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, 2017, 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 Lustre, https://wiki.whamcloud.com/
  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 struct ksock_tx *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         struct ksock_tx *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next,
  42                                         struct ksock_tx, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         list_del(&tx->tx_list);
  45                 }
  46
  47                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_hstatus = LNET_MSG_STATUS_OK;
  61         tx->tx_desc_size  = size;
  62
  63         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  64
  65         return tx;
  66 }
  67
  68 struct ksock_tx *
  69 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  70 {
  71         struct ksock_tx *tx;
  72
  73         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  74         if (tx == NULL) {
  75                 CERROR("Can't allocate noop tx desc\n");
  76                 return NULL;
  77         }
  78
  79         tx->tx_conn     = NULL;
  80         tx->tx_lnetmsg  = NULL;
  81         tx->tx_kiov     = NULL;
  82         tx->tx_nkiov    = 0;
  83         tx->tx_iov      = tx->tx_frags.virt.iov;
  84         tx->tx_niov     = 1;
  85         tx->tx_nonblk   = nonblk;
  86
  87         tx->tx_msg.ksm_csum = 0;
  88         tx->tx_msg.ksm_type = KSOCK_MSG_NOOP;
  89         tx->tx_msg.ksm_zc_cookies[0] = 0;
  90         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  91
  92         return tx;
  93 }
  94
  95
  96 void
  97 ksocknal_free_tx(struct ksock_tx *tx)
  98 {
  99         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
 100
 101         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
 102                 /* it's a noop tx */
 103                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 104
 105                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 106
 107                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 108         } else {
 109                 LIBCFS_FREE(tx, tx->tx_desc_size);
 110         }
 111 }
 112
 113 static int
 114 ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx,
 115                   struct kvec *scratch_iov)
 116 {
 117         struct kvec *iov = tx->tx_iov;
 118         int    nob;
 119         int    rc;
 120
 121         LASSERT(tx->tx_niov > 0);
 122
 123         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 124         rc = ksocknal_lib_send_iov(conn, tx, scratch_iov);
 125
 126         if (rc <= 0)                            /* sent nothing? */
 127                 return rc;
 128
 129         nob = rc;
 130         LASSERT(nob <= tx->tx_resid);
 131         tx->tx_resid -= nob;
 132
 133         /* "consume" iov */
 134         do {
 135                 LASSERT(tx->tx_niov > 0);
 136
 137                 if (nob < (int) iov->iov_len) {
 138                         iov->iov_base += nob;
 139                         iov->iov_len -= nob;
 140                         return rc;
 141                 }
 142
 143                 nob -= iov->iov_len;
 144                 tx->tx_iov = ++iov;
 145                 tx->tx_niov--;
 146         } while (nob != 0);
 147
 148         return rc;
 149 }
 150
 151 static int
 152 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
 153                    struct kvec *scratch_iov)
 154 {
 155         lnet_kiov_t *kiov = tx->tx_kiov;
 156         int nob;
 157         int rc;
 158
 159         LASSERT(tx->tx_niov == 0);
 160         LASSERT(tx->tx_nkiov > 0);
 161
 162         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 163         rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov);
 164
 165         if (rc <= 0)                            /* sent nothing? */
 166                 return rc;
 167
 168         nob = rc;
 169         LASSERT(nob <= tx->tx_resid);
 170         tx->tx_resid -= nob;
 171
 172         /* "consume" kiov */
 173         do {
 174                 LASSERT(tx->tx_nkiov > 0);
 175
 176                 if (nob < (int)kiov->kiov_len) {
 177                         kiov->kiov_offset += nob;
 178                         kiov->kiov_len -= nob;
 179                         return rc;
 180                 }
 181
 182                 nob -= (int)kiov->kiov_len;
 183                 tx->tx_kiov = ++kiov;
 184                 tx->tx_nkiov--;
 185         } while (nob != 0);
 186
 187         return rc;
 188 }
 189
 190 static int
 191 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 192                   struct kvec *scratch_iov)
 193 {
 194         int     rc;
 195         int     bufnob;
 196
 197         if (ksocknal_data.ksnd_stall_tx != 0)
 198                 schedule_timeout_uninterruptible(
 199                         cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 200
 201         LASSERT(tx->tx_resid != 0);
 202
 203         rc = ksocknal_connsock_addref(conn);
 204         if (rc != 0) {
 205                 LASSERT(conn->ksnc_closing);
 206                 return -ESHUTDOWN;
 207         }
 208
 209         do {
 210                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 211                         /* testing... */
 212                         ksocknal_data.ksnd_enomem_tx--;
 213                         rc = -EAGAIN;
 214                 } else if (tx->tx_niov != 0) {
 215                         rc = ksocknal_send_iov(conn, tx, scratch_iov);
 216                 } else {
 217                         rc = ksocknal_send_kiov(conn, tx, scratch_iov);
 218                 }
 219
 220                 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 221                 if (rc > 0)                     /* sent something? */
 222                         conn->ksnc_tx_bufnob += rc; /* account it */
 223
 224                 if (bufnob < conn->ksnc_tx_bufnob) {
 225                         /* allocated send buffer bytes < computed; infer
 226                          * something got ACKed */
 227                         conn->ksnc_tx_deadline = ktime_get_seconds() +
 228                                                  lnet_get_lnd_timeout();
 229                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 230                         conn->ksnc_tx_bufnob = bufnob;
 231                         smp_mb();
 232                 }
 233
 234                 if (rc <= 0) { /* Didn't write anything? */
 235                         /* some stacks return 0 instead of -EAGAIN */
 236                         if (rc == 0)
 237                                 rc = -EAGAIN;
 238
 239                         /* Check if EAGAIN is due to memory pressure */
 240                         if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 241                                 rc = -ENOMEM;
 242
 243                         break;
 244                 }
 245
 246                 /* socket's wmem_queued now includes 'rc' bytes */
 247                 atomic_sub (rc, &conn->ksnc_tx_nob);
 248                 rc = 0;
 249
 250         } while (tx->tx_resid != 0);
 251
 252         ksocknal_connsock_decref(conn);
 253         return rc;
 254 }
 255
 256 static int
 257 ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
 258 {
 259         struct kvec *iov = conn->ksnc_rx_iov;
 260         int     nob;
 261         int     rc;
 262
 263         LASSERT(conn->ksnc_rx_niov > 0);
 264
 265         /* Never touch conn->ksnc_rx_iov or change connection
 266          * status inside ksocknal_lib_recv_iov */
 267         rc = ksocknal_lib_recv_iov(conn, scratchiov);
 268
 269         if (rc <= 0)
 270                 return rc;
 271
 272         /* received something... */
 273         nob = rc;
 274
 275         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 276         conn->ksnc_rx_deadline = ktime_get_seconds() +
 277                                  lnet_get_lnd_timeout();
 278         smp_mb();                       /* order with setting rx_started */
 279         conn->ksnc_rx_started = 1;
 280
 281         conn->ksnc_rx_nob_wanted -= nob;
 282         conn->ksnc_rx_nob_left -= nob;
 283
 284         do {
 285                 LASSERT(conn->ksnc_rx_niov > 0);
 286
 287                 if (nob < (int)iov->iov_len) {
 288                         iov->iov_len -= nob;
 289                         iov->iov_base += nob;
 290                         return -EAGAIN;
 291                 }
 292
 293                 nob -= iov->iov_len;
 294                 conn->ksnc_rx_iov = ++iov;
 295                 conn->ksnc_rx_niov--;
 296         } while (nob != 0);
 297
 298         return rc;
 299 }
 300
 301 static int
 302 ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 303                    struct kvec *scratch_iov)
 304 {
 305         lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
 306         int nob;
 307         int rc;
 308         LASSERT(conn->ksnc_rx_nkiov > 0);
 309
 310         /* Never touch conn->ksnc_rx_kiov or change connection
 311          * status inside ksocknal_lib_recv_iov */
 312         rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov);
 313
 314         if (rc <= 0)
 315                 return rc;
 316
 317         /* received something... */
 318         nob = rc;
 319
 320         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 321         conn->ksnc_rx_deadline = ktime_get_seconds() +
 322                                  lnet_get_lnd_timeout();
 323         smp_mb();                       /* order with setting rx_started */
 324         conn->ksnc_rx_started = 1;
 325
 326         conn->ksnc_rx_nob_wanted -= nob;
 327         conn->ksnc_rx_nob_left -= nob;
 328
 329         do {
 330                 LASSERT(conn->ksnc_rx_nkiov > 0);
 331
 332                 if (nob < (int) kiov->kiov_len) {
 333                         kiov->kiov_offset += nob;
 334                         kiov->kiov_len -= nob;
 335                         return -EAGAIN;
 336                 }
 337
 338                 nob -= kiov->kiov_len;
 339                 conn->ksnc_rx_kiov = ++kiov;
 340                 conn->ksnc_rx_nkiov--;
 341         } while (nob != 0);
 342
 343         return 1;
 344 }
 345
 346 static int
 347 ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 348                  struct kvec *scratch_iov)
 349 {
 350         /* Return 1 on success, 0 on EOF, < 0 on error.
 351          * Caller checks ksnc_rx_nob_wanted to determine
 352          * progress/completion. */
 353         int     rc;
 354         ENTRY;
 355
 356         if (ksocknal_data.ksnd_stall_rx != 0)
 357                 schedule_timeout_uninterruptible(
 358                         cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
 359
 360         rc = ksocknal_connsock_addref(conn);
 361         if (rc != 0) {
 362                 LASSERT(conn->ksnc_closing);
 363                 return -ESHUTDOWN;
 364         }
 365
 366         for (;;) {
 367                 if (conn->ksnc_rx_niov != 0)
 368                         rc = ksocknal_recv_iov(conn, scratch_iov);
 369                 else
 370                         rc = ksocknal_recv_kiov(conn, rx_scratch_pgs,
 371                                                  scratch_iov);
 372
 373                 if (rc <= 0) {
 374                         /* error/EOF or partial receive */
 375                         if (rc == -EAGAIN) {
 376                                 rc = 1;
 377                         } else if (rc == 0 && conn->ksnc_rx_started) {
 378                                 /* EOF in the middle of a message */
 379                                 rc = -EPROTO;
 380                         }
 381                         break;
 382                 }
 383
 384                 /* Completed a fragment */
 385
 386                 if (conn->ksnc_rx_nob_wanted == 0) {
 387                         rc = 1;
 388                         break;
 389                 }
 390         }
 391
 392         ksocknal_connsock_decref(conn);
 393         RETURN(rc);
 394 }
 395
 396 void
 397 ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc)
 398 {
 399         struct lnet_msg *lnetmsg = tx->tx_lnetmsg;
 400         enum lnet_msg_hstatus hstatus = tx->tx_hstatus;
 401         ENTRY;
 402
 403         LASSERT(ni != NULL || tx->tx_conn != NULL);
 404
 405         if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
 406                 rc = -EIO;
 407                 if (hstatus == LNET_MSG_STATUS_OK)
 408                         hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 409         }
 410
 411         if (tx->tx_conn != NULL)
 412                 ksocknal_conn_decref(tx->tx_conn);
 413
 414         ksocknal_free_tx(tx);
 415         if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
 416                 lnetmsg->msg_health_status = hstatus;
 417                 lnet_finalize(lnetmsg, rc);
 418         }
 419
 420         EXIT;
 421 }
 422
 423 void
 424 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
 425 {
 426         struct ksock_tx *tx;
 427
 428         while (!list_empty(txlist)) {
 429                 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
 430
 431                 if (error && tx->tx_lnetmsg != NULL) {
 432                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 433                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
 434                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
 435                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 436                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 437                 } else if (error) {
 438                         CNETERR("Deleting noop packet\n");
 439                 }
 440
 441                 list_del(&tx->tx_list);
 442
 443                 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
 444                         if (error == -ETIMEDOUT)
 445                                 tx->tx_hstatus =
 446                                   LNET_MSG_STATUS_LOCAL_TIMEOUT;
 447                         else if (error == -ENETDOWN ||
 448                                  error == -EHOSTUNREACH ||
 449                                  error == -ENETUNREACH ||
 450                                  error == -ECONNREFUSED ||
 451                                  error == -ECONNRESET)
 452                                 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
 453                         /*
 454                          * for all other errors we don't want to
 455                          * retransmit
 456                          */
 457                         else if (error)
 458                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 459                 }
 460
 461                 LASSERT(atomic_read(&tx->tx_refcount) == 1);
 462                 ksocknal_tx_done(ni, tx, error);
 463         }
 464 }
 465
 466 static void
 467 ksocknal_check_zc_req(struct ksock_tx *tx)
 468 {
 469         struct ksock_conn *conn = tx->tx_conn;
 470         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
 471
 472         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 473          * to ksnp_zc_req_list if some fragment of this message should be sent
 474          * zero-copy.  Our peer_ni will send an ACK containing this cookie when
 475          * she has received this message to tell us we can signal completion.
 476          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 477          * ksnp_zc_req_list. */
 478         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 479         LASSERT (tx->tx_zc_capable);
 480
 481         tx->tx_zc_checked = 1;
 482
 483         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 484             !conn->ksnc_zc_capable)
 485                 return;
 486
 487         /* assign cookie and queue tx to pending list, it will be released when
 488          * a matching ack is received. See ksocknal_handle_zcack() */
 489
 490         ksocknal_tx_addref(tx);
 491
 492         spin_lock(&peer_ni->ksnp_lock);
 493
 494         /* ZC_REQ is going to be pinned to the peer_ni */
 495         tx->tx_deadline = ktime_get_seconds() +
 496                           lnet_get_lnd_timeout();
 497
 498         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 499
 500         tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
 501
 502         if (peer_ni->ksnp_zc_next_cookie == 0)
 503                 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 504
 505         list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
 506
 507         spin_unlock(&peer_ni->ksnp_lock);
 508 }
 509
 510 static void
 511 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
 512 {
 513         struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
 514
 515         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 516         LASSERT(tx->tx_zc_capable);
 517
 518         tx->tx_zc_checked = 0;
 519
 520         spin_lock(&peer_ni->ksnp_lock);
 521
 522         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 523                 /* Not waiting for an ACK */
 524                 spin_unlock(&peer_ni->ksnp_lock);
 525                 return;
 526         }
 527
 528         tx->tx_msg.ksm_zc_cookies[0] = 0;
 529         list_del(&tx->tx_zc_list);
 530
 531         spin_unlock(&peer_ni->ksnp_lock);
 532
 533         ksocknal_tx_decref(tx);
 534 }
 535
 536 static int
 537 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 538                           struct kvec *scratch_iov)
 539 {
 540         int rc;
 541         bool error_sim = false;
 542
 543         if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
 544                 error_sim = true;
 545                 rc = -EINVAL;
 546                 goto simulate_error;
 547         }
 548
 549         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 550                 ksocknal_check_zc_req(tx);
 551
 552         rc = ksocknal_transmit(conn, tx, scratch_iov);
 553
 554         CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 555
 556         if (tx->tx_resid == 0) {
 557                 /* Sent everything OK */
 558                 LASSERT(rc == 0);
 559
 560                 return 0;
 561         }
 562
 563         if (rc == -EAGAIN)
 564                 return rc;
 565
 566         if (rc == -ENOMEM) {
 567                 static int counter;
 568
 569                 counter++;   /* exponential backoff warnings */
 570                 if ((counter & (-counter)) == counter)
 571                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 572                               counter, conn, atomic_read(&libcfs_kmemory));
 573
 574                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 575                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 576
 577                 /* enomem list takes over scheduler's ref... */
 578                 LASSERT(conn->ksnc_tx_scheduled);
 579                 list_add_tail(&conn->ksnc_tx_list,
 580                                   &ksocknal_data.ksnd_enomem_conns);
 581                 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
 582                     ksocknal_data.ksnd_reaper_waketime)
 583                         wake_up(&ksocknal_data.ksnd_reaper_waitq);
 584
 585                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 586
 587                 /*
 588                  * set the health status of the message which determines
 589                  * whether we should retry the transmit
 590                  */
 591                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 592                 return (rc);
 593         }
 594
 595 simulate_error:
 596
 597         /* Actual error */
 598         LASSERT(rc < 0);
 599
 600         if (!error_sim) {
 601                 /*
 602                 * set the health status of the message which determines
 603                 * whether we should retry the transmit
 604                 */
 605                 if (rc == -ETIMEDOUT)
 606                         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
 607                 else
 608                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 609         }
 610
 611         if (!conn->ksnc_closing) {
 612                 switch (rc) {
 613                 case -ECONNRESET:
 614                         LCONSOLE_WARN("Host %pI4h reset our connection "
 615                                       "while we were sending data; it may have "
 616                                       "rebooted.\n",
 617                                       &conn->ksnc_ipaddr);
 618                         break;
 619                 default:
 620                         LCONSOLE_WARN("There was an unexpected network error "
 621                                       "while writing to %pI4h: %d.\n",
 622                                       &conn->ksnc_ipaddr, rc);
 623                         break;
 624                 }
 625                 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
 626                        conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id),
 627                        &conn->ksnc_ipaddr, conn->ksnc_port);
 628         }
 629
 630         if (tx->tx_zc_checked)
 631                 ksocknal_uncheck_zc_req(tx);
 632
 633         /* it's not an error if conn is being closed */
 634         ksocknal_close_conn_and_siblings(conn,
 635                                           (conn->ksnc_closing) ? 0 : rc);
 636
 637         return rc;
 638 }
 639
 640 static void
 641 ksocknal_launch_connection_locked(struct ksock_route *route)
 642 {
 643
 644         /* called holding write lock on ksnd_global_lock */
 645
 646         LASSERT (!route->ksnr_scheduled);
 647         LASSERT (!route->ksnr_connecting);
 648         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 649
 650         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 651         ksocknal_route_addref(route);           /* extra ref for connd */
 652
 653         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 654
 655         list_add_tail(&route->ksnr_connd_list,
 656                           &ksocknal_data.ksnd_connd_routes);
 657         wake_up(&ksocknal_data.ksnd_connd_waitq);
 658
 659         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 660 }
 661
 662 void
 663 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
 664 {
 665         struct ksock_route *route;
 666
 667         /* called holding write lock on ksnd_global_lock */
 668         for (;;) {
 669                 /* launch any/all connections that need it */
 670                 route = ksocknal_find_connectable_route_locked(peer_ni);
 671                 if (route == NULL)
 672                         return;
 673
 674                 ksocknal_launch_connection_locked(route);
 675         }
 676 }
 677
 678 struct ksock_conn *
 679 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
 680 {
 681         struct list_head *tmp;
 682         struct ksock_conn *conn;
 683         struct ksock_conn *typed = NULL;
 684         struct ksock_conn *fallback = NULL;
 685         int tnob = 0;
 686         int fnob = 0;
 687
 688         list_for_each(tmp, &peer_ni->ksnp_conns) {
 689                 struct ksock_conn *c = list_entry(tmp, struct ksock_conn,
 690                                                   ksnc_list);
 691                 int nob = atomic_read(&c->ksnc_tx_nob) +
 692                           c->ksnc_sock->sk->sk_wmem_queued;
 693                 int rc;
 694
 695                 LASSERT (!c->ksnc_closing);
 696                 LASSERT (c->ksnc_proto != NULL &&
 697                          c->ksnc_proto->pro_match_tx != NULL);
 698
 699                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 700
 701                 switch (rc) {
 702                 default:
 703                         LBUG();
 704                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 705                         continue;
 706
 707                 case SOCKNAL_MATCH_YES: /* typed connection */
 708                         if (typed == NULL || tnob > nob ||
 709                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 710                              typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 711                                 typed = c;
 712                                 tnob  = nob;
 713                         }
 714                         break;
 715
 716                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 717                         if (fallback == NULL || fnob > nob ||
 718                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 719                              fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 720                                 fallback = c;
 721                                 fnob     = nob;
 722                         }
 723                         break;
 724                 }
 725         }
 726
 727         /* prefer the typed selection */
 728         conn = (typed != NULL) ? typed : fallback;
 729
 730         if (conn != NULL)
 731                 conn->ksnc_tx_last_post = ktime_get_seconds();
 732
 733         return conn;
 734 }
 735
 736 void
 737 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
 738 {
 739         conn->ksnc_proto->pro_pack(tx);
 740
 741         atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 742         ksocknal_conn_addref(conn); /* +1 ref for tx */
 743         tx->tx_conn = conn;
 744 }
 745
 746 void
 747 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
 748 {
 749         struct ksock_sched *sched = conn->ksnc_scheduler;
 750         struct ksock_msg *msg = &tx->tx_msg;
 751         struct ksock_tx *ztx = NULL;
 752         int bufnob = 0;
 753
 754         /* called holding global lock (read or irq-write) and caller may
 755          * not have dropped this lock between finding conn and calling me,
 756          * so we don't need the {get,put}connsock dance to deref
 757          * ksnc_sock... */
 758         LASSERT(!conn->ksnc_closing);
 759
 760         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 761                libcfs_id2str(conn->ksnc_peer->ksnp_id),
 762                &conn->ksnc_ipaddr, conn->ksnc_port);
 763
 764         ksocknal_tx_prep(conn, tx);
 765
 766         /* Ensure the frags we've been given EXACTLY match the number of
 767          * bytes we want to send.  Many TCP/IP stacks disregard any total
 768          * size parameters passed to them and just look at the frags.
 769          *
 770          * We always expect at least 1 mapped fragment containing the
 771          * complete ksocknal message header. */
 772         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 773                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 774                  (unsigned int)tx->tx_nob);
 775         LASSERT (tx->tx_niov >= 1);
 776         LASSERT (tx->tx_resid == tx->tx_nob);
 777
 778         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 779                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 780                                                KSOCK_MSG_NOOP,
 781                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 782
 783         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 784         spin_lock_bh(&sched->kss_lock);
 785
 786         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 787                 /* First packet starts the timeout */
 788                 conn->ksnc_tx_deadline = ktime_get_seconds() +
 789                                          lnet_get_lnd_timeout();
 790                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 791                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 792                 conn->ksnc_tx_bufnob = 0;
 793                 smp_mb(); /* order with adding to tx_queue */
 794         }
 795
 796         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 797                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 798                  * on a normal packet so I don't need to send it */
 799                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 800                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 801
 802                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 803                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 804
 805         } else {
 806                 /* It's a normal packet - can it piggback a noop zc-ack that
 807                  * has been queued already? */
 808                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 809                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 810
 811                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 812                 /* ztx will be released later */
 813         }
 814
 815         if (ztx != NULL) {
 816                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 817                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 818         }
 819
 820         if (conn->ksnc_tx_ready &&      /* able to send */
 821             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 822                 /* +1 ref for scheduler */
 823                 ksocknal_conn_addref(conn);
 824                 list_add_tail(&conn->ksnc_tx_list,
 825                                    &sched->kss_tx_conns);
 826                 conn->ksnc_tx_scheduled = 1;
 827                 wake_up(&sched->kss_waitq);
 828         }
 829
 830         spin_unlock_bh(&sched->kss_lock);
 831 }
 832
 833
 834 struct ksock_route *
 835 ksocknal_find_connectable_route_locked(struct ksock_peer_ni *peer_ni)
 836 {
 837         time64_t now = ktime_get_seconds();
 838         struct list_head *tmp;
 839         struct ksock_route *route;
 840
 841         list_for_each(tmp, &peer_ni->ksnp_routes) {
 842                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 843
 844                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 845
 846                 if (route->ksnr_scheduled)      /* connections being established */
 847                         continue;
 848
 849                 /* all route types connected ? */
 850                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 851                         continue;
 852
 853                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 854                       now >= route->ksnr_timeout)) {
 855                         CDEBUG(D_NET,
 856                                "Too soon to retry route %pI4h "
 857                                "(cnted %d, interval %lld, %lld secs later)\n",
 858                                &route->ksnr_ipaddr,
 859                                route->ksnr_connected,
 860                                route->ksnr_retry_interval,
 861                                route->ksnr_timeout - now);
 862                         continue;
 863                 }
 864
 865                 return (route);
 866         }
 867
 868         return (NULL);
 869 }
 870
 871 struct ksock_route *
 872 ksocknal_find_connecting_route_locked(struct ksock_peer_ni *peer_ni)
 873 {
 874         struct list_head *tmp;
 875         struct ksock_route *route;
 876
 877         list_for_each(tmp, &peer_ni->ksnp_routes) {
 878                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 879
 880                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 881
 882                 if (route->ksnr_scheduled)
 883                         return (route);
 884         }
 885
 886         return (NULL);
 887 }
 888
 889 int
 890 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
 891                        struct lnet_process_id id)
 892 {
 893         struct ksock_peer_ni *peer_ni;
 894         struct ksock_conn *conn;
 895         rwlock_t *g_lock;
 896         int retry;
 897         int rc;
 898
 899         LASSERT (tx->tx_conn == NULL);
 900
 901         g_lock = &ksocknal_data.ksnd_global_lock;
 902
 903         for (retry = 0;; retry = 1) {
 904                 read_lock(g_lock);
 905                 peer_ni = ksocknal_find_peer_locked(ni, id);
 906                 if (peer_ni != NULL) {
 907                         if (ksocknal_find_connectable_route_locked(peer_ni) == NULL) {
 908                                 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 909                                 if (conn != NULL) {
 910                                         /* I've got no routes that need to be
 911                                          * connecting and I do have an actual
 912                                          * connection... */
 913                                         ksocknal_queue_tx_locked (tx, conn);
 914                                         read_unlock(g_lock);
 915                                         return (0);
 916                                 }
 917                         }
 918                 }
 919
 920                 /* I'll need a write lock... */
 921                 read_unlock(g_lock);
 922
 923                 write_lock_bh(g_lock);
 924
 925                 peer_ni = ksocknal_find_peer_locked(ni, id);
 926                 if (peer_ni != NULL)
 927                         break;
 928
 929                 write_unlock_bh(g_lock);
 930
 931                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 932                         CERROR("Refusing to create a connection to "
 933                                "userspace process %s\n", libcfs_id2str(id));
 934                         return -EHOSTUNREACH;
 935                 }
 936
 937                 if (retry) {
 938                         CERROR("Can't find peer_ni %s\n", libcfs_id2str(id));
 939                         return -EHOSTUNREACH;
 940                 }
 941
 942                 rc = ksocknal_add_peer(ni, id,
 943                                        LNET_NIDADDR(id.nid),
 944                                        lnet_acceptor_port());
 945                 if (rc != 0) {
 946                         CERROR("Can't add peer_ni %s: %d\n",
 947                                libcfs_id2str(id), rc);
 948                         return rc;
 949                 }
 950         }
 951
 952         ksocknal_launch_all_connections_locked(peer_ni);
 953
 954         conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 955         if (conn != NULL) {
 956                 /* Connection exists; queue message on it */
 957                 ksocknal_queue_tx_locked (tx, conn);
 958                 write_unlock_bh(g_lock);
 959                 return (0);
 960         }
 961
 962         if (peer_ni->ksnp_accepting > 0 ||
 963             ksocknal_find_connecting_route_locked (peer_ni) != NULL) {
 964                 /* the message is going to be pinned to the peer_ni */
 965                 tx->tx_deadline = ktime_get_seconds() +
 966                                   lnet_get_lnd_timeout();
 967
 968                 /* Queue the message until a connection is established */
 969                 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
 970                 write_unlock_bh(g_lock);
 971                 return 0;
 972         }
 973
 974         write_unlock_bh(g_lock);
 975
 976         /* NB Routes may be ignored if connections to them failed recently */
 977         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 978         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 979         return (-EHOSTUNREACH);
 980 }
 981
 982 int
 983 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
 984 {
 985         int mpflag = 1;
 986         int type = lntmsg->msg_type;
 987         struct lnet_process_id target = lntmsg->msg_target;
 988         unsigned int      payload_niov = lntmsg->msg_niov;
 989         struct kvec *payload_iov = lntmsg->msg_iov;
 990         lnet_kiov_t      *payload_kiov = lntmsg->msg_kiov;
 991         unsigned int      payload_offset = lntmsg->msg_offset;
 992         unsigned int      payload_nob = lntmsg->msg_len;
 993         struct ksock_tx *tx;
 994         int               desc_size;
 995         int               rc;
 996
 997         /* NB 'private' is different depending on what we're sending.
 998          * Just ignore it... */
 999
1000         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
1001                payload_nob, payload_niov, libcfs_id2str(target));
1002
1003         LASSERT (payload_nob == 0 || payload_niov > 0);
1004         LASSERT (payload_niov <= LNET_MAX_IOV);
1005         /* payload is either all vaddrs or all pages */
1006         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1007         LASSERT (!in_interrupt ());
1008
1009         if (payload_iov != NULL)
1010                 desc_size = offsetof(struct ksock_tx,
1011                                      tx_frags.virt.iov[1 + payload_niov]);
1012         else
1013                 desc_size = offsetof(struct ksock_tx,
1014                                      tx_frags.paged.kiov[payload_niov]);
1015
1016         if (lntmsg->msg_vmflush)
1017                 mpflag = cfs_memory_pressure_get_and_set();
1018         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1019         if (tx == NULL) {
1020                 CERROR("Can't allocate tx desc type %d size %d\n",
1021                        type, desc_size);
1022                 if (lntmsg->msg_vmflush)
1023                         cfs_memory_pressure_restore(mpflag);
1024                 return (-ENOMEM);
1025         }
1026
1027         tx->tx_conn = NULL;                     /* set when assigned a conn */
1028         tx->tx_lnetmsg = lntmsg;
1029
1030         if (payload_iov != NULL) {
1031                 tx->tx_kiov = NULL;
1032                 tx->tx_nkiov = 0;
1033                 tx->tx_iov = tx->tx_frags.virt.iov;
1034                 tx->tx_niov = 1 +
1035                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
1036                                                payload_niov, payload_iov,
1037                                                payload_offset, payload_nob);
1038         } else {
1039                 tx->tx_niov = 1;
1040                 tx->tx_iov = &tx->tx_frags.paged.iov;
1041                 tx->tx_kiov = tx->tx_frags.paged.kiov;
1042                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1043                                                  payload_niov, payload_kiov,
1044                                                  payload_offset, payload_nob);
1045
1046                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1047                         tx->tx_zc_capable = 1;
1048         }
1049
1050         tx->tx_msg.ksm_csum = 0;
1051         tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1052         tx->tx_msg.ksm_zc_cookies[0] = 0;
1053         tx->tx_msg.ksm_zc_cookies[1] = 0;
1054
1055         /* The first fragment will be set later in pro_pack */
1056         rc = ksocknal_launch_packet(ni, tx, target);
1057         if (!mpflag)
1058                 cfs_memory_pressure_restore(mpflag);
1059
1060         if (rc == 0)
1061                 return (0);
1062
1063         lntmsg->msg_health_status = tx->tx_hstatus;
1064         ksocknal_free_tx(tx);
1065         return (-EIO);
1066 }
1067
1068 int
1069 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1070 {
1071         struct task_struct *task = kthread_run(fn, arg, name);
1072
1073         if (IS_ERR(task))
1074                 return PTR_ERR(task);
1075
1076         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1077         ksocknal_data.ksnd_nthreads++;
1078         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1079         return 0;
1080 }
1081
1082 void
1083 ksocknal_thread_fini (void)
1084 {
1085         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1086         ksocknal_data.ksnd_nthreads--;
1087         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1088 }
1089
1090 int
1091 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1092 {
1093         static char ksocknal_slop_buffer[4096];
1094         int nob;
1095         unsigned int niov;
1096         int skipped;
1097
1098         LASSERT(conn->ksnc_proto != NULL);
1099
1100         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1101                 /* Remind the socket to ack eagerly... */
1102                 ksocknal_lib_eager_ack(conn);
1103         }
1104
1105         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1106                 conn->ksnc_rx_started = 0;
1107                 smp_mb();                       /* racing with timeout thread */
1108
1109                 switch (conn->ksnc_proto->pro_version) {
1110                 case  KSOCK_PROTO_V2:
1111                 case  KSOCK_PROTO_V3:
1112                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1113                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1114                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1115
1116                         conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u);
1117                         conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u);
1118                         conn->ksnc_rx_iov[0].iov_len  = offsetof(struct ksock_msg, ksm_u);
1119                         break;
1120
1121                 case KSOCK_PROTO_V1:
1122                         /* Receiving bare struct lnet_hdr */
1123                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1124                         conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr);
1125                         conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr);
1126
1127                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1128                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1129                         conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr);
1130                         break;
1131
1132                 default:
1133                         LBUG ();
1134                 }
1135                 conn->ksnc_rx_niov = 1;
1136
1137                 conn->ksnc_rx_kiov = NULL;
1138                 conn->ksnc_rx_nkiov = 0;
1139                 conn->ksnc_rx_csum = ~0;
1140                 return (1);
1141         }
1142
1143         /* Set up to skip as much as possible now.  If there's more left
1144          * (ran out of iov entries) we'll get called again */
1145
1146         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1147         conn->ksnc_rx_nob_left = nob_to_skip;
1148         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1149         skipped = 0;
1150         niov = 0;
1151
1152         do {
1153                 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1154
1155                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1156                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1157                 niov++;
1158                 skipped += nob;
1159                 nob_to_skip -= nob;
1160
1161         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1162                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1163
1164         conn->ksnc_rx_niov = niov;
1165         conn->ksnc_rx_kiov = NULL;
1166         conn->ksnc_rx_nkiov = 0;
1167         conn->ksnc_rx_nob_wanted = skipped;
1168         return (0);
1169 }
1170
1171 static int
1172 ksocknal_process_receive(struct ksock_conn *conn,
1173                          struct page **rx_scratch_pgs,
1174                          struct kvec *scratch_iov)
1175 {
1176         struct lnet_hdr *lhdr;
1177         struct lnet_process_id *id;
1178         int rc;
1179
1180         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1181
1182         /* NB: sched lock NOT held */
1183         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1184         LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1185                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1186                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1187                 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1188  again:
1189         if (conn->ksnc_rx_nob_wanted != 0) {
1190                 rc = ksocknal_receive(conn, rx_scratch_pgs,
1191                                       scratch_iov);
1192
1193                 if (rc <= 0) {
1194                         struct lnet_process_id ksnp_id;
1195
1196                         ksnp_id = conn->ksnc_peer->ksnp_id;
1197
1198                         LASSERT(rc != -EAGAIN);
1199                         if (rc == 0)
1200                                 CDEBUG(D_NET, "[%p] EOF from %s "
1201                                        "ip %pI4h:%d\n", conn,
1202                                         libcfs_id2str(ksnp_id),
1203                                         &conn->ksnc_ipaddr,
1204                                         conn->ksnc_port);
1205                         else if (!conn->ksnc_closing)
1206                                 CERROR("[%p] Error %d on read from %s "
1207                                        "ip %pI4h:%d\n", conn, rc,
1208                                        libcfs_id2str(ksnp_id),
1209                                        &conn->ksnc_ipaddr,
1210                                        conn->ksnc_port);
1211
1212                         /* it's not an error if conn is being closed */
1213                         ksocknal_close_conn_and_siblings (conn,
1214                                                           (conn->ksnc_closing) ? 0 : rc);
1215                         return (rc == 0 ? -ESHUTDOWN : rc);
1216                 }
1217
1218                 if (conn->ksnc_rx_nob_wanted != 0) {
1219                         /* short read */
1220                         return (-EAGAIN);
1221                 }
1222         }
1223         switch (conn->ksnc_rx_state) {
1224         case SOCKNAL_RX_KSM_HEADER:
1225                 if (conn->ksnc_flip) {
1226                         __swab32s(&conn->ksnc_msg.ksm_type);
1227                         __swab32s(&conn->ksnc_msg.ksm_csum);
1228                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1229                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1230                 }
1231
1232                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1233                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1234                         CERROR("%s: Unknown message type: %x\n",
1235                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1236                                conn->ksnc_msg.ksm_type);
1237                         ksocknal_new_packet(conn, 0);
1238                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1239                         return (-EPROTO);
1240                 }
1241
1242                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1243                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1244                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1245                         /* NOOP Checksum error */
1246                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1247                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1248                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1249                         ksocknal_new_packet(conn, 0);
1250                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1251                         return (-EIO);
1252                 }
1253
1254                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1255                         __u64 cookie = 0;
1256
1257                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1258
1259                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1260                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1261
1262                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1263                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1264
1265                         if (rc != 0) {
1266                                 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1267                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1268                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1269                                 ksocknal_new_packet(conn, 0);
1270                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1271                                 return (rc);
1272                         }
1273                 }
1274
1275                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1276                         ksocknal_new_packet (conn, 0);
1277                         return 0;       /* NOOP is done and just return */
1278                 }
1279
1280                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1281                 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg);
1282                 conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg);
1283
1284                 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1285                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1286                 conn->ksnc_rx_iov[0].iov_len  = sizeof(struct ksock_lnet_msg);
1287
1288                 conn->ksnc_rx_niov = 1;
1289                 conn->ksnc_rx_kiov = NULL;
1290                 conn->ksnc_rx_nkiov = 0;
1291
1292                 goto again;     /* read lnet header now */
1293
1294         case SOCKNAL_RX_LNET_HEADER:
1295                 /* unpack message header */
1296                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1297
1298                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1299                         /* Userspace peer_ni */
1300                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1301                         id   = &conn->ksnc_peer->ksnp_id;
1302
1303                         /* Substitute process ID assigned at connection time */
1304                         lhdr->src_pid = cpu_to_le32(id->pid);
1305                         lhdr->src_nid = cpu_to_le64(id->nid);
1306                 }
1307
1308                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1309                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1310
1311                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1312                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1313                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1314                 if (rc < 0) {
1315                         /* I just received garbage: give up on this conn */
1316                         ksocknal_new_packet(conn, 0);
1317                         ksocknal_close_conn_and_siblings (conn, rc);
1318                         ksocknal_conn_decref(conn);
1319                         return (-EPROTO);
1320                 }
1321
1322                 /* I'm racing with ksocknal_recv() */
1323                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1324                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1325
1326                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1327                         return 0;
1328
1329                 /* ksocknal_recv() got called */
1330                 goto again;
1331
1332         case SOCKNAL_RX_LNET_PAYLOAD:
1333                 /* payload all received */
1334                 rc = 0;
1335
1336                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1337                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1338                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1339                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1340                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1341                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1342                         rc = -EIO;
1343                 }
1344
1345                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1346                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1347
1348                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1349                         id   = &conn->ksnc_peer->ksnp_id;
1350
1351                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1352                                         conn->ksnc_msg.ksm_zc_cookies[0],
1353                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1354                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1355                 }
1356
1357                 if (rc && conn->ksnc_lnet_msg)
1358                         conn->ksnc_lnet_msg->msg_health_status =
1359                                 LNET_MSG_STATUS_REMOTE_ERROR;
1360                 lnet_finalize(conn->ksnc_lnet_msg, rc);
1361
1362                 if (rc != 0) {
1363                         ksocknal_new_packet(conn, 0);
1364                         ksocknal_close_conn_and_siblings (conn, rc);
1365                         return (-EPROTO);
1366                 }
1367                 /* Fall through */
1368
1369         case SOCKNAL_RX_SLOP:
1370                 /* starting new packet? */
1371                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1372                         return 0;       /* come back later */
1373                 goto again;             /* try to finish reading slop now */
1374
1375         default:
1376                 break;
1377         }
1378
1379         /* Not Reached */
1380         LBUG ();
1381         return (-EINVAL);                       /* keep gcc happy */
1382 }
1383
1384 int
1385 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1386               int delayed, unsigned int niov, struct kvec *iov,
1387               lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,
1388               unsigned int rlen)
1389 {
1390         struct ksock_conn *conn = private;
1391         struct ksock_sched *sched = conn->ksnc_scheduler;
1392
1393         LASSERT (mlen <= rlen);
1394         LASSERT (niov <= LNET_MAX_IOV);
1395
1396         conn->ksnc_lnet_msg = msg;
1397         conn->ksnc_rx_nob_wanted = mlen;
1398         conn->ksnc_rx_nob_left   = rlen;
1399
1400         if (mlen == 0 || iov != NULL) {
1401                 conn->ksnc_rx_nkiov = 0;
1402                 conn->ksnc_rx_kiov = NULL;
1403                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1404                 conn->ksnc_rx_niov =
1405                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1406                                          niov, iov, offset, mlen);
1407         } else {
1408                 conn->ksnc_rx_niov = 0;
1409                 conn->ksnc_rx_iov  = NULL;
1410                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1411                 conn->ksnc_rx_nkiov =
1412                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1413                                           niov, kiov, offset, mlen);
1414         }
1415
1416         LASSERT (mlen ==
1417                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1418                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1419
1420         LASSERT (conn->ksnc_rx_scheduled);
1421
1422         spin_lock_bh(&sched->kss_lock);
1423
1424         switch (conn->ksnc_rx_state) {
1425         case SOCKNAL_RX_PARSE_WAIT:
1426                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1427                 wake_up(&sched->kss_waitq);
1428                 LASSERT(conn->ksnc_rx_ready);
1429                 break;
1430
1431         case SOCKNAL_RX_PARSE:
1432                 /* scheduler hasn't noticed I'm parsing yet */
1433                 break;
1434         }
1435
1436         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1437
1438         spin_unlock_bh(&sched->kss_lock);
1439         ksocknal_conn_decref(conn);
1440         return 0;
1441 }
1442
1443 static inline int
1444 ksocknal_sched_cansleep(struct ksock_sched *sched)
1445 {
1446         int           rc;
1447
1448         spin_lock_bh(&sched->kss_lock);
1449
1450         rc = (!ksocknal_data.ksnd_shuttingdown &&
1451               list_empty(&sched->kss_rx_conns) &&
1452               list_empty(&sched->kss_tx_conns));
1453
1454         spin_unlock_bh(&sched->kss_lock);
1455         return rc;
1456 }
1457
1458 int ksocknal_scheduler(void *arg)
1459 {
1460         struct ksock_sched *sched;
1461         struct ksock_conn *conn;
1462         struct ksock_tx *tx;
1463         int rc;
1464         int nloops = 0;
1465         long id = (long)arg;
1466         struct page **rx_scratch_pgs;
1467         struct kvec *scratch_iov;
1468
1469         sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1470
1471         LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1472                          sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1473         if (!rx_scratch_pgs) {
1474                 CERROR("Unable to allocate scratch pages\n");
1475                 return -ENOMEM;
1476         }
1477
1478         LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1479                          sizeof(*scratch_iov) * LNET_MAX_IOV);
1480         if (!scratch_iov) {
1481                 CERROR("Unable to allocate scratch iov\n");
1482                 return -ENOMEM;
1483         }
1484
1485         rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1486         if (rc != 0) {
1487                 CWARN("Can't set CPU partition affinity to %d: %d\n",
1488                         sched->kss_cpt, rc);
1489         }
1490
1491         spin_lock_bh(&sched->kss_lock);
1492
1493         while (!ksocknal_data.ksnd_shuttingdown) {
1494                 int did_something = 0;
1495
1496                 /* Ensure I progress everything semi-fairly */
1497
1498                 if (!list_empty(&sched->kss_rx_conns)) {
1499                         conn = list_entry(sched->kss_rx_conns.next,
1500                                           struct ksock_conn, ksnc_rx_list);
1501                         list_del(&conn->ksnc_rx_list);
1502
1503                         LASSERT(conn->ksnc_rx_scheduled);
1504                         LASSERT(conn->ksnc_rx_ready);
1505
1506                         /* clear rx_ready in case receive isn't complete.
1507                          * Do it BEFORE we call process_recv, since
1508                          * data_ready can set it any time after we release
1509                          * kss_lock. */
1510                         conn->ksnc_rx_ready = 0;
1511                         spin_unlock_bh(&sched->kss_lock);
1512
1513                         rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1514                                                       scratch_iov);
1515
1516                         spin_lock_bh(&sched->kss_lock);
1517
1518                         /* I'm the only one that can clear this flag */
1519                         LASSERT(conn->ksnc_rx_scheduled);
1520
1521                         /* Did process_receive get everything it wanted? */
1522                         if (rc == 0)
1523                                 conn->ksnc_rx_ready = 1;
1524
1525                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1526                                 /* Conn blocked waiting for ksocknal_recv()
1527                                  * I change its state (under lock) to signal
1528                                  * it can be rescheduled */
1529                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1530                         } else if (conn->ksnc_rx_ready) {
1531                                 /* reschedule for rx */
1532                                 list_add_tail(&conn->ksnc_rx_list,
1533                                                    &sched->kss_rx_conns);
1534                         } else {
1535                                 conn->ksnc_rx_scheduled = 0;
1536                                 /* drop my ref */
1537                                 ksocknal_conn_decref(conn);
1538                         }
1539
1540                         did_something = 1;
1541                 }
1542
1543                 if (!list_empty(&sched->kss_tx_conns)) {
1544                         LIST_HEAD(zlist);
1545
1546                         list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1547
1548                         conn = list_entry(sched->kss_tx_conns.next,
1549                                           struct ksock_conn, ksnc_tx_list);
1550                         list_del(&conn->ksnc_tx_list);
1551
1552                         LASSERT(conn->ksnc_tx_scheduled);
1553                         LASSERT(conn->ksnc_tx_ready);
1554                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1555
1556                         tx = list_entry(conn->ksnc_tx_queue.next,
1557                                         struct ksock_tx, tx_list);
1558
1559                         if (conn->ksnc_tx_carrier == tx)
1560                                 ksocknal_next_tx_carrier(conn);
1561
1562                         /* dequeue now so empty list => more to send */
1563                         list_del(&tx->tx_list);
1564
1565                         /* Clear tx_ready in case send isn't complete.  Do
1566                          * it BEFORE we call process_transmit, since
1567                          * write_space can set it any time after we release
1568                          * kss_lock. */
1569                         conn->ksnc_tx_ready = 0;
1570                         spin_unlock_bh(&sched->kss_lock);
1571
1572                         if (!list_empty(&zlist)) {
1573                                 /* free zombie noop txs, it's fast because
1574                                  * noop txs are just put in freelist */
1575                                 ksocknal_txlist_done(NULL, &zlist, 0);
1576                         }
1577
1578                         rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1579
1580                         if (rc == -ENOMEM || rc == -EAGAIN) {
1581                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1582                                 spin_lock_bh(&sched->kss_lock);
1583                                 list_add(&tx->tx_list,
1584                                          &conn->ksnc_tx_queue);
1585                         } else {
1586                                 /* Complete send; tx -ref */
1587                                 ksocknal_tx_decref(tx);
1588
1589                                 spin_lock_bh(&sched->kss_lock);
1590                                 /* assume space for more */
1591                                 conn->ksnc_tx_ready = 1;
1592                         }
1593
1594                         if (rc == -ENOMEM) {
1595                                 /* Do nothing; after a short timeout, this
1596                                  * conn will be reposted on kss_tx_conns. */
1597                         } else if (conn->ksnc_tx_ready &&
1598                                    !list_empty(&conn->ksnc_tx_queue)) {
1599                                 /* reschedule for tx */
1600                                 list_add_tail(&conn->ksnc_tx_list,
1601                                               &sched->kss_tx_conns);
1602                         } else {
1603                                 conn->ksnc_tx_scheduled = 0;
1604                                 /* drop my ref */
1605                                 ksocknal_conn_decref(conn);
1606                         }
1607
1608                         did_something = 1;
1609                 }
1610                 if (!did_something ||           /* nothing to do */
1611                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1612                         spin_unlock_bh(&sched->kss_lock);
1613
1614                         nloops = 0;
1615
1616                         if (!did_something) {   /* wait for something to do */
1617                                 rc = wait_event_interruptible_exclusive(
1618                                         sched->kss_waitq,
1619                                         !ksocknal_sched_cansleep(sched));
1620                                 LASSERT (rc == 0);
1621                         } else {
1622                                 cond_resched();
1623                         }
1624
1625                         spin_lock_bh(&sched->kss_lock);
1626                 }
1627         }
1628
1629         spin_unlock_bh(&sched->kss_lock);
1630         CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV);
1631         CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV);
1632         ksocknal_thread_fini();
1633         return 0;
1634 }
1635
1636 /*
1637  * Add connection to kss_rx_conns of scheduler
1638  * and wakeup the scheduler.
1639  */
1640 void ksocknal_read_callback(struct ksock_conn *conn)
1641 {
1642         struct ksock_sched *sched;
1643         ENTRY;
1644
1645         sched = conn->ksnc_scheduler;
1646
1647         spin_lock_bh(&sched->kss_lock);
1648
1649         conn->ksnc_rx_ready = 1;
1650
1651         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1652                 list_add_tail(&conn->ksnc_rx_list,
1653                                   &sched->kss_rx_conns);
1654                 conn->ksnc_rx_scheduled = 1;
1655                 /* extra ref for scheduler */
1656                 ksocknal_conn_addref(conn);
1657
1658                 wake_up (&sched->kss_waitq);
1659         }
1660         spin_unlock_bh(&sched->kss_lock);
1661
1662         EXIT;
1663 }
1664
1665 /*
1666  * Add connection to kss_tx_conns of scheduler
1667  * and wakeup the scheduler.
1668  */
1669 void ksocknal_write_callback(struct ksock_conn *conn)
1670 {
1671         struct ksock_sched *sched;
1672         ENTRY;
1673
1674         sched = conn->ksnc_scheduler;
1675
1676         spin_lock_bh(&sched->kss_lock);
1677
1678         conn->ksnc_tx_ready = 1;
1679
1680         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1681             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1682                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1683                 conn->ksnc_tx_scheduled = 1;
1684                 /* extra ref for scheduler */
1685                 ksocknal_conn_addref(conn);
1686
1687                 wake_up(&sched->kss_waitq);
1688         }
1689
1690         spin_unlock_bh(&sched->kss_lock);
1691
1692         EXIT;
1693 }
1694
1695 static const struct ksock_proto *
1696 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1697 {
1698         __u32   version = 0;
1699
1700         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1701                 version = hello->kshm_version;
1702         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1703                 version = __swab32(hello->kshm_version);
1704
1705         if (version != 0) {
1706 #if SOCKNAL_VERSION_DEBUG
1707                 if (*ksocknal_tunables.ksnd_protocol == 1)
1708                         return NULL;
1709
1710                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1711                     version == KSOCK_PROTO_V3)
1712                         return NULL;
1713 #endif
1714                 if (version == KSOCK_PROTO_V2)
1715                         return &ksocknal_protocol_v2x;
1716
1717                 if (version == KSOCK_PROTO_V3)
1718                         return &ksocknal_protocol_v3x;
1719
1720                 return NULL;
1721         }
1722
1723         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1724                 struct lnet_magicversion *hmv;
1725
1726                 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1727                          offsetof(struct ksock_hello_msg, kshm_src_nid));
1728
1729                 hmv = (struct lnet_magicversion *)hello;
1730
1731                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1732                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1733                         return &ksocknal_protocol_v1x;
1734         }
1735
1736         return NULL;
1737 }
1738
1739 int
1740 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1741                     lnet_nid_t peer_nid, struct ksock_hello_msg *hello)
1742 {
1743         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1744         struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1745
1746         LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1747
1748         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1749         LASSERT(conn->ksnc_proto != NULL);
1750
1751         hello->kshm_src_nid         = ni->ni_nid;
1752         hello->kshm_dst_nid         = peer_nid;
1753         hello->kshm_src_pid         = the_lnet.ln_pid;
1754
1755         hello->kshm_src_incarnation = net->ksnn_incarnation;
1756         hello->kshm_ctype           = conn->ksnc_type;
1757
1758         return conn->ksnc_proto->pro_send_hello(conn, hello);
1759 }
1760
1761 static int
1762 ksocknal_invert_type(int type)
1763 {
1764         switch (type)
1765         {
1766         case SOCKLND_CONN_ANY:
1767         case SOCKLND_CONN_CONTROL:
1768                 return (type);
1769         case SOCKLND_CONN_BULK_IN:
1770                 return SOCKLND_CONN_BULK_OUT;
1771         case SOCKLND_CONN_BULK_OUT:
1772                 return SOCKLND_CONN_BULK_IN;
1773         default:
1774                 return (SOCKLND_CONN_NONE);
1775         }
1776 }
1777
1778 int
1779 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1780                     struct ksock_hello_msg *hello,
1781                     struct lnet_process_id *peerid,
1782                     __u64 *incarnation)
1783 {
1784         /* Return < 0        fatal error
1785          *        0          success
1786          *        EALREADY   lost connection race
1787          *        EPROTO     protocol version mismatch
1788          */
1789         struct socket        *sock = conn->ksnc_sock;
1790         int                  active = (conn->ksnc_proto != NULL);
1791         int                  timeout;
1792         int                  proto_match;
1793         int                  rc;
1794         const struct ksock_proto *proto;
1795         struct lnet_process_id recv_id;
1796
1797         /* socket type set on active connections - not set on passive */
1798         LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1799
1800         timeout = active ? lnet_get_lnd_timeout() :
1801                             lnet_acceptor_timeout();
1802
1803         rc = lnet_sock_read(sock, &hello->kshm_magic,
1804                             sizeof(hello->kshm_magic), timeout);
1805         if (rc != 0) {
1806                 CERROR("Error %d reading HELLO from %pI4h\n",
1807                        rc, &conn->ksnc_ipaddr);
1808                 LASSERT (rc < 0);
1809                 return rc;
1810         }
1811
1812         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1813             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1814             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1815                 /* Unexpected magic! */
1816                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1817                         "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1818                         LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr);
1819                 return -EPROTO;
1820         }
1821
1822         rc = lnet_sock_read(sock, &hello->kshm_version,
1823                             sizeof(hello->kshm_version), timeout);
1824         if (rc != 0) {
1825                 CERROR("Error %d reading HELLO from %pI4h\n",
1826                        rc, &conn->ksnc_ipaddr);
1827                 LASSERT(rc < 0);
1828                 return rc;
1829         }
1830
1831         proto = ksocknal_parse_proto_version(hello);
1832         if (proto == NULL) {
1833                 if (!active) {
1834                         /* unknown protocol from peer_ni, tell peer_ni my protocol */
1835                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1836 #if SOCKNAL_VERSION_DEBUG
1837                         if (*ksocknal_tunables.ksnd_protocol == 2)
1838                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1839                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1840                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1841 #endif
1842                         hello->kshm_nips = 0;
1843                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1844                 }
1845
1846                 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1847                        conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr);
1848
1849                 return -EPROTO;
1850         }
1851
1852         proto_match = (conn->ksnc_proto == proto);
1853         conn->ksnc_proto = proto;
1854
1855         /* receive the rest of hello message anyway */
1856         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1857         if (rc != 0) {
1858                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1859                        rc, &conn->ksnc_ipaddr);
1860                 LASSERT (rc < 0);
1861                 return rc;
1862         }
1863
1864         *incarnation = hello->kshm_src_incarnation;
1865
1866         if (hello->kshm_src_nid == LNET_NID_ANY) {
1867                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1868                        &conn->ksnc_ipaddr);
1869                 return -EPROTO;
1870         }
1871
1872         if (!active &&
1873             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1874                 /* Userspace NAL assigns peer_ni process ID from socket */
1875                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1876                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1877         } else {
1878                 recv_id.nid = hello->kshm_src_nid;
1879                 recv_id.pid = hello->kshm_src_pid;
1880         }
1881
1882         if (!active) {
1883                 *peerid = recv_id;
1884
1885                 /* peer_ni determines type */
1886                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1887                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1888                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1889                                hello->kshm_ctype, libcfs_id2str(*peerid),
1890                                &conn->ksnc_ipaddr);
1891                         return -EPROTO;
1892                 }
1893                 return 0;
1894         }
1895
1896         if (peerid->pid != recv_id.pid ||
1897             peerid->nid != recv_id.nid) {
1898                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1899                                    " %pI4h, but they claimed they were "
1900                                    "%s; please check your Lustre "
1901                                    "configuration.\n",
1902                                    libcfs_id2str(*peerid),
1903                                    &conn->ksnc_ipaddr,
1904                                    libcfs_id2str(recv_id));
1905                 return -EPROTO;
1906         }
1907
1908         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1909                 /* Possible protocol mismatch or I lost the connection race */
1910                 return proto_match ? EALREADY : EPROTO;
1911         }
1912
1913         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1914                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1915                        conn->ksnc_type, libcfs_id2str(*peerid),
1916                        &conn->ksnc_ipaddr,
1917                        hello->kshm_ctype);
1918                 return -EPROTO;
1919         }
1920         return 0;
1921 }
1922
1923 static int
1924 ksocknal_connect(struct ksock_route *route)
1925 {
1926         LIST_HEAD(zombies);
1927         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
1928         int               type;
1929         int               wanted;
1930         struct socket     *sock;
1931         time64_t deadline;
1932         int               retry_later = 0;
1933         int               rc = 0;
1934
1935         deadline = ktime_get_seconds() + lnet_get_lnd_timeout();
1936
1937         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1938
1939         LASSERT (route->ksnr_scheduled);
1940         LASSERT (!route->ksnr_connecting);
1941
1942         route->ksnr_connecting = 1;
1943
1944         for (;;) {
1945                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1946
1947                 /* stop connecting if peer_ni/route got closed under me, or
1948                  * route got connected while queued */
1949                 if (peer_ni->ksnp_closing || route->ksnr_deleted ||
1950                     wanted == 0) {
1951                         retry_later = 0;
1952                         break;
1953                 }
1954
1955                 /* reschedule if peer_ni is connecting to me */
1956                 if (peer_ni->ksnp_accepting > 0) {
1957                         CDEBUG(D_NET,
1958                                "peer_ni %s(%d) already connecting to me, retry later.\n",
1959                                libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting);
1960                         retry_later = 1;
1961                 }
1962
1963                 if (retry_later) /* needs reschedule */
1964                         break;
1965
1966                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1967                         type = SOCKLND_CONN_ANY;
1968                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1969                         type = SOCKLND_CONN_CONTROL;
1970                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1971                         type = SOCKLND_CONN_BULK_IN;
1972                 } else {
1973                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1974                         type = SOCKLND_CONN_BULK_OUT;
1975                 }
1976
1977                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1978
1979                 if (ktime_get_seconds() >= deadline) {
1980                         rc = -ETIMEDOUT;
1981                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1982                                                    route->ksnr_ipaddr,
1983                                                    route->ksnr_port);
1984                         goto failed;
1985                 }
1986
1987                 rc = lnet_connect(&sock, peer_ni->ksnp_id.nid,
1988                                   route->ksnr_myipaddr,
1989                                   route->ksnr_ipaddr, route->ksnr_port,
1990                                   peer_ni->ksnp_ni->ni_net_ns);
1991                 if (rc != 0)
1992                         goto failed;
1993
1994                 rc = ksocknal_create_conn(peer_ni->ksnp_ni, route, sock, type);
1995                 if (rc < 0) {
1996                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1997                                                    route->ksnr_ipaddr,
1998                                                    route->ksnr_port);
1999                         goto failed;
2000                 }
2001
2002                 /* A +ve RC means I have to retry because I lost the connection
2003                  * race or I have to renegotiate protocol version */
2004                 retry_later = (rc != 0);
2005                 if (retry_later)
2006                         CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
2007                                libcfs_nid2str(peer_ni->ksnp_id.nid));
2008
2009                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2010         }
2011
2012         route->ksnr_scheduled = 0;
2013         route->ksnr_connecting = 0;
2014
2015         if (retry_later) {
2016                 /* re-queue for attention; this frees me up to handle
2017                  * the peer_ni's incoming connection request */
2018
2019                 if (rc == EALREADY ||
2020                     (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2021                         /* We want to introduce a delay before next
2022                          * attempt to connect if we lost conn race,
2023                          * but the race is resolved quickly usually,
2024                          * so min_reconnectms should be good heuristic */
2025                         route->ksnr_retry_interval = *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2026                         route->ksnr_timeout = ktime_get_seconds() +
2027                                               route->ksnr_retry_interval;
2028                 }
2029
2030                 ksocknal_launch_connection_locked(route);
2031         }
2032
2033         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2034         return retry_later;
2035
2036  failed:
2037         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2038
2039         route->ksnr_scheduled = 0;
2040         route->ksnr_connecting = 0;
2041
2042         /* This is a retry rather than a new connection */
2043         route->ksnr_retry_interval *= 2;
2044         route->ksnr_retry_interval =
2045                 max_t(time64_t, route->ksnr_retry_interval,
2046                       *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2047         route->ksnr_retry_interval =
2048                 min_t(time64_t, route->ksnr_retry_interval,
2049                       *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2050
2051         LASSERT(route->ksnr_retry_interval);
2052         route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval;
2053
2054         if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2055             peer_ni->ksnp_accepting == 0 &&
2056             ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
2057                 struct ksock_conn *conn;
2058
2059                 /* ksnp_tx_queue is queued on a conn on successful
2060                  * connection for V1.x and V2.x */
2061                 if (!list_empty(&peer_ni->ksnp_conns)) {
2062                         conn = list_entry(peer_ni->ksnp_conns.next,
2063                                           struct ksock_conn, ksnc_list);
2064                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
2065                 }
2066
2067                 /* take all the blocked packets while I've got the lock and
2068                  * complete below... */
2069                 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2070         }
2071
2072         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2073
2074         ksocknal_peer_failed(peer_ni);
2075         ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2076         return 0;
2077 }
2078
2079 /*
2080  * check whether we need to create more connds.
2081  * It will try to create new thread if it's necessary, @timeout can
2082  * be updated if failed to create, so caller wouldn't keep try while
2083  * running out of resource.
2084  */
2085 static int
2086 ksocknal_connd_check_start(time64_t sec, long *timeout)
2087 {
2088         char name[16];
2089         int rc;
2090         int total = ksocknal_data.ksnd_connd_starting +
2091                     ksocknal_data.ksnd_connd_running;
2092
2093         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2094                 /* still in initializing */
2095                 return 0;
2096         }
2097
2098         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2099             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2100                 /* can't create more connd, or still have enough
2101                  * threads to handle more connecting */
2102                 return 0;
2103         }
2104
2105         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2106                 /* no pending connecting request */
2107                 return 0;
2108         }
2109
2110         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2111                 /* may run out of resource, retry later */
2112                 *timeout = cfs_time_seconds(1);
2113                 return 0;
2114         }
2115
2116         if (ksocknal_data.ksnd_connd_starting > 0) {
2117                 /* serialize starting to avoid flood */
2118                 return 0;
2119         }
2120
2121         ksocknal_data.ksnd_connd_starting_stamp = sec;
2122         ksocknal_data.ksnd_connd_starting++;
2123         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2124
2125         /* NB: total is the next id */
2126         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2127         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2128
2129         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2130         if (rc == 0)
2131                 return 1;
2132
2133         /* we tried ... */
2134         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2135         ksocknal_data.ksnd_connd_starting--;
2136         ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2137
2138         return 1;
2139 }
2140
2141 /*
2142  * check whether current thread can exit, it will return 1 if there are too
2143  * many threads and no creating in past 120 seconds.
2144  * Also, this function may update @timeout to make caller come back
2145  * again to recheck these conditions.
2146  */
2147 static int
2148 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2149 {
2150         int val;
2151
2152         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2153                 /* still in initializing */
2154                 return 0;
2155         }
2156
2157         if (ksocknal_data.ksnd_connd_starting > 0) {
2158                 /* in progress of starting new thread */
2159                 return 0;
2160         }
2161
2162         if (ksocknal_data.ksnd_connd_running <=
2163             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2164                 return 0;
2165         }
2166
2167         /* created thread in past 120 seconds? */
2168         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2169                     SOCKNAL_CONND_TIMEOUT - sec);
2170
2171         *timeout = (val > 0) ? cfs_time_seconds(val) :
2172                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2173         if (val > 0)
2174                 return 0;
2175
2176         /* no creating in past 120 seconds */
2177
2178         return ksocknal_data.ksnd_connd_running >
2179                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2180 }
2181
2182 /* Go through connd_routes queue looking for a route that we can process
2183  * right now, @timeout_p can be updated if we need to come back later */
2184 static struct ksock_route *
2185 ksocknal_connd_get_route_locked(signed long *timeout_p)
2186 {
2187         time64_t now = ktime_get_seconds();
2188         struct ksock_route *route;
2189
2190         /* connd_routes can contain both pending and ordinary routes */
2191         list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2192                                  ksnr_connd_list) {
2193
2194                 if (route->ksnr_retry_interval == 0 ||
2195                     now >= route->ksnr_timeout)
2196                         return route;
2197
2198                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2199                     *timeout_p > cfs_time_seconds(route->ksnr_timeout - now))
2200                         *timeout_p = cfs_time_seconds(route->ksnr_timeout - now);
2201         }
2202
2203         return NULL;
2204 }
2205
2206 int
2207 ksocknal_connd(void *arg)
2208 {
2209         spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2210         struct ksock_connreq *cr;
2211         wait_queue_entry_t wait;
2212         int nloops = 0;
2213         int cons_retry = 0;
2214
2215         init_waitqueue_entry(&wait, current);
2216
2217         spin_lock_bh(connd_lock);
2218
2219         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2220         ksocknal_data.ksnd_connd_starting--;
2221         ksocknal_data.ksnd_connd_running++;
2222
2223         while (!ksocknal_data.ksnd_shuttingdown) {
2224                 struct ksock_route *route = NULL;
2225                 time64_t sec = ktime_get_real_seconds();
2226                 long timeout = MAX_SCHEDULE_TIMEOUT;
2227                 int  dropped_lock = 0;
2228
2229                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2230                         /* wakeup another one to check stop */
2231                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2232                         break;
2233                 }
2234
2235                 if (ksocknal_connd_check_start(sec, &timeout)) {
2236                         /* created new thread */
2237                         dropped_lock = 1;
2238                 }
2239
2240                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2241                         /* Connection accepted by the listener */
2242                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2243                                         struct ksock_connreq, ksncr_list);
2244
2245                         list_del(&cr->ksncr_list);
2246                         spin_unlock_bh(connd_lock);
2247                         dropped_lock = 1;
2248
2249                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2250                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2251                         lnet_ni_decref(cr->ksncr_ni);
2252                         LIBCFS_FREE(cr, sizeof(*cr));
2253
2254                         spin_lock_bh(connd_lock);
2255                 }
2256
2257                 /* Only handle an outgoing connection request if there
2258                  * is a thread left to handle incoming connections and
2259                  * create new connd */
2260                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2261                     ksocknal_data.ksnd_connd_running) {
2262                         route = ksocknal_connd_get_route_locked(&timeout);
2263                 }
2264                 if (route != NULL) {
2265                         list_del(&route->ksnr_connd_list);
2266                         ksocknal_data.ksnd_connd_connecting++;
2267                         spin_unlock_bh(connd_lock);
2268                         dropped_lock = 1;
2269
2270                         if (ksocknal_connect(route)) {
2271                                 /* consecutive retry */
2272                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2273                                         CWARN("massive consecutive "
2274                                               "re-connecting to %pI4h\n",
2275                                               &route->ksnr_ipaddr);
2276                                         cons_retry = 0;
2277                                 }
2278                         } else {
2279                                 cons_retry = 0;
2280                         }
2281
2282                         ksocknal_route_decref(route);
2283
2284                         spin_lock_bh(connd_lock);
2285                         ksocknal_data.ksnd_connd_connecting--;
2286                 }
2287
2288                 if (dropped_lock) {
2289                         if (++nloops < SOCKNAL_RESCHED)
2290                                 continue;
2291                         spin_unlock_bh(connd_lock);
2292                         nloops = 0;
2293                         cond_resched();
2294                         spin_lock_bh(connd_lock);
2295                         continue;
2296                 }
2297
2298                 /* Nothing to do for 'timeout'  */
2299                 set_current_state(TASK_INTERRUPTIBLE);
2300                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2301                 spin_unlock_bh(connd_lock);
2302
2303                 nloops = 0;
2304                 schedule_timeout(timeout);
2305
2306                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2307                 spin_lock_bh(connd_lock);
2308         }
2309         ksocknal_data.ksnd_connd_running--;
2310         spin_unlock_bh(connd_lock);
2311
2312         ksocknal_thread_fini();
2313         return 0;
2314 }
2315
2316 static struct ksock_conn *
2317 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2318 {
2319         /* We're called with a shared lock on ksnd_global_lock */
2320         struct ksock_conn *conn;
2321         struct list_head *ctmp;
2322         struct ksock_tx *tx;
2323
2324         list_for_each(ctmp, &peer_ni->ksnp_conns) {
2325                 int error;
2326
2327                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2328
2329                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2330                 LASSERT (!conn->ksnc_closing);
2331
2332                 error = conn->ksnc_sock->sk->sk_err;
2333                 if (error != 0) {
2334                         ksocknal_conn_addref(conn);
2335
2336                         switch (error) {
2337                         case ECONNRESET:
2338                                 CNETERR("A connection with %s "
2339                                         "(%pI4h:%d) was reset; "
2340                                         "it may have rebooted.\n",
2341                                         libcfs_id2str(peer_ni->ksnp_id),
2342                                         &conn->ksnc_ipaddr,
2343                                         conn->ksnc_port);
2344                                 break;
2345                         case ETIMEDOUT:
2346                                 CNETERR("A connection with %s "
2347                                         "(%pI4h:%d) timed out; the "
2348                                         "network or node may be down.\n",
2349                                         libcfs_id2str(peer_ni->ksnp_id),
2350                                         &conn->ksnc_ipaddr,
2351                                         conn->ksnc_port);
2352                                 break;
2353                         default:
2354                                 CNETERR("An unexpected network error %d "
2355                                         "occurred with %s "
2356                                         "(%pI4h:%d\n", error,
2357                                         libcfs_id2str(peer_ni->ksnp_id),
2358                                         &conn->ksnc_ipaddr,
2359                                         conn->ksnc_port);
2360                                 break;
2361                         }
2362
2363                         return (conn);
2364                 }
2365
2366                 if (conn->ksnc_rx_started &&
2367                     ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2368                         /* Timed out incomplete incoming message */
2369                         ksocknal_conn_addref(conn);
2370                         CNETERR("Timeout receiving from %s (%pI4h:%d), "
2371                                 "state %d wanted %d left %d\n",
2372                                 libcfs_id2str(peer_ni->ksnp_id),
2373                                 &conn->ksnc_ipaddr,
2374                                 conn->ksnc_port,
2375                                 conn->ksnc_rx_state,
2376                                 conn->ksnc_rx_nob_wanted,
2377                                 conn->ksnc_rx_nob_left);
2378                         return (conn);
2379                 }
2380
2381                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2382                      conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2383                     ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2384                         /* Timed out messages queued for sending or
2385                          * buffered in the socket's send buffer */
2386                         ksocknal_conn_addref(conn);
2387                         list_for_each_entry(tx, &conn->ksnc_tx_queue,
2388                                             tx_list)
2389                                 tx->tx_hstatus =
2390                                         LNET_MSG_STATUS_LOCAL_TIMEOUT;
2391                         CNETERR("Timeout sending data to %s (%pI4h:%d) "
2392                                 "the network or that node may be down.\n",
2393                                 libcfs_id2str(peer_ni->ksnp_id),
2394                                 &conn->ksnc_ipaddr, conn->ksnc_port);
2395                         return (conn);
2396                 }
2397         }
2398
2399         return (NULL);
2400 }
2401
2402 static inline void
2403 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2404 {
2405         struct ksock_tx *tx;
2406         LIST_HEAD(stale_txs);
2407
2408         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2409
2410         while (!list_empty(&peer_ni->ksnp_tx_queue)) {
2411                 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2412                                 struct ksock_tx, tx_list);
2413
2414                 if (ktime_get_seconds() < tx->tx_deadline)
2415                         break;
2416
2417                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2418
2419                 list_move_tail(&tx->tx_list, &stale_txs);
2420         }
2421
2422         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2423
2424         ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2425 }
2426
2427 static int
2428 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2429 __must_hold(&ksocknal_data.ksnd_global_lock)
2430 {
2431         struct ksock_sched *sched;
2432         struct ksock_conn *conn;
2433         struct ksock_tx *tx;
2434
2435         /* last_alive will be updated by create_conn */
2436         if (list_empty(&peer_ni->ksnp_conns))
2437                 return 0;
2438
2439         if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x)
2440                 return 0;
2441
2442         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2443             ktime_get_seconds() < peer_ni->ksnp_last_alive +
2444                                   *ksocknal_tunables.ksnd_keepalive)
2445                 return 0;
2446
2447         if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2448                 return 0;
2449
2450         /* retry 10 secs later, so we wouldn't put pressure
2451          * on this peer_ni if we failed to send keepalive this time */
2452         peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2453
2454         conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2455         if (conn != NULL) {
2456                 sched = conn->ksnc_scheduler;
2457
2458                 spin_lock_bh(&sched->kss_lock);
2459                 if (!list_empty(&conn->ksnc_tx_queue)) {
2460                         spin_unlock_bh(&sched->kss_lock);
2461                         /* there is an queued ACK, don't need keepalive */
2462                         return 0;
2463                 }
2464
2465                 spin_unlock_bh(&sched->kss_lock);
2466         }
2467
2468         read_unlock(&ksocknal_data.ksnd_global_lock);
2469
2470         /* cookie = 1 is reserved for keepalive PING */
2471         tx = ksocknal_alloc_tx_noop(1, 1);
2472         if (tx == NULL) {
2473                 read_lock(&ksocknal_data.ksnd_global_lock);
2474                 return -ENOMEM;
2475         }
2476
2477         if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) {
2478                 read_lock(&ksocknal_data.ksnd_global_lock);
2479                 return 1;
2480         }
2481
2482         ksocknal_free_tx(tx);
2483         read_lock(&ksocknal_data.ksnd_global_lock);
2484
2485         return -EIO;
2486 }
2487
2488
2489 static void
2490 ksocknal_check_peer_timeouts(int idx)
2491 {
2492         struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2493         struct ksock_peer_ni *peer_ni;
2494         struct ksock_conn *conn;
2495         struct ksock_tx *tx;
2496
2497  again:
2498         /* NB. We expect to have a look at all the peers and not find any
2499          * connections to time out, so we just use a shared lock while we
2500          * take a look...
2501          */
2502         read_lock(&ksocknal_data.ksnd_global_lock);
2503
2504         hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2505                 struct ksock_tx *tx_stale;
2506                 time64_t deadline = 0;
2507                 int resid = 0;
2508                 int n = 0;
2509
2510                 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2511                         read_unlock(&ksocknal_data.ksnd_global_lock);
2512                         goto again;
2513                 }
2514
2515                 conn = ksocknal_find_timed_out_conn(peer_ni);
2516
2517                 if (conn != NULL) {
2518                         read_unlock(&ksocknal_data.ksnd_global_lock);
2519
2520                         ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2521
2522                         /* NB we won't find this one again, but we can't
2523                          * just proceed with the next peer_ni, since we dropped
2524                          * ksnd_global_lock and it might be dead already!
2525                          */
2526                         ksocknal_conn_decref(conn);
2527                         goto again;
2528                 }
2529
2530                 /* we can't process stale txs right here because we're
2531                  * holding only shared lock
2532                  */
2533                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
2534                         struct ksock_tx *tx;
2535
2536                         tx = list_entry(peer_ni->ksnp_tx_queue.next,
2537                                         struct ksock_tx, tx_list);
2538                         if (ktime_get_seconds() >= tx->tx_deadline) {
2539                                 ksocknal_peer_addref(peer_ni);
2540                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2541
2542                                 ksocknal_flush_stale_txs(peer_ni);
2543
2544                                 ksocknal_peer_decref(peer_ni);
2545                                 goto again;
2546                         }
2547                 }
2548
2549                 if (list_empty(&peer_ni->ksnp_zc_req_list))
2550                         continue;
2551
2552                 tx_stale = NULL;
2553                 spin_lock(&peer_ni->ksnp_lock);
2554                 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2555                         if (ktime_get_seconds() < tx->tx_deadline)
2556                                 break;
2557                         /* ignore the TX if connection is being closed */
2558                         if (tx->tx_conn->ksnc_closing)
2559                                 continue;
2560                         n++;
2561                         if (tx_stale == NULL)
2562                                 tx_stale = tx;
2563                 }
2564
2565                 if (tx_stale == NULL) {
2566                         spin_unlock(&peer_ni->ksnp_lock);
2567                         continue;
2568                 }
2569
2570                 deadline = tx_stale->tx_deadline;
2571                 resid    = tx_stale->tx_resid;
2572                 conn     = tx_stale->tx_conn;
2573                 ksocknal_conn_addref(conn);
2574
2575                 spin_unlock(&peer_ni->ksnp_lock);
2576                 read_unlock(&ksocknal_data.ksnd_global_lock);
2577
2578                 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2579                        "oldest(%p) timed out %lld secs ago, "
2580                        "resid: %d, wmem: %d\n",
2581                        n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale,
2582                        ktime_get_seconds() - deadline,
2583                        resid, conn->ksnc_sock->sk->sk_wmem_queued);
2584
2585                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2586                 ksocknal_conn_decref(conn);
2587                 goto again;
2588         }
2589
2590         read_unlock(&ksocknal_data.ksnd_global_lock);
2591 }
2592
2593 int ksocknal_reaper(void *arg)
2594 {
2595         wait_queue_entry_t wait;
2596         struct ksock_conn *conn;
2597         struct ksock_sched *sched;
2598         LIST_HEAD(enomem_conns);
2599         int nenomem_conns;
2600         time64_t timeout;
2601         int i;
2602         int peer_index = 0;
2603         time64_t deadline = ktime_get_seconds();
2604
2605         init_waitqueue_entry(&wait, current);
2606
2607         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2608
2609         while (!ksocknal_data.ksnd_shuttingdown) {
2610                 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2611                         conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2612                                           struct ksock_conn, ksnc_list);
2613                         list_del(&conn->ksnc_list);
2614
2615                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2616
2617                         ksocknal_terminate_conn(conn);
2618                         ksocknal_conn_decref(conn);
2619
2620                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2621                         continue;
2622                 }
2623
2624                 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2625                         conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2626                                           struct ksock_conn, ksnc_list);
2627                         list_del(&conn->ksnc_list);
2628
2629                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2630
2631                         ksocknal_destroy_conn(conn);
2632
2633                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2634                         continue;
2635                 }
2636
2637                 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2638                                  &enomem_conns);
2639
2640                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2641
2642                 /* reschedule all the connections that stalled with ENOMEM... */
2643                 nenomem_conns = 0;
2644                 while (!list_empty(&enomem_conns)) {
2645                         conn = list_entry(enomem_conns.next,
2646                                           struct ksock_conn, ksnc_tx_list);
2647                         list_del(&conn->ksnc_tx_list);
2648
2649                         sched = conn->ksnc_scheduler;
2650
2651                         spin_lock_bh(&sched->kss_lock);
2652
2653                         LASSERT(conn->ksnc_tx_scheduled);
2654                         conn->ksnc_tx_ready = 1;
2655                         list_add_tail(&conn->ksnc_tx_list,
2656                                           &sched->kss_tx_conns);
2657                         wake_up(&sched->kss_waitq);
2658
2659                         spin_unlock_bh(&sched->kss_lock);
2660                         nenomem_conns++;
2661                 }
2662
2663                 /* careful with the jiffy wrap... */
2664                 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2665                         const int n = 4;
2666                         const int p = 1;
2667                         int  chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2668                         unsigned int lnd_timeout;
2669
2670                         /* Time to check for timeouts on a few more peers: I
2671                          * do checks every 'p' seconds on a proportion of the
2672                          * peer_ni table and I need to check every connection
2673                          * 'n' times within a timeout interval, to ensure I
2674                          * detect a timeout on any connection within (n+1)/n
2675                          * times the timeout interval.
2676                          */
2677
2678                         lnd_timeout = lnet_get_lnd_timeout();
2679                         if (lnd_timeout > n * p)
2680                                 chunk = (chunk * n * p) / lnd_timeout;
2681                         if (chunk == 0)
2682                                 chunk = 1;
2683
2684                         for (i = 0; i < chunk; i++) {
2685                                 ksocknal_check_peer_timeouts(peer_index);
2686                                 peer_index = (peer_index + 1) %
2687                                         HASH_SIZE(ksocknal_data.ksnd_peers);
2688                         }
2689
2690                         deadline += p;
2691                 }
2692
2693                 if (nenomem_conns != 0) {
2694                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2695                          * This also prevents me getting woken immediately
2696                          * if any go back on my enomem list. */
2697                         timeout = SOCKNAL_ENOMEM_RETRY;
2698                 }
2699                 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2700                                                      timeout;
2701
2702                 set_current_state(TASK_INTERRUPTIBLE);
2703                 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2704
2705                 if (!ksocknal_data.ksnd_shuttingdown &&
2706                     list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2707                     list_empty(&ksocknal_data.ksnd_zombie_conns))
2708                         schedule_timeout(cfs_time_seconds(timeout));
2709
2710                 set_current_state(TASK_RUNNING);
2711                 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2712
2713                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2714         }
2715
2716         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2717
2718         ksocknal_thread_fini();
2719         return 0;
2720 }