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 <libcfs/linux/linux-mem.h>
  28 #include "socklnd.h"
  29
  30 struct ksock_tx *
  31 ksocknal_alloc_tx(int type, int size)
  32 {
  33         struct ksock_tx *tx = NULL;
  34
  35         if (type == KSOCK_MSG_NOOP) {
  36                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  37
  38                 /* searching for a noop tx in free list */
  39                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  40
  41                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  42                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next,
  43                                         struct ksock_tx, tx_list);
  44                         LASSERT(tx->tx_desc_size == size);
  45                         list_del(&tx->tx_list);
  46                 }
  47
  48                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  49         }
  50
  51         if (tx == NULL)
  52                 LIBCFS_ALLOC(tx, size);
  53
  54         if (tx == NULL)
  55                 return NULL;
  56
  57         atomic_set(&tx->tx_refcount, 1);
  58         tx->tx_zc_aborted = 0;
  59         tx->tx_zc_capable = 0;
  60         tx->tx_zc_checked = 0;
  61         tx->tx_hstatus = LNET_MSG_STATUS_OK;
  62         tx->tx_desc_size  = size;
  63
  64         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  65
  66         return tx;
  67 }
  68
  69 struct ksock_tx *
  70 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  71 {
  72         struct ksock_tx *tx;
  73
  74         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  75         if (tx == NULL) {
  76                 CERROR("Can't allocate noop tx desc\n");
  77                 return NULL;
  78         }
  79
  80         tx->tx_conn     = NULL;
  81         tx->tx_lnetmsg  = NULL;
  82         tx->tx_kiov     = NULL;
  83         tx->tx_nkiov    = 0;
  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_hdr(struct ksock_conn *conn, struct ksock_tx *tx,
 115                   struct kvec *scratch_iov)
 116 {
 117         struct kvec *iov = &tx->tx_hdr;
 118         int    nob;
 119         int    rc;
 120
 121         LASSERT(tx->tx_niov > 0);
 122
 123         /* Never touch tx->tx_hdr inside ksocknal_lib_send_hdr() */
 124         rc = ksocknal_lib_send_hdr(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         LASSERT(tx->tx_niov == 1);
 135
 136         if (nob < (int) iov->iov_len) {
 137                 iov->iov_base += nob;
 138                 iov->iov_len -= nob;
 139                 return rc;
 140         }
 141
 142         LASSERT(nob == iov->iov_len);
 143         tx->tx_niov--;
 144
 145         return rc;
 146 }
 147
 148 static int
 149 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
 150                    struct kvec *scratch_iov)
 151 {
 152         struct bio_vec *kiov = tx->tx_kiov;
 153         int nob;
 154         int rc;
 155
 156         LASSERT(tx->tx_niov == 0);
 157         LASSERT(tx->tx_nkiov > 0);
 158
 159         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 160         rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov);
 161
 162         if (rc <= 0)                            /* sent nothing? */
 163                 return rc;
 164
 165         nob = rc;
 166         LASSERT(nob <= tx->tx_resid);
 167         tx->tx_resid -= nob;
 168
 169         /* "consume" kiov */
 170         do {
 171                 LASSERT(tx->tx_nkiov > 0);
 172
 173                 if (nob < (int)kiov->bv_len) {
 174                         kiov->bv_offset += nob;
 175                         kiov->bv_len -= nob;
 176                         return rc;
 177                 }
 178
 179                 nob -= (int)kiov->bv_len;
 180                 tx->tx_kiov = ++kiov;
 181                 tx->tx_nkiov--;
 182         } while (nob != 0);
 183
 184         return rc;
 185 }
 186
 187 static int
 188 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 189                   struct kvec *scratch_iov)
 190 {
 191         int     rc;
 192         int     bufnob;
 193
 194         if (ksocknal_data.ksnd_stall_tx != 0)
 195                 schedule_timeout_uninterruptible(
 196                         cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 197
 198         LASSERT(tx->tx_resid != 0);
 199
 200         rc = ksocknal_connsock_addref(conn);
 201         if (rc != 0) {
 202                 LASSERT(conn->ksnc_closing);
 203                 return -ESHUTDOWN;
 204         }
 205
 206         do {
 207                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 208                         /* testing... */
 209                         ksocknal_data.ksnd_enomem_tx--;
 210                         rc = -EAGAIN;
 211                 } else if (tx->tx_niov != 0) {
 212                         rc = ksocknal_send_hdr(conn, tx, scratch_iov);
 213                 } else {
 214                         rc = ksocknal_send_kiov(conn, tx, scratch_iov);
 215                 }
 216
 217                 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 218                 if (rc > 0)                     /* sent something? */
 219                         conn->ksnc_tx_bufnob += rc; /* account it */
 220
 221                 if (bufnob < conn->ksnc_tx_bufnob) {
 222                         /* allocated send buffer bytes < computed; infer
 223                          * something got ACKed */
 224                         conn->ksnc_tx_deadline = ktime_get_seconds() +
 225                                                  ksocknal_timeout();
 226                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 227                         conn->ksnc_tx_bufnob = bufnob;
 228                         smp_mb();
 229                 }
 230
 231                 if (rc <= 0) { /* Didn't write anything? */
 232                         /* some stacks return 0 instead of -EAGAIN */
 233                         if (rc == 0)
 234                                 rc = -EAGAIN;
 235
 236                         /* Check if EAGAIN is due to memory pressure */
 237                         if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 238                                 rc = -ENOMEM;
 239
 240                         break;
 241                 }
 242
 243                 /* socket's wmem_queued now includes 'rc' bytes */
 244                 atomic_sub (rc, &conn->ksnc_tx_nob);
 245                 rc = 0;
 246
 247         } while (tx->tx_resid != 0);
 248
 249         ksocknal_connsock_decref(conn);
 250         return rc;
 251 }
 252
 253 static int
 254 ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
 255 {
 256         struct kvec *iov = conn->ksnc_rx_iov;
 257         int     nob;
 258         int     rc;
 259
 260         LASSERT(conn->ksnc_rx_niov > 0);
 261
 262         /* Never touch conn->ksnc_rx_iov or change connection
 263          * status inside ksocknal_lib_recv_iov */
 264         rc = ksocknal_lib_recv_iov(conn, scratchiov);
 265
 266         if (rc <= 0)
 267                 return rc;
 268
 269         /* received something... */
 270         nob = rc;
 271
 272         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 273         conn->ksnc_rx_deadline = ktime_get_seconds() +
 274                                  ksocknal_timeout();
 275         smp_mb();                       /* order with setting rx_started */
 276         conn->ksnc_rx_started = 1;
 277
 278         conn->ksnc_rx_nob_wanted -= nob;
 279         conn->ksnc_rx_nob_left -= nob;
 280
 281         do {
 282                 LASSERT(conn->ksnc_rx_niov > 0);
 283
 284                 if (nob < (int)iov->iov_len) {
 285                         iov->iov_len -= nob;
 286                         iov->iov_base += nob;
 287                         return -EAGAIN;
 288                 }
 289
 290                 nob -= iov->iov_len;
 291                 conn->ksnc_rx_iov = ++iov;
 292                 conn->ksnc_rx_niov--;
 293         } while (nob != 0);
 294
 295         return rc;
 296 }
 297
 298 static int
 299 ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 300                    struct kvec *scratch_iov)
 301 {
 302         struct bio_vec *kiov = conn->ksnc_rx_kiov;
 303         int nob;
 304         int rc;
 305         LASSERT(conn->ksnc_rx_nkiov > 0);
 306
 307         /* Never touch conn->ksnc_rx_kiov or change connection
 308          * status inside ksocknal_lib_recv_iov */
 309         rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov);
 310
 311         if (rc <= 0)
 312                 return rc;
 313
 314         /* received something... */
 315         nob = rc;
 316
 317         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 318         conn->ksnc_rx_deadline = ktime_get_seconds() +
 319                                  ksocknal_timeout();
 320         smp_mb();                       /* order with setting rx_started */
 321         conn->ksnc_rx_started = 1;
 322
 323         conn->ksnc_rx_nob_wanted -= nob;
 324         conn->ksnc_rx_nob_left -= nob;
 325
 326         do {
 327                 LASSERT(conn->ksnc_rx_nkiov > 0);
 328
 329                 if (nob < (int) kiov->bv_len) {
 330                         kiov->bv_offset += nob;
 331                         kiov->bv_len -= nob;
 332                         return -EAGAIN;
 333                 }
 334
 335                 nob -= kiov->bv_len;
 336                 conn->ksnc_rx_kiov = ++kiov;
 337                 conn->ksnc_rx_nkiov--;
 338         } while (nob != 0);
 339
 340         return 1;
 341 }
 342
 343 static int
 344 ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 345                  struct kvec *scratch_iov)
 346 {
 347         /* Return 1 on success, 0 on EOF, < 0 on error.
 348          * Caller checks ksnc_rx_nob_wanted to determine
 349          * progress/completion. */
 350         int     rc;
 351         ENTRY;
 352
 353         if (ksocknal_data.ksnd_stall_rx != 0)
 354                 schedule_timeout_uninterruptible(
 355                         cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
 356
 357         rc = ksocknal_connsock_addref(conn);
 358         if (rc != 0) {
 359                 LASSERT(conn->ksnc_closing);
 360                 return -ESHUTDOWN;
 361         }
 362
 363         for (;;) {
 364                 if (conn->ksnc_rx_niov != 0)
 365                         rc = ksocknal_recv_iov(conn, scratch_iov);
 366                 else
 367                         rc = ksocknal_recv_kiov(conn, rx_scratch_pgs,
 368                                                  scratch_iov);
 369
 370                 if (rc <= 0) {
 371                         /* error/EOF or partial receive */
 372                         if (rc == -EAGAIN) {
 373                                 rc = 1;
 374                         } else if (rc == 0 && conn->ksnc_rx_started) {
 375                                 /* EOF in the middle of a message */
 376                                 rc = -EPROTO;
 377                         }
 378                         break;
 379                 }
 380
 381                 /* Completed a fragment */
 382
 383                 if (conn->ksnc_rx_nob_wanted == 0) {
 384                         rc = 1;
 385                         break;
 386                 }
 387         }
 388
 389         ksocknal_connsock_decref(conn);
 390         RETURN(rc);
 391 }
 392
 393 void
 394 ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc)
 395 {
 396         struct lnet_msg *lnetmsg = tx->tx_lnetmsg;
 397         enum lnet_msg_hstatus hstatus = tx->tx_hstatus;
 398         ENTRY;
 399
 400         LASSERT(ni != NULL || tx->tx_conn != NULL);
 401
 402         if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
 403                 rc = -EIO;
 404                 if (hstatus == LNET_MSG_STATUS_OK)
 405                         hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 406         }
 407
 408         if (tx->tx_conn != NULL)
 409                 ksocknal_conn_decref(tx->tx_conn);
 410
 411         ksocknal_free_tx(tx);
 412         if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
 413                 lnetmsg->msg_health_status = hstatus;
 414                 lnet_finalize(lnetmsg, rc);
 415         }
 416
 417         EXIT;
 418 }
 419
 420 void
 421 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
 422 {
 423         struct ksock_tx *tx;
 424
 425         while (!list_empty(txlist)) {
 426                 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
 427
 428                 if (error && tx->tx_lnetmsg != NULL) {
 429                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 430                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
 431                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
 432                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 433                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 434                 } else if (error) {
 435                         CNETERR("Deleting noop packet\n");
 436                 }
 437
 438                 list_del(&tx->tx_list);
 439
 440                 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
 441                         if (error == -ETIMEDOUT)
 442                                 tx->tx_hstatus =
 443                                   LNET_MSG_STATUS_LOCAL_TIMEOUT;
 444                         else if (error == -ENETDOWN ||
 445                                  error == -EHOSTUNREACH ||
 446                                  error == -ENETUNREACH ||
 447                                  error == -ECONNREFUSED ||
 448                                  error == -ECONNRESET)
 449                                 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
 450                         /*
 451                          * for all other errors we don't want to
 452                          * retransmit
 453                          */
 454                         else if (error)
 455                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 456                 }
 457
 458                 LASSERT(atomic_read(&tx->tx_refcount) == 1);
 459                 ksocknal_tx_done(ni, tx, error);
 460         }
 461 }
 462
 463 static void
 464 ksocknal_check_zc_req(struct ksock_tx *tx)
 465 {
 466         struct ksock_conn *conn = tx->tx_conn;
 467         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
 468
 469         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 470          * to ksnp_zc_req_list if some fragment of this message should be sent
 471          * zero-copy.  Our peer_ni will send an ACK containing this cookie when
 472          * she has received this message to tell us we can signal completion.
 473          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 474          * ksnp_zc_req_list. */
 475         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 476         LASSERT (tx->tx_zc_capable);
 477
 478         tx->tx_zc_checked = 1;
 479
 480         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 481             !conn->ksnc_zc_capable)
 482                 return;
 483
 484         /* assign cookie and queue tx to pending list, it will be released when
 485          * a matching ack is received. See ksocknal_handle_zcack() */
 486
 487         ksocknal_tx_addref(tx);
 488
 489         spin_lock(&peer_ni->ksnp_lock);
 490
 491         /* ZC_REQ is going to be pinned to the peer_ni */
 492         tx->tx_deadline = ktime_get_seconds() +
 493                           ksocknal_timeout();
 494
 495         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 496
 497         tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
 498
 499         if (peer_ni->ksnp_zc_next_cookie == 0)
 500                 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 501
 502         list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
 503
 504         spin_unlock(&peer_ni->ksnp_lock);
 505 }
 506
 507 static void
 508 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
 509 {
 510         struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
 511
 512         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 513         LASSERT(tx->tx_zc_capable);
 514
 515         tx->tx_zc_checked = 0;
 516
 517         spin_lock(&peer_ni->ksnp_lock);
 518
 519         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 520                 /* Not waiting for an ACK */
 521                 spin_unlock(&peer_ni->ksnp_lock);
 522                 return;
 523         }
 524
 525         tx->tx_msg.ksm_zc_cookies[0] = 0;
 526         list_del(&tx->tx_zc_list);
 527
 528         spin_unlock(&peer_ni->ksnp_lock);
 529
 530         ksocknal_tx_decref(tx);
 531 }
 532
 533 static int
 534 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 535                           struct kvec *scratch_iov)
 536 {
 537         int rc;
 538         bool error_sim = false;
 539
 540         if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
 541                 error_sim = true;
 542                 rc = -EINVAL;
 543                 goto simulate_error;
 544         }
 545
 546         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 547                 ksocknal_check_zc_req(tx);
 548
 549         rc = ksocknal_transmit(conn, tx, scratch_iov);
 550
 551         CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 552
 553         if (tx->tx_resid == 0) {
 554                 /* Sent everything OK */
 555                 LASSERT(rc == 0);
 556
 557                 return 0;
 558         }
 559
 560         if (rc == -EAGAIN)
 561                 return rc;
 562
 563         if (rc == -ENOMEM) {
 564                 static int counter;
 565
 566                 counter++;   /* exponential backoff warnings */
 567                 if ((counter & (-counter)) == counter)
 568                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 569                               counter, conn, atomic_read(&libcfs_kmemory));
 570
 571                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 572                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 573
 574                 /* enomem list takes over scheduler's ref... */
 575                 LASSERT(conn->ksnc_tx_scheduled);
 576                 list_add_tail(&conn->ksnc_tx_list,
 577                                   &ksocknal_data.ksnd_enomem_conns);
 578                 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
 579                     ksocknal_data.ksnd_reaper_waketime)
 580                         wake_up(&ksocknal_data.ksnd_reaper_waitq);
 581
 582                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 583
 584                 /*
 585                  * set the health status of the message which determines
 586                  * whether we should retry the transmit
 587                  */
 588                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 589                 return (rc);
 590         }
 591
 592 simulate_error:
 593
 594         /* Actual error */
 595         LASSERT(rc < 0);
 596
 597         if (!error_sim) {
 598                 /*
 599                 * set the health status of the message which determines
 600                 * whether we should retry the transmit
 601                 */
 602                 if (rc == -ETIMEDOUT)
 603                         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
 604                 else
 605                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 606         }
 607
 608         if (!conn->ksnc_closing) {
 609                 switch (rc) {
 610                 case -ECONNRESET:
 611                         LCONSOLE_WARN("Host %pI4h reset our connection "
 612                                       "while we were sending data; it may have "
 613                                       "rebooted.\n",
 614                                       &conn->ksnc_ipaddr);
 615                         break;
 616                 default:
 617                         LCONSOLE_WARN("There was an unexpected network error "
 618                                       "while writing to %pI4h: %d.\n",
 619                                       &conn->ksnc_ipaddr, rc);
 620                         break;
 621                 }
 622                 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
 623                        conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id),
 624                        &conn->ksnc_ipaddr, conn->ksnc_port);
 625         }
 626
 627         if (tx->tx_zc_checked)
 628                 ksocknal_uncheck_zc_req(tx);
 629
 630         /* it's not an error if conn is being closed */
 631         ksocknal_close_conn_and_siblings(conn,
 632                                           (conn->ksnc_closing) ? 0 : rc);
 633
 634         return rc;
 635 }
 636
 637 static void
 638 ksocknal_launch_connection_locked(struct ksock_route *route)
 639 {
 640
 641         /* called holding write lock on ksnd_global_lock */
 642
 643         LASSERT (!route->ksnr_scheduled);
 644         LASSERT (!route->ksnr_connecting);
 645         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 646
 647         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 648         ksocknal_route_addref(route);           /* extra ref for connd */
 649
 650         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 651
 652         list_add_tail(&route->ksnr_connd_list,
 653                           &ksocknal_data.ksnd_connd_routes);
 654         wake_up(&ksocknal_data.ksnd_connd_waitq);
 655
 656         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 657 }
 658
 659 void
 660 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
 661 {
 662         struct ksock_route *route;
 663
 664         /* called holding write lock on ksnd_global_lock */
 665         for (;;) {
 666                 /* launch any/all connections that need it */
 667                 route = ksocknal_find_connectable_route_locked(peer_ni);
 668                 if (route == NULL)
 669                         return;
 670
 671                 ksocknal_launch_connection_locked(route);
 672         }
 673 }
 674
 675 struct ksock_conn *
 676 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
 677 {
 678         struct list_head *tmp;
 679         struct ksock_conn *conn;
 680         struct ksock_conn *typed = NULL;
 681         struct ksock_conn *fallback = NULL;
 682         int tnob = 0;
 683         int fnob = 0;
 684
 685         list_for_each(tmp, &peer_ni->ksnp_conns) {
 686                 struct ksock_conn *c = list_entry(tmp, struct ksock_conn,
 687                                                   ksnc_list);
 688                 int nob = atomic_read(&c->ksnc_tx_nob) +
 689                           c->ksnc_sock->sk->sk_wmem_queued;
 690                 int rc;
 691
 692                 LASSERT (!c->ksnc_closing);
 693                 LASSERT (c->ksnc_proto != NULL &&
 694                          c->ksnc_proto->pro_match_tx != NULL);
 695
 696                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 697
 698                 switch (rc) {
 699                 default:
 700                         LBUG();
 701                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 702                         continue;
 703
 704                 case SOCKNAL_MATCH_YES: /* typed connection */
 705                         if (typed == NULL || tnob > nob ||
 706                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 707                              typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 708                                 typed = c;
 709                                 tnob  = nob;
 710                         }
 711                         break;
 712
 713                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 714                         if (fallback == NULL || fnob > nob ||
 715                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 716                              fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 717                                 fallback = c;
 718                                 fnob     = nob;
 719                         }
 720                         break;
 721                 }
 722         }
 723
 724         /* prefer the typed selection */
 725         conn = (typed != NULL) ? typed : fallback;
 726
 727         if (conn != NULL)
 728                 conn->ksnc_tx_last_post = ktime_get_seconds();
 729
 730         return conn;
 731 }
 732
 733 void
 734 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
 735 {
 736         conn->ksnc_proto->pro_pack(tx);
 737
 738         atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 739         ksocknal_conn_addref(conn); /* +1 ref for tx */
 740         tx->tx_conn = conn;
 741 }
 742
 743 void
 744 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
 745 {
 746         struct ksock_sched *sched = conn->ksnc_scheduler;
 747         struct ksock_msg *msg = &tx->tx_msg;
 748         struct ksock_tx *ztx = NULL;
 749         int bufnob = 0;
 750
 751         /* called holding global lock (read or irq-write) and caller may
 752          * not have dropped this lock between finding conn and calling me,
 753          * so we don't need the {get,put}connsock dance to deref
 754          * ksnc_sock... */
 755         LASSERT(!conn->ksnc_closing);
 756
 757         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 758                libcfs_id2str(conn->ksnc_peer->ksnp_id),
 759                &conn->ksnc_ipaddr, conn->ksnc_port);
 760
 761         ksocknal_tx_prep(conn, tx);
 762
 763         /* Ensure the frags we've been given EXACTLY match the number of
 764          * bytes we want to send.  Many TCP/IP stacks disregard any total
 765          * size parameters passed to them and just look at the frags.
 766          *
 767          * We always expect at least 1 mapped fragment containing the
 768          * complete ksocknal message header.
 769          */
 770         LASSERT(lnet_iov_nob(tx->tx_niov, &tx->tx_hdr) +
 771                 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 772                 (unsigned int)tx->tx_nob);
 773         LASSERT(tx->tx_niov >= 1);
 774         LASSERT(tx->tx_resid == tx->tx_nob);
 775
 776         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 777                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 778                                                KSOCK_MSG_NOOP,
 779                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 780
 781         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 782         spin_lock_bh(&sched->kss_lock);
 783
 784         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 785                 /* First packet starts the timeout */
 786                 conn->ksnc_tx_deadline = ktime_get_seconds() +
 787                                          ksocknal_timeout();
 788                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 789                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 790                 conn->ksnc_tx_bufnob = 0;
 791                 smp_mb(); /* order with adding to tx_queue */
 792         }
 793
 794         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 795                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 796                  * on a normal packet so I don't need to send it */
 797                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 798                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 799
 800                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 801                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 802
 803         } else {
 804                 /* It's a normal packet - can it piggback a noop zc-ack that
 805                  * has been queued already? */
 806                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 807                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 808
 809                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 810                 /* ztx will be released later */
 811         }
 812
 813         if (ztx != NULL) {
 814                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 815                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 816         }
 817
 818         if (conn->ksnc_tx_ready &&      /* able to send */
 819             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 820                 /* +1 ref for scheduler */
 821                 ksocknal_conn_addref(conn);
 822                 list_add_tail(&conn->ksnc_tx_list,
 823                                    &sched->kss_tx_conns);
 824                 conn->ksnc_tx_scheduled = 1;
 825                 wake_up(&sched->kss_waitq);
 826         }
 827
 828         spin_unlock_bh(&sched->kss_lock);
 829 }
 830
 831
 832 struct ksock_route *
 833 ksocknal_find_connectable_route_locked(struct ksock_peer_ni *peer_ni)
 834 {
 835         time64_t now = ktime_get_seconds();
 836         struct list_head *tmp;
 837         struct ksock_route *route;
 838
 839         list_for_each(tmp, &peer_ni->ksnp_routes) {
 840                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 841
 842                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 843
 844                 if (route->ksnr_scheduled)      /* connections being established */
 845                         continue;
 846
 847                 /* all route types connected ? */
 848                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 849                         continue;
 850
 851                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 852                       now >= route->ksnr_timeout)) {
 853                         CDEBUG(D_NET,
 854                                "Too soon to retry route %pI4h "
 855                                "(cnted %d, interval %lld, %lld secs later)\n",
 856                                &route->ksnr_ipaddr,
 857                                route->ksnr_connected,
 858                                route->ksnr_retry_interval,
 859                                route->ksnr_timeout - now);
 860                         continue;
 861                 }
 862
 863                 return (route);
 864         }
 865
 866         return (NULL);
 867 }
 868
 869 struct ksock_route *
 870 ksocknal_find_connecting_route_locked(struct ksock_peer_ni *peer_ni)
 871 {
 872         struct list_head *tmp;
 873         struct ksock_route *route;
 874
 875         list_for_each(tmp, &peer_ni->ksnp_routes) {
 876                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 877
 878                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 879
 880                 if (route->ksnr_scheduled)
 881                         return (route);
 882         }
 883
 884         return (NULL);
 885 }
 886
 887 int
 888 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
 889                        struct lnet_process_id id)
 890 {
 891         struct ksock_peer_ni *peer_ni;
 892         struct ksock_conn *conn;
 893         rwlock_t *g_lock;
 894         int retry;
 895         int rc;
 896
 897         LASSERT (tx->tx_conn == NULL);
 898
 899         g_lock = &ksocknal_data.ksnd_global_lock;
 900
 901         for (retry = 0;; retry = 1) {
 902                 read_lock(g_lock);
 903                 peer_ni = ksocknal_find_peer_locked(ni, id);
 904                 if (peer_ni != NULL) {
 905                         if (ksocknal_find_connectable_route_locked(peer_ni) == NULL) {
 906                                 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 907                                 if (conn != NULL) {
 908                                         /* I've got no routes that need to be
 909                                          * connecting and I do have an actual
 910                                          * connection... */
 911                                         ksocknal_queue_tx_locked (tx, conn);
 912                                         read_unlock(g_lock);
 913                                         return (0);
 914                                 }
 915                         }
 916                 }
 917
 918                 /* I'll need a write lock... */
 919                 read_unlock(g_lock);
 920
 921                 write_lock_bh(g_lock);
 922
 923                 peer_ni = ksocknal_find_peer_locked(ni, id);
 924                 if (peer_ni != NULL)
 925                         break;
 926
 927                 write_unlock_bh(g_lock);
 928
 929                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 930                         CERROR("Refusing to create a connection to "
 931                                "userspace process %s\n", libcfs_id2str(id));
 932                         return -EHOSTUNREACH;
 933                 }
 934
 935                 if (retry) {
 936                         CERROR("Can't find peer_ni %s\n", libcfs_id2str(id));
 937                         return -EHOSTUNREACH;
 938                 }
 939
 940                 rc = ksocknal_add_peer(ni, id,
 941                                        LNET_NIDADDR(id.nid),
 942                                        lnet_acceptor_port());
 943                 if (rc != 0) {
 944                         CERROR("Can't add peer_ni %s: %d\n",
 945                                libcfs_id2str(id), rc);
 946                         return rc;
 947                 }
 948         }
 949
 950         ksocknal_launch_all_connections_locked(peer_ni);
 951
 952         conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 953         if (conn != NULL) {
 954                 /* Connection exists; queue message on it */
 955                 ksocknal_queue_tx_locked (tx, conn);
 956                 write_unlock_bh(g_lock);
 957                 return (0);
 958         }
 959
 960         if (peer_ni->ksnp_accepting > 0 ||
 961             ksocknal_find_connecting_route_locked (peer_ni) != NULL) {
 962                 /* the message is going to be pinned to the peer_ni */
 963                 tx->tx_deadline = ktime_get_seconds() +
 964                                   ksocknal_timeout();
 965
 966                 /* Queue the message until a connection is established */
 967                 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
 968                 write_unlock_bh(g_lock);
 969                 return 0;
 970         }
 971
 972         write_unlock_bh(g_lock);
 973
 974         /* NB Routes may be ignored if connections to them failed recently */
 975         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 976         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 977         return (-EHOSTUNREACH);
 978 }
 979
 980 int
 981 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
 982 {
 983         /* '1' for consistency with code that checks !mpflag to restore */
 984         unsigned int mpflag = 1;
 985         int type = lntmsg->msg_type;
 986         struct lnet_process_id target = lntmsg->msg_target;
 987         unsigned int payload_niov = lntmsg->msg_niov;
 988         struct bio_vec *payload_kiov = lntmsg->msg_kiov;
 989         unsigned int payload_offset = lntmsg->msg_offset;
 990         unsigned int payload_nob = lntmsg->msg_len;
 991         struct ksock_tx *tx;
 992         int desc_size;
 993         int rc;
 994
 995         /* NB 'private' is different depending on what we're sending.
 996          * Just ignore it... */
 997
 998         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
 999                payload_nob, payload_niov, libcfs_id2str(target));
1000
1001         LASSERT (payload_nob == 0 || payload_niov > 0);
1002         LASSERT (payload_niov <= LNET_MAX_IOV);
1003         LASSERT (!in_interrupt ());
1004
1005         desc_size = offsetof(struct ksock_tx,
1006                              tx_payload[payload_niov]);
1007
1008         if (lntmsg->msg_vmflush)
1009                 mpflag = memalloc_noreclaim_save();
1010
1011         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1012         if (tx == NULL) {
1013                 CERROR("Can't allocate tx desc type %d size %d\n",
1014                        type, desc_size);
1015                 if (lntmsg->msg_vmflush)
1016                         memalloc_noreclaim_restore(mpflag);
1017                 return -ENOMEM;
1018         }
1019
1020         tx->tx_conn = NULL;                     /* set when assigned a conn */
1021         tx->tx_lnetmsg = lntmsg;
1022
1023         tx->tx_niov = 1;
1024         tx->tx_kiov = tx->tx_payload;
1025         tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1026                                          payload_niov, payload_kiov,
1027                                          payload_offset, payload_nob);
1028
1029         if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1030                 tx->tx_zc_capable = 1;
1031
1032         tx->tx_msg.ksm_csum = 0;
1033         tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1034         tx->tx_msg.ksm_zc_cookies[0] = 0;
1035         tx->tx_msg.ksm_zc_cookies[1] = 0;
1036
1037         /* The first fragment will be set later in pro_pack */
1038         rc = ksocknal_launch_packet(ni, tx, target);
1039         /*
1040          * We can't test lntsmg->msg_vmflush again as lntmsg may
1041          * have been freed.
1042          */
1043         if (!mpflag)
1044                 memalloc_noreclaim_restore(mpflag);
1045
1046         if (rc == 0)
1047                 return (0);
1048
1049         lntmsg->msg_health_status = tx->tx_hstatus;
1050         ksocknal_free_tx(tx);
1051         return (-EIO);
1052 }
1053
1054 int
1055 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1056 {
1057         struct task_struct *task = kthread_run(fn, arg, name);
1058
1059         if (IS_ERR(task))
1060                 return PTR_ERR(task);
1061
1062         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1063         ksocknal_data.ksnd_nthreads++;
1064         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1065         return 0;
1066 }
1067
1068 void
1069 ksocknal_thread_fini (void)
1070 {
1071         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1072         if (--ksocknal_data.ksnd_nthreads == 0)
1073                 wake_up_var(&ksocknal_data.ksnd_nthreads);
1074         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1075 }
1076
1077 int
1078 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1079 {
1080         static char ksocknal_slop_buffer[4096];
1081         int nob;
1082         unsigned int niov;
1083         int skipped;
1084
1085         LASSERT(conn->ksnc_proto != NULL);
1086
1087         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1088                 /* Remind the socket to ack eagerly... */
1089                 ksocknal_lib_eager_ack(conn);
1090         }
1091
1092         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1093                 conn->ksnc_rx_started = 0;
1094                 smp_mb();                       /* racing with timeout thread */
1095
1096                 switch (conn->ksnc_proto->pro_version) {
1097                 case  KSOCK_PROTO_V2:
1098                 case  KSOCK_PROTO_V3:
1099                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1100                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1101                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1102
1103                         conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u);
1104                         conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u);
1105                         conn->ksnc_rx_iov[0].iov_len  = offsetof(struct ksock_msg, ksm_u);
1106                         break;
1107
1108                 case KSOCK_PROTO_V1:
1109                         /* Receiving bare struct lnet_hdr */
1110                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1111                         conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr);
1112                         conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr);
1113
1114                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1115                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1116                         conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr);
1117                         break;
1118
1119                 default:
1120                         LBUG ();
1121                 }
1122                 conn->ksnc_rx_niov = 1;
1123
1124                 conn->ksnc_rx_kiov = NULL;
1125                 conn->ksnc_rx_nkiov = 0;
1126                 conn->ksnc_rx_csum = ~0;
1127                 return (1);
1128         }
1129
1130         /* Set up to skip as much as possible now.  If there's more left
1131          * (ran out of iov entries) we'll get called again */
1132
1133         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1134         conn->ksnc_rx_nob_left = nob_to_skip;
1135         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1136         skipped = 0;
1137         niov = 0;
1138
1139         do {
1140                 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1141
1142                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1143                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1144                 niov++;
1145                 skipped += nob;
1146                 nob_to_skip -= nob;
1147
1148         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1149                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1150
1151         conn->ksnc_rx_niov = niov;
1152         conn->ksnc_rx_kiov = NULL;
1153         conn->ksnc_rx_nkiov = 0;
1154         conn->ksnc_rx_nob_wanted = skipped;
1155         return (0);
1156 }
1157
1158 static int
1159 ksocknal_process_receive(struct ksock_conn *conn,
1160                          struct page **rx_scratch_pgs,
1161                          struct kvec *scratch_iov)
1162 {
1163         struct lnet_hdr *lhdr;
1164         struct lnet_process_id *id;
1165         int rc;
1166
1167         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1168
1169         /* NB: sched lock NOT held */
1170         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1171         LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1172                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1173                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1174                 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1175  again:
1176         if (conn->ksnc_rx_nob_wanted != 0) {
1177                 rc = ksocknal_receive(conn, rx_scratch_pgs,
1178                                       scratch_iov);
1179
1180                 if (rc <= 0) {
1181                         struct lnet_process_id ksnp_id;
1182
1183                         ksnp_id = conn->ksnc_peer->ksnp_id;
1184
1185                         LASSERT(rc != -EAGAIN);
1186                         if (rc == 0)
1187                                 CDEBUG(D_NET, "[%p] EOF from %s "
1188                                        "ip %pI4h:%d\n", conn,
1189                                         libcfs_id2str(ksnp_id),
1190                                         &conn->ksnc_ipaddr,
1191                                         conn->ksnc_port);
1192                         else if (!conn->ksnc_closing)
1193                                 CERROR("[%p] Error %d on read from %s "
1194                                        "ip %pI4h:%d\n", conn, rc,
1195                                        libcfs_id2str(ksnp_id),
1196                                        &conn->ksnc_ipaddr,
1197                                        conn->ksnc_port);
1198
1199                         /* it's not an error if conn is being closed */
1200                         ksocknal_close_conn_and_siblings (conn,
1201                                                           (conn->ksnc_closing) ? 0 : rc);
1202                         return (rc == 0 ? -ESHUTDOWN : rc);
1203                 }
1204
1205                 if (conn->ksnc_rx_nob_wanted != 0) {
1206                         /* short read */
1207                         return (-EAGAIN);
1208                 }
1209         }
1210         switch (conn->ksnc_rx_state) {
1211         case SOCKNAL_RX_KSM_HEADER:
1212                 if (conn->ksnc_flip) {
1213                         __swab32s(&conn->ksnc_msg.ksm_type);
1214                         __swab32s(&conn->ksnc_msg.ksm_csum);
1215                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1216                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1217                 }
1218
1219                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1220                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1221                         CERROR("%s: Unknown message type: %x\n",
1222                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1223                                conn->ksnc_msg.ksm_type);
1224                         ksocknal_new_packet(conn, 0);
1225                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1226                         return (-EPROTO);
1227                 }
1228
1229                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1230                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1231                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1232                         /* NOOP Checksum error */
1233                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1234                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1235                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1236                         ksocknal_new_packet(conn, 0);
1237                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1238                         return (-EIO);
1239                 }
1240
1241                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1242                         __u64 cookie = 0;
1243
1244                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1245
1246                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1247                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1248
1249                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1250                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1251
1252                         if (rc != 0) {
1253                                 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1254                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1255                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1256                                 ksocknal_new_packet(conn, 0);
1257                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1258                                 return (rc);
1259                         }
1260                 }
1261
1262                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1263                         ksocknal_new_packet (conn, 0);
1264                         return 0;       /* NOOP is done and just return */
1265                 }
1266
1267                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1268                 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg);
1269                 conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg);
1270
1271                 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1272                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1273                 conn->ksnc_rx_iov[0].iov_len  = sizeof(struct ksock_lnet_msg);
1274
1275                 conn->ksnc_rx_niov = 1;
1276                 conn->ksnc_rx_kiov = NULL;
1277                 conn->ksnc_rx_nkiov = 0;
1278
1279                 goto again;     /* read lnet header now */
1280
1281         case SOCKNAL_RX_LNET_HEADER:
1282                 /* unpack message header */
1283                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1284
1285                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1286                         /* Userspace peer_ni */
1287                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1288                         id   = &conn->ksnc_peer->ksnp_id;
1289
1290                         /* Substitute process ID assigned at connection time */
1291                         lhdr->src_pid = cpu_to_le32(id->pid);
1292                         lhdr->src_nid = cpu_to_le64(id->nid);
1293                 }
1294
1295                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1296                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1297
1298                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1299                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1300                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1301                 if (rc < 0) {
1302                         /* I just received garbage: give up on this conn */
1303                         ksocknal_new_packet(conn, 0);
1304                         ksocknal_close_conn_and_siblings (conn, rc);
1305                         ksocknal_conn_decref(conn);
1306                         return (-EPROTO);
1307                 }
1308
1309                 /* I'm racing with ksocknal_recv() */
1310                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1311                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1312
1313                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1314                         return 0;
1315
1316                 /* ksocknal_recv() got called */
1317                 goto again;
1318
1319         case SOCKNAL_RX_LNET_PAYLOAD:
1320                 /* payload all received */
1321                 rc = 0;
1322
1323                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1324                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1325                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1326                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1327                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1328                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1329                         rc = -EIO;
1330                 }
1331
1332                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1333                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1334
1335                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1336                         id   = &conn->ksnc_peer->ksnp_id;
1337
1338                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1339                                         conn->ksnc_msg.ksm_zc_cookies[0],
1340                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1341                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1342                 }
1343
1344                 if (rc && conn->ksnc_lnet_msg)
1345                         conn->ksnc_lnet_msg->msg_health_status =
1346                                 LNET_MSG_STATUS_REMOTE_ERROR;
1347                 lnet_finalize(conn->ksnc_lnet_msg, rc);
1348
1349                 if (rc != 0) {
1350                         ksocknal_new_packet(conn, 0);
1351                         ksocknal_close_conn_and_siblings (conn, rc);
1352                         return (-EPROTO);
1353                 }
1354                 /* Fall through */
1355
1356         case SOCKNAL_RX_SLOP:
1357                 /* starting new packet? */
1358                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1359                         return 0;       /* come back later */
1360                 goto again;             /* try to finish reading slop now */
1361
1362         default:
1363                 break;
1364         }
1365
1366         /* Not Reached */
1367         LBUG ();
1368         return (-EINVAL);                       /* keep gcc happy */
1369 }
1370
1371 int
1372 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1373               int delayed, unsigned int niov,
1374               struct bio_vec *kiov, unsigned int offset, unsigned int mlen,
1375               unsigned int rlen)
1376 {
1377         struct ksock_conn *conn = private;
1378         struct ksock_sched *sched = conn->ksnc_scheduler;
1379
1380         LASSERT (mlen <= rlen);
1381         LASSERT (niov <= LNET_MAX_IOV);
1382
1383         conn->ksnc_lnet_msg = msg;
1384         conn->ksnc_rx_nob_wanted = mlen;
1385         conn->ksnc_rx_nob_left   = rlen;
1386
1387         if (mlen == 0) {
1388                 conn->ksnc_rx_nkiov = 0;
1389                 conn->ksnc_rx_kiov = NULL;
1390                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1391                 conn->ksnc_rx_niov = 0;
1392         } else {
1393                 conn->ksnc_rx_niov = 0;
1394                 conn->ksnc_rx_iov  = NULL;
1395                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1396                 conn->ksnc_rx_nkiov =
1397                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1398                                           niov, kiov, offset, mlen);
1399         }
1400
1401         LASSERT (mlen ==
1402                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1403                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1404
1405         LASSERT (conn->ksnc_rx_scheduled);
1406
1407         spin_lock_bh(&sched->kss_lock);
1408
1409         switch (conn->ksnc_rx_state) {
1410         case SOCKNAL_RX_PARSE_WAIT:
1411                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1412                 wake_up(&sched->kss_waitq);
1413                 LASSERT(conn->ksnc_rx_ready);
1414                 break;
1415
1416         case SOCKNAL_RX_PARSE:
1417                 /* scheduler hasn't noticed I'm parsing yet */
1418                 break;
1419         }
1420
1421         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1422
1423         spin_unlock_bh(&sched->kss_lock);
1424         ksocknal_conn_decref(conn);
1425         return 0;
1426 }
1427
1428 static inline int
1429 ksocknal_sched_cansleep(struct ksock_sched *sched)
1430 {
1431         int           rc;
1432
1433         spin_lock_bh(&sched->kss_lock);
1434
1435         rc = (!ksocknal_data.ksnd_shuttingdown &&
1436               list_empty(&sched->kss_rx_conns) &&
1437               list_empty(&sched->kss_tx_conns));
1438
1439         spin_unlock_bh(&sched->kss_lock);
1440         return rc;
1441 }
1442
1443 int ksocknal_scheduler(void *arg)
1444 {
1445         struct ksock_sched *sched;
1446         struct ksock_conn *conn;
1447         struct ksock_tx *tx;
1448         int rc;
1449         int nloops = 0;
1450         long id = (long)arg;
1451         struct page **rx_scratch_pgs;
1452         struct kvec *scratch_iov;
1453
1454         sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1455
1456         LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1457                          sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1458         if (!rx_scratch_pgs) {
1459                 CERROR("Unable to allocate scratch pages\n");
1460                 return -ENOMEM;
1461         }
1462
1463         LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1464                          sizeof(*scratch_iov) * LNET_MAX_IOV);
1465         if (!scratch_iov) {
1466                 CERROR("Unable to allocate scratch iov\n");
1467                 return -ENOMEM;
1468         }
1469
1470         rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1471         if (rc != 0) {
1472                 CWARN("Can't set CPU partition affinity to %d: %d\n",
1473                         sched->kss_cpt, rc);
1474         }
1475
1476         spin_lock_bh(&sched->kss_lock);
1477
1478         while (!ksocknal_data.ksnd_shuttingdown) {
1479                 int did_something = 0;
1480
1481                 /* Ensure I progress everything semi-fairly */
1482
1483                 if (!list_empty(&sched->kss_rx_conns)) {
1484                         conn = list_entry(sched->kss_rx_conns.next,
1485                                           struct ksock_conn, ksnc_rx_list);
1486                         list_del(&conn->ksnc_rx_list);
1487
1488                         LASSERT(conn->ksnc_rx_scheduled);
1489                         LASSERT(conn->ksnc_rx_ready);
1490
1491                         /* clear rx_ready in case receive isn't complete.
1492                          * Do it BEFORE we call process_recv, since
1493                          * data_ready can set it any time after we release
1494                          * kss_lock. */
1495                         conn->ksnc_rx_ready = 0;
1496                         spin_unlock_bh(&sched->kss_lock);
1497
1498                         rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1499                                                       scratch_iov);
1500
1501                         spin_lock_bh(&sched->kss_lock);
1502
1503                         /* I'm the only one that can clear this flag */
1504                         LASSERT(conn->ksnc_rx_scheduled);
1505
1506                         /* Did process_receive get everything it wanted? */
1507                         if (rc == 0)
1508                                 conn->ksnc_rx_ready = 1;
1509
1510                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1511                                 /* Conn blocked waiting for ksocknal_recv()
1512                                  * I change its state (under lock) to signal
1513                                  * it can be rescheduled */
1514                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1515                         } else if (conn->ksnc_rx_ready) {
1516                                 /* reschedule for rx */
1517                                 list_add_tail(&conn->ksnc_rx_list,
1518                                                    &sched->kss_rx_conns);
1519                         } else {
1520                                 conn->ksnc_rx_scheduled = 0;
1521                                 /* drop my ref */
1522                                 ksocknal_conn_decref(conn);
1523                         }
1524
1525                         did_something = 1;
1526                 }
1527
1528                 if (!list_empty(&sched->kss_tx_conns)) {
1529                         LIST_HEAD(zlist);
1530
1531                         list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1532
1533                         conn = list_entry(sched->kss_tx_conns.next,
1534                                           struct ksock_conn, ksnc_tx_list);
1535                         list_del(&conn->ksnc_tx_list);
1536
1537                         LASSERT(conn->ksnc_tx_scheduled);
1538                         LASSERT(conn->ksnc_tx_ready);
1539                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1540
1541                         tx = list_entry(conn->ksnc_tx_queue.next,
1542                                         struct ksock_tx, tx_list);
1543
1544                         if (conn->ksnc_tx_carrier == tx)
1545                                 ksocknal_next_tx_carrier(conn);
1546
1547                         /* dequeue now so empty list => more to send */
1548                         list_del(&tx->tx_list);
1549
1550                         /* Clear tx_ready in case send isn't complete.  Do
1551                          * it BEFORE we call process_transmit, since
1552                          * write_space can set it any time after we release
1553                          * kss_lock. */
1554                         conn->ksnc_tx_ready = 0;
1555                         spin_unlock_bh(&sched->kss_lock);
1556
1557                         if (!list_empty(&zlist)) {
1558                                 /* free zombie noop txs, it's fast because
1559                                  * noop txs are just put in freelist */
1560                                 ksocknal_txlist_done(NULL, &zlist, 0);
1561                         }
1562
1563                         rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1564
1565                         if (rc == -ENOMEM || rc == -EAGAIN) {
1566                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1567                                 spin_lock_bh(&sched->kss_lock);
1568                                 list_add(&tx->tx_list,
1569                                          &conn->ksnc_tx_queue);
1570                         } else {
1571                                 /* Complete send; tx -ref */
1572                                 ksocknal_tx_decref(tx);
1573
1574                                 spin_lock_bh(&sched->kss_lock);
1575                                 /* assume space for more */
1576                                 conn->ksnc_tx_ready = 1;
1577                         }
1578
1579                         if (rc == -ENOMEM) {
1580                                 /* Do nothing; after a short timeout, this
1581                                  * conn will be reposted on kss_tx_conns. */
1582                         } else if (conn->ksnc_tx_ready &&
1583                                    !list_empty(&conn->ksnc_tx_queue)) {
1584                                 /* reschedule for tx */
1585                                 list_add_tail(&conn->ksnc_tx_list,
1586                                               &sched->kss_tx_conns);
1587                         } else {
1588                                 conn->ksnc_tx_scheduled = 0;
1589                                 /* drop my ref */
1590                                 ksocknal_conn_decref(conn);
1591                         }
1592
1593                         did_something = 1;
1594                 }
1595                 if (!did_something ||           /* nothing to do */
1596                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1597                         spin_unlock_bh(&sched->kss_lock);
1598
1599                         nloops = 0;
1600
1601                         if (!did_something) {   /* wait for something to do */
1602                                 rc = wait_event_interruptible_exclusive(
1603                                         sched->kss_waitq,
1604                                         !ksocknal_sched_cansleep(sched));
1605                                 LASSERT (rc == 0);
1606                         } else {
1607                                 cond_resched();
1608                         }
1609
1610                         spin_lock_bh(&sched->kss_lock);
1611                 }
1612         }
1613
1614         spin_unlock_bh(&sched->kss_lock);
1615         CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV);
1616         CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV);
1617         ksocknal_thread_fini();
1618         return 0;
1619 }
1620
1621 /*
1622  * Add connection to kss_rx_conns of scheduler
1623  * and wakeup the scheduler.
1624  */
1625 void ksocknal_read_callback(struct ksock_conn *conn)
1626 {
1627         struct ksock_sched *sched;
1628         ENTRY;
1629
1630         sched = conn->ksnc_scheduler;
1631
1632         spin_lock_bh(&sched->kss_lock);
1633
1634         conn->ksnc_rx_ready = 1;
1635
1636         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1637                 list_add_tail(&conn->ksnc_rx_list,
1638                                   &sched->kss_rx_conns);
1639                 conn->ksnc_rx_scheduled = 1;
1640                 /* extra ref for scheduler */
1641                 ksocknal_conn_addref(conn);
1642
1643                 wake_up (&sched->kss_waitq);
1644         }
1645         spin_unlock_bh(&sched->kss_lock);
1646
1647         EXIT;
1648 }
1649
1650 /*
1651  * Add connection to kss_tx_conns of scheduler
1652  * and wakeup the scheduler.
1653  */
1654 void ksocknal_write_callback(struct ksock_conn *conn)
1655 {
1656         struct ksock_sched *sched;
1657         ENTRY;
1658
1659         sched = conn->ksnc_scheduler;
1660
1661         spin_lock_bh(&sched->kss_lock);
1662
1663         conn->ksnc_tx_ready = 1;
1664
1665         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1666             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1667                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1668                 conn->ksnc_tx_scheduled = 1;
1669                 /* extra ref for scheduler */
1670                 ksocknal_conn_addref(conn);
1671
1672                 wake_up(&sched->kss_waitq);
1673         }
1674
1675         spin_unlock_bh(&sched->kss_lock);
1676
1677         EXIT;
1678 }
1679
1680 static const struct ksock_proto *
1681 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1682 {
1683         __u32   version = 0;
1684
1685         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1686                 version = hello->kshm_version;
1687         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1688                 version = __swab32(hello->kshm_version);
1689
1690         if (version != 0) {
1691 #if SOCKNAL_VERSION_DEBUG
1692                 if (*ksocknal_tunables.ksnd_protocol == 1)
1693                         return NULL;
1694
1695                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1696                     version == KSOCK_PROTO_V3)
1697                         return NULL;
1698 #endif
1699                 if (version == KSOCK_PROTO_V2)
1700                         return &ksocknal_protocol_v2x;
1701
1702                 if (version == KSOCK_PROTO_V3)
1703                         return &ksocknal_protocol_v3x;
1704
1705                 return NULL;
1706         }
1707
1708         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1709                 struct lnet_magicversion *hmv;
1710
1711                 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1712                          offsetof(struct ksock_hello_msg, kshm_src_nid));
1713
1714                 hmv = (struct lnet_magicversion *)hello;
1715
1716                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1717                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1718                         return &ksocknal_protocol_v1x;
1719         }
1720
1721         return NULL;
1722 }
1723
1724 int
1725 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1726                     lnet_nid_t peer_nid, struct ksock_hello_msg *hello)
1727 {
1728         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1729         struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1730
1731         LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1732
1733         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1734         LASSERT(conn->ksnc_proto != NULL);
1735
1736         hello->kshm_src_nid         = ni->ni_nid;
1737         hello->kshm_dst_nid         = peer_nid;
1738         hello->kshm_src_pid         = the_lnet.ln_pid;
1739
1740         hello->kshm_src_incarnation = net->ksnn_incarnation;
1741         hello->kshm_ctype           = conn->ksnc_type;
1742
1743         return conn->ksnc_proto->pro_send_hello(conn, hello);
1744 }
1745
1746 static int
1747 ksocknal_invert_type(int type)
1748 {
1749         switch (type)
1750         {
1751         case SOCKLND_CONN_ANY:
1752         case SOCKLND_CONN_CONTROL:
1753                 return (type);
1754         case SOCKLND_CONN_BULK_IN:
1755                 return SOCKLND_CONN_BULK_OUT;
1756         case SOCKLND_CONN_BULK_OUT:
1757                 return SOCKLND_CONN_BULK_IN;
1758         default:
1759                 return (SOCKLND_CONN_NONE);
1760         }
1761 }
1762
1763 int
1764 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1765                     struct ksock_hello_msg *hello,
1766                     struct lnet_process_id *peerid,
1767                     __u64 *incarnation)
1768 {
1769         /* Return < 0        fatal error
1770          *        0          success
1771          *        EALREADY   lost connection race
1772          *        EPROTO     protocol version mismatch
1773          */
1774         struct socket        *sock = conn->ksnc_sock;
1775         int                  active = (conn->ksnc_proto != NULL);
1776         int                  timeout;
1777         int                  proto_match;
1778         int                  rc;
1779         const struct ksock_proto *proto;
1780         struct lnet_process_id recv_id;
1781
1782         /* socket type set on active connections - not set on passive */
1783         LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1784
1785         timeout = active ? ksocknal_timeout() :
1786                             lnet_acceptor_timeout();
1787
1788         rc = lnet_sock_read(sock, &hello->kshm_magic,
1789                             sizeof(hello->kshm_magic), timeout);
1790         if (rc != 0) {
1791                 CERROR("Error %d reading HELLO from %pI4h\n",
1792                        rc, &conn->ksnc_ipaddr);
1793                 LASSERT (rc < 0);
1794                 return rc;
1795         }
1796
1797         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1798             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1799             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1800                 /* Unexpected magic! */
1801                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1802                         "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1803                         LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr);
1804                 return -EPROTO;
1805         }
1806
1807         rc = lnet_sock_read(sock, &hello->kshm_version,
1808                             sizeof(hello->kshm_version), timeout);
1809         if (rc != 0) {
1810                 CERROR("Error %d reading HELLO from %pI4h\n",
1811                        rc, &conn->ksnc_ipaddr);
1812                 LASSERT(rc < 0);
1813                 return rc;
1814         }
1815
1816         proto = ksocknal_parse_proto_version(hello);
1817         if (proto == NULL) {
1818                 if (!active) {
1819                         /* unknown protocol from peer_ni, tell peer_ni my protocol */
1820                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1821 #if SOCKNAL_VERSION_DEBUG
1822                         if (*ksocknal_tunables.ksnd_protocol == 2)
1823                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1824                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1825                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1826 #endif
1827                         hello->kshm_nips = 0;
1828                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1829                 }
1830
1831                 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1832                        conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr);
1833
1834                 return -EPROTO;
1835         }
1836
1837         proto_match = (conn->ksnc_proto == proto);
1838         conn->ksnc_proto = proto;
1839
1840         /* receive the rest of hello message anyway */
1841         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1842         if (rc != 0) {
1843                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1844                        rc, &conn->ksnc_ipaddr);
1845                 LASSERT (rc < 0);
1846                 return rc;
1847         }
1848
1849         *incarnation = hello->kshm_src_incarnation;
1850
1851         if (hello->kshm_src_nid == LNET_NID_ANY) {
1852                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1853                        &conn->ksnc_ipaddr);
1854                 return -EPROTO;
1855         }
1856
1857         if (!active &&
1858             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1859                 /* Userspace NAL assigns peer_ni process ID from socket */
1860                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1861                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1862         } else {
1863                 recv_id.nid = hello->kshm_src_nid;
1864                 recv_id.pid = hello->kshm_src_pid;
1865         }
1866
1867         if (!active) {
1868                 *peerid = recv_id;
1869
1870                 /* peer_ni determines type */
1871                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1872                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1873                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1874                                hello->kshm_ctype, libcfs_id2str(*peerid),
1875                                &conn->ksnc_ipaddr);
1876                         return -EPROTO;
1877                 }
1878                 return 0;
1879         }
1880
1881         if (peerid->pid != recv_id.pid ||
1882             peerid->nid != recv_id.nid) {
1883                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1884                                    " %pI4h, but they claimed they were "
1885                                    "%s; please check your Lustre "
1886                                    "configuration.\n",
1887                                    libcfs_id2str(*peerid),
1888                                    &conn->ksnc_ipaddr,
1889                                    libcfs_id2str(recv_id));
1890                 return -EPROTO;
1891         }
1892
1893         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1894                 /* Possible protocol mismatch or I lost the connection race */
1895                 return proto_match ? EALREADY : EPROTO;
1896         }
1897
1898         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1899                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1900                        conn->ksnc_type, libcfs_id2str(*peerid),
1901                        &conn->ksnc_ipaddr,
1902                        hello->kshm_ctype);
1903                 return -EPROTO;
1904         }
1905         return 0;
1906 }
1907
1908 static int
1909 ksocknal_connect(struct ksock_route *route)
1910 {
1911         LIST_HEAD(zombies);
1912         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
1913         int               type;
1914         int               wanted;
1915         struct socket     *sock;
1916         time64_t deadline;
1917         int               retry_later = 0;
1918         int               rc = 0;
1919
1920         deadline = ktime_get_seconds() + ksocknal_timeout();
1921
1922         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1923
1924         LASSERT (route->ksnr_scheduled);
1925         LASSERT (!route->ksnr_connecting);
1926
1927         route->ksnr_connecting = 1;
1928
1929         for (;;) {
1930                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1931
1932                 /* stop connecting if peer_ni/route got closed under me, or
1933                  * route got connected while queued */
1934                 if (peer_ni->ksnp_closing || route->ksnr_deleted ||
1935                     wanted == 0) {
1936                         retry_later = 0;
1937                         break;
1938                 }
1939
1940                 /* reschedule if peer_ni is connecting to me */
1941                 if (peer_ni->ksnp_accepting > 0) {
1942                         CDEBUG(D_NET,
1943                                "peer_ni %s(%d) already connecting to me, retry later.\n",
1944                                libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting);
1945                         retry_later = 1;
1946                 }
1947
1948                 if (retry_later) /* needs reschedule */
1949                         break;
1950
1951                 if ((wanted & BIT(SOCKLND_CONN_ANY)) != 0) {
1952                         type = SOCKLND_CONN_ANY;
1953                 } else if ((wanted & BIT(SOCKLND_CONN_CONTROL)) != 0) {
1954                         type = SOCKLND_CONN_CONTROL;
1955                 } else if ((wanted & BIT(SOCKLND_CONN_BULK_IN)) != 0) {
1956                         type = SOCKLND_CONN_BULK_IN;
1957                 } else {
1958                         LASSERT ((wanted & BIT(SOCKLND_CONN_BULK_OUT)) != 0);
1959                         type = SOCKLND_CONN_BULK_OUT;
1960                 }
1961
1962                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1963
1964                 if (ktime_get_seconds() >= deadline) {
1965                         rc = -ETIMEDOUT;
1966                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1967                                                    route->ksnr_ipaddr,
1968                                                    route->ksnr_port);
1969                         goto failed;
1970                 }
1971
1972                 sock = lnet_connect(peer_ni->ksnp_id.nid,
1973                                     route->ksnr_myiface,
1974                                     route->ksnr_ipaddr, route->ksnr_port,
1975                                     peer_ni->ksnp_ni->ni_net_ns);
1976                 if (IS_ERR(sock)) {
1977                         rc = PTR_ERR(sock);
1978                         goto failed;
1979                 }
1980
1981                 rc = ksocknal_create_conn(peer_ni->ksnp_ni, route, sock, type);
1982                 if (rc < 0) {
1983                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1984                                                    route->ksnr_ipaddr,
1985                                                    route->ksnr_port);
1986                         goto failed;
1987                 }
1988
1989                 /* A +ve RC means I have to retry because I lost the connection
1990                  * race or I have to renegotiate protocol version */
1991                 retry_later = (rc != 0);
1992                 if (retry_later)
1993                         CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
1994                                libcfs_nid2str(peer_ni->ksnp_id.nid));
1995
1996                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1997         }
1998
1999         route->ksnr_scheduled = 0;
2000         route->ksnr_connecting = 0;
2001
2002         if (retry_later) {
2003                 /* re-queue for attention; this frees me up to handle
2004                  * the peer_ni's incoming connection request */
2005
2006                 if (rc == EALREADY ||
2007                     (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2008                         /* We want to introduce a delay before next
2009                          * attempt to connect if we lost conn race,
2010                          * but the race is resolved quickly usually,
2011                          * so min_reconnectms should be good heuristic */
2012                         route->ksnr_retry_interval = *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2013                         route->ksnr_timeout = ktime_get_seconds() +
2014                                               route->ksnr_retry_interval;
2015                 }
2016
2017                 ksocknal_launch_connection_locked(route);
2018         }
2019
2020         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2021         return retry_later;
2022
2023  failed:
2024         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2025
2026         route->ksnr_scheduled = 0;
2027         route->ksnr_connecting = 0;
2028
2029         /* This is a retry rather than a new connection */
2030         route->ksnr_retry_interval *= 2;
2031         route->ksnr_retry_interval =
2032                 max_t(time64_t, route->ksnr_retry_interval,
2033                       *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2034         route->ksnr_retry_interval =
2035                 min_t(time64_t, route->ksnr_retry_interval,
2036                       *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2037
2038         LASSERT(route->ksnr_retry_interval);
2039         route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval;
2040
2041         if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2042             peer_ni->ksnp_accepting == 0 &&
2043             ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
2044                 struct ksock_conn *conn;
2045
2046                 /* ksnp_tx_queue is queued on a conn on successful
2047                  * connection for V1.x and V2.x */
2048                 if (!list_empty(&peer_ni->ksnp_conns)) {
2049                         conn = list_entry(peer_ni->ksnp_conns.next,
2050                                           struct ksock_conn, ksnc_list);
2051                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
2052                 }
2053
2054                 /* take all the blocked packets while I've got the lock and
2055                  * complete below... */
2056                 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2057         }
2058
2059         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2060
2061         ksocknal_peer_failed(peer_ni);
2062         ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2063         return 0;
2064 }
2065
2066 /*
2067  * check whether we need to create more connds.
2068  * It will try to create new thread if it's necessary, @timeout can
2069  * be updated if failed to create, so caller wouldn't keep try while
2070  * running out of resource.
2071  */
2072 static int
2073 ksocknal_connd_check_start(time64_t sec, long *timeout)
2074 {
2075         char name[16];
2076         int rc;
2077         int total = ksocknal_data.ksnd_connd_starting +
2078                     ksocknal_data.ksnd_connd_running;
2079
2080         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2081                 /* still in initializing */
2082                 return 0;
2083         }
2084
2085         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2086             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2087                 /* can't create more connd, or still have enough
2088                  * threads to handle more connecting */
2089                 return 0;
2090         }
2091
2092         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2093                 /* no pending connecting request */
2094                 return 0;
2095         }
2096
2097         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2098                 /* may run out of resource, retry later */
2099                 *timeout = cfs_time_seconds(1);
2100                 return 0;
2101         }
2102
2103         if (ksocknal_data.ksnd_connd_starting > 0) {
2104                 /* serialize starting to avoid flood */
2105                 return 0;
2106         }
2107
2108         ksocknal_data.ksnd_connd_starting_stamp = sec;
2109         ksocknal_data.ksnd_connd_starting++;
2110         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2111
2112         /* NB: total is the next id */
2113         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2114         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2115
2116         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2117         if (rc == 0)
2118                 return 1;
2119
2120         /* we tried ... */
2121         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2122         ksocknal_data.ksnd_connd_starting--;
2123         ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2124
2125         return 1;
2126 }
2127
2128 /*
2129  * check whether current thread can exit, it will return 1 if there are too
2130  * many threads and no creating in past 120 seconds.
2131  * Also, this function may update @timeout to make caller come back
2132  * again to recheck these conditions.
2133  */
2134 static int
2135 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2136 {
2137         int val;
2138
2139         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2140                 /* still in initializing */
2141                 return 0;
2142         }
2143
2144         if (ksocknal_data.ksnd_connd_starting > 0) {
2145                 /* in progress of starting new thread */
2146                 return 0;
2147         }
2148
2149         if (ksocknal_data.ksnd_connd_running <=
2150             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2151                 return 0;
2152         }
2153
2154         /* created thread in past 120 seconds? */
2155         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2156                     SOCKNAL_CONND_TIMEOUT - sec);
2157
2158         *timeout = (val > 0) ? cfs_time_seconds(val) :
2159                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2160         if (val > 0)
2161                 return 0;
2162
2163         /* no creating in past 120 seconds */
2164
2165         return ksocknal_data.ksnd_connd_running >
2166                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2167 }
2168
2169 /* Go through connd_routes queue looking for a route that we can process
2170  * right now, @timeout_p can be updated if we need to come back later */
2171 static struct ksock_route *
2172 ksocknal_connd_get_route_locked(signed long *timeout_p)
2173 {
2174         time64_t now = ktime_get_seconds();
2175         struct ksock_route *route;
2176
2177         /* connd_routes can contain both pending and ordinary routes */
2178         list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2179                                  ksnr_connd_list) {
2180
2181                 if (route->ksnr_retry_interval == 0 ||
2182                     now >= route->ksnr_timeout)
2183                         return route;
2184
2185                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2186                     *timeout_p > cfs_time_seconds(route->ksnr_timeout - now))
2187                         *timeout_p = cfs_time_seconds(route->ksnr_timeout - now);
2188         }
2189
2190         return NULL;
2191 }
2192
2193 int
2194 ksocknal_connd(void *arg)
2195 {
2196         spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2197         struct ksock_connreq *cr;
2198         wait_queue_entry_t wait;
2199         int nloops = 0;
2200         int cons_retry = 0;
2201
2202         init_waitqueue_entry(&wait, current);
2203
2204         spin_lock_bh(connd_lock);
2205
2206         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2207         ksocknal_data.ksnd_connd_starting--;
2208         ksocknal_data.ksnd_connd_running++;
2209
2210         while (!ksocknal_data.ksnd_shuttingdown) {
2211                 struct ksock_route *route = NULL;
2212                 time64_t sec = ktime_get_real_seconds();
2213                 long timeout = MAX_SCHEDULE_TIMEOUT;
2214                 int  dropped_lock = 0;
2215
2216                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2217                         /* wakeup another one to check stop */
2218                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2219                         break;
2220                 }
2221
2222                 if (ksocknal_connd_check_start(sec, &timeout)) {
2223                         /* created new thread */
2224                         dropped_lock = 1;
2225                 }
2226
2227                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2228                         /* Connection accepted by the listener */
2229                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2230                                         struct ksock_connreq, ksncr_list);
2231
2232                         list_del(&cr->ksncr_list);
2233                         spin_unlock_bh(connd_lock);
2234                         dropped_lock = 1;
2235
2236                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2237                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2238                         lnet_ni_decref(cr->ksncr_ni);
2239                         LIBCFS_FREE(cr, sizeof(*cr));
2240
2241                         spin_lock_bh(connd_lock);
2242                 }
2243
2244                 /* Only handle an outgoing connection request if there
2245                  * is a thread left to handle incoming connections and
2246                  * create new connd */
2247                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2248                     ksocknal_data.ksnd_connd_running) {
2249                         route = ksocknal_connd_get_route_locked(&timeout);
2250                 }
2251                 if (route != NULL) {
2252                         list_del(&route->ksnr_connd_list);
2253                         ksocknal_data.ksnd_connd_connecting++;
2254                         spin_unlock_bh(connd_lock);
2255                         dropped_lock = 1;
2256
2257                         if (ksocknal_connect(route)) {
2258                                 /* consecutive retry */
2259                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2260                                         CWARN("massive consecutive "
2261                                               "re-connecting to %pI4h\n",
2262                                               &route->ksnr_ipaddr);
2263                                         cons_retry = 0;
2264                                 }
2265                         } else {
2266                                 cons_retry = 0;
2267                         }
2268
2269                         ksocknal_route_decref(route);
2270
2271                         spin_lock_bh(connd_lock);
2272                         ksocknal_data.ksnd_connd_connecting--;
2273                 }
2274
2275                 if (dropped_lock) {
2276                         if (++nloops < SOCKNAL_RESCHED)
2277                                 continue;
2278                         spin_unlock_bh(connd_lock);
2279                         nloops = 0;
2280                         cond_resched();
2281                         spin_lock_bh(connd_lock);
2282                         continue;
2283                 }
2284
2285                 /* Nothing to do for 'timeout'  */
2286                 set_current_state(TASK_INTERRUPTIBLE);
2287                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2288                 spin_unlock_bh(connd_lock);
2289
2290                 nloops = 0;
2291                 schedule_timeout(timeout);
2292
2293                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2294                 spin_lock_bh(connd_lock);
2295         }
2296         ksocknal_data.ksnd_connd_running--;
2297         spin_unlock_bh(connd_lock);
2298
2299         ksocknal_thread_fini();
2300         return 0;
2301 }
2302
2303 static struct ksock_conn *
2304 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2305 {
2306         /* We're called with a shared lock on ksnd_global_lock */
2307         struct ksock_conn *conn;
2308         struct list_head *ctmp;
2309         struct ksock_tx *tx;
2310
2311         list_for_each(ctmp, &peer_ni->ksnp_conns) {
2312                 int error;
2313
2314                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2315
2316                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2317                 LASSERT (!conn->ksnc_closing);
2318
2319                 error = conn->ksnc_sock->sk->sk_err;
2320                 if (error != 0) {
2321                         ksocknal_conn_addref(conn);
2322
2323                         switch (error) {
2324                         case ECONNRESET:
2325                                 CNETERR("A connection with %s "
2326                                         "(%pI4h:%d) was reset; "
2327                                         "it may have rebooted.\n",
2328                                         libcfs_id2str(peer_ni->ksnp_id),
2329                                         &conn->ksnc_ipaddr,
2330                                         conn->ksnc_port);
2331                                 break;
2332                         case ETIMEDOUT:
2333                                 CNETERR("A connection with %s "
2334                                         "(%pI4h:%d) timed out; the "
2335                                         "network or node may be down.\n",
2336                                         libcfs_id2str(peer_ni->ksnp_id),
2337                                         &conn->ksnc_ipaddr,
2338                                         conn->ksnc_port);
2339                                 break;
2340                         default:
2341                                 CNETERR("An unexpected network error %d "
2342                                         "occurred with %s "
2343                                         "(%pI4h:%d\n", error,
2344                                         libcfs_id2str(peer_ni->ksnp_id),
2345                                         &conn->ksnc_ipaddr,
2346                                         conn->ksnc_port);
2347                                 break;
2348                         }
2349
2350                         return (conn);
2351                 }
2352
2353                 if (conn->ksnc_rx_started &&
2354                     ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2355                         /* Timed out incomplete incoming message */
2356                         ksocknal_conn_addref(conn);
2357                         CNETERR("Timeout receiving from %s (%pI4h:%d), "
2358                                 "state %d wanted %d left %d\n",
2359                                 libcfs_id2str(peer_ni->ksnp_id),
2360                                 &conn->ksnc_ipaddr,
2361                                 conn->ksnc_port,
2362                                 conn->ksnc_rx_state,
2363                                 conn->ksnc_rx_nob_wanted,
2364                                 conn->ksnc_rx_nob_left);
2365                         return (conn);
2366                 }
2367
2368                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2369                      conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2370                     ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2371                         /* Timed out messages queued for sending or
2372                          * buffered in the socket's send buffer */
2373                         ksocknal_conn_addref(conn);
2374                         list_for_each_entry(tx, &conn->ksnc_tx_queue,
2375                                             tx_list)
2376                                 tx->tx_hstatus =
2377                                         LNET_MSG_STATUS_LOCAL_TIMEOUT;
2378                         CNETERR("Timeout sending data to %s (%pI4h:%d) "
2379                                 "the network or that node may be down.\n",
2380                                 libcfs_id2str(peer_ni->ksnp_id),
2381                                 &conn->ksnc_ipaddr, conn->ksnc_port);
2382                         return (conn);
2383                 }
2384         }
2385
2386         return (NULL);
2387 }
2388
2389 static inline void
2390 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2391 {
2392         struct ksock_tx *tx;
2393         LIST_HEAD(stale_txs);
2394
2395         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2396
2397         while (!list_empty(&peer_ni->ksnp_tx_queue)) {
2398                 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2399                                 struct ksock_tx, tx_list);
2400
2401                 if (ktime_get_seconds() < tx->tx_deadline)
2402                         break;
2403
2404                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2405
2406                 list_move_tail(&tx->tx_list, &stale_txs);
2407         }
2408
2409         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2410
2411         ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2412 }
2413
2414 static int
2415 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2416 __must_hold(&ksocknal_data.ksnd_global_lock)
2417 {
2418         struct ksock_sched *sched;
2419         struct ksock_conn *conn;
2420         struct ksock_tx *tx;
2421
2422         /* last_alive will be updated by create_conn */
2423         if (list_empty(&peer_ni->ksnp_conns))
2424                 return 0;
2425
2426         if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x)
2427                 return 0;
2428
2429         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2430             ktime_get_seconds() < peer_ni->ksnp_last_alive +
2431                                   *ksocknal_tunables.ksnd_keepalive)
2432                 return 0;
2433
2434         if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2435                 return 0;
2436
2437         /* retry 10 secs later, so we wouldn't put pressure
2438          * on this peer_ni if we failed to send keepalive this time */
2439         peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2440
2441         conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2442         if (conn != NULL) {
2443                 sched = conn->ksnc_scheduler;
2444
2445                 spin_lock_bh(&sched->kss_lock);
2446                 if (!list_empty(&conn->ksnc_tx_queue)) {
2447                         spin_unlock_bh(&sched->kss_lock);
2448                         /* there is an queued ACK, don't need keepalive */
2449                         return 0;
2450                 }
2451
2452                 spin_unlock_bh(&sched->kss_lock);
2453         }
2454
2455         read_unlock(&ksocknal_data.ksnd_global_lock);
2456
2457         /* cookie = 1 is reserved for keepalive PING */
2458         tx = ksocknal_alloc_tx_noop(1, 1);
2459         if (tx == NULL) {
2460                 read_lock(&ksocknal_data.ksnd_global_lock);
2461                 return -ENOMEM;
2462         }
2463
2464         if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) {
2465                 read_lock(&ksocknal_data.ksnd_global_lock);
2466                 return 1;
2467         }
2468
2469         ksocknal_free_tx(tx);
2470         read_lock(&ksocknal_data.ksnd_global_lock);
2471
2472         return -EIO;
2473 }
2474
2475
2476 static void
2477 ksocknal_check_peer_timeouts(int idx)
2478 {
2479         struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2480         struct ksock_peer_ni *peer_ni;
2481         struct ksock_conn *conn;
2482         struct ksock_tx *tx;
2483
2484  again:
2485         /* NB. We expect to have a look at all the peers and not find any
2486          * connections to time out, so we just use a shared lock while we
2487          * take a look...
2488          */
2489         read_lock(&ksocknal_data.ksnd_global_lock);
2490
2491         hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2492                 struct ksock_tx *tx_stale;
2493                 time64_t deadline = 0;
2494                 int resid = 0;
2495                 int n = 0;
2496
2497                 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2498                         read_unlock(&ksocknal_data.ksnd_global_lock);
2499                         goto again;
2500                 }
2501
2502                 conn = ksocknal_find_timed_out_conn(peer_ni);
2503
2504                 if (conn != NULL) {
2505                         read_unlock(&ksocknal_data.ksnd_global_lock);
2506
2507                         ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2508
2509                         /* NB we won't find this one again, but we can't
2510                          * just proceed with the next peer_ni, since we dropped
2511                          * ksnd_global_lock and it might be dead already!
2512                          */
2513                         ksocknal_conn_decref(conn);
2514                         goto again;
2515                 }
2516
2517                 /* we can't process stale txs right here because we're
2518                  * holding only shared lock
2519                  */
2520                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
2521                         struct ksock_tx *tx;
2522
2523                         tx = list_entry(peer_ni->ksnp_tx_queue.next,
2524                                         struct ksock_tx, tx_list);
2525                         if (ktime_get_seconds() >= tx->tx_deadline) {
2526                                 ksocknal_peer_addref(peer_ni);
2527                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2528
2529                                 ksocknal_flush_stale_txs(peer_ni);
2530
2531                                 ksocknal_peer_decref(peer_ni);
2532                                 goto again;
2533                         }
2534                 }
2535
2536                 if (list_empty(&peer_ni->ksnp_zc_req_list))
2537                         continue;
2538
2539                 tx_stale = NULL;
2540                 spin_lock(&peer_ni->ksnp_lock);
2541                 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2542                         if (ktime_get_seconds() < tx->tx_deadline)
2543                                 break;
2544                         /* ignore the TX if connection is being closed */
2545                         if (tx->tx_conn->ksnc_closing)
2546                                 continue;
2547                         n++;
2548                         if (tx_stale == NULL)
2549                                 tx_stale = tx;
2550                 }
2551
2552                 if (tx_stale == NULL) {
2553                         spin_unlock(&peer_ni->ksnp_lock);
2554                         continue;
2555                 }
2556
2557                 deadline = tx_stale->tx_deadline;
2558                 resid    = tx_stale->tx_resid;
2559                 conn     = tx_stale->tx_conn;
2560                 ksocknal_conn_addref(conn);
2561
2562                 spin_unlock(&peer_ni->ksnp_lock);
2563                 read_unlock(&ksocknal_data.ksnd_global_lock);
2564
2565                 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2566                        "oldest(%p) timed out %lld secs ago, "
2567                        "resid: %d, wmem: %d\n",
2568                        n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale,
2569                        ktime_get_seconds() - deadline,
2570                        resid, conn->ksnc_sock->sk->sk_wmem_queued);
2571
2572                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2573                 ksocknal_conn_decref(conn);
2574                 goto again;
2575         }
2576
2577         read_unlock(&ksocknal_data.ksnd_global_lock);
2578 }
2579
2580 int ksocknal_reaper(void *arg)
2581 {
2582         wait_queue_entry_t wait;
2583         struct ksock_conn *conn;
2584         struct ksock_sched *sched;
2585         LIST_HEAD(enomem_conns);
2586         int nenomem_conns;
2587         time64_t timeout;
2588         int i;
2589         int peer_index = 0;
2590         time64_t deadline = ktime_get_seconds();
2591
2592         init_waitqueue_entry(&wait, current);
2593
2594         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2595
2596         while (!ksocknal_data.ksnd_shuttingdown) {
2597                 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2598                         conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2599                                           struct ksock_conn, ksnc_list);
2600                         list_del(&conn->ksnc_list);
2601
2602                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2603
2604                         ksocknal_terminate_conn(conn);
2605                         ksocknal_conn_decref(conn);
2606
2607                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2608                         continue;
2609                 }
2610
2611                 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2612                         conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2613                                           struct ksock_conn, ksnc_list);
2614                         list_del(&conn->ksnc_list);
2615
2616                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2617
2618                         ksocknal_destroy_conn(conn);
2619
2620                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2621                         continue;
2622                 }
2623
2624                 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2625                                  &enomem_conns);
2626
2627                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2628
2629                 /* reschedule all the connections that stalled with ENOMEM... */
2630                 nenomem_conns = 0;
2631                 while (!list_empty(&enomem_conns)) {
2632                         conn = list_entry(enomem_conns.next,
2633                                           struct ksock_conn, ksnc_tx_list);
2634                         list_del(&conn->ksnc_tx_list);
2635
2636                         sched = conn->ksnc_scheduler;
2637
2638                         spin_lock_bh(&sched->kss_lock);
2639
2640                         LASSERT(conn->ksnc_tx_scheduled);
2641                         conn->ksnc_tx_ready = 1;
2642                         list_add_tail(&conn->ksnc_tx_list,
2643                                           &sched->kss_tx_conns);
2644                         wake_up(&sched->kss_waitq);
2645
2646                         spin_unlock_bh(&sched->kss_lock);
2647                         nenomem_conns++;
2648                 }
2649
2650                 /* careful with the jiffy wrap... */
2651                 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2652                         const int n = 4;
2653                         const int p = 1;
2654                         int  chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2655                         unsigned int lnd_timeout;
2656
2657                         /* Time to check for timeouts on a few more peers: I
2658                          * do checks every 'p' seconds on a proportion of the
2659                          * peer_ni table and I need to check every connection
2660                          * 'n' times within a timeout interval, to ensure I
2661                          * detect a timeout on any connection within (n+1)/n
2662                          * times the timeout interval.
2663                          */
2664
2665                         lnd_timeout = ksocknal_timeout();
2666                         if (lnd_timeout > n * p)
2667                                 chunk = (chunk * n * p) / lnd_timeout;
2668                         if (chunk == 0)
2669                                 chunk = 1;
2670
2671                         for (i = 0; i < chunk; i++) {
2672                                 ksocknal_check_peer_timeouts(peer_index);
2673                                 peer_index = (peer_index + 1) %
2674                                         HASH_SIZE(ksocknal_data.ksnd_peers);
2675                         }
2676
2677                         deadline += p;
2678                 }
2679
2680                 if (nenomem_conns != 0) {
2681                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2682                          * This also prevents me getting woken immediately
2683                          * if any go back on my enomem list. */
2684                         timeout = SOCKNAL_ENOMEM_RETRY;
2685                 }
2686                 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2687                                                      timeout;
2688
2689                 set_current_state(TASK_INTERRUPTIBLE);
2690                 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2691
2692                 if (!ksocknal_data.ksnd_shuttingdown &&
2693                     list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2694                     list_empty(&ksocknal_data.ksnd_zombie_conns))
2695                         schedule_timeout(cfs_time_seconds(timeout));
2696
2697                 set_current_state(TASK_RUNNING);
2698                 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2699
2700                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2701         }
2702
2703         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2704
2705         ksocknal_thread_fini();
2706         return 0;
2707 }