if (rc != 0) {
 199                 LASSERT (conn->ksnc_closing);
 200                 return (-ESHUTDOWN);
 201         }
 202
 203         do {
 204                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 205                         /* testing... */
 206                         ksocknal_data.ksnd_enomem_tx--;
 207                         rc = -EAGAIN;
 208                 } else if (tx->tx_niov != 0) {
 209                         rc = ksocknal_send_iov (conn, tx);
 210                 } else {
 211                         rc = ksocknal_send_kiov (conn, tx);
 212                 }
 213
 214                 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 215                 if (rc > 0)                     /* sent something? */
 216                         conn->ksnc_tx_bufnob += rc; /* account it */
 217
 218                 if (bufnob < conn->ksnc_tx_bufnob) {
 219                         /* allocated send buffer bytes < computed; infer
 220                          * something got ACKed */
 221                         conn->ksnc_tx_deadline =
 222                                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 223                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 224                         conn->ksnc_tx_bufnob = bufnob;
 225                         cfs_mb();
 226                 }
 227
 228                 if (rc <= 0) { /* Didn't write anything? */
 229
 230                         if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
 231                                 rc = -EAGAIN;
 232
 233                         /* Check if EAGAIN is due to memory pressure */
 234                         if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 235                                 rc = -ENOMEM;
 236
 237                         break;
 238                 }
 239
 240                 /* socket's wmem_queued now includes 'rc' bytes */
 241                 cfs_atomic_sub (rc, &conn->ksnc_tx_nob);
 242                 rc = 0;
 243
 244         } while (tx->tx_resid != 0);
 245
 246         ksocknal_connsock_decref(conn);
 247         return (rc);
 248 }
 249
 250 int
 251 ksocknal_recv_iov (ksock_conn_t *conn)
 252 {
 253         struct iovec *iov = conn->ksnc_rx_iov;
 254         int     nob;
 255         int     rc;
 256
 257         LASSERT (conn->ksnc_rx_niov > 0);
 258
 259         /* Never touch conn->ksnc_rx_iov or change connection
 260          * status inside ksocknal_lib_recv_iov */
 261         rc = ksocknal_lib_recv_iov(conn);
 262
 263         if (rc <= 0)
 264                 return (rc);
 265
 266         /* received something... */
 267         nob = rc;
 268
 269         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 270         conn->ksnc_rx_deadline =
 271                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 272         cfs_mb();                       /* order with setting rx_started */
 273         conn->ksnc_rx_started = 1;
 274
 275         conn->ksnc_rx_nob_wanted -= nob;
 276         conn->ksnc_rx_nob_left -= nob;
 277
 278         do {
 279                 LASSERT (conn->ksnc_rx_niov > 0);
 280
 281                 if (nob < (int)iov->iov_len) {
 282                         iov->iov_len -= nob;
 283                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
 284                         return (-EAGAIN);
 285                 }
 286
 287                 nob -= iov->iov_len;
 288                 conn->ksnc_rx_iov = ++iov;
 289                 conn->ksnc_rx_niov--;
 290         } while (nob != 0);
 291
 292         return (rc);
 293 }
 294
 295 int
 296 ksocknal_recv_kiov (ksock_conn_t *conn)
 297 {
 298         lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
 299         int     nob;
 300         int     rc;
 301         LASSERT (conn->ksnc_rx_nkiov > 0);
 302
 303         /* Never touch conn->ksnc_rx_kiov or change connection
 304          * status inside ksocknal_lib_recv_iov */
 305         rc = ksocknal_lib_recv_kiov(conn);
 306
 307         if (rc <= 0)
 308                 return (rc);
 309
 310         /* received something... */
 311         nob = rc;
 312
 313         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 314         conn->ksnc_rx_deadline =
 315                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 316         cfs_mb();                       /* order with setting rx_started */
 317         conn->ksnc_rx_started = 1;
 318
 319         conn->ksnc_rx_nob_wanted -= nob;
 320         conn->ksnc_rx_nob_left -= nob;
 321
 322         do {
 323                 LASSERT (conn->ksnc_rx_nkiov > 0);
 324
 325                 if (nob < (int) kiov->kiov_len) {
 326                         kiov->kiov_offset += nob;
 327                         kiov->kiov_len -= nob;
 328                         return -EAGAIN;
 329                 }
 330
 331                 nob -= kiov->kiov_len;
 332                 conn->ksnc_rx_kiov = ++kiov;
 333                 conn->ksnc_rx_nkiov--;
 334         } while (nob != 0);
 335
 336         return 1;
 337 }
 338
 339 int
 340 ksocknal_receive (ksock_conn_t *conn)
 341 {
 342         /* Return 1 on success, 0 on EOF, < 0 on error.
 343          * Caller checks ksnc_rx_nob_wanted to determine
 344          * progress/completion. */
 345         int     rc;
 346         ENTRY;
 347
 348         if (ksocknal_data.ksnd_stall_rx != 0) {
 349                 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
 350         }
 351
 352         rc = ksocknal_connsock_addref(conn);
 353         if (rc != 0) {
 354                 LASSERT (conn->ksnc_closing);
 355                 return (-ESHUTDOWN);
 356         }
 357
 358         for (;;) {
 359                 if (conn->ksnc_rx_niov != 0)
 360                         rc = ksocknal_recv_iov (conn);
 361                 else
 362                         rc = ksocknal_recv_kiov (conn);
 363
 364                 if (rc <= 0) {
 365                         /* error/EOF or partial receive */
 366                         if (rc == -EAGAIN) {
 367                                 rc = 1;
 368                         } else if (rc == 0 && conn->ksnc_rx_started) {
 369                                 /* EOF in the middle of a message */
 370                                 rc = -EPROTO;
 371                         }
 372                         break;
 373                 }
 374
 375                 /* Completed a fragment */
 376
 377                 if (conn->ksnc_rx_nob_wanted == 0) {
 378                         rc = 1;
 379                         break;
 380                 }
 381         }
 382
 383         ksocknal_connsock_decref(conn);
 384         RETURN (rc);
 385 }
 386
 387 void
 388 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
 389 {
 390         lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
 391         int          rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
 392         ENTRY;
 393
 394         LASSERT(ni != NULL || tx->tx_conn != NULL);
 395
 396         if (tx->tx_conn != NULL)
 397                 ksocknal_conn_decref(tx->tx_conn);
 398
 399         if (ni == NULL && tx->tx_conn != NULL)
 400                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
 401
 402         ksocknal_free_tx (tx);
 403         if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
 404                 lnet_finalize (ni, lnetmsg, rc);
 405
 406         EXIT;
 407 }
 408
 409 void
 410 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
 411 {
 412         ksock_tx_t *tx;
 413
 414         while (!cfs_list_empty (txlist)) {
 415                 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
 416
 417                 if (error && tx->tx_lnetmsg != NULL) {
 418                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 419                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
 420                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
 421                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 422                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 423                 } else if (error) {
 424                         CNETERR("Deleting noop packet\n");
 425                 }
 426
 427                 cfs_list_del (&tx->tx_list);
 428
 429                 LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1);
 430                 ksocknal_tx_done (ni, tx);
 431         }
 432 }
 433
 434 static void
 435 ksocknal_check_zc_req(ksock_tx_t *tx)
 436 {
 437         ksock_conn_t   *conn = tx->tx_conn;
 438         ksock_peer_t   *peer = conn->ksnc_peer;
 439
 440         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 441          * to ksnp_zc_req_list if some fragment of this message should be sent
 442          * zero-copy.  Our peer will send an ACK containing this cookie when
 443          * she has received this message to tell us we can signal completion.
 444          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 445          * ksnp_zc_req_list. */
 446         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 447         LASSERT (tx->tx_zc_capable);
 448
 449         tx->tx_zc_checked = 1;
 450
 451         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 452             !conn->ksnc_zc_capable)
 453                 return;
 454
 455         /* assign cookie and queue tx to pending list, it will be released when
 456          * a matching ack is received. See ksocknal_handle_zcack() */
 457
 458         ksocknal_tx_addref(tx);
 459
 460         spin_lock(&peer->ksnp_lock);
 461
 462         /* ZC_REQ is going to be pinned to the peer */
 463         tx->tx_deadline =
 464                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 465
 466         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 467
 468         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
 469
 470         if (peer->ksnp_zc_next_cookie == 0)
 471                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 472
 473         cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
 474
 475         spin_unlock(&peer->ksnp_lock);
 476 }
 477
 478 static void
 479 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
 480 {
 481         ksock_peer_t   *peer = tx->tx_conn->ksnc_peer;
 482
 483         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 484         LASSERT(tx->tx_zc_capable);
 485
 486         tx->tx_zc_checked = 0;
 487
 488         spin_lock(&peer->ksnp_lock);
 489
 490         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 491                 /* Not waiting for an ACK */
 492                 spin_unlock(&peer->ksnp_lock);
 493                 return;
 494         }
 495
 496         tx->tx_msg.ksm_zc_cookies[0] = 0;
 497         cfs_list_del(&tx->tx_zc_list);
 498
 499         spin_unlock(&peer->ksnp_lock);
 500
 501         ksocknal_tx_decref(tx);
 502 }
 503
 504 int
 505 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
 506 {
 507         int            rc;
 508
 509         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 510                 ksocknal_check_zc_req(tx);
 511
 512         rc = ksocknal_transmit (conn, tx);
 513
 514         CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 515
 516         if (tx->tx_resid == 0) {
 517                 /* Sent everything OK */
 518                 LASSERT (rc == 0);
 519
 520                 return (0);
 521         }
 522
 523         if (rc == -EAGAIN)
 524                 return (rc);
 525
 526         if (rc == -ENOMEM) {
 527                 static int counter;
 528
 529                 counter++;   /* exponential backoff warnings */
 530                 if ((counter & (-counter)) == counter)
 531                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 532                               counter, conn, cfs_atomic_read(&libcfs_kmemory));
 533
 534                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 535                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 536
 537                 /* enomem list takes over scheduler's ref... */
 538                 LASSERT (conn->ksnc_tx_scheduled);
 539                 cfs_list_add_tail(&conn->ksnc_tx_list,
 540                                   &ksocknal_data.ksnd_enomem_conns);
 541                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
 542                                                    SOCKNAL_ENOMEM_RETRY),
 543                                    ksocknal_data.ksnd_reaper_waketime))
 544                         cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
 545
 546                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 547                 return (rc);
 548         }
 549
 550         /* Actual error */
 551         LASSERT (rc < 0);
 552
 553         if (!conn->ksnc_closing) {
 554                 switch (rc) {
 555                 case -ECONNRESET:
 556                         LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
 557                                       "while we were sending data; it may have "
 558                                       "rebooted.\n",
 559                                       HIPQUAD(conn->ksnc_ipaddr));
 560                         break;
 561                 default:
 562                         LCONSOLE_WARN("There was an unexpected network error "
 563                                       "while writing to %u.%u.%u.%u: %d.\n",
 564                                       HIPQUAD(conn->ksnc_ipaddr), rc);
 565                         break;
 566                 }
 567                 CDEBUG(D_NET, "[%p] Error %d on write to %s"
 568                        " ip %d.%d.%d.%d:%d\n", conn, rc,
 569                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
 570                        HIPQUAD(conn->ksnc_ipaddr),
 571                        conn->ksnc_port);
 572         }
 573
 574         if (tx->tx_zc_checked)
 575                 ksocknal_uncheck_zc_req(tx);
 576
 577         /* it's not an error if conn is being closed */
 578         ksocknal_close_conn_and_siblings (conn,
 579                                           (conn->ksnc_closing) ? 0 : rc);
 580
 581         return (rc);
 582 }
 583
 584 void
 585 ksocknal_launch_connection_locked (ksock_route_t *route)
 586 {
 587
 588         /* called holding write lock on ksnd_global_lock */
 589
 590         LASSERT (!route->ksnr_scheduled);
 591         LASSERT (!route->ksnr_connecting);
 592         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 593
 594         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 595         ksocknal_route_addref(route);           /* extra ref for connd */
 596
 597         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 598
 599         cfs_list_add_tail(&route->ksnr_connd_list,
 600                           &ksocknal_data.ksnd_connd_routes);
 601         cfs_waitq_signal(&ksocknal_data.ksnd_connd_waitq);
 602
 603         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 604 }
 605
 606 void
 607 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
 608 {
 609         ksock_route_t *route;
 610
 611         /* called holding write lock on ksnd_global_lock */
 612         for (;;) {
 613                 /* launch any/all connections that need it */
 614                 route = ksocknal_find_connectable_route_locked(peer);
 615                 if (route == NULL)
 616                         return;
 617
 618                 ksocknal_launch_connection_locked(route);
 619         }
 620 }
 621
 622 ksock_conn_t *
 623 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
 624 {
 625         cfs_list_t       *tmp;
 626         ksock_conn_t     *conn;
 627         ksock_conn_t     *typed = NULL;
 628         ksock_conn_t     *fallback = NULL;
 629         int               tnob     = 0;
 630         int               fnob     = 0;
 631
 632         cfs_list_for_each (tmp, &peer->ksnp_conns) {
 633                 ksock_conn_t *c  = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
 634                 int           nob = cfs_atomic_read(&c->ksnc_tx_nob) +
 635                                     libcfs_sock_wmem_queued(c->ksnc_sock);
 636                 int           rc;
 637
 638                 LASSERT (!c->ksnc_closing);
 639                 LASSERT (c->ksnc_proto != NULL &&
 640                          c->ksnc_proto->pro_match_tx != NULL);
 641
 642                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 643
 644                 switch (rc) {
 645                 default:
 646                         LBUG();
 647                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 648                         continue;
 649
 650                 case SOCKNAL_MATCH_YES: /* typed connection */
 651                         if (typed == NULL || tnob > nob ||
 652                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 653                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 654                                 typed = c;
 655                                 tnob  = nob;
 656                         }
 657                         break;
 658
 659                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 660                         if (fallback == NULL || fnob > nob ||
 661                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 662                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
 663                                 fallback = c;
 664                                 fnob     = nob;
 665                         }
 666                         break;
 667                 }
 668         }
 669
 670         /* prefer the typed selection */
 671         conn = (typed != NULL) ? typed : fallback;
 672
 673         if (conn != NULL)
 674                 conn->ksnc_tx_last_post = cfs_time_current();
 675
 676         return conn;
 677 }
 678
 679 void
 680 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
 681 {
 682         conn->ksnc_proto->pro_pack(tx);
 683
 684         cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 685         ksocknal_conn_addref(conn); /* +1 ref for tx */
 686         tx->tx_conn = conn;
 687 }
 688
 689 void
 690 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
 691 {
 692         ksock_sched_t *sched = conn->ksnc_scheduler;
 693         ksock_msg_t   *msg = &tx->tx_msg;
 694         ksock_tx_t    *ztx = NULL;
 695         int            bufnob = 0;
 696
 697         /* called holding global lock (read or irq-write) and caller may
 698          * not have dropped this lock between finding conn and calling me,
 699          * so we don't need the {get,put}connsock dance to deref
 700          * ksnc_sock... */
 701         LASSERT(!conn->ksnc_closing);
 702
 703         CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n",
 704                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
 705                 HIPQUAD(conn->ksnc_ipaddr),
 706                 conn->ksnc_port);
 707
 708         ksocknal_tx_prep(conn, tx);
 709
 710         /* Ensure the frags we've been given EXACTLY match the number of
 711          * bytes we want to send.  Many TCP/IP stacks disregard any total
 712          * size parameters passed to them and just look at the frags.
 713          *
 714          * We always expect at least 1 mapped fragment containing the
 715          * complete ksocknal message header. */
 716         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 717                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 718                  (unsigned int)tx->tx_nob);
 719         LASSERT (tx->tx_niov >= 1);
 720         LASSERT (tx->tx_resid == tx->tx_nob);
 721
 722         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 723                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 724                                                KSOCK_MSG_NOOP,
 725                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 726
 727         /*
 728          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
 729          * but they're used inside spinlocks a lot.
 730          */
 731         bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
 732         spin_lock_bh(&sched->kss_lock);
 733
 734         if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 735                 /* First packet starts the timeout */
 736                 conn->ksnc_tx_deadline =
 737                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 738                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 739                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
 740                 conn->ksnc_tx_bufnob = 0;
 741                 cfs_mb(); /* order with adding to tx_queue */
 742         }
 743
 744         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 745                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 746                  * on a normal packet so I don't need to send it */
 747                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 748                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 749
 750                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 751                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 752
 753         } else {
 754                 /* It's a normal packet - can it piggback a noop zc-ack that
 755                  * has been queued already? */
 756                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 757                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 758
 759                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 760                 /* ztx will be released later */
 761         }
 762
 763         if (ztx != NULL) {
 764                 cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 765                 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 766         }
 767
 768         if (conn->ksnc_tx_ready &&      /* able to send */
 769             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 770                 /* +1 ref for scheduler */
 771                 ksocknal_conn_addref(conn);
 772                 cfs_list_add_tail (&conn->ksnc_tx_list,
 773                                    &sched->kss_tx_conns);
 774                 conn->ksnc_tx_scheduled = 1;
 775                 cfs_waitq_signal (&sched->kss_waitq);
 776         }
 777
 778         spin_unlock_bh(&sched->kss_lock);
 779 }
 780
 781
 782 ksock_route_t *
 783 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
 784 {
 785         cfs_time_t     now = cfs_time_current();
 786         cfs_list_t    *tmp;
 787         ksock_route_t *route;
 788
 789         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 790                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 791
 792                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 793
 794                 if (route->ksnr_scheduled)      /* connections being established */
 795                         continue;
 796
 797                 /* all route types connected ? */
 798                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 799                         continue;
 800
 801                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 802                       cfs_time_aftereq(now, route->ksnr_timeout))) {
 803                         CDEBUG(D_NET,
 804                                "Too soon to retry route %u.%u.%u.%u "
 805                                "(cnted %d, interval %ld, %ld secs later)\n",
 806                                HIPQUAD(route->ksnr_ipaddr),
 807                                route->ksnr_connected,
 808                                route->ksnr_retry_interval,
 809                                cfs_duration_sec(route->ksnr_timeout - now));
 810                         continue;
 811                 }
 812
 813                 return (route);
 814         }
 815
 816         return (NULL);
 817 }
 818
 819 ksock_route_t *
 820 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
 821 {
 822         cfs_list_t        *tmp;
 823         ksock_route_t     *route;
 824
 825         cfs_list_for_each (tmp, &peer->ksnp_routes) {
 826                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
 827
 828                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 829
 830                 if (route->ksnr_scheduled)
 831                         return (route);
 832         }
 833
 834         return (NULL);
 835 }
 836
 837 int
 838 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
 839 {
 840         ksock_peer_t     *peer;
 841         ksock_conn_t     *conn;
 842         rwlock_t     *g_lock;
 843         int               retry;
 844         int               rc;
 845
 846         LASSERT (tx->tx_conn == NULL);
 847
 848         g_lock = &ksocknal_data.ksnd_global_lock;
 849
 850         for (retry = 0;; retry = 1) {
 851                 read_lock(g_lock);
 852                 peer = ksocknal_find_peer_locked(ni, id);
 853                 if (peer != NULL) {
 854                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
 855                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 856                                 if (conn != NULL) {
 857                                         /* I've got no routes that need to be
 858                                          * connecting and I do have an actual
 859                                          * connection... */
 860                                         ksocknal_queue_tx_locked (tx, conn);
 861                                         read_unlock(g_lock);
 862                                         return (0);
 863                                 }
 864                         }
 865                 }
 866
 867                 /* I'll need a write lock... */
 868                 read_unlock(g_lock);
 869
 870                 write_lock_bh(g_lock);
 871
 872                 peer = ksocknal_find_peer_locked(ni, id);
 873                 if (peer != NULL)
 874                         break;
 875
 876                 write_unlock_bh(g_lock);
 877
 878                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 879                         CERROR("Refusing to create a connection to "
 880                                "userspace process %s\n", libcfs_id2str(id));
 881                         return -EHOSTUNREACH;
 882                 }
 883
 884                 if (retry) {
 885                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
 886                         return -EHOSTUNREACH;
 887                 }
 888
 889                 rc = ksocknal_add_peer(ni, id,
 890                                        LNET_NIDADDR(id.nid),
 891                                        lnet_acceptor_port());
 892                 if (rc != 0) {
 893                         CERROR("Can't add peer %s: %d\n",
 894                                libcfs_id2str(id), rc);
 895                         return rc;
 896                 }
 897         }
 898
 899         ksocknal_launch_all_connections_locked(peer);
 900
 901         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
 902         if (conn != NULL) {
 903                 /* Connection exists; queue message on it */
 904                 ksocknal_queue_tx_locked (tx, conn);
 905                 write_unlock_bh(g_lock);
 906                 return (0);
 907         }
 908
 909         if (peer->ksnp_accepting > 0 ||
 910             ksocknal_find_connecting_route_locked (peer) != NULL) {
 911                 /* the message is going to be pinned to the peer */
 912                 tx->tx_deadline =
 913                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
 914
 915                 /* Queue the message until a connection is established */
 916                 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
 917                 write_unlock_bh(g_lock);
 918                 return 0;
 919         }
 920
 921         write_unlock_bh(g_lock);
 922
 923         /* NB Routes may be ignored if connections to them failed recently */
 924         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 925         return (-EHOSTUNREACH);
 926 }
 927
 928 int
 929 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
 930 {
 931         int               mpflag = 0;
 932         int               type = lntmsg->msg_type;
 933         lnet_process_id_t target = lntmsg->msg_target;
 934         unsigned int      payload_niov = lntmsg->msg_niov;
 935         struct iovec     *payload_iov = lntmsg->msg_iov;
 936         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 937         unsigned int      payload_offset = lntmsg->msg_offset;
 938         unsigned int      payload_nob = lntmsg->msg_len;
 939         ksock_tx_t       *tx;
 940         int               desc_size;
 941         int               rc;
 942
 943         /* NB 'private' is different depending on what we're sending.
 944          * Just ignore it... */
 945
 946         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 947                payload_nob, payload_niov, libcfs_id2str(target));
 948
 949         LASSERT (payload_nob == 0 || payload_niov > 0);
 950         LASSERT (payload_niov <= LNET_MAX_IOV);
 951         /* payload is either all vaddrs or all pages */
 952         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
 953         LASSERT (!cfs_in_interrupt ());
 954
 955         if (payload_iov != NULL)
 956                 desc_size = offsetof(ksock_tx_t,
 957                                      tx_frags.virt.iov[1 + payload_niov]);
 958         else
 959                 desc_size = offsetof(ksock_tx_t,
 960                                      tx_frags.paged.kiov[payload_niov]);
 961
 962         if (lntmsg->msg_vmflush)
 963                 mpflag = cfs_memory_pressure_get_and_set();
 964         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
 965         if (tx == NULL) {
 966                 CERROR("Can't allocate tx desc type %d size %d\n",
 967                        type, desc_size);
 968                 if (lntmsg->msg_vmflush)
 969                         cfs_memory_pressure_restore(mpflag);
 970                 return (-ENOMEM);
 971         }
 972
 973         tx->tx_conn = NULL;                     /* set when assigned a conn */
 974         tx->tx_lnetmsg = lntmsg;
 975
 976         if (payload_iov != NULL) {
 977                 tx->tx_kiov = NULL;
 978                 tx->tx_nkiov = 0;
 979                 tx->tx_iov = tx->tx_frags.virt.iov;
 980                 tx->tx_niov = 1 +
 981                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
 982                                                payload_niov, payload_iov,
 983                                                payload_offset, payload_nob);
 984         } else {
 985                 tx->tx_niov = 1;
 986                 tx->tx_iov = &tx->tx_frags.paged.iov;
 987                 tx->tx_kiov = tx->tx_frags.paged.kiov;
 988                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
 989                                                  payload_niov, payload_kiov,
 990                                                  payload_offset, payload_nob);
 991
 992                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
 993                         tx->tx_zc_capable = 1;
 994         }
 995
 996         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
 997
 998         /* The first fragment will be set later in pro_pack */
 999         rc = ksocknal_launch_packet(ni, tx, target);
1000         if (lntmsg->msg_vmflush)
1001                 cfs_memory_pressure_restore(mpflag);
1002         if (rc == 0)
1003                 return (0);
1004
1005         ksocknal_free_tx(tx);
1006         return (-EIO);
1007 }
1008
1009 int
1010 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1011 {
1012         long          pid = cfs_create_thread (fn, arg, 0);
1013
1014         if (pid < 0)
1015                 return ((int)pid);
1016
1017         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1018         ksocknal_data.ksnd_nthreads++;
1019         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1020         return (0);
1021 }
1022
1023 void
1024 ksocknal_thread_fini (void)
1025 {
1026         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1027         ksocknal_data.ksnd_nthreads--;
1028         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1029 }
1030
1031 int
1032 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1033 {
1034         static char ksocknal_slop_buffer[4096];
1035
1036         int            nob;
1037         unsigned int   niov;
1038         int            skipped;
1039
1040         LASSERT(conn->ksnc_proto != NULL);
1041
1042         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1043                 /* Remind the socket to ack eagerly... */
1044                 ksocknal_lib_eager_ack(conn);
1045         }
1046
1047         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1048                 conn->ksnc_rx_started = 0;
1049                 cfs_mb();                       /* racing with timeout thread */
1050
1051                 switch (conn->ksnc_proto->pro_version) {
1052                 case  KSOCK_PROTO_V2:
1053                 case  KSOCK_PROTO_V3:
1054                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1055                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1056                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1057
1058                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1059                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1060                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1061                         break;
1062
1063                 case KSOCK_PROTO_V1:
1064                         /* Receiving bare lnet_hdr_t */
1065                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1066                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1067                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1068
1069                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1070                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1071                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1072                         break;
1073
1074                 default:
1075                         LBUG ();
1076                 }
1077                 conn->ksnc_rx_niov = 1;
1078
1079                 conn->ksnc_rx_kiov = NULL;
1080                 conn->ksnc_rx_nkiov = 0;
1081                 conn->ksnc_rx_csum = ~0;
1082                 return (1);
1083         }
1084
1085         /* Set up to skip as much as possible now.  If there's more left
1086          * (ran out of iov entries) we'll get called again */
1087
1088         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1089         conn->ksnc_rx_nob_left = nob_to_skip;
1090         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1091         skipped = 0;
1092         niov = 0;
1093
1094         do {
1095                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1096
1097                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1098                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1099                 niov++;
1100                 skipped += nob;
1101                 nob_to_skip -=nob;
1102
1103         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1104                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1105
1106         conn->ksnc_rx_niov = niov;
1107         conn->ksnc_rx_kiov = NULL;
1108         conn->ksnc_rx_nkiov = 0;
1109         conn->ksnc_rx_nob_wanted = skipped;
1110         return (0);
1111 }
1112
1113 int
1114 ksocknal_process_receive (ksock_conn_t *conn)
1115 {
1116         lnet_hdr_t        *lhdr;
1117         lnet_process_id_t *id;
1118         int                rc;
1119
1120         LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0);
1121
1122         /* NB: sched lock NOT held */
1123         /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1124         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1125                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1126                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1127                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1128  again:
1129         if (conn->ksnc_rx_nob_wanted != 0) {
1130                 rc = ksocknal_receive(conn);
1131
1132                 if (rc <= 0) {
1133                         LASSERT (rc != -EAGAIN);
1134
1135                         if (rc == 0)
1136                                 CDEBUG (D_NET, "[%p] EOF from %s"
1137                                         " ip %d.%d.%d.%d:%d\n", conn,
1138                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1139                                         HIPQUAD(conn->ksnc_ipaddr),
1140                                         conn->ksnc_port);
1141                         else if (!conn->ksnc_closing)
1142                                 CERROR ("[%p] Error %d on read from %s"
1143                                         " ip %d.%d.%d.%d:%d\n",
1144                                         conn, rc,
1145                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1146                                         HIPQUAD(conn->ksnc_ipaddr),
1147                                         conn->ksnc_port);
1148
1149                         /* it's not an error if conn is being closed */
1150                         ksocknal_close_conn_and_siblings (conn,
1151                                                           (conn->ksnc_closing) ? 0 : rc);
1152                         return (rc == 0 ? -ESHUTDOWN : rc);
1153                 }
1154
1155                 if (conn->ksnc_rx_nob_wanted != 0) {
1156                         /* short read */
1157                         return (-EAGAIN);
1158                 }
1159         }
1160         switch (conn->ksnc_rx_state) {
1161         case SOCKNAL_RX_KSM_HEADER:
1162                 if (conn->ksnc_flip) {
1163                         __swab32s(&conn->ksnc_msg.ksm_type);
1164                         __swab32s(&conn->ksnc_msg.ksm_csum);
1165                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1166                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1167                 }
1168
1169                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1170                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1171                         CERROR("%s: Unknown message type: %x\n",
1172                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1173                                conn->ksnc_msg.ksm_type);
1174                         ksocknal_new_packet(conn, 0);
1175                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1176                         return (-EPROTO);
1177                 }
1178
1179                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1180                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1181                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1182                         /* NOOP Checksum error */
1183                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1184                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1185                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1186                         ksocknal_new_packet(conn, 0);
1187                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1188                         return (-EIO);
1189                 }
1190
1191                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1192                         __u64 cookie = 0;
1193
1194                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1195
1196                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1197                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1198
1199                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1200                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1201
1202                         if (rc != 0) {
1203                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1204                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1205                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1206                                 ksocknal_new_packet(conn, 0);
1207                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1208                                 return (rc);
1209                         }
1210                 }
1211
1212                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1213                         ksocknal_new_packet (conn, 0);
1214                         return 0;       /* NOOP is done and just return */
1215                 }
1216
1217                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1218                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1219                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1220
1221                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1222                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1223                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1224
1225                 conn->ksnc_rx_niov = 1;
1226                 conn->ksnc_rx_kiov = NULL;
1227                 conn->ksnc_rx_nkiov = 0;
1228
1229                 goto again;     /* read lnet header now */
1230
1231         case SOCKNAL_RX_LNET_HEADER:
1232                 /* unpack message header */
1233                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1234
1235                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1236                         /* Userspace peer */
1237                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1238                         id   = &conn->ksnc_peer->ksnp_id;
1239
1240                         /* Substitute process ID assigned at connection time */
1241                         lhdr->src_pid = cpu_to_le32(id->pid);
1242                         lhdr->src_nid = cpu_to_le64(id->nid);
1243                 }
1244
1245                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1246                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1247
1248                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1249                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1250                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1251                 if (rc < 0) {
1252                         /* I just received garbage: give up on this conn */
1253                         ksocknal_new_packet(conn, 0);
1254                         ksocknal_close_conn_and_siblings (conn, rc);
1255                         ksocknal_conn_decref(conn);
1256                         return (-EPROTO);
1257                 }
1258
1259                 /* I'm racing with ksocknal_recv() */
1260                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1261                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1262
1263                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1264                         return 0;
1265
1266                 /* ksocknal_recv() got called */
1267                 goto again;
1268
1269         case SOCKNAL_RX_LNET_PAYLOAD:
1270                 /* payload all received */
1271                 rc = 0;
1272
1273                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1274                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1275                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1276                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1277                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1278                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1279                         rc = -EIO;
1280                 }
1281
1282                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1283                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1284
1285                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1286                         id   = &conn->ksnc_peer->ksnp_id;
1287
1288                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1289                                         conn->ksnc_msg.ksm_zc_cookies[0],
1290                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1291                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1292                 }
1293
1294                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1295
1296                 if (rc != 0) {
1297                         ksocknal_new_packet(conn, 0);
1298                         ksocknal_close_conn_and_siblings (conn, rc);
1299                         return (-EPROTO);
1300                 }
1301                 /* Fall through */
1302
1303         case SOCKNAL_RX_SLOP:
1304                 /* starting new packet? */
1305                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1306                         return 0;       /* come back later */
1307                 goto again;             /* try to finish reading slop now */
1308
1309         default:
1310                 break;
1311         }
1312
1313         /* Not Reached */
1314         LBUG ();
1315         return (-EINVAL);                       /* keep gcc happy */
1316 }
1317
1318 int
1319 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1320                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1321                unsigned int offset, unsigned int mlen, unsigned int rlen)
1322 {
1323         ksock_conn_t  *conn = (ksock_conn_t *)private;
1324         ksock_sched_t *sched = conn->ksnc_scheduler;
1325
1326         LASSERT (mlen <= rlen);
1327         LASSERT (niov <= LNET_MAX_IOV);
1328
1329         conn->ksnc_cookie = msg;
1330         conn->ksnc_rx_nob_wanted = mlen;
1331         conn->ksnc_rx_nob_left   = rlen;
1332
1333         if (mlen == 0 || iov != NULL) {
1334                 conn->ksnc_rx_nkiov = 0;
1335                 conn->ksnc_rx_kiov = NULL;
1336                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1337                 conn->ksnc_rx_niov =
1338                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1339                                          niov, iov, offset, mlen);
1340         } else {
1341                 conn->ksnc_rx_niov = 0;
1342                 conn->ksnc_rx_iov  = NULL;
1343                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1344                 conn->ksnc_rx_nkiov =
1345                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1346                                           niov, kiov, offset, mlen);
1347         }
1348
1349         LASSERT (mlen ==
1350                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1351                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1352
1353         LASSERT (conn->ksnc_rx_scheduled);
1354
1355         spin_lock_bh(&sched->kss_lock);
1356
1357         switch (conn->ksnc_rx_state) {
1358         case SOCKNAL_RX_PARSE_WAIT:
1359                 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1360                 cfs_waitq_signal (&sched->kss_waitq);
1361                 LASSERT (conn->ksnc_rx_ready);
1362                 break;
1363
1364         case SOCKNAL_RX_PARSE:
1365                 /* scheduler hasn't noticed I'm parsing yet */
1366                 break;
1367         }
1368
1369         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1370
1371         spin_unlock_bh(&sched->kss_lock);
1372         ksocknal_conn_decref(conn);
1373         return 0;
1374 }
1375
1376 static inline int
1377 ksocknal_sched_cansleep(ksock_sched_t *sched)
1378 {
1379         int           rc;
1380
1381         spin_lock_bh(&sched->kss_lock);
1382
1383         rc = (!ksocknal_data.ksnd_shuttingdown &&
1384               cfs_list_empty(&sched->kss_rx_conns) &&
1385               cfs_list_empty(&sched->kss_tx_conns));
1386
1387         spin_unlock_bh(&sched->kss_lock);
1388         return rc;
1389 }
1390
1391 int ksocknal_scheduler(void *arg)
1392 {
1393         struct ksock_sched_info *info;
1394         ksock_sched_t           *sched;
1395         ksock_conn_t            *conn;
1396         ksock_tx_t              *tx;
1397         int                     rc;
1398         int                     nloops = 0;
1399         char                    name[20];
1400         long                    id = (long)arg;
1401
1402         info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1403         sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1404
1405         snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
1406                  info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
1407
1408         cfs_daemonize(name);
1409         cfs_block_allsigs();
1410
1411         rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1412         if (rc != 0) {
1413                 CERROR("Can't set CPT affinity for %s to %d: %d\n",
1414                        name, info->ksi_cpt, rc);
1415         }
1416
1417         spin_lock_bh(&sched->kss_lock);
1418
1419         while (!ksocknal_data.ksnd_shuttingdown) {
1420                 int did_something = 0;
1421
1422                 /* Ensure I progress everything semi-fairly */
1423
1424                 if (!cfs_list_empty (&sched->kss_rx_conns)) {
1425                         conn = cfs_list_entry(sched->kss_rx_conns.next,
1426                                               ksock_conn_t, ksnc_rx_list);
1427                         cfs_list_del(&conn->ksnc_rx_list);
1428
1429                         LASSERT(conn->ksnc_rx_scheduled);
1430                         LASSERT(conn->ksnc_rx_ready);
1431
1432                         /* clear rx_ready in case receive isn't complete.
1433                          * Do it BEFORE we call process_recv, since
1434                          * data_ready can set it any time after we release
1435                          * kss_lock. */
1436                         conn->ksnc_rx_ready = 0;
1437                         spin_unlock_bh(&sched->kss_lock);
1438
1439                         rc = ksocknal_process_receive(conn);
1440
1441                         spin_lock_bh(&sched->kss_lock);
1442
1443                         /* I'm the only one that can clear this flag */
1444                         LASSERT(conn->ksnc_rx_scheduled);
1445
1446                         /* Did process_receive get everything it wanted? */
1447                         if (rc == 0)
1448                                 conn->ksnc_rx_ready = 1;
1449
1450                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1451                                 /* Conn blocked waiting for ksocknal_recv()
1452                                  * I change its state (under lock) to signal
1453                                  * it can be rescheduled */
1454                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1455                         } else if (conn->ksnc_rx_ready) {
1456                                 /* reschedule for rx */
1457                                 cfs_list_add_tail (&conn->ksnc_rx_list,
1458                                                    &sched->kss_rx_conns);
1459                         } else {
1460                                 conn->ksnc_rx_scheduled = 0;
1461                                 /* drop my ref */
1462                                 ksocknal_conn_decref(conn);
1463                         }
1464
1465                         did_something = 1;
1466                 }
1467
1468                 if (!cfs_list_empty (&sched->kss_tx_conns)) {
1469                         CFS_LIST_HEAD    (zlist);
1470
1471                         if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1472                                 cfs_list_add(&zlist,
1473                                              &sched->kss_zombie_noop_txs);
1474                                 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1475                         }
1476
1477                         conn = cfs_list_entry(sched->kss_tx_conns.next,
1478                                               ksock_conn_t, ksnc_tx_list);
1479                         cfs_list_del (&conn->ksnc_tx_list);
1480
1481                         LASSERT(conn->ksnc_tx_scheduled);
1482                         LASSERT(conn->ksnc_tx_ready);
1483                         LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1484
1485                         tx = cfs_list_entry(conn->ksnc_tx_queue.next,
1486                                             ksock_tx_t, tx_list);
1487
1488                         if (conn->ksnc_tx_carrier == tx)
1489                                 ksocknal_next_tx_carrier(conn);
1490
1491                         /* dequeue now so empty list => more to send */
1492                         cfs_list_del(&tx->tx_list);
1493
1494                         /* Clear tx_ready in case send isn't complete.  Do
1495                          * it BEFORE we call process_transmit, since
1496                          * write_space can set it any time after we release
1497                          * kss_lock. */
1498                         conn->ksnc_tx_ready = 0;
1499                         spin_unlock_bh(&sched->kss_lock);
1500
1501                         if (!cfs_list_empty(&zlist)) {
1502                                 /* free zombie noop txs, it's fast because
1503                                  * noop txs are just put in freelist */
1504                                 ksocknal_txlist_done(NULL, &zlist, 0);
1505                         }
1506
1507                         rc = ksocknal_process_transmit(conn, tx);
1508
1509                         if (rc == -ENOMEM || rc == -EAGAIN) {
1510                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1511                                 spin_lock_bh(&sched->kss_lock);
1512                                 cfs_list_add(&tx->tx_list,
1513                                              &conn->ksnc_tx_queue);
1514                         } else {
1515                                 /* Complete send; tx -ref */
1516                                 ksocknal_tx_decref(tx);
1517
1518                                 spin_lock_bh(&sched->kss_lock);
1519                                 /* assume space for more */
1520                                 conn->ksnc_tx_ready = 1;
1521                         }
1522
1523                         if (rc == -ENOMEM) {
1524                                 /* Do nothing; after a short timeout, this
1525                                  * conn will be reposted on kss_tx_conns. */
1526                         } else if (conn->ksnc_tx_ready &&
1527                                    !cfs_list_empty (&conn->ksnc_tx_queue)) {
1528                                 /* reschedule for tx */
1529                                 cfs_list_add_tail (&conn->ksnc_tx_list,
1530                                                    &sched->kss_tx_conns);
1531                         } else {
1532                                 conn->ksnc_tx_scheduled = 0;
1533                                 /* drop my ref */
1534                                 ksocknal_conn_decref(conn);
1535                         }
1536
1537                         did_something = 1;
1538                 }
1539                 if (!did_something ||           /* nothing to do */
1540                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1541                         spin_unlock_bh(&sched->kss_lock);
1542
1543                         nloops = 0;
1544
1545                         if (!did_something) {   /* wait for something to do */
1546                                 cfs_wait_event_interruptible_exclusive(
1547                                         sched->kss_waitq,
1548                                         !ksocknal_sched_cansleep(sched), rc);
1549                                 LASSERT (rc == 0);
1550                         } else {
1551                                 cfs_cond_resched();
1552                         }
1553
1554                         spin_lock_bh(&sched->kss_lock);
1555                 }
1556         }
1557
1558         spin_unlock_bh(&sched->kss_lock);
1559         ksocknal_thread_fini();
1560         return 0;
1561 }
1562
1563 /*
1564  * Add connection to kss_rx_conns of scheduler
1565  * and wakeup the scheduler.
1566  */
1567 void ksocknal_read_callback (ksock_conn_t *conn)
1568 {
1569         ksock_sched_t *sched;
1570         ENTRY;
1571
1572         sched = conn->ksnc_scheduler;
1573
1574         spin_lock_bh(&sched->kss_lock);
1575
1576         conn->ksnc_rx_ready = 1;
1577
1578         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1579                 cfs_list_add_tail(&conn->ksnc_rx_list,
1580                                   &sched->kss_rx_conns);
1581                 conn->ksnc_rx_scheduled = 1;
1582                 /* extra ref for scheduler */
1583                 ksocknal_conn_addref(conn);
1584
1585                 cfs_waitq_signal (&sched->kss_waitq);
1586         }
1587         spin_unlock_bh(&sched->kss_lock);
1588
1589         EXIT;
1590 }
1591
1592 /*
1593  * Add connection to kss_tx_conns of scheduler
1594  * and wakeup the scheduler.
1595  */
1596 void ksocknal_write_callback (ksock_conn_t *conn)
1597 {
1598         ksock_sched_t *sched;
1599         ENTRY;
1600
1601         sched = conn->ksnc_scheduler;
1602
1603         spin_lock_bh(&sched->kss_lock);
1604
1605         conn->ksnc_tx_ready = 1;
1606
1607         if (!conn->ksnc_tx_scheduled && // not being progressed
1608             !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1609                 cfs_list_add_tail (&conn->ksnc_tx_list,
1610                                    &sched->kss_tx_conns);
1611                 conn->ksnc_tx_scheduled = 1;
1612                 /* extra ref for scheduler */
1613                 ksocknal_conn_addref(conn);
1614
1615                 cfs_waitq_signal (&sched->kss_waitq);
1616         }
1617
1618         spin_unlock_bh(&sched->kss_lock);
1619
1620         EXIT;
1621 }
1622
1623 ksock_proto_t *
1624 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1625 {
1626         __u32   version = 0;
1627
1628         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1629                 version = hello->kshm_version;
1630         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1631                 version = __swab32(hello->kshm_version);
1632
1633         if (version != 0) {
1634 #if SOCKNAL_VERSION_DEBUG
1635                 if (*ksocknal_tunables.ksnd_protocol == 1)
1636                         return NULL;
1637
1638                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1639                     version == KSOCK_PROTO_V3)
1640                         return NULL;
1641 #endif
1642                 if (version == KSOCK_PROTO_V2)
1643                         return &ksocknal_protocol_v2x;
1644
1645                 if (version == KSOCK_PROTO_V3)
1646                         return &ksocknal_protocol_v3x;
1647
1648                 return NULL;
1649         }
1650
1651         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1652                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1653
1654                 CLASSERT (sizeof (lnet_magicversion_t) ==
1655                           offsetof (ksock_hello_msg_t, kshm_src_nid));
1656
1657                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1658                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1659                         return &ksocknal_protocol_v1x;
1660         }
1661
1662         return NULL;
1663 }
1664
1665 int
1666 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1667                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1668 {
1669         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1670         ksock_net_t         *net = (ksock_net_t *)ni->ni_data;
1671
1672         LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1673
1674         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1675         LASSERT (conn->ksnc_proto != NULL);
1676
1677         hello->kshm_src_nid         = ni->ni_nid;
1678         hello->kshm_dst_nid         = peer_nid;
1679         hello->kshm_src_pid         = the_lnet.ln_pid;
1680
1681         hello->kshm_src_incarnation = net->ksnn_incarnation;
1682         hello->kshm_ctype           = conn->ksnc_type;
1683
1684         return conn->ksnc_proto->pro_send_hello(conn, hello);
1685 }
1686
1687 int
1688 ksocknal_invert_type(int type)
1689 {
1690         switch (type)
1691         {
1692         case SOCKLND_CONN_ANY:
1693         case SOCKLND_CONN_CONTROL:
1694                 return (type);
1695         case SOCKLND_CONN_BULK_IN:
1696                 return SOCKLND_CONN_BULK_OUT;
1697         case SOCKLND_CONN_BULK_OUT:
1698                 return SOCKLND_CONN_BULK_IN;
1699         default:
1700                 return (SOCKLND_CONN_NONE);
1701         }
1702 }
1703
1704 int
1705 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1706                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1707                      __u64 *incarnation)
1708 {
1709         /* Return < 0        fatal error
1710          *        0          success
1711          *        EALREADY   lost connection race
1712          *        EPROTO     protocol version mismatch
1713          */
1714         cfs_socket_t        *sock = conn->ksnc_sock;
1715         int                  active = (conn->ksnc_proto != NULL);
1716         int                  timeout;
1717         int                  proto_match;
1718         int                  rc;
1719         ksock_proto_t       *proto;
1720         lnet_process_id_t    recv_id;
1721
1722         /* socket type set on active connections - not set on passive */
1723         LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1724
1725         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1726                             lnet_acceptor_timeout();
1727
1728         rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1729         if (rc != 0) {
1730                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1731                         rc, HIPQUAD(conn->ksnc_ipaddr));
1732                 LASSERT (rc < 0);
1733                 return rc;
1734         }
1735
1736         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1737             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1738             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1739                 /* Unexpected magic! */
1740                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1741                         "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1742                         LNET_PROTO_TCP_MAGIC,
1743                         HIPQUAD(conn->ksnc_ipaddr));
1744                 return -EPROTO;
1745         }
1746
1747         rc = libcfs_sock_read(sock, &hello->kshm_version,
1748                               sizeof(hello->kshm_version), timeout);
1749         if (rc != 0) {
1750                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1751                         rc, HIPQUAD(conn->ksnc_ipaddr));
1752                 LASSERT (rc < 0);
1753                 return rc;
1754         }
1755
1756         proto = ksocknal_parse_proto_version(hello);
1757         if (proto == NULL) {
1758                 if (!active) {
1759                         /* unknown protocol from peer, tell peer my protocol */
1760                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1761 #if SOCKNAL_VERSION_DEBUG
1762                         if (*ksocknal_tunables.ksnd_protocol == 2)
1763                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1764                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1765                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1766 #endif
1767                         hello->kshm_nips = 0;
1768                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1769                 }
1770
1771                 CERROR ("Unknown protocol version (%d.x expected)"
1772                         " from %u.%u.%u.%u\n",
1773                         conn->ksnc_proto->pro_version,
1774                         HIPQUAD(conn->ksnc_ipaddr));
1775
1776                 return -EPROTO;
1777         }
1778
1779         proto_match = (conn->ksnc_proto == proto);
1780         conn->ksnc_proto = proto;
1781
1782         /* receive the rest of hello message anyway */
1783         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1784         if (rc != 0) {
1785                 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1786                        rc, HIPQUAD(conn->ksnc_ipaddr));
1787                 LASSERT (rc < 0);
1788                 return rc;
1789         }
1790
1791         *incarnation = hello->kshm_src_incarnation;
1792
1793         if (hello->kshm_src_nid == LNET_NID_ANY) {
1794                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1795                        "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1796                 return -EPROTO;
1797         }
1798
1799         if (!active &&
1800             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1801                 /* Userspace NAL assigns peer process ID from socket */
1802                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1803                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1804         } else {
1805                 recv_id.nid = hello->kshm_src_nid;
1806                 recv_id.pid = hello->kshm_src_pid;
1807         }
1808
1809         if (!active) {
1810                 *peerid = recv_id;
1811
1812                 /* peer determines type */
1813                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1814                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1815                         CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1816                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1817                                 HIPQUAD(conn->ksnc_ipaddr));
1818                         return -EPROTO;
1819                 }
1820
1821                 return 0;
1822         }
1823
1824         if (peerid->pid != recv_id.pid ||
1825             peerid->nid != recv_id.nid) {
1826                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1827                                    " %u.%u.%u.%u, but they claimed they were "
1828                                    "%s; please check your Lustre "
1829                                    "configuration.\n",
1830                                    libcfs_id2str(*peerid),
1831                                    HIPQUAD(conn->ksnc_ipaddr),
1832                                    libcfs_id2str(recv_id));
1833                 return -EPROTO;
1834         }
1835
1836         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1837                 /* Possible protocol mismatch or I lost the connection race */
1838                 return proto_match ? EALREADY : EPROTO;
1839         }
1840
1841         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1842                 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1843                         conn->ksnc_type, libcfs_id2str(*peerid),
1844                         HIPQUAD(conn->ksnc_ipaddr),
1845                         hello->kshm_ctype);
1846                 return -EPROTO;
1847         }
1848
1849         return 0;
1850 }
1851
1852 int
1853 ksocknal_connect (ksock_route_t *route)
1854 {
1855         CFS_LIST_HEAD    (zombies);
1856         ksock_peer_t     *peer = route->ksnr_peer;
1857         int               type;
1858         int               wanted;
1859         cfs_socket_t     *sock;
1860         cfs_time_t        deadline;
1861         int               retry_later = 0;
1862         int               rc = 0;
1863
1864         deadline = cfs_time_add(cfs_time_current(),
1865                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1866
1867         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1868
1869         LASSERT (route->ksnr_scheduled);
1870         LASSERT (!route->ksnr_connecting);
1871
1872         route->ksnr_connecting = 1;
1873
1874         for (;;) {
1875                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1876
1877                 /* stop connecting if peer/route got closed under me, or
1878                  * route got connected while queued */
1879                 if (peer->ksnp_closing || route->ksnr_deleted ||
1880                     wanted == 0) {
1881                         retry_later = 0;
1882                         break;
1883                 }
1884
1885                 /* reschedule if peer is connecting to me */
1886                 if (peer->ksnp_accepting > 0) {
1887                         CDEBUG(D_NET,
1888                                "peer %s(%d) already connecting to me, retry later.\n",
1889                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1890                         retry_later = 1;
1891                 }
1892
1893                 if (retry_later) /* needs reschedule */
1894                         break;
1895
1896                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1897                         type = SOCKLND_CONN_ANY;
1898                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1899                         type = SOCKLND_CONN_CONTROL;
1900                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1901                         type = SOCKLND_CONN_BULK_IN;
1902                 } else {
1903                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1904                         type = SOCKLND_CONN_BULK_OUT;
1905                 }
1906
1907                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1908
1909                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1910                         rc = -ETIMEDOUT;
1911                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1912                                                    route->ksnr_ipaddr,
1913                                                    route->ksnr_port);
1914                         goto failed;
1915                 }
1916
1917                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1918                                   route->ksnr_myipaddr,
1919                                   route->ksnr_ipaddr, route->ksnr_port);
1920                 if (rc != 0)
1921                         goto failed;
1922
1923                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1924                 if (rc < 0) {
1925                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1926                                                    route->ksnr_ipaddr,
1927                                                    route->ksnr_port);
1928                         goto failed;
1929                 }
1930
1931                 /* A +ve RC means I have to retry because I lost the connection
1932                  * race or I have to renegotiate protocol version */
1933                 retry_later = (rc != 0);
1934                 if (retry_later)
1935                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1936                                libcfs_nid2str(peer->ksnp_id.nid));
1937
1938                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1939         }
1940
1941         route->ksnr_scheduled = 0;
1942         route->ksnr_connecting = 0;
1943
1944         if (retry_later) {
1945                 /* re-queue for attention; this frees me up to handle
1946                  * the peer's incoming connection request */
1947
1948                 if (rc == EALREADY ||
1949                     (rc == 0 && peer->ksnp_accepting > 0)) {
1950                         /* We want to introduce a delay before next
1951                          * attempt to connect if we lost conn race,
1952                          * but the race is resolved quickly usually,
1953                          * so min_reconnectms should be good heuristic */
1954                         route->ksnr_retry_interval =
1955                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1956                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1957                                                            route->ksnr_retry_interval);
1958                 }
1959
1960                 ksocknal_launch_connection_locked(route);
1961         }
1962
1963         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1964         return retry_later;
1965
1966  failed:
1967         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1968
1969         route->ksnr_scheduled = 0;
1970         route->ksnr_connecting = 0;
1971
1972         /* This is a retry rather than a new connection */
1973         route->ksnr_retry_interval *= 2;
1974         route->ksnr_retry_interval =
1975                 MAX(route->ksnr_retry_interval,
1976                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1977         route->ksnr_retry_interval =
1978                 MIN(route->ksnr_retry_interval,
1979                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1980
1981         LASSERT (route->ksnr_retry_interval != 0);
1982         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1983                                            route->ksnr_retry_interval);
1984
1985         if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1986             peer->ksnp_accepting == 0 &&
1987             ksocknal_find_connecting_route_locked(peer) == NULL) {
1988                 ksock_conn_t *conn;
1989
1990                 /* ksnp_tx_queue is queued on a conn on successful
1991                  * connection for V1.x and V2.x */
1992                 if (!cfs_list_empty (&peer->ksnp_conns)) {
1993                         conn = cfs_list_entry(peer->ksnp_conns.next,
1994                                               ksock_conn_t, ksnc_list);
1995                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1996                 }
1997
1998                 /* take all the blocked packets while I've got the lock and
1999                  * complete below... */
2000                 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
2001         }
2002
2003 #if 0           /* irrelevent with only eager routes */
2004         if (!route->ksnr_deleted) {
2005                 /* make this route least-favourite for re-selection */
2006                 cfs_list_del(&route->ksnr_list);
2007                 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2008         }
2009 #endif
2010         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2011
2012         ksocknal_peer_failed(peer);
2013         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2014         return 0;
2015 }
2016
2017 /*
2018  * check whether we need to create more connds.
2019  * It will try to create new thread if it's necessary, @timeout can
2020  * be updated if failed to create, so caller wouldn't keep try while
2021  * running out of resource.
2022  */
2023 static int
2024 ksocknal_connd_check_start(long sec, long *timeout)
2025 {
2026         int rc;
2027         int total = ksocknal_data.ksnd_connd_starting +
2028                     ksocknal_data.ksnd_connd_running;
2029
2030         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2031                 /* still in initializing */
2032                 return 0;
2033         }
2034
2035         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2036             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2037                 /* can't create more connd, or still have enough
2038                  * threads to handle more connecting */
2039                 return 0;
2040         }
2041
2042         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2043                 /* no pending connecting request */
2044                 return 0;
2045         }
2046
2047         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2048                 /* may run out of resource, retry later */
2049                 *timeout = cfs_time_seconds(1);
2050                 return 0;
2051         }
2052
2053         if (ksocknal_data.ksnd_connd_starting > 0) {
2054                 /* serialize starting to avoid flood */
2055                 return 0;
2056         }
2057
2058         ksocknal_data.ksnd_connd_starting_stamp = sec;
2059         ksocknal_data.ksnd_connd_starting++;
2060         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2061
2062         /* NB: total is the next id */
2063         rc = ksocknal_thread_start(ksocknal_connd, (void *)((long)total));
2064
2065         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2066         if (rc == 0)
2067                 return 1;
2068
2069         /* we tried ... */
2070         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2071         ksocknal_data.ksnd_connd_starting--;
2072         ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2073
2074         return 1;
2075 }
2076
2077 /*
2078  * check whether current thread can exit, it will return 1 if there are too
2079  * many threads and no creating in past 120 seconds.
2080  * Also, this function may update @timeout to make caller come back
2081  * again to recheck these conditions.
2082  */
2083 static int
2084 ksocknal_connd_check_stop(long sec, long *timeout)
2085 {
2086         int val;
2087
2088         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2089                 /* still in initializing */
2090                 return 0;
2091         }
2092
2093         if (ksocknal_data.ksnd_connd_starting > 0) {
2094                 /* in progress of starting new thread */
2095                 return 0;
2096         }
2097
2098         if (ksocknal_data.ksnd_connd_running <=
2099             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2100                 return 0;
2101         }
2102
2103         /* created thread in past 120 seconds? */
2104         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2105                     SOCKNAL_CONND_TIMEOUT - sec);
2106
2107         *timeout = (val > 0) ? cfs_time_seconds(val) :
2108                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2109         if (val > 0)
2110                 return 0;
2111
2112         /* no creating in past 120 seconds */
2113
2114         return ksocknal_data.ksnd_connd_running >
2115                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2116 }
2117
2118 /* Go through connd_routes queue looking for a route that we can process
2119  * right now, @timeout_p can be updated if we need to come back later */
2120 static ksock_route_t *
2121 ksocknal_connd_get_route_locked(signed long *timeout_p)
2122 {
2123         ksock_route_t *route;
2124         cfs_time_t     now;
2125
2126         now = cfs_time_current();
2127
2128         /* connd_routes can contain both pending and ordinary routes */
2129         cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2130                                  ksnr_connd_list) {
2131
2132                 if (route->ksnr_retry_interval == 0 ||
2133                     cfs_time_aftereq(now, route->ksnr_timeout))
2134                         return route;
2135
2136                 if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT ||
2137                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2138                         *timeout_p = (int)(route->ksnr_timeout - now);
2139         }
2140
2141         return NULL;
2142 }
2143
2144 int
2145 ksocknal_connd (void *arg)
2146 {
2147         spinlock_t    *connd_lock = &ksocknal_data.ksnd_connd_lock;
2148         long               id = (long)(long_ptr_t)arg;
2149         char               name[16];
2150         ksock_connreq_t   *cr;
2151         cfs_waitlink_t     wait;
2152         int                nloops = 0;
2153         int                cons_retry = 0;
2154
2155         snprintf (name, sizeof (name), "socknal_cd%02ld", id);
2156         cfs_daemonize (name);
2157         cfs_block_allsigs ();
2158
2159         cfs_waitlink_init (&wait);
2160
2161         spin_lock_bh(connd_lock);
2162
2163         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2164         ksocknal_data.ksnd_connd_starting--;
2165         ksocknal_data.ksnd_connd_running++;
2166
2167         while (!ksocknal_data.ksnd_shuttingdown) {
2168                 ksock_route_t *route = NULL;
2169                 long sec = cfs_time_current_sec();
2170                 long timeout = CFS_MAX_SCHEDULE_TIMEOUT;
2171                 int  dropped_lock = 0;
2172
2173                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2174                         /* wakeup another one to check stop */
2175                         cfs_waitq_signal(&ksocknal_data.ksnd_connd_waitq);
2176                         break;
2177                 }
2178
2179                 if (ksocknal_connd_check_start(sec, &timeout)) {
2180                         /* created new thread */
2181                         dropped_lock = 1;
2182                 }
2183
2184                 if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2185                         /* Connection accepted by the listener */
2186                         cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \
2187                                             next, ksock_connreq_t, ksncr_list);
2188
2189                         cfs_list_del(&cr->ksncr_list);
2190                         spin_unlock_bh(connd_lock);
2191                         dropped_lock = 1;
2192
2193                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2194                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2195                         lnet_ni_decref(cr->ksncr_ni);
2196                         LIBCFS_FREE(cr, sizeof(*cr));
2197
2198                         spin_lock_bh(connd_lock);
2199                 }
2200
2201                 /* Only handle an outgoing connection request if there
2202                  * is a thread left to handle incoming connections and
2203                  * create new connd */
2204                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2205                     ksocknal_data.ksnd_connd_running) {
2206                         route = ksocknal_connd_get_route_locked(&timeout);
2207                 }
2208                 if (route != NULL) {
2209                         cfs_list_del (&route->ksnr_connd_list);
2210                         ksocknal_data.ksnd_connd_connecting++;
2211                         spin_unlock_bh(connd_lock);
2212                         dropped_lock = 1;
2213
2214                         if (ksocknal_connect(route)) {
2215                                 /* consecutive retry */
2216                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2217                                         CWARN("massive consecutive "
2218                                               "re-connecting to %u.%u.%u.%u\n",
2219                                               HIPQUAD(route->ksnr_ipaddr));
2220                                         cons_retry = 0;
2221                                 }
2222                         } else {
2223                                 cons_retry = 0;
2224                         }
2225
2226                         ksocknal_route_decref(route);
2227
2228                         spin_lock_bh(connd_lock);
2229                         ksocknal_data.ksnd_connd_connecting--;
2230                 }
2231
2232                 if (dropped_lock) {
2233                         if (++nloops < SOCKNAL_RESCHED)
2234                                 continue;
2235                         spin_unlock_bh(connd_lock);
2236                         nloops = 0;
2237                         cfs_cond_resched();
2238                         spin_lock_bh(connd_lock);
2239                         continue;
2240                 }
2241
2242                 /* Nothing to do for 'timeout'  */
2243                 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
2244                 cfs_waitq_add_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2245                 spin_unlock_bh(connd_lock);
2246
2247                 nloops = 0;
2248                 cfs_waitq_timedwait(&wait, CFS_TASK_INTERRUPTIBLE, timeout);
2249
2250                 cfs_set_current_state(CFS_TASK_RUNNING);
2251                 cfs_waitq_del(&ksocknal_data.ksnd_connd_waitq, &wait);
2252                 spin_lock_bh(connd_lock);
2253         }
2254         ksocknal_data.ksnd_connd_running--;
2255         spin_unlock_bh(connd_lock);
2256
2257         ksocknal_thread_fini();
2258         return 0;
2259 }
2260
2261 ksock_conn_t *
2262 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2263 {
2264         /* We're called with a shared lock on ksnd_global_lock */
2265         ksock_conn_t      *conn;
2266         cfs_list_t        *ctmp;
2267
2268         cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2269                 int     error;
2270                 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2271
2272                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2273                 LASSERT (!conn->ksnc_closing);
2274
2275                 /* SOCK_ERROR will reset error code of socket in
2276                  * some platform (like Darwin8.x) */
2277                 error = libcfs_sock_error(conn->ksnc_sock);
2278                 if (error != 0) {
2279                         ksocknal_conn_addref(conn);
2280
2281                         switch (error) {
2282                         case ECONNRESET:
2283                                 CNETERR("A connection with %s "
2284                                         "(%u.%u.%u.%u:%d) was reset; "
2285                                         "it may have rebooted.\n",
2286                                         libcfs_id2str(peer->ksnp_id),
2287                                         HIPQUAD(conn->ksnc_ipaddr),
2288                                         conn->ksnc_port);
2289                                 break;
2290                         case ETIMEDOUT:
2291                                 CNETERR("A connection with %s "
2292                                         "(%u.%u.%u.%u:%d) timed out; the "
2293                                         "network or node may be down.\n",
2294                                         libcfs_id2str(peer->ksnp_id),
2295                                         HIPQUAD(conn->ksnc_ipaddr),
2296                                         conn->ksnc_port);
2297                                 break;
2298                         default:
2299                                 CNETERR("An unexpected network error %d "
2300                                         "occurred with %s "
2301                                         "(%u.%u.%u.%u:%d\n", error,
2302                                         libcfs_id2str(peer->ksnp_id),
2303                                         HIPQUAD(conn->ksnc_ipaddr),
2304                                         conn->ksnc_port);
2305                                 break;
2306                         }
2307
2308                         return (conn);
2309                 }
2310
2311                 if (conn->ksnc_rx_started &&
2312                     cfs_time_aftereq(cfs_time_current(),
2313                                      conn->ksnc_rx_deadline)) {
2314                         /* Timed out incomplete incoming message */
2315                         ksocknal_conn_addref(conn);
2316                         CNETERR("Timeout receiving from %s (%u.%u.%u.%u:%d), "
2317                                 "state %d wanted %d left %d\n",
2318                                 libcfs_id2str(peer->ksnp_id),
2319                                 HIPQUAD(conn->ksnc_ipaddr),
2320                                 conn->ksnc_port,
2321                                 conn->ksnc_rx_state,
2322                                 conn->ksnc_rx_nob_wanted,
2323                                 conn->ksnc_rx_nob_left);
2324                         return (conn);
2325                 }
2326
2327                 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2328                      libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2329                     cfs_time_aftereq(cfs_time_current(),
2330                                      conn->ksnc_tx_deadline)) {
2331                         /* Timed out messages queued for sending or
2332                          * buffered in the socket's send buffer */
2333                         ksocknal_conn_addref(conn);
2334                         CNETERR("Timeout sending data to %s (%u.%u.%u.%u:%d) "
2335                                 "the network or that node may be down.\n",
2336                                 libcfs_id2str(peer->ksnp_id),
2337                                 HIPQUAD(conn->ksnc_ipaddr),
2338                                 conn->ksnc_port);
2339                         return (conn);
2340                 }
2341         }
2342
2343         return (NULL);
2344 }
2345
2346 static inline void
2347 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2348 {
2349         ksock_tx_t        *tx;
2350         CFS_LIST_HEAD      (stale_txs);
2351
2352         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2353
2354         while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2355                 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2356                                      ksock_tx_t, tx_list);
2357
2358                 if (!cfs_time_aftereq(cfs_time_current(),
2359                                       tx->tx_deadline))
2360                         break;
2361
2362                 cfs_list_del (&tx->tx_list);
2363                 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2364         }
2365
2366         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2367
2368         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2369 }
2370
2371 int
2372 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2373 {
2374         ksock_sched_t  *sched;
2375         ksock_conn_t   *conn;
2376         ksock_tx_t     *tx;
2377
2378         if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2379                 return 0;
2380
2381         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2382                 return 0;
2383
2384         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2385             cfs_time_before(cfs_time_current(),
2386                             cfs_time_add(peer->ksnp_last_alive,
2387                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2388                 return 0;
2389
2390         if (cfs_time_before(cfs_time_current(),
2391                             peer->ksnp_send_keepalive))
2392                 return 0;
2393
2394         /* retry 10 secs later, so we wouldn't put pressure
2395          * on this peer if we failed to send keepalive this time */
2396         peer->ksnp_send_keepalive = cfs_time_shift(10);
2397
2398         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2399         if (conn != NULL) {
2400                 sched = conn->ksnc_scheduler;
2401
2402                 spin_lock_bh(&sched->kss_lock);
2403                 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2404                         spin_unlock_bh(&sched->kss_lock);
2405                         /* there is an queued ACK, don't need keepalive */
2406                         return 0;
2407                 }
2408
2409                 spin_unlock_bh(&sched->kss_lock);
2410         }
2411
2412         read_unlock(&ksocknal_data.ksnd_global_lock);
2413
2414         /* cookie = 1 is reserved for keepalive PING */
2415         tx = ksocknal_alloc_tx_noop(1, 1);
2416         if (tx == NULL) {
2417                 read_lock(&ksocknal_data.ksnd_global_lock);
2418                 return -ENOMEM;
2419         }
2420
2421         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2422                 read_lock(&ksocknal_data.ksnd_global_lock);
2423                 return 1;
2424         }
2425
2426         ksocknal_free_tx(tx);
2427         read_lock(&ksocknal_data.ksnd_global_lock);
2428
2429         return -EIO;
2430 }
2431
2432
2433 void
2434 ksocknal_check_peer_timeouts (int idx)
2435 {
2436         cfs_list_t       *peers = &ksocknal_data.ksnd_peers[idx];
2437         ksock_peer_t     *peer;
2438         ksock_conn_t     *conn;
2439         ksock_tx_t       *tx;
2440
2441  again:
2442         /* NB. We expect to have a look at all the peers and not find any
2443          * connections to time out, so we just use a shared lock while we
2444          * take a look... */
2445         read_lock(&ksocknal_data.ksnd_global_lock);
2446
2447         cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2448                 cfs_time_t  deadline = 0;
2449                 int         resid = 0;
2450                 int         n     = 0;
2451
2452                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2453                         read_unlock(&ksocknal_data.ksnd_global_lock);
2454                         goto again;
2455                 }
2456
2457                 conn = ksocknal_find_timed_out_conn (peer);
2458
2459                 if (conn != NULL) {
2460                         read_unlock(&ksocknal_data.ksnd_global_lock);
2461
2462                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2463
2464                         /* NB we won't find this one again, but we can't
2465                          * just proceed with the next peer, since we dropped
2466                          * ksnd_global_lock and it might be dead already! */
2467                         ksocknal_conn_decref(conn);
2468                         goto again;
2469                 }
2470
2471                 /* we can't process stale txs right here because we're
2472                  * holding only shared lock */
2473                 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2474                         ksock_tx_t *tx =
2475                                 cfs_list_entry (peer->ksnp_tx_queue.next,
2476                                                 ksock_tx_t, tx_list);
2477
2478                         if (cfs_time_aftereq(cfs_time_current(),
2479                                              tx->tx_deadline)) {
2480
2481                                 ksocknal_peer_addref(peer);
2482                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2483
2484                                 ksocknal_flush_stale_txs(peer);
2485
2486                                 ksocknal_peer_decref(peer);
2487                                 goto again;
2488                         }
2489                 }
2490
2491                 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2492                         continue;
2493
2494                 spin_lock(&peer->ksnp_lock);
2495                 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2496                                               ksock_tx_t, tx_zc_list) {
2497                         if (!cfs_time_aftereq(cfs_time_current(),
2498                                               tx->tx_deadline))
2499                                 break;
2500                         /* ignore the TX if connection is being closed */
2501                         if (tx->tx_conn->ksnc_closing)
2502                                 continue;
2503                         n++;
2504                 }
2505
2506                 if (n == 0) {
2507                         spin_unlock(&peer->ksnp_lock);
2508                         continue;
2509                 }
2510
2511                 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2512                                     ksock_tx_t, tx_zc_list);
2513                 deadline = tx->tx_deadline;
2514                 resid    = tx->tx_resid;
2515                 conn     = tx->tx_conn;
2516                 ksocknal_conn_addref(conn);
2517
2518                 spin_unlock(&peer->ksnp_lock);
2519                 read_unlock(&ksocknal_data.ksnd_global_lock);
2520
2521                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2522                        "oldest(%p) timed out %ld secs ago, "
2523                        "resid: %d, wmem: %d\n",
2524                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2525                        cfs_duration_sec(cfs_time_current() - deadline),
2526                        resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2527
2528                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2529                 ksocknal_conn_decref(conn);
2530                 goto again;
2531         }
2532
2533         read_unlock(&ksocknal_data.ksnd_global_lock);
2534 }
2535
2536 int
2537 ksocknal_reaper (void *arg)
2538 {
2539         cfs_waitlink_t     wait;
2540         ksock_conn_t      *conn;
2541         ksock_sched_t     *sched;
2542         cfs_list_t         enomem_conns;
2543         int                nenomem_conns;
2544         cfs_duration_t     timeout;
2545         int                i;
2546         int                peer_index = 0;
2547         cfs_time_t         deadline = cfs_time_current();
2548
2549         cfs_daemonize ("socknal_reaper");
2550         cfs_block_allsigs ();
2551
2552         CFS_INIT_LIST_HEAD(&enomem_conns);
2553         cfs_waitlink_init (&wait);
2554
2555         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2556
2557         while (!ksocknal_data.ksnd_shuttingdown) {
2558
2559                 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2560                         conn = cfs_list_entry (ksocknal_data. \
2561                                                ksnd_deathrow_conns.next,
2562                                                ksock_conn_t, ksnc_list);
2563                         cfs_list_del (&conn->ksnc_list);
2564
2565                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2566
2567                         ksocknal_terminate_conn(conn);
2568                         ksocknal_conn_decref(conn);
2569
2570                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2571                         continue;
2572                 }
2573
2574                 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2575                         conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2576                                                next, ksock_conn_t, ksnc_list);
2577                         cfs_list_del (&conn->ksnc_list);
2578
2579                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2580
2581                         ksocknal_destroy_conn(conn);
2582
2583                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2584                         continue;
2585                 }
2586
2587                 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2588                         cfs_list_add(&enomem_conns,
2589                                      &ksocknal_data.ksnd_enomem_conns);
2590                         cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2591                 }
2592
2593                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2594
2595                 /* reschedule all the connections that stalled with ENOMEM... */
2596                 nenomem_conns = 0;
2597                 while (!cfs_list_empty (&enomem_conns)) {
2598                         conn = cfs_list_entry (enomem_conns.next,
2599                                                ksock_conn_t, ksnc_tx_list);
2600                         cfs_list_del (&conn->ksnc_tx_list);
2601
2602                         sched = conn->ksnc_scheduler;
2603
2604                         spin_lock_bh(&sched->kss_lock);
2605
2606                         LASSERT(conn->ksnc_tx_scheduled);
2607                         conn->ksnc_tx_ready = 1;
2608                         cfs_list_add_tail(&conn->ksnc_tx_list,
2609                                           &sched->kss_tx_conns);
2610                         cfs_waitq_signal(&sched->kss_waitq);
2611
2612                         spin_unlock_bh(&sched->kss_lock);
2613                         nenomem_conns++;
2614                 }
2615
2616                 /* careful with the jiffy wrap... */
2617                 while ((timeout = cfs_time_sub(deadline,
2618                                                cfs_time_current())) <= 0) {
2619                         const int n = 4;
2620                         const int p = 1;
2621                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2622
2623                         /* Time to check for timeouts on a few more peers: I do
2624                          * checks every 'p' seconds on a proportion of the peer
2625                          * table and I need to check every connection 'n' times
2626                          * within a timeout interval, to ensure I detect a
2627                          * timeout on any connection within (n+1)/n times the
2628                          * timeout interval. */
2629
2630                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2631                                 chunk = (chunk * n * p) /
2632                                         *ksocknal_tunables.ksnd_timeout;
2633                         if (chunk == 0)
2634                                 chunk = 1;
2635
2636                         for (i = 0; i < chunk; i++) {
2637                                 ksocknal_check_peer_timeouts (peer_index);
2638                                 peer_index = (peer_index + 1) %
2639                                              ksocknal_data.ksnd_peer_hash_size;
2640                         }
2641
2642                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2643                 }
2644
2645                 if (nenomem_conns != 0) {
2646                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2647                          * This also prevents me getting woken immediately
2648                          * if any go back on my enomem list. */
2649                         timeout = SOCKNAL_ENOMEM_RETRY;
2650                 }
2651                 ksocknal_data.ksnd_reaper_waketime =
2652                         cfs_time_add(cfs_time_current(), timeout);
2653
2654                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2655                 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2656
2657                 if (!ksocknal_data.ksnd_shuttingdown &&
2658                     cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2659                     cfs_list_empty (&ksocknal_data.ksnd_zombie_conns))
2660                         cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE,
2661                                              timeout);
2662
2663                 cfs_set_current_state (CFS_TASK_RUNNING);
2664                 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2665
2666                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2667         }
2668
2669         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2670
2671         ksocknal_thread_fini();
2672         return 0;
2673 }