lnet/klnds/socklnd/socklnd_cb.c

   1 /*
   2  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2017, Intel Corporation.
   5  *
   6  *   Author: Zach Brown <zab@zabbo.net>
   7  *   Author: Peter J. Braam <braam@clusterfs.com>
   8  *   Author: Phil Schwan <phil@clusterfs.com>
   9  *   Author: Eric Barton <eric@bartonsoftware.com>
  10  *
  11  *   This file is part of Lustre, https://wiki.whamcloud.com/
  12  *
  13  *   Portals is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Portals is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Portals; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #include "socklnd.h"
  28
  29 struct ksock_tx *
  30 ksocknal_alloc_tx(int type, int size)
  31 {
  32         struct ksock_tx *tx = NULL;
  33
  34         if (type == KSOCK_MSG_NOOP) {
  35                 LASSERT(size == KSOCK_NOOP_TX_SIZE);
  36
  37                 /* searching for a noop tx in free list */
  38                 spin_lock(&ksocknal_data.ksnd_tx_lock);
  39
  40                 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
  41                         tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next,
  42                                         struct ksock_tx, tx_list);
  43                         LASSERT(tx->tx_desc_size == size);
  44                         list_del(&tx->tx_list);
  45                 }
  46
  47                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
  48         }
  49
  50         if (tx == NULL)
  51                 LIBCFS_ALLOC(tx, size);
  52
  53         if (tx == NULL)
  54                 return NULL;
  55
  56         atomic_set(&tx->tx_refcount, 1);
  57         tx->tx_zc_aborted = 0;
  58         tx->tx_zc_capable = 0;
  59         tx->tx_zc_checked = 0;
  60         tx->tx_hstatus = LNET_MSG_STATUS_OK;
  61         tx->tx_desc_size  = size;
  62
  63         atomic_inc(&ksocknal_data.ksnd_nactive_txs);
  64
  65         return tx;
  66 }
  67
  68 struct ksock_tx *
  69 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
  70 {
  71         struct ksock_tx *tx;
  72
  73         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
  74         if (tx == NULL) {
  75                 CERROR("Can't allocate noop tx desc\n");
  76                 return NULL;
  77         }
  78
  79         tx->tx_conn     = NULL;
  80         tx->tx_lnetmsg  = NULL;
  81         tx->tx_kiov     = NULL;
  82         tx->tx_nkiov    = 0;
  83         tx->tx_iov      = tx->tx_frags.virt.iov;
  84         tx->tx_niov     = 1;
  85         tx->tx_nonblk   = nonblk;
  86
  87         tx->tx_msg.ksm_csum = 0;
  88         tx->tx_msg.ksm_type = KSOCK_MSG_NOOP;
  89         tx->tx_msg.ksm_zc_cookies[0] = 0;
  90         tx->tx_msg.ksm_zc_cookies[1] = cookie;
  91
  92         return tx;
  93 }
  94
  95
  96 void
  97 ksocknal_free_tx(struct ksock_tx *tx)
  98 {
  99         atomic_dec(&ksocknal_data.ksnd_nactive_txs);
 100
 101         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
 102                 /* it's a noop tx */
 103                 spin_lock(&ksocknal_data.ksnd_tx_lock);
 104
 105                 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
 106
 107                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
 108         } else {
 109                 LIBCFS_FREE(tx, tx->tx_desc_size);
 110         }
 111 }
 112
 113 static int
 114 ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx,
 115                   struct kvec *scratch_iov)
 116 {
 117         struct kvec *iov = tx->tx_iov;
 118         int    nob;
 119         int    rc;
 120
 121         LASSERT(tx->tx_niov > 0);
 122
 123         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
 124         rc = ksocknal_lib_send_iov(conn, tx, scratch_iov);
 125
 126         if (rc <= 0)                            /* sent nothing? */
 127                 return rc;
 128
 129         nob = rc;
 130         LASSERT(nob <= tx->tx_resid);
 131         tx->tx_resid -= nob;
 132
 133         /* "consume" iov */
 134         do {
 135                 LASSERT(tx->tx_niov > 0);
 136
 137                 if (nob < (int) iov->iov_len) {
 138                         iov->iov_base += nob;
 139                         iov->iov_len -= nob;
 140                         return rc;
 141                 }
 142
 143                 nob -= iov->iov_len;
 144                 tx->tx_iov = ++iov;
 145                 tx->tx_niov--;
 146         } while (nob != 0);
 147
 148         return rc;
 149 }
 150
 151 static int
 152 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
 153                    struct kvec *scratch_iov)
 154 {
 155         struct bio_vec *kiov = tx->tx_kiov;
 156         int nob;
 157         int rc;
 158
 159         LASSERT(tx->tx_niov == 0);
 160         LASSERT(tx->tx_nkiov > 0);
 161
 162         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
 163         rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov);
 164
 165         if (rc <= 0)                            /* sent nothing? */
 166                 return rc;
 167
 168         nob = rc;
 169         LASSERT(nob <= tx->tx_resid);
 170         tx->tx_resid -= nob;
 171
 172         /* "consume" kiov */
 173         do {
 174                 LASSERT(tx->tx_nkiov > 0);
 175
 176                 if (nob < (int)kiov->bv_len) {
 177                         kiov->bv_offset += nob;
 178                         kiov->bv_len -= nob;
 179                         return rc;
 180                 }
 181
 182                 nob -= (int)kiov->bv_len;
 183                 tx->tx_kiov = ++kiov;
 184                 tx->tx_nkiov--;
 185         } while (nob != 0);
 186
 187         return rc;
 188 }
 189
 190 static int
 191 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 192                   struct kvec *scratch_iov)
 193 {
 194         int     rc;
 195         int     bufnob;
 196
 197         if (ksocknal_data.ksnd_stall_tx != 0)
 198                 schedule_timeout_uninterruptible(
 199                         cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
 200
 201         LASSERT(tx->tx_resid != 0);
 202
 203         rc = ksocknal_connsock_addref(conn);
 204         if (rc != 0) {
 205                 LASSERT(conn->ksnc_closing);
 206                 return -ESHUTDOWN;
 207         }
 208
 209         do {
 210                 if (ksocknal_data.ksnd_enomem_tx > 0) {
 211                         /* testing... */
 212                         ksocknal_data.ksnd_enomem_tx--;
 213                         rc = -EAGAIN;
 214                 } else if (tx->tx_niov != 0) {
 215                         rc = ksocknal_send_iov(conn, tx, scratch_iov);
 216                 } else {
 217                         rc = ksocknal_send_kiov(conn, tx, scratch_iov);
 218                 }
 219
 220                 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 221                 if (rc > 0)                     /* sent something? */
 222                         conn->ksnc_tx_bufnob += rc; /* account it */
 223
 224                 if (bufnob < conn->ksnc_tx_bufnob) {
 225                         /* allocated send buffer bytes < computed; infer
 226                          * something got ACKed */
 227                         conn->ksnc_tx_deadline = ktime_get_seconds() +
 228                                                  lnet_get_lnd_timeout();
 229                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 230                         conn->ksnc_tx_bufnob = bufnob;
 231                         smp_mb();
 232                 }
 233
 234                 if (rc <= 0) { /* Didn't write anything? */
 235                         /* some stacks return 0 instead of -EAGAIN */
 236                         if (rc == 0)
 237                                 rc = -EAGAIN;
 238
 239                         /* Check if EAGAIN is due to memory pressure */
 240                         if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
 241                                 rc = -ENOMEM;
 242
 243                         break;
 244                 }
 245
 246                 /* socket's wmem_queued now includes 'rc' bytes */
 247                 atomic_sub (rc, &conn->ksnc_tx_nob);
 248                 rc = 0;
 249
 250         } while (tx->tx_resid != 0);
 251
 252         ksocknal_connsock_decref(conn);
 253         return rc;
 254 }
 255
 256 static int
 257 ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
 258 {
 259         struct kvec *iov = conn->ksnc_rx_iov;
 260         int     nob;
 261         int     rc;
 262
 263         LASSERT(conn->ksnc_rx_niov > 0);
 264
 265         /* Never touch conn->ksnc_rx_iov or change connection
 266          * status inside ksocknal_lib_recv_iov */
 267         rc = ksocknal_lib_recv_iov(conn, scratchiov);
 268
 269         if (rc <= 0)
 270                 return rc;
 271
 272         /* received something... */
 273         nob = rc;
 274
 275         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 276         conn->ksnc_rx_deadline = ktime_get_seconds() +
 277                                  lnet_get_lnd_timeout();
 278         smp_mb();                       /* order with setting rx_started */
 279         conn->ksnc_rx_started = 1;
 280
 281         conn->ksnc_rx_nob_wanted -= nob;
 282         conn->ksnc_rx_nob_left -= nob;
 283
 284         do {
 285                 LASSERT(conn->ksnc_rx_niov > 0);
 286
 287                 if (nob < (int)iov->iov_len) {
 288                         iov->iov_len -= nob;
 289                         iov->iov_base += nob;
 290                         return -EAGAIN;
 291                 }
 292
 293                 nob -= iov->iov_len;
 294                 conn->ksnc_rx_iov = ++iov;
 295                 conn->ksnc_rx_niov--;
 296         } while (nob != 0);
 297
 298         return rc;
 299 }
 300
 301 static int
 302 ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 303                    struct kvec *scratch_iov)
 304 {
 305         struct bio_vec *kiov = conn->ksnc_rx_kiov;
 306         int nob;
 307         int rc;
 308         LASSERT(conn->ksnc_rx_nkiov > 0);
 309
 310         /* Never touch conn->ksnc_rx_kiov or change connection
 311          * status inside ksocknal_lib_recv_iov */
 312         rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov);
 313
 314         if (rc <= 0)
 315                 return rc;
 316
 317         /* received something... */
 318         nob = rc;
 319
 320         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 321         conn->ksnc_rx_deadline = ktime_get_seconds() +
 322                                  lnet_get_lnd_timeout();
 323         smp_mb();                       /* order with setting rx_started */
 324         conn->ksnc_rx_started = 1;
 325
 326         conn->ksnc_rx_nob_wanted -= nob;
 327         conn->ksnc_rx_nob_left -= nob;
 328
 329         do {
 330                 LASSERT(conn->ksnc_rx_nkiov > 0);
 331
 332                 if (nob < (int) kiov->bv_len) {
 333                         kiov->bv_offset += nob;
 334                         kiov->bv_len -= nob;
 335                         return -EAGAIN;
 336                 }
 337
 338                 nob -= kiov->bv_len;
 339                 conn->ksnc_rx_kiov = ++kiov;
 340                 conn->ksnc_rx_nkiov--;
 341         } while (nob != 0);
 342
 343         return 1;
 344 }
 345
 346 static int
 347 ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs,
 348                  struct kvec *scratch_iov)
 349 {
 350         /* Return 1 on success, 0 on EOF, < 0 on error.
 351          * Caller checks ksnc_rx_nob_wanted to determine
 352          * progress/completion. */
 353         int     rc;
 354         ENTRY;
 355
 356         if (ksocknal_data.ksnd_stall_rx != 0)
 357                 schedule_timeout_uninterruptible(
 358                         cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
 359
 360         rc = ksocknal_connsock_addref(conn);
 361         if (rc != 0) {
 362                 LASSERT(conn->ksnc_closing);
 363                 return -ESHUTDOWN;
 364         }
 365
 366         for (;;) {
 367                 if (conn->ksnc_rx_niov != 0)
 368                         rc = ksocknal_recv_iov(conn, scratch_iov);
 369                 else
 370                         rc = ksocknal_recv_kiov(conn, rx_scratch_pgs,
 371                                                  scratch_iov);
 372
 373                 if (rc <= 0) {
 374                         /* error/EOF or partial receive */
 375                         if (rc == -EAGAIN) {
 376                                 rc = 1;
 377                         } else if (rc == 0 && conn->ksnc_rx_started) {
 378                                 /* EOF in the middle of a message */
 379                                 rc = -EPROTO;
 380                         }
 381                         break;
 382                 }
 383
 384                 /* Completed a fragment */
 385
 386                 if (conn->ksnc_rx_nob_wanted == 0) {
 387                         rc = 1;
 388                         break;
 389                 }
 390         }
 391
 392         ksocknal_connsock_decref(conn);
 393         RETURN(rc);
 394 }
 395
 396 void
 397 ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc)
 398 {
 399         struct lnet_msg *lnetmsg = tx->tx_lnetmsg;
 400         enum lnet_msg_hstatus hstatus = tx->tx_hstatus;
 401         ENTRY;
 402
 403         LASSERT(ni != NULL || tx->tx_conn != NULL);
 404
 405         if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
 406                 rc = -EIO;
 407                 if (hstatus == LNET_MSG_STATUS_OK)
 408                         hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 409         }
 410
 411         if (tx->tx_conn != NULL)
 412                 ksocknal_conn_decref(tx->tx_conn);
 413
 414         ksocknal_free_tx(tx);
 415         if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
 416                 lnetmsg->msg_health_status = hstatus;
 417                 lnet_finalize(lnetmsg, rc);
 418         }
 419
 420         EXIT;
 421 }
 422
 423 void
 424 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
 425 {
 426         struct ksock_tx *tx;
 427
 428         while (!list_empty(txlist)) {
 429                 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
 430
 431                 if (error && tx->tx_lnetmsg != NULL) {
 432                         CNETERR("Deleting packet type %d len %d %s->%s\n",
 433                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
 434                                 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
 435                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
 436                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
 437                 } else if (error) {
 438                         CNETERR("Deleting noop packet\n");
 439                 }
 440
 441                 list_del(&tx->tx_list);
 442
 443                 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
 444                         if (error == -ETIMEDOUT)
 445                                 tx->tx_hstatus =
 446                                   LNET_MSG_STATUS_LOCAL_TIMEOUT;
 447                         else if (error == -ENETDOWN ||
 448                                  error == -EHOSTUNREACH ||
 449                                  error == -ENETUNREACH ||
 450                                  error == -ECONNREFUSED ||
 451                                  error == -ECONNRESET)
 452                                 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
 453                         /*
 454                          * for all other errors we don't want to
 455                          * retransmit
 456                          */
 457                         else if (error)
 458                                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 459                 }
 460
 461                 LASSERT(atomic_read(&tx->tx_refcount) == 1);
 462                 ksocknal_tx_done(ni, tx, error);
 463         }
 464 }
 465
 466 static void
 467 ksocknal_check_zc_req(struct ksock_tx *tx)
 468 {
 469         struct ksock_conn *conn = tx->tx_conn;
 470         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
 471
 472         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
 473          * to ksnp_zc_req_list if some fragment of this message should be sent
 474          * zero-copy.  Our peer_ni will send an ACK containing this cookie when
 475          * she has received this message to tell us we can signal completion.
 476          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
 477          * ksnp_zc_req_list. */
 478         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 479         LASSERT (tx->tx_zc_capable);
 480
 481         tx->tx_zc_checked = 1;
 482
 483         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
 484             !conn->ksnc_zc_capable)
 485                 return;
 486
 487         /* assign cookie and queue tx to pending list, it will be released when
 488          * a matching ack is received. See ksocknal_handle_zcack() */
 489
 490         ksocknal_tx_addref(tx);
 491
 492         spin_lock(&peer_ni->ksnp_lock);
 493
 494         /* ZC_REQ is going to be pinned to the peer_ni */
 495         tx->tx_deadline = ktime_get_seconds() +
 496                           lnet_get_lnd_timeout();
 497
 498         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
 499
 500         tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
 501
 502         if (peer_ni->ksnp_zc_next_cookie == 0)
 503                 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 504
 505         list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
 506
 507         spin_unlock(&peer_ni->ksnp_lock);
 508 }
 509
 510 static void
 511 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
 512 {
 513         struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
 514
 515         LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
 516         LASSERT(tx->tx_zc_capable);
 517
 518         tx->tx_zc_checked = 0;
 519
 520         spin_lock(&peer_ni->ksnp_lock);
 521
 522         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
 523                 /* Not waiting for an ACK */
 524                 spin_unlock(&peer_ni->ksnp_lock);
 525                 return;
 526         }
 527
 528         tx->tx_msg.ksm_zc_cookies[0] = 0;
 529         list_del(&tx->tx_zc_list);
 530
 531         spin_unlock(&peer_ni->ksnp_lock);
 532
 533         ksocknal_tx_decref(tx);
 534 }
 535
 536 static int
 537 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
 538                           struct kvec *scratch_iov)
 539 {
 540         int rc;
 541         bool error_sim = false;
 542
 543         if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
 544                 error_sim = true;
 545                 rc = -EINVAL;
 546                 goto simulate_error;
 547         }
 548
 549         if (tx->tx_zc_capable && !tx->tx_zc_checked)
 550                 ksocknal_check_zc_req(tx);
 551
 552         rc = ksocknal_transmit(conn, tx, scratch_iov);
 553
 554         CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
 555
 556         if (tx->tx_resid == 0) {
 557                 /* Sent everything OK */
 558                 LASSERT(rc == 0);
 559
 560                 return 0;
 561         }
 562
 563         if (rc == -EAGAIN)
 564                 return rc;
 565
 566         if (rc == -ENOMEM) {
 567                 static int counter;
 568
 569                 counter++;   /* exponential backoff warnings */
 570                 if ((counter & (-counter)) == counter)
 571                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
 572                               counter, conn, atomic_read(&libcfs_kmemory));
 573
 574                 /* Queue on ksnd_enomem_conns for retry after a timeout */
 575                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
 576
 577                 /* enomem list takes over scheduler's ref... */
 578                 LASSERT(conn->ksnc_tx_scheduled);
 579                 list_add_tail(&conn->ksnc_tx_list,
 580                                   &ksocknal_data.ksnd_enomem_conns);
 581                 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
 582                     ksocknal_data.ksnd_reaper_waketime)
 583                         wake_up(&ksocknal_data.ksnd_reaper_waitq);
 584
 585                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
 586
 587                 /*
 588                  * set the health status of the message which determines
 589                  * whether we should retry the transmit
 590                  */
 591                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 592                 return (rc);
 593         }
 594
 595 simulate_error:
 596
 597         /* Actual error */
 598         LASSERT(rc < 0);
 599
 600         if (!error_sim) {
 601                 /*
 602                 * set the health status of the message which determines
 603                 * whether we should retry the transmit
 604                 */
 605                 if (rc == -ETIMEDOUT)
 606                         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
 607                 else
 608                         tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
 609         }
 610
 611         if (!conn->ksnc_closing) {
 612                 switch (rc) {
 613                 case -ECONNRESET:
 614                         LCONSOLE_WARN("Host %pI4h reset our connection "
 615                                       "while we were sending data; it may have "
 616                                       "rebooted.\n",
 617                                       &conn->ksnc_ipaddr);
 618                         break;
 619                 default:
 620                         LCONSOLE_WARN("There was an unexpected network error "
 621                                       "while writing to %pI4h: %d.\n",
 622                                       &conn->ksnc_ipaddr, rc);
 623                         break;
 624                 }
 625                 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
 626                        conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id),
 627                        &conn->ksnc_ipaddr, conn->ksnc_port);
 628         }
 629
 630         if (tx->tx_zc_checked)
 631                 ksocknal_uncheck_zc_req(tx);
 632
 633         /* it's not an error if conn is being closed */
 634         ksocknal_close_conn_and_siblings(conn,
 635                                           (conn->ksnc_closing) ? 0 : rc);
 636
 637         return rc;
 638 }
 639
 640 static void
 641 ksocknal_launch_connection_locked(struct ksock_route *route)
 642 {
 643
 644         /* called holding write lock on ksnd_global_lock */
 645
 646         LASSERT (!route->ksnr_scheduled);
 647         LASSERT (!route->ksnr_connecting);
 648         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
 649
 650         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
 651         ksocknal_route_addref(route);           /* extra ref for connd */
 652
 653         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 654
 655         list_add_tail(&route->ksnr_connd_list,
 656                           &ksocknal_data.ksnd_connd_routes);
 657         wake_up(&ksocknal_data.ksnd_connd_waitq);
 658
 659         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 660 }
 661
 662 void
 663 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
 664 {
 665         struct ksock_route *route;
 666
 667         /* called holding write lock on ksnd_global_lock */
 668         for (;;) {
 669                 /* launch any/all connections that need it */
 670                 route = ksocknal_find_connectable_route_locked(peer_ni);
 671                 if (route == NULL)
 672                         return;
 673
 674                 ksocknal_launch_connection_locked(route);
 675         }
 676 }
 677
 678 struct ksock_conn *
 679 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
 680 {
 681         struct list_head *tmp;
 682         struct ksock_conn *conn;
 683         struct ksock_conn *typed = NULL;
 684         struct ksock_conn *fallback = NULL;
 685         int tnob = 0;
 686         int fnob = 0;
 687
 688         list_for_each(tmp, &peer_ni->ksnp_conns) {
 689                 struct ksock_conn *c = list_entry(tmp, struct ksock_conn,
 690                                                   ksnc_list);
 691                 int nob = atomic_read(&c->ksnc_tx_nob) +
 692                           c->ksnc_sock->sk->sk_wmem_queued;
 693                 int rc;
 694
 695                 LASSERT (!c->ksnc_closing);
 696                 LASSERT (c->ksnc_proto != NULL &&
 697                          c->ksnc_proto->pro_match_tx != NULL);
 698
 699                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
 700
 701                 switch (rc) {
 702                 default:
 703                         LBUG();
 704                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
 705                         continue;
 706
 707                 case SOCKNAL_MATCH_YES: /* typed connection */
 708                         if (typed == NULL || tnob > nob ||
 709                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 710                              typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 711                                 typed = c;
 712                                 tnob  = nob;
 713                         }
 714                         break;
 715
 716                 case SOCKNAL_MATCH_MAY: /* fallback connection */
 717                         if (fallback == NULL || fnob > nob ||
 718                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
 719                              fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
 720                                 fallback = c;
 721                                 fnob     = nob;
 722                         }
 723                         break;
 724                 }
 725         }
 726
 727         /* prefer the typed selection */
 728         conn = (typed != NULL) ? typed : fallback;
 729
 730         if (conn != NULL)
 731                 conn->ksnc_tx_last_post = ktime_get_seconds();
 732
 733         return conn;
 734 }
 735
 736 void
 737 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
 738 {
 739         conn->ksnc_proto->pro_pack(tx);
 740
 741         atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
 742         ksocknal_conn_addref(conn); /* +1 ref for tx */
 743         tx->tx_conn = conn;
 744 }
 745
 746 void
 747 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
 748 {
 749         struct ksock_sched *sched = conn->ksnc_scheduler;
 750         struct ksock_msg *msg = &tx->tx_msg;
 751         struct ksock_tx *ztx = NULL;
 752         int bufnob = 0;
 753
 754         /* called holding global lock (read or irq-write) and caller may
 755          * not have dropped this lock between finding conn and calling me,
 756          * so we don't need the {get,put}connsock dance to deref
 757          * ksnc_sock... */
 758         LASSERT(!conn->ksnc_closing);
 759
 760         CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
 761                libcfs_id2str(conn->ksnc_peer->ksnp_id),
 762                &conn->ksnc_ipaddr, conn->ksnc_port);
 763
 764         ksocknal_tx_prep(conn, tx);
 765
 766         /* Ensure the frags we've been given EXACTLY match the number of
 767          * bytes we want to send.  Many TCP/IP stacks disregard any total
 768          * size parameters passed to them and just look at the frags.
 769          *
 770          * We always expect at least 1 mapped fragment containing the
 771          * complete ksocknal message header. */
 772         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
 773                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
 774                  (unsigned int)tx->tx_nob);
 775         LASSERT (tx->tx_niov >= 1);
 776         LASSERT (tx->tx_resid == tx->tx_nob);
 777
 778         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
 779                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
 780                                                KSOCK_MSG_NOOP,
 781                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
 782
 783         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
 784         spin_lock_bh(&sched->kss_lock);
 785
 786         if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
 787                 /* First packet starts the timeout */
 788                 conn->ksnc_tx_deadline = ktime_get_seconds() +
 789                                          lnet_get_lnd_timeout();
 790                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
 791                         conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
 792                 conn->ksnc_tx_bufnob = 0;
 793                 smp_mb(); /* order with adding to tx_queue */
 794         }
 795
 796         if (msg->ksm_type == KSOCK_MSG_NOOP) {
 797                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
 798                  * on a normal packet so I don't need to send it */
 799                 LASSERT (msg->ksm_zc_cookies[1] != 0);
 800                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
 801
 802                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
 803                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
 804
 805         } else {
 806                 /* It's a normal packet - can it piggback a noop zc-ack that
 807                  * has been queued already? */
 808                 LASSERT (msg->ksm_zc_cookies[1] == 0);
 809                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
 810
 811                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
 812                 /* ztx will be released later */
 813         }
 814
 815         if (ztx != NULL) {
 816                 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
 817                 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
 818         }
 819
 820         if (conn->ksnc_tx_ready &&      /* able to send */
 821             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
 822                 /* +1 ref for scheduler */
 823                 ksocknal_conn_addref(conn);
 824                 list_add_tail(&conn->ksnc_tx_list,
 825                                    &sched->kss_tx_conns);
 826                 conn->ksnc_tx_scheduled = 1;
 827                 wake_up(&sched->kss_waitq);
 828         }
 829
 830         spin_unlock_bh(&sched->kss_lock);
 831 }
 832
 833
 834 struct ksock_route *
 835 ksocknal_find_connectable_route_locked(struct ksock_peer_ni *peer_ni)
 836 {
 837         time64_t now = ktime_get_seconds();
 838         struct list_head *tmp;
 839         struct ksock_route *route;
 840
 841         list_for_each(tmp, &peer_ni->ksnp_routes) {
 842                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 843
 844                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 845
 846                 if (route->ksnr_scheduled)      /* connections being established */
 847                         continue;
 848
 849                 /* all route types connected ? */
 850                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
 851                         continue;
 852
 853                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
 854                       now >= route->ksnr_timeout)) {
 855                         CDEBUG(D_NET,
 856                                "Too soon to retry route %pI4h "
 857                                "(cnted %d, interval %lld, %lld secs later)\n",
 858                                &route->ksnr_ipaddr,
 859                                route->ksnr_connected,
 860                                route->ksnr_retry_interval,
 861                                route->ksnr_timeout - now);
 862                         continue;
 863                 }
 864
 865                 return (route);
 866         }
 867
 868         return (NULL);
 869 }
 870
 871 struct ksock_route *
 872 ksocknal_find_connecting_route_locked(struct ksock_peer_ni *peer_ni)
 873 {
 874         struct list_head *tmp;
 875         struct ksock_route *route;
 876
 877         list_for_each(tmp, &peer_ni->ksnp_routes) {
 878                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 879
 880                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
 881
 882                 if (route->ksnr_scheduled)
 883                         return (route);
 884         }
 885
 886         return (NULL);
 887 }
 888
 889 int
 890 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
 891                        struct lnet_process_id id)
 892 {
 893         struct ksock_peer_ni *peer_ni;
 894         struct ksock_conn *conn;
 895         rwlock_t *g_lock;
 896         int retry;
 897         int rc;
 898
 899         LASSERT (tx->tx_conn == NULL);
 900
 901         g_lock = &ksocknal_data.ksnd_global_lock;
 902
 903         for (retry = 0;; retry = 1) {
 904                 read_lock(g_lock);
 905                 peer_ni = ksocknal_find_peer_locked(ni, id);
 906                 if (peer_ni != NULL) {
 907                         if (ksocknal_find_connectable_route_locked(peer_ni) == NULL) {
 908                                 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 909                                 if (conn != NULL) {
 910                                         /* I've got no routes that need to be
 911                                          * connecting and I do have an actual
 912                                          * connection... */
 913                                         ksocknal_queue_tx_locked (tx, conn);
 914                                         read_unlock(g_lock);
 915                                         return (0);
 916                                 }
 917                         }
 918                 }
 919
 920                 /* I'll need a write lock... */
 921                 read_unlock(g_lock);
 922
 923                 write_lock_bh(g_lock);
 924
 925                 peer_ni = ksocknal_find_peer_locked(ni, id);
 926                 if (peer_ni != NULL)
 927                         break;
 928
 929                 write_unlock_bh(g_lock);
 930
 931                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
 932                         CERROR("Refusing to create a connection to "
 933                                "userspace process %s\n", libcfs_id2str(id));
 934                         return -EHOSTUNREACH;
 935                 }
 936
 937                 if (retry) {
 938                         CERROR("Can't find peer_ni %s\n", libcfs_id2str(id));
 939                         return -EHOSTUNREACH;
 940                 }
 941
 942                 rc = ksocknal_add_peer(ni, id,
 943                                        LNET_NIDADDR(id.nid),
 944                                        lnet_acceptor_port());
 945                 if (rc != 0) {
 946                         CERROR("Can't add peer_ni %s: %d\n",
 947                                libcfs_id2str(id), rc);
 948                         return rc;
 949                 }
 950         }
 951
 952         ksocknal_launch_all_connections_locked(peer_ni);
 953
 954         conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
 955         if (conn != NULL) {
 956                 /* Connection exists; queue message on it */
 957                 ksocknal_queue_tx_locked (tx, conn);
 958                 write_unlock_bh(g_lock);
 959                 return (0);
 960         }
 961
 962         if (peer_ni->ksnp_accepting > 0 ||
 963             ksocknal_find_connecting_route_locked (peer_ni) != NULL) {
 964                 /* the message is going to be pinned to the peer_ni */
 965                 tx->tx_deadline = ktime_get_seconds() +
 966                                   lnet_get_lnd_timeout();
 967
 968                 /* Queue the message until a connection is established */
 969                 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
 970                 write_unlock_bh(g_lock);
 971                 return 0;
 972         }
 973
 974         write_unlock_bh(g_lock);
 975
 976         /* NB Routes may be ignored if connections to them failed recently */
 977         CNETERR("No usable routes to %s\n", libcfs_id2str(id));
 978         tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
 979         return (-EHOSTUNREACH);
 980 }
 981
 982 int
 983 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
 984 {
 985         int mpflag = 1;
 986         int type = lntmsg->msg_type;
 987         struct lnet_process_id target = lntmsg->msg_target;
 988         unsigned int     payload_niov = lntmsg->msg_niov;
 989         struct kvec *payload_iov = lntmsg->msg_iov;
 990         struct bio_vec  *payload_kiov = lntmsg->msg_kiov;
 991         unsigned int     payload_offset = lntmsg->msg_offset;
 992         unsigned int     payload_nob = lntmsg->msg_len;
 993         struct ksock_tx *tx;
 994         int              desc_size;
 995         int              rc;
 996
 997         /* NB 'private' is different depending on what we're sending.
 998          * Just ignore it... */
 999
1000         CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
1001                payload_nob, payload_niov, libcfs_id2str(target));
1002
1003         LASSERT (payload_nob == 0 || payload_niov > 0);
1004         LASSERT (payload_niov <= LNET_MAX_IOV);
1005         /* payload is either all vaddrs or all pages */
1006         LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1007         LASSERT (!in_interrupt ());
1008
1009         if (payload_iov != NULL)
1010                 desc_size = offsetof(struct ksock_tx,
1011                                      tx_frags.virt.iov[1 + payload_niov]);
1012         else
1013                 desc_size = offsetof(struct ksock_tx,
1014                                      tx_frags.paged.kiov[payload_niov]);
1015
1016         if (lntmsg->msg_vmflush)
1017                 mpflag = cfs_memory_pressure_get_and_set();
1018         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1019         if (tx == NULL) {
1020                 CERROR("Can't allocate tx desc type %d size %d\n",
1021                        type, desc_size);
1022                 if (lntmsg->msg_vmflush)
1023                         cfs_memory_pressure_restore(mpflag);
1024                 return (-ENOMEM);
1025         }
1026
1027         tx->tx_conn = NULL;                     /* set when assigned a conn */
1028         tx->tx_lnetmsg = lntmsg;
1029
1030         if (payload_iov != NULL) {
1031                 tx->tx_kiov = NULL;
1032                 tx->tx_nkiov = 0;
1033                 tx->tx_iov = tx->tx_frags.virt.iov;
1034                 tx->tx_niov = 1 +
1035                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
1036                                                payload_niov, payload_iov,
1037                                                payload_offset, payload_nob);
1038         } else {
1039                 tx->tx_niov = 1;
1040                 tx->tx_iov = &tx->tx_frags.paged.iov;
1041                 tx->tx_kiov = tx->tx_frags.paged.kiov;
1042                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1043                                                  payload_niov, payload_kiov,
1044                                                  payload_offset, payload_nob);
1045
1046                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1047                         tx->tx_zc_capable = 1;
1048         }
1049
1050         tx->tx_msg.ksm_csum = 0;
1051         tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1052         tx->tx_msg.ksm_zc_cookies[0] = 0;
1053         tx->tx_msg.ksm_zc_cookies[1] = 0;
1054
1055         /* The first fragment will be set later in pro_pack */
1056         rc = ksocknal_launch_packet(ni, tx, target);
1057         if (!mpflag)
1058                 cfs_memory_pressure_restore(mpflag);
1059
1060         if (rc == 0)
1061                 return (0);
1062
1063         lntmsg->msg_health_status = tx->tx_hstatus;
1064         ksocknal_free_tx(tx);
1065         return (-EIO);
1066 }
1067
1068 int
1069 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1070 {
1071         struct task_struct *task = kthread_run(fn, arg, name);
1072
1073         if (IS_ERR(task))
1074                 return PTR_ERR(task);
1075
1076         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1077         ksocknal_data.ksnd_nthreads++;
1078         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1079         return 0;
1080 }
1081
1082 void
1083 ksocknal_thread_fini (void)
1084 {
1085         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1086         if (--ksocknal_data.ksnd_nthreads == 0)
1087                 wake_up_var(&ksocknal_data.ksnd_nthreads);
1088         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1089 }
1090
1091 int
1092 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1093 {
1094         static char ksocknal_slop_buffer[4096];
1095         int nob;
1096         unsigned int niov;
1097         int skipped;
1098
1099         LASSERT(conn->ksnc_proto != NULL);
1100
1101         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1102                 /* Remind the socket to ack eagerly... */
1103                 ksocknal_lib_eager_ack(conn);
1104         }
1105
1106         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1107                 conn->ksnc_rx_started = 0;
1108                 smp_mb();                       /* racing with timeout thread */
1109
1110                 switch (conn->ksnc_proto->pro_version) {
1111                 case  KSOCK_PROTO_V2:
1112                 case  KSOCK_PROTO_V3:
1113                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1114                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1115                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1116
1117                         conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u);
1118                         conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u);
1119                         conn->ksnc_rx_iov[0].iov_len  = offsetof(struct ksock_msg, ksm_u);
1120                         break;
1121
1122                 case KSOCK_PROTO_V1:
1123                         /* Receiving bare struct lnet_hdr */
1124                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1125                         conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr);
1126                         conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr);
1127
1128                         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1129                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1130                         conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr);
1131                         break;
1132
1133                 default:
1134                         LBUG ();
1135                 }
1136                 conn->ksnc_rx_niov = 1;
1137
1138                 conn->ksnc_rx_kiov = NULL;
1139                 conn->ksnc_rx_nkiov = 0;
1140                 conn->ksnc_rx_csum = ~0;
1141                 return (1);
1142         }
1143
1144         /* Set up to skip as much as possible now.  If there's more left
1145          * (ran out of iov entries) we'll get called again */
1146
1147         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1148         conn->ksnc_rx_nob_left = nob_to_skip;
1149         conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1150         skipped = 0;
1151         niov = 0;
1152
1153         do {
1154                 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1155
1156                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1157                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1158                 niov++;
1159                 skipped += nob;
1160                 nob_to_skip -= nob;
1161
1162         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1163                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1164
1165         conn->ksnc_rx_niov = niov;
1166         conn->ksnc_rx_kiov = NULL;
1167         conn->ksnc_rx_nkiov = 0;
1168         conn->ksnc_rx_nob_wanted = skipped;
1169         return (0);
1170 }
1171
1172 static int
1173 ksocknal_process_receive(struct ksock_conn *conn,
1174                          struct page **rx_scratch_pgs,
1175                          struct kvec *scratch_iov)
1176 {
1177         struct lnet_hdr *lhdr;
1178         struct lnet_process_id *id;
1179         int rc;
1180
1181         LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1182
1183         /* NB: sched lock NOT held */
1184         /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1185         LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1186                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1187                 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1188                 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1189  again:
1190         if (conn->ksnc_rx_nob_wanted != 0) {
1191                 rc = ksocknal_receive(conn, rx_scratch_pgs,
1192                                       scratch_iov);
1193
1194                 if (rc <= 0) {
1195                         struct lnet_process_id ksnp_id;
1196
1197                         ksnp_id = conn->ksnc_peer->ksnp_id;
1198
1199                         LASSERT(rc != -EAGAIN);
1200                         if (rc == 0)
1201                                 CDEBUG(D_NET, "[%p] EOF from %s "
1202                                        "ip %pI4h:%d\n", conn,
1203                                         libcfs_id2str(ksnp_id),
1204                                         &conn->ksnc_ipaddr,
1205                                         conn->ksnc_port);
1206                         else if (!conn->ksnc_closing)
1207                                 CERROR("[%p] Error %d on read from %s "
1208                                        "ip %pI4h:%d\n", conn, rc,
1209                                        libcfs_id2str(ksnp_id),
1210                                        &conn->ksnc_ipaddr,
1211                                        conn->ksnc_port);
1212
1213                         /* it's not an error if conn is being closed */
1214                         ksocknal_close_conn_and_siblings (conn,
1215                                                           (conn->ksnc_closing) ? 0 : rc);
1216                         return (rc == 0 ? -ESHUTDOWN : rc);
1217                 }
1218
1219                 if (conn->ksnc_rx_nob_wanted != 0) {
1220                         /* short read */
1221                         return (-EAGAIN);
1222                 }
1223         }
1224         switch (conn->ksnc_rx_state) {
1225         case SOCKNAL_RX_KSM_HEADER:
1226                 if (conn->ksnc_flip) {
1227                         __swab32s(&conn->ksnc_msg.ksm_type);
1228                         __swab32s(&conn->ksnc_msg.ksm_csum);
1229                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1230                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1231                 }
1232
1233                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1234                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1235                         CERROR("%s: Unknown message type: %x\n",
1236                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1237                                conn->ksnc_msg.ksm_type);
1238                         ksocknal_new_packet(conn, 0);
1239                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1240                         return (-EPROTO);
1241                 }
1242
1243                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1244                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1245                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1246                         /* NOOP Checksum error */
1247                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1248                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1249                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1250                         ksocknal_new_packet(conn, 0);
1251                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1252                         return (-EIO);
1253                 }
1254
1255                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1256                         __u64 cookie = 0;
1257
1258                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1259
1260                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1261                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1262
1263                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1264                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1265
1266                         if (rc != 0) {
1267                                 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1268                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1269                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1270                                 ksocknal_new_packet(conn, 0);
1271                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1272                                 return (rc);
1273                         }
1274                 }
1275
1276                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1277                         ksocknal_new_packet (conn, 0);
1278                         return 0;       /* NOOP is done and just return */
1279                 }
1280
1281                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1282                 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg);
1283                 conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg);
1284
1285                 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1286                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1287                 conn->ksnc_rx_iov[0].iov_len  = sizeof(struct ksock_lnet_msg);
1288
1289                 conn->ksnc_rx_niov = 1;
1290                 conn->ksnc_rx_kiov = NULL;
1291                 conn->ksnc_rx_nkiov = 0;
1292
1293                 goto again;     /* read lnet header now */
1294
1295         case SOCKNAL_RX_LNET_HEADER:
1296                 /* unpack message header */
1297                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1298
1299                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1300                         /* Userspace peer_ni */
1301                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1302                         id   = &conn->ksnc_peer->ksnp_id;
1303
1304                         /* Substitute process ID assigned at connection time */
1305                         lhdr->src_pid = cpu_to_le32(id->pid);
1306                         lhdr->src_nid = cpu_to_le64(id->nid);
1307                 }
1308
1309                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1310                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1311
1312                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1313                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1314                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1315                 if (rc < 0) {
1316                         /* I just received garbage: give up on this conn */
1317                         ksocknal_new_packet(conn, 0);
1318                         ksocknal_close_conn_and_siblings (conn, rc);
1319                         ksocknal_conn_decref(conn);
1320                         return (-EPROTO);
1321                 }
1322
1323                 /* I'm racing with ksocknal_recv() */
1324                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1325                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1326
1327                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1328                         return 0;
1329
1330                 /* ksocknal_recv() got called */
1331                 goto again;
1332
1333         case SOCKNAL_RX_LNET_PAYLOAD:
1334                 /* payload all received */
1335                 rc = 0;
1336
1337                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1338                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1339                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1340                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1341                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1342                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1343                         rc = -EIO;
1344                 }
1345
1346                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1347                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1348
1349                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1350                         id   = &conn->ksnc_peer->ksnp_id;
1351
1352                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1353                                         conn->ksnc_msg.ksm_zc_cookies[0],
1354                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1355                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1356                 }
1357
1358                 if (rc && conn->ksnc_lnet_msg)
1359                         conn->ksnc_lnet_msg->msg_health_status =
1360                                 LNET_MSG_STATUS_REMOTE_ERROR;
1361                 lnet_finalize(conn->ksnc_lnet_msg, rc);
1362
1363                 if (rc != 0) {
1364                         ksocknal_new_packet(conn, 0);
1365                         ksocknal_close_conn_and_siblings (conn, rc);
1366                         return (-EPROTO);
1367                 }
1368                 /* Fall through */
1369
1370         case SOCKNAL_RX_SLOP:
1371                 /* starting new packet? */
1372                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1373                         return 0;       /* come back later */
1374                 goto again;             /* try to finish reading slop now */
1375
1376         default:
1377                 break;
1378         }
1379
1380         /* Not Reached */
1381         LBUG ();
1382         return (-EINVAL);                       /* keep gcc happy */
1383 }
1384
1385 int
1386 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1387               int delayed, unsigned int niov, struct kvec *iov,
1388               struct bio_vec *kiov, unsigned int offset, unsigned int mlen,
1389               unsigned int rlen)
1390 {
1391         struct ksock_conn *conn = private;
1392         struct ksock_sched *sched = conn->ksnc_scheduler;
1393
1394         LASSERT (mlen <= rlen);
1395         LASSERT (niov <= LNET_MAX_IOV);
1396
1397         conn->ksnc_lnet_msg = msg;
1398         conn->ksnc_rx_nob_wanted = mlen;
1399         conn->ksnc_rx_nob_left   = rlen;
1400
1401         if (mlen == 0 || iov != NULL) {
1402                 conn->ksnc_rx_nkiov = 0;
1403                 conn->ksnc_rx_kiov = NULL;
1404                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1405                 conn->ksnc_rx_niov =
1406                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1407                                          niov, iov, offset, mlen);
1408         } else {
1409                 conn->ksnc_rx_niov = 0;
1410                 conn->ksnc_rx_iov  = NULL;
1411                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1412                 conn->ksnc_rx_nkiov =
1413                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1414                                           niov, kiov, offset, mlen);
1415         }
1416
1417         LASSERT (mlen ==
1418                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1419                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1420
1421         LASSERT (conn->ksnc_rx_scheduled);
1422
1423         spin_lock_bh(&sched->kss_lock);
1424
1425         switch (conn->ksnc_rx_state) {
1426         case SOCKNAL_RX_PARSE_WAIT:
1427                 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1428                 wake_up(&sched->kss_waitq);
1429                 LASSERT(conn->ksnc_rx_ready);
1430                 break;
1431
1432         case SOCKNAL_RX_PARSE:
1433                 /* scheduler hasn't noticed I'm parsing yet */
1434                 break;
1435         }
1436
1437         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1438
1439         spin_unlock_bh(&sched->kss_lock);
1440         ksocknal_conn_decref(conn);
1441         return 0;
1442 }
1443
1444 static inline int
1445 ksocknal_sched_cansleep(struct ksock_sched *sched)
1446 {
1447         int           rc;
1448
1449         spin_lock_bh(&sched->kss_lock);
1450
1451         rc = (!ksocknal_data.ksnd_shuttingdown &&
1452               list_empty(&sched->kss_rx_conns) &&
1453               list_empty(&sched->kss_tx_conns));
1454
1455         spin_unlock_bh(&sched->kss_lock);
1456         return rc;
1457 }
1458
1459 int ksocknal_scheduler(void *arg)
1460 {
1461         struct ksock_sched *sched;
1462         struct ksock_conn *conn;
1463         struct ksock_tx *tx;
1464         int rc;
1465         int nloops = 0;
1466         long id = (long)arg;
1467         struct page **rx_scratch_pgs;
1468         struct kvec *scratch_iov;
1469
1470         sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1471
1472         LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1473                          sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1474         if (!rx_scratch_pgs) {
1475                 CERROR("Unable to allocate scratch pages\n");
1476                 return -ENOMEM;
1477         }
1478
1479         LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1480                          sizeof(*scratch_iov) * LNET_MAX_IOV);
1481         if (!scratch_iov) {
1482                 CERROR("Unable to allocate scratch iov\n");
1483                 return -ENOMEM;
1484         }
1485
1486         rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1487         if (rc != 0) {
1488                 CWARN("Can't set CPU partition affinity to %d: %d\n",
1489                         sched->kss_cpt, rc);
1490         }
1491
1492         spin_lock_bh(&sched->kss_lock);
1493
1494         while (!ksocknal_data.ksnd_shuttingdown) {
1495                 int did_something = 0;
1496
1497                 /* Ensure I progress everything semi-fairly */
1498
1499                 if (!list_empty(&sched->kss_rx_conns)) {
1500                         conn = list_entry(sched->kss_rx_conns.next,
1501                                           struct ksock_conn, ksnc_rx_list);
1502                         list_del(&conn->ksnc_rx_list);
1503
1504                         LASSERT(conn->ksnc_rx_scheduled);
1505                         LASSERT(conn->ksnc_rx_ready);
1506
1507                         /* clear rx_ready in case receive isn't complete.
1508                          * Do it BEFORE we call process_recv, since
1509                          * data_ready can set it any time after we release
1510                          * kss_lock. */
1511                         conn->ksnc_rx_ready = 0;
1512                         spin_unlock_bh(&sched->kss_lock);
1513
1514                         rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1515                                                       scratch_iov);
1516
1517                         spin_lock_bh(&sched->kss_lock);
1518
1519                         /* I'm the only one that can clear this flag */
1520                         LASSERT(conn->ksnc_rx_scheduled);
1521
1522                         /* Did process_receive get everything it wanted? */
1523                         if (rc == 0)
1524                                 conn->ksnc_rx_ready = 1;
1525
1526                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1527                                 /* Conn blocked waiting for ksocknal_recv()
1528                                  * I change its state (under lock) to signal
1529                                  * it can be rescheduled */
1530                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1531                         } else if (conn->ksnc_rx_ready) {
1532                                 /* reschedule for rx */
1533                                 list_add_tail(&conn->ksnc_rx_list,
1534                                                    &sched->kss_rx_conns);
1535                         } else {
1536                                 conn->ksnc_rx_scheduled = 0;
1537                                 /* drop my ref */
1538                                 ksocknal_conn_decref(conn);
1539                         }
1540
1541                         did_something = 1;
1542                 }
1543
1544                 if (!list_empty(&sched->kss_tx_conns)) {
1545                         LIST_HEAD(zlist);
1546
1547                         list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1548
1549                         conn = list_entry(sched->kss_tx_conns.next,
1550                                           struct ksock_conn, ksnc_tx_list);
1551                         list_del(&conn->ksnc_tx_list);
1552
1553                         LASSERT(conn->ksnc_tx_scheduled);
1554                         LASSERT(conn->ksnc_tx_ready);
1555                         LASSERT(!list_empty(&conn->ksnc_tx_queue));
1556
1557                         tx = list_entry(conn->ksnc_tx_queue.next,
1558                                         struct ksock_tx, tx_list);
1559
1560                         if (conn->ksnc_tx_carrier == tx)
1561                                 ksocknal_next_tx_carrier(conn);
1562
1563                         /* dequeue now so empty list => more to send */
1564                         list_del(&tx->tx_list);
1565
1566                         /* Clear tx_ready in case send isn't complete.  Do
1567                          * it BEFORE we call process_transmit, since
1568                          * write_space can set it any time after we release
1569                          * kss_lock. */
1570                         conn->ksnc_tx_ready = 0;
1571                         spin_unlock_bh(&sched->kss_lock);
1572
1573                         if (!list_empty(&zlist)) {
1574                                 /* free zombie noop txs, it's fast because
1575                                  * noop txs are just put in freelist */
1576                                 ksocknal_txlist_done(NULL, &zlist, 0);
1577                         }
1578
1579                         rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1580
1581                         if (rc == -ENOMEM || rc == -EAGAIN) {
1582                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1583                                 spin_lock_bh(&sched->kss_lock);
1584                                 list_add(&tx->tx_list,
1585                                          &conn->ksnc_tx_queue);
1586                         } else {
1587                                 /* Complete send; tx -ref */
1588                                 ksocknal_tx_decref(tx);
1589
1590                                 spin_lock_bh(&sched->kss_lock);
1591                                 /* assume space for more */
1592                                 conn->ksnc_tx_ready = 1;
1593                         }
1594
1595                         if (rc == -ENOMEM) {
1596                                 /* Do nothing; after a short timeout, this
1597                                  * conn will be reposted on kss_tx_conns. */
1598                         } else if (conn->ksnc_tx_ready &&
1599                                    !list_empty(&conn->ksnc_tx_queue)) {
1600                                 /* reschedule for tx */
1601                                 list_add_tail(&conn->ksnc_tx_list,
1602                                               &sched->kss_tx_conns);
1603                         } else {
1604                                 conn->ksnc_tx_scheduled = 0;
1605                                 /* drop my ref */
1606                                 ksocknal_conn_decref(conn);
1607                         }
1608
1609                         did_something = 1;
1610                 }
1611                 if (!did_something ||           /* nothing to do */
1612                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1613                         spin_unlock_bh(&sched->kss_lock);
1614
1615                         nloops = 0;
1616
1617                         if (!did_something) {   /* wait for something to do */
1618                                 rc = wait_event_interruptible_exclusive(
1619                                         sched->kss_waitq,
1620                                         !ksocknal_sched_cansleep(sched));
1621                                 LASSERT (rc == 0);
1622                         } else {
1623                                 cond_resched();
1624                         }
1625
1626                         spin_lock_bh(&sched->kss_lock);
1627                 }
1628         }
1629
1630         spin_unlock_bh(&sched->kss_lock);
1631         CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV);
1632         CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV);
1633         ksocknal_thread_fini();
1634         return 0;
1635 }
1636
1637 /*
1638  * Add connection to kss_rx_conns of scheduler
1639  * and wakeup the scheduler.
1640  */
1641 void ksocknal_read_callback(struct ksock_conn *conn)
1642 {
1643         struct ksock_sched *sched;
1644         ENTRY;
1645
1646         sched = conn->ksnc_scheduler;
1647
1648         spin_lock_bh(&sched->kss_lock);
1649
1650         conn->ksnc_rx_ready = 1;
1651
1652         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1653                 list_add_tail(&conn->ksnc_rx_list,
1654                                   &sched->kss_rx_conns);
1655                 conn->ksnc_rx_scheduled = 1;
1656                 /* extra ref for scheduler */
1657                 ksocknal_conn_addref(conn);
1658
1659                 wake_up (&sched->kss_waitq);
1660         }
1661         spin_unlock_bh(&sched->kss_lock);
1662
1663         EXIT;
1664 }
1665
1666 /*
1667  * Add connection to kss_tx_conns of scheduler
1668  * and wakeup the scheduler.
1669  */
1670 void ksocknal_write_callback(struct ksock_conn *conn)
1671 {
1672         struct ksock_sched *sched;
1673         ENTRY;
1674
1675         sched = conn->ksnc_scheduler;
1676
1677         spin_lock_bh(&sched->kss_lock);
1678
1679         conn->ksnc_tx_ready = 1;
1680
1681         if (!conn->ksnc_tx_scheduled && /* not being progressed */
1682             !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1683                 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1684                 conn->ksnc_tx_scheduled = 1;
1685                 /* extra ref for scheduler */
1686                 ksocknal_conn_addref(conn);
1687
1688                 wake_up(&sched->kss_waitq);
1689         }
1690
1691         spin_unlock_bh(&sched->kss_lock);
1692
1693         EXIT;
1694 }
1695
1696 static const struct ksock_proto *
1697 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1698 {
1699         __u32   version = 0;
1700
1701         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1702                 version = hello->kshm_version;
1703         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1704                 version = __swab32(hello->kshm_version);
1705
1706         if (version != 0) {
1707 #if SOCKNAL_VERSION_DEBUG
1708                 if (*ksocknal_tunables.ksnd_protocol == 1)
1709                         return NULL;
1710
1711                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1712                     version == KSOCK_PROTO_V3)
1713                         return NULL;
1714 #endif
1715                 if (version == KSOCK_PROTO_V2)
1716                         return &ksocknal_protocol_v2x;
1717
1718                 if (version == KSOCK_PROTO_V3)
1719                         return &ksocknal_protocol_v3x;
1720
1721                 return NULL;
1722         }
1723
1724         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1725                 struct lnet_magicversion *hmv;
1726
1727                 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1728                          offsetof(struct ksock_hello_msg, kshm_src_nid));
1729
1730                 hmv = (struct lnet_magicversion *)hello;
1731
1732                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1733                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1734                         return &ksocknal_protocol_v1x;
1735         }
1736
1737         return NULL;
1738 }
1739
1740 int
1741 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1742                     lnet_nid_t peer_nid, struct ksock_hello_msg *hello)
1743 {
1744         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1745         struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1746
1747         LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1748
1749         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1750         LASSERT(conn->ksnc_proto != NULL);
1751
1752         hello->kshm_src_nid         = ni->ni_nid;
1753         hello->kshm_dst_nid         = peer_nid;
1754         hello->kshm_src_pid         = the_lnet.ln_pid;
1755
1756         hello->kshm_src_incarnation = net->ksnn_incarnation;
1757         hello->kshm_ctype           = conn->ksnc_type;
1758
1759         return conn->ksnc_proto->pro_send_hello(conn, hello);
1760 }
1761
1762 static int
1763 ksocknal_invert_type(int type)
1764 {
1765         switch (type)
1766         {
1767         case SOCKLND_CONN_ANY:
1768         case SOCKLND_CONN_CONTROL:
1769                 return (type);
1770         case SOCKLND_CONN_BULK_IN:
1771                 return SOCKLND_CONN_BULK_OUT;
1772         case SOCKLND_CONN_BULK_OUT:
1773                 return SOCKLND_CONN_BULK_IN;
1774         default:
1775                 return (SOCKLND_CONN_NONE);
1776         }
1777 }
1778
1779 int
1780 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1781                     struct ksock_hello_msg *hello,
1782                     struct lnet_process_id *peerid,
1783                     __u64 *incarnation)
1784 {
1785         /* Return < 0        fatal error
1786          *        0          success
1787          *        EALREADY   lost connection race
1788          *        EPROTO     protocol version mismatch
1789          */
1790         struct socket        *sock = conn->ksnc_sock;
1791         int                  active = (conn->ksnc_proto != NULL);
1792         int                  timeout;
1793         int                  proto_match;
1794         int                  rc;
1795         const struct ksock_proto *proto;
1796         struct lnet_process_id recv_id;
1797
1798         /* socket type set on active connections - not set on passive */
1799         LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1800
1801         timeout = active ? lnet_get_lnd_timeout() :
1802                             lnet_acceptor_timeout();
1803
1804         rc = lnet_sock_read(sock, &hello->kshm_magic,
1805                             sizeof(hello->kshm_magic), timeout);
1806         if (rc != 0) {
1807                 CERROR("Error %d reading HELLO from %pI4h\n",
1808                        rc, &conn->ksnc_ipaddr);
1809                 LASSERT (rc < 0);
1810                 return rc;
1811         }
1812
1813         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1814             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1815             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1816                 /* Unexpected magic! */
1817                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1818                         "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1819                         LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr);
1820                 return -EPROTO;
1821         }
1822
1823         rc = lnet_sock_read(sock, &hello->kshm_version,
1824                             sizeof(hello->kshm_version), timeout);
1825         if (rc != 0) {
1826                 CERROR("Error %d reading HELLO from %pI4h\n",
1827                        rc, &conn->ksnc_ipaddr);
1828                 LASSERT(rc < 0);
1829                 return rc;
1830         }
1831
1832         proto = ksocknal_parse_proto_version(hello);
1833         if (proto == NULL) {
1834                 if (!active) {
1835                         /* unknown protocol from peer_ni, tell peer_ni my protocol */
1836                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1837 #if SOCKNAL_VERSION_DEBUG
1838                         if (*ksocknal_tunables.ksnd_protocol == 2)
1839                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1840                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1841                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1842 #endif
1843                         hello->kshm_nips = 0;
1844                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1845                 }
1846
1847                 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1848                        conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr);
1849
1850                 return -EPROTO;
1851         }
1852
1853         proto_match = (conn->ksnc_proto == proto);
1854         conn->ksnc_proto = proto;
1855
1856         /* receive the rest of hello message anyway */
1857         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1858         if (rc != 0) {
1859                 CERROR("Error %d reading or checking hello from from %pI4h\n",
1860                        rc, &conn->ksnc_ipaddr);
1861                 LASSERT (rc < 0);
1862                 return rc;
1863         }
1864
1865         *incarnation = hello->kshm_src_incarnation;
1866
1867         if (hello->kshm_src_nid == LNET_NID_ANY) {
1868                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1869                        &conn->ksnc_ipaddr);
1870                 return -EPROTO;
1871         }
1872
1873         if (!active &&
1874             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1875                 /* Userspace NAL assigns peer_ni process ID from socket */
1876                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1877                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1878         } else {
1879                 recv_id.nid = hello->kshm_src_nid;
1880                 recv_id.pid = hello->kshm_src_pid;
1881         }
1882
1883         if (!active) {
1884                 *peerid = recv_id;
1885
1886                 /* peer_ni determines type */
1887                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1888                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1889                         CERROR("Unexpected type %d from %s ip %pI4h\n",
1890                                hello->kshm_ctype, libcfs_id2str(*peerid),
1891                                &conn->ksnc_ipaddr);
1892                         return -EPROTO;
1893                 }
1894                 return 0;
1895         }
1896
1897         if (peerid->pid != recv_id.pid ||
1898             peerid->nid != recv_id.nid) {
1899                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1900                                    " %pI4h, but they claimed they were "
1901                                    "%s; please check your Lustre "
1902                                    "configuration.\n",
1903                                    libcfs_id2str(*peerid),
1904                                    &conn->ksnc_ipaddr,
1905                                    libcfs_id2str(recv_id));
1906                 return -EPROTO;
1907         }
1908
1909         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1910                 /* Possible protocol mismatch or I lost the connection race */
1911                 return proto_match ? EALREADY : EPROTO;
1912         }
1913
1914         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1915                 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1916                        conn->ksnc_type, libcfs_id2str(*peerid),
1917                        &conn->ksnc_ipaddr,
1918                        hello->kshm_ctype);
1919                 return -EPROTO;
1920         }
1921         return 0;
1922 }
1923
1924 static int
1925 ksocknal_connect(struct ksock_route *route)
1926 {
1927         LIST_HEAD(zombies);
1928         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
1929         int               type;
1930         int               wanted;
1931         struct socket     *sock;
1932         time64_t deadline;
1933         int               retry_later = 0;
1934         int               rc = 0;
1935
1936         deadline = ktime_get_seconds() + lnet_get_lnd_timeout();
1937
1938         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1939
1940         LASSERT (route->ksnr_scheduled);
1941         LASSERT (!route->ksnr_connecting);
1942
1943         route->ksnr_connecting = 1;
1944
1945         for (;;) {
1946                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1947
1948                 /* stop connecting if peer_ni/route got closed under me, or
1949                  * route got connected while queued */
1950                 if (peer_ni->ksnp_closing || route->ksnr_deleted ||
1951                     wanted == 0) {
1952                         retry_later = 0;
1953                         break;
1954                 }
1955
1956                 /* reschedule if peer_ni is connecting to me */
1957                 if (peer_ni->ksnp_accepting > 0) {
1958                         CDEBUG(D_NET,
1959                                "peer_ni %s(%d) already connecting to me, retry later.\n",
1960                                libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting);
1961                         retry_later = 1;
1962                 }
1963
1964                 if (retry_later) /* needs reschedule */
1965                         break;
1966
1967                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1968                         type = SOCKLND_CONN_ANY;
1969                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1970                         type = SOCKLND_CONN_CONTROL;
1971                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1972                         type = SOCKLND_CONN_BULK_IN;
1973                 } else {
1974                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1975                         type = SOCKLND_CONN_BULK_OUT;
1976                 }
1977
1978                 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1979
1980                 if (ktime_get_seconds() >= deadline) {
1981                         rc = -ETIMEDOUT;
1982                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1983                                                    route->ksnr_ipaddr,
1984                                                    route->ksnr_port);
1985                         goto failed;
1986                 }
1987
1988                 sock = lnet_connect(peer_ni->ksnp_id.nid,
1989                                     route->ksnr_myipaddr,
1990                                     route->ksnr_ipaddr, route->ksnr_port,
1991                                     peer_ni->ksnp_ni->ni_net_ns);
1992                 if (IS_ERR(sock)) {
1993                         rc = PTR_ERR(sock);
1994                         goto failed;
1995                 }
1996
1997                 rc = ksocknal_create_conn(peer_ni->ksnp_ni, route, sock, type);
1998                 if (rc < 0) {
1999                         lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
2000                                                    route->ksnr_ipaddr,
2001                                                    route->ksnr_port);
2002                         goto failed;
2003                 }
2004
2005                 /* A +ve RC means I have to retry because I lost the connection
2006                  * race or I have to renegotiate protocol version */
2007                 retry_later = (rc != 0);
2008                 if (retry_later)
2009                         CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
2010                                libcfs_nid2str(peer_ni->ksnp_id.nid));
2011
2012                 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2013         }
2014
2015         route->ksnr_scheduled = 0;
2016         route->ksnr_connecting = 0;
2017
2018         if (retry_later) {
2019                 /* re-queue for attention; this frees me up to handle
2020                  * the peer_ni's incoming connection request */
2021
2022                 if (rc == EALREADY ||
2023                     (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2024                         /* We want to introduce a delay before next
2025                          * attempt to connect if we lost conn race,
2026                          * but the race is resolved quickly usually,
2027                          * so min_reconnectms should be good heuristic */
2028                         route->ksnr_retry_interval = *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2029                         route->ksnr_timeout = ktime_get_seconds() +
2030                                               route->ksnr_retry_interval;
2031                 }
2032
2033                 ksocknal_launch_connection_locked(route);
2034         }
2035
2036         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2037         return retry_later;
2038
2039  failed:
2040         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2041
2042         route->ksnr_scheduled = 0;
2043         route->ksnr_connecting = 0;
2044
2045         /* This is a retry rather than a new connection */
2046         route->ksnr_retry_interval *= 2;
2047         route->ksnr_retry_interval =
2048                 max_t(time64_t, route->ksnr_retry_interval,
2049                       *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2050         route->ksnr_retry_interval =
2051                 min_t(time64_t, route->ksnr_retry_interval,
2052                       *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2053
2054         LASSERT(route->ksnr_retry_interval);
2055         route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval;
2056
2057         if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2058             peer_ni->ksnp_accepting == 0 &&
2059             ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
2060                 struct ksock_conn *conn;
2061
2062                 /* ksnp_tx_queue is queued on a conn on successful
2063                  * connection for V1.x and V2.x */
2064                 if (!list_empty(&peer_ni->ksnp_conns)) {
2065                         conn = list_entry(peer_ni->ksnp_conns.next,
2066                                           struct ksock_conn, ksnc_list);
2067                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
2068                 }
2069
2070                 /* take all the blocked packets while I've got the lock and
2071                  * complete below... */
2072                 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2073         }
2074
2075         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2076
2077         ksocknal_peer_failed(peer_ni);
2078         ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2079         return 0;
2080 }
2081
2082 /*
2083  * check whether we need to create more connds.
2084  * It will try to create new thread if it's necessary, @timeout can
2085  * be updated if failed to create, so caller wouldn't keep try while
2086  * running out of resource.
2087  */
2088 static int
2089 ksocknal_connd_check_start(time64_t sec, long *timeout)
2090 {
2091         char name[16];
2092         int rc;
2093         int total = ksocknal_data.ksnd_connd_starting +
2094                     ksocknal_data.ksnd_connd_running;
2095
2096         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2097                 /* still in initializing */
2098                 return 0;
2099         }
2100
2101         if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2102             total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2103                 /* can't create more connd, or still have enough
2104                  * threads to handle more connecting */
2105                 return 0;
2106         }
2107
2108         if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2109                 /* no pending connecting request */
2110                 return 0;
2111         }
2112
2113         if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2114                 /* may run out of resource, retry later */
2115                 *timeout = cfs_time_seconds(1);
2116                 return 0;
2117         }
2118
2119         if (ksocknal_data.ksnd_connd_starting > 0) {
2120                 /* serialize starting to avoid flood */
2121                 return 0;
2122         }
2123
2124         ksocknal_data.ksnd_connd_starting_stamp = sec;
2125         ksocknal_data.ksnd_connd_starting++;
2126         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2127
2128         /* NB: total is the next id */
2129         snprintf(name, sizeof(name), "socknal_cd%02d", total);
2130         rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2131
2132         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2133         if (rc == 0)
2134                 return 1;
2135
2136         /* we tried ... */
2137         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2138         ksocknal_data.ksnd_connd_starting--;
2139         ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2140
2141         return 1;
2142 }
2143
2144 /*
2145  * check whether current thread can exit, it will return 1 if there are too
2146  * many threads and no creating in past 120 seconds.
2147  * Also, this function may update @timeout to make caller come back
2148  * again to recheck these conditions.
2149  */
2150 static int
2151 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2152 {
2153         int val;
2154
2155         if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2156                 /* still in initializing */
2157                 return 0;
2158         }
2159
2160         if (ksocknal_data.ksnd_connd_starting > 0) {
2161                 /* in progress of starting new thread */
2162                 return 0;
2163         }
2164
2165         if (ksocknal_data.ksnd_connd_running <=
2166             *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2167                 return 0;
2168         }
2169
2170         /* created thread in past 120 seconds? */
2171         val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2172                     SOCKNAL_CONND_TIMEOUT - sec);
2173
2174         *timeout = (val > 0) ? cfs_time_seconds(val) :
2175                                cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2176         if (val > 0)
2177                 return 0;
2178
2179         /* no creating in past 120 seconds */
2180
2181         return ksocknal_data.ksnd_connd_running >
2182                ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2183 }
2184
2185 /* Go through connd_routes queue looking for a route that we can process
2186  * right now, @timeout_p can be updated if we need to come back later */
2187 static struct ksock_route *
2188 ksocknal_connd_get_route_locked(signed long *timeout_p)
2189 {
2190         time64_t now = ktime_get_seconds();
2191         struct ksock_route *route;
2192
2193         /* connd_routes can contain both pending and ordinary routes */
2194         list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2195                                  ksnr_connd_list) {
2196
2197                 if (route->ksnr_retry_interval == 0 ||
2198                     now >= route->ksnr_timeout)
2199                         return route;
2200
2201                 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2202                     *timeout_p > cfs_time_seconds(route->ksnr_timeout - now))
2203                         *timeout_p = cfs_time_seconds(route->ksnr_timeout - now);
2204         }
2205
2206         return NULL;
2207 }
2208
2209 int
2210 ksocknal_connd(void *arg)
2211 {
2212         spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2213         struct ksock_connreq *cr;
2214         wait_queue_entry_t wait;
2215         int nloops = 0;
2216         int cons_retry = 0;
2217
2218         init_waitqueue_entry(&wait, current);
2219
2220         spin_lock_bh(connd_lock);
2221
2222         LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2223         ksocknal_data.ksnd_connd_starting--;
2224         ksocknal_data.ksnd_connd_running++;
2225
2226         while (!ksocknal_data.ksnd_shuttingdown) {
2227                 struct ksock_route *route = NULL;
2228                 time64_t sec = ktime_get_real_seconds();
2229                 long timeout = MAX_SCHEDULE_TIMEOUT;
2230                 int  dropped_lock = 0;
2231
2232                 if (ksocknal_connd_check_stop(sec, &timeout)) {
2233                         /* wakeup another one to check stop */
2234                         wake_up(&ksocknal_data.ksnd_connd_waitq);
2235                         break;
2236                 }
2237
2238                 if (ksocknal_connd_check_start(sec, &timeout)) {
2239                         /* created new thread */
2240                         dropped_lock = 1;
2241                 }
2242
2243                 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2244                         /* Connection accepted by the listener */
2245                         cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2246                                         struct ksock_connreq, ksncr_list);
2247
2248                         list_del(&cr->ksncr_list);
2249                         spin_unlock_bh(connd_lock);
2250                         dropped_lock = 1;
2251
2252                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2253                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2254                         lnet_ni_decref(cr->ksncr_ni);
2255                         LIBCFS_FREE(cr, sizeof(*cr));
2256
2257                         spin_lock_bh(connd_lock);
2258                 }
2259
2260                 /* Only handle an outgoing connection request if there
2261                  * is a thread left to handle incoming connections and
2262                  * create new connd */
2263                 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2264                     ksocknal_data.ksnd_connd_running) {
2265                         route = ksocknal_connd_get_route_locked(&timeout);
2266                 }
2267                 if (route != NULL) {
2268                         list_del(&route->ksnr_connd_list);
2269                         ksocknal_data.ksnd_connd_connecting++;
2270                         spin_unlock_bh(connd_lock);
2271                         dropped_lock = 1;
2272
2273                         if (ksocknal_connect(route)) {
2274                                 /* consecutive retry */
2275                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2276                                         CWARN("massive consecutive "
2277                                               "re-connecting to %pI4h\n",
2278                                               &route->ksnr_ipaddr);
2279                                         cons_retry = 0;
2280                                 }
2281                         } else {
2282                                 cons_retry = 0;
2283                         }
2284
2285                         ksocknal_route_decref(route);
2286
2287                         spin_lock_bh(connd_lock);
2288                         ksocknal_data.ksnd_connd_connecting--;
2289                 }
2290
2291                 if (dropped_lock) {
2292                         if (++nloops < SOCKNAL_RESCHED)
2293                                 continue;
2294                         spin_unlock_bh(connd_lock);
2295                         nloops = 0;
2296                         cond_resched();
2297                         spin_lock_bh(connd_lock);
2298                         continue;
2299                 }
2300
2301                 /* Nothing to do for 'timeout'  */
2302                 set_current_state(TASK_INTERRUPTIBLE);
2303                 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2304                 spin_unlock_bh(connd_lock);
2305
2306                 nloops = 0;
2307                 schedule_timeout(timeout);
2308
2309                 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2310                 spin_lock_bh(connd_lock);
2311         }
2312         ksocknal_data.ksnd_connd_running--;
2313         spin_unlock_bh(connd_lock);
2314
2315         ksocknal_thread_fini();
2316         return 0;
2317 }
2318
2319 static struct ksock_conn *
2320 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2321 {
2322         /* We're called with a shared lock on ksnd_global_lock */
2323         struct ksock_conn *conn;
2324         struct list_head *ctmp;
2325         struct ksock_tx *tx;
2326
2327         list_for_each(ctmp, &peer_ni->ksnp_conns) {
2328                 int error;
2329
2330                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2331
2332                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2333                 LASSERT (!conn->ksnc_closing);
2334
2335                 error = conn->ksnc_sock->sk->sk_err;
2336                 if (error != 0) {
2337                         ksocknal_conn_addref(conn);
2338
2339                         switch (error) {
2340                         case ECONNRESET:
2341                                 CNETERR("A connection with %s "
2342                                         "(%pI4h:%d) was reset; "
2343                                         "it may have rebooted.\n",
2344                                         libcfs_id2str(peer_ni->ksnp_id),
2345                                         &conn->ksnc_ipaddr,
2346                                         conn->ksnc_port);
2347                                 break;
2348                         case ETIMEDOUT:
2349                                 CNETERR("A connection with %s "
2350                                         "(%pI4h:%d) timed out; the "
2351                                         "network or node may be down.\n",
2352                                         libcfs_id2str(peer_ni->ksnp_id),
2353                                         &conn->ksnc_ipaddr,
2354                                         conn->ksnc_port);
2355                                 break;
2356                         default:
2357                                 CNETERR("An unexpected network error %d "
2358                                         "occurred with %s "
2359                                         "(%pI4h:%d\n", error,
2360                                         libcfs_id2str(peer_ni->ksnp_id),
2361                                         &conn->ksnc_ipaddr,
2362                                         conn->ksnc_port);
2363                                 break;
2364                         }
2365
2366                         return (conn);
2367                 }
2368
2369                 if (conn->ksnc_rx_started &&
2370                     ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2371                         /* Timed out incomplete incoming message */
2372                         ksocknal_conn_addref(conn);
2373                         CNETERR("Timeout receiving from %s (%pI4h:%d), "
2374                                 "state %d wanted %d left %d\n",
2375                                 libcfs_id2str(peer_ni->ksnp_id),
2376                                 &conn->ksnc_ipaddr,
2377                                 conn->ksnc_port,
2378                                 conn->ksnc_rx_state,
2379                                 conn->ksnc_rx_nob_wanted,
2380                                 conn->ksnc_rx_nob_left);
2381                         return (conn);
2382                 }
2383
2384                 if ((!list_empty(&conn->ksnc_tx_queue) ||
2385                      conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2386                     ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2387                         /* Timed out messages queued for sending or
2388                          * buffered in the socket's send buffer */
2389                         ksocknal_conn_addref(conn);
2390                         list_for_each_entry(tx, &conn->ksnc_tx_queue,
2391                                             tx_list)
2392                                 tx->tx_hstatus =
2393                                         LNET_MSG_STATUS_LOCAL_TIMEOUT;
2394                         CNETERR("Timeout sending data to %s (%pI4h:%d) "
2395                                 "the network or that node may be down.\n",
2396                                 libcfs_id2str(peer_ni->ksnp_id),
2397                                 &conn->ksnc_ipaddr, conn->ksnc_port);
2398                         return (conn);
2399                 }
2400         }
2401
2402         return (NULL);
2403 }
2404
2405 static inline void
2406 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2407 {
2408         struct ksock_tx *tx;
2409         LIST_HEAD(stale_txs);
2410
2411         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2412
2413         while (!list_empty(&peer_ni->ksnp_tx_queue)) {
2414                 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2415                                 struct ksock_tx, tx_list);
2416
2417                 if (ktime_get_seconds() < tx->tx_deadline)
2418                         break;
2419
2420                 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2421
2422                 list_move_tail(&tx->tx_list, &stale_txs);
2423         }
2424
2425         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2426
2427         ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2428 }
2429
2430 static int
2431 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2432 __must_hold(&ksocknal_data.ksnd_global_lock)
2433 {
2434         struct ksock_sched *sched;
2435         struct ksock_conn *conn;
2436         struct ksock_tx *tx;
2437
2438         /* last_alive will be updated by create_conn */
2439         if (list_empty(&peer_ni->ksnp_conns))
2440                 return 0;
2441
2442         if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x)
2443                 return 0;
2444
2445         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2446             ktime_get_seconds() < peer_ni->ksnp_last_alive +
2447                                   *ksocknal_tunables.ksnd_keepalive)
2448                 return 0;
2449
2450         if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2451                 return 0;
2452
2453         /* retry 10 secs later, so we wouldn't put pressure
2454          * on this peer_ni if we failed to send keepalive this time */
2455         peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2456
2457         conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2458         if (conn != NULL) {
2459                 sched = conn->ksnc_scheduler;
2460
2461                 spin_lock_bh(&sched->kss_lock);
2462                 if (!list_empty(&conn->ksnc_tx_queue)) {
2463                         spin_unlock_bh(&sched->kss_lock);
2464                         /* there is an queued ACK, don't need keepalive */
2465                         return 0;
2466                 }
2467
2468                 spin_unlock_bh(&sched->kss_lock);
2469         }
2470
2471         read_unlock(&ksocknal_data.ksnd_global_lock);
2472
2473         /* cookie = 1 is reserved for keepalive PING */
2474         tx = ksocknal_alloc_tx_noop(1, 1);
2475         if (tx == NULL) {
2476                 read_lock(&ksocknal_data.ksnd_global_lock);
2477                 return -ENOMEM;
2478         }
2479
2480         if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) {
2481                 read_lock(&ksocknal_data.ksnd_global_lock);
2482                 return 1;
2483         }
2484
2485         ksocknal_free_tx(tx);
2486         read_lock(&ksocknal_data.ksnd_global_lock);
2487
2488         return -EIO;
2489 }
2490
2491
2492 static void
2493 ksocknal_check_peer_timeouts(int idx)
2494 {
2495         struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2496         struct ksock_peer_ni *peer_ni;
2497         struct ksock_conn *conn;
2498         struct ksock_tx *tx;
2499
2500  again:
2501         /* NB. We expect to have a look at all the peers and not find any
2502          * connections to time out, so we just use a shared lock while we
2503          * take a look...
2504          */
2505         read_lock(&ksocknal_data.ksnd_global_lock);
2506
2507         hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2508                 struct ksock_tx *tx_stale;
2509                 time64_t deadline = 0;
2510                 int resid = 0;
2511                 int n = 0;
2512
2513                 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2514                         read_unlock(&ksocknal_data.ksnd_global_lock);
2515                         goto again;
2516                 }
2517
2518                 conn = ksocknal_find_timed_out_conn(peer_ni);
2519
2520                 if (conn != NULL) {
2521                         read_unlock(&ksocknal_data.ksnd_global_lock);
2522
2523                         ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2524
2525                         /* NB we won't find this one again, but we can't
2526                          * just proceed with the next peer_ni, since we dropped
2527                          * ksnd_global_lock and it might be dead already!
2528                          */
2529                         ksocknal_conn_decref(conn);
2530                         goto again;
2531                 }
2532
2533                 /* we can't process stale txs right here because we're
2534                  * holding only shared lock
2535                  */
2536                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
2537                         struct ksock_tx *tx;
2538
2539                         tx = list_entry(peer_ni->ksnp_tx_queue.next,
2540                                         struct ksock_tx, tx_list);
2541                         if (ktime_get_seconds() >= tx->tx_deadline) {
2542                                 ksocknal_peer_addref(peer_ni);
2543                                 read_unlock(&ksocknal_data.ksnd_global_lock);
2544
2545                                 ksocknal_flush_stale_txs(peer_ni);
2546
2547                                 ksocknal_peer_decref(peer_ni);
2548                                 goto again;
2549                         }
2550                 }
2551
2552                 if (list_empty(&peer_ni->ksnp_zc_req_list))
2553                         continue;
2554
2555                 tx_stale = NULL;
2556                 spin_lock(&peer_ni->ksnp_lock);
2557                 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2558                         if (ktime_get_seconds() < tx->tx_deadline)
2559                                 break;
2560                         /* ignore the TX if connection is being closed */
2561                         if (tx->tx_conn->ksnc_closing)
2562                                 continue;
2563                         n++;
2564                         if (tx_stale == NULL)
2565                                 tx_stale = tx;
2566                 }
2567
2568                 if (tx_stale == NULL) {
2569                         spin_unlock(&peer_ni->ksnp_lock);
2570                         continue;
2571                 }
2572
2573                 deadline = tx_stale->tx_deadline;
2574                 resid    = tx_stale->tx_resid;
2575                 conn     = tx_stale->tx_conn;
2576                 ksocknal_conn_addref(conn);
2577
2578                 spin_unlock(&peer_ni->ksnp_lock);
2579                 read_unlock(&ksocknal_data.ksnd_global_lock);
2580
2581                 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2582                        "oldest(%p) timed out %lld secs ago, "
2583                        "resid: %d, wmem: %d\n",
2584                        n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale,
2585                        ktime_get_seconds() - deadline,
2586                        resid, conn->ksnc_sock->sk->sk_wmem_queued);
2587
2588                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2589                 ksocknal_conn_decref(conn);
2590                 goto again;
2591         }
2592
2593         read_unlock(&ksocknal_data.ksnd_global_lock);
2594 }
2595
2596 int ksocknal_reaper(void *arg)
2597 {
2598         wait_queue_entry_t wait;
2599         struct ksock_conn *conn;
2600         struct ksock_sched *sched;
2601         LIST_HEAD(enomem_conns);
2602         int nenomem_conns;
2603         time64_t timeout;
2604         int i;
2605         int peer_index = 0;
2606         time64_t deadline = ktime_get_seconds();
2607
2608         init_waitqueue_entry(&wait, current);
2609
2610         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2611
2612         while (!ksocknal_data.ksnd_shuttingdown) {
2613                 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2614                         conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2615                                           struct ksock_conn, ksnc_list);
2616                         list_del(&conn->ksnc_list);
2617
2618                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2619
2620                         ksocknal_terminate_conn(conn);
2621                         ksocknal_conn_decref(conn);
2622
2623                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2624                         continue;
2625                 }
2626
2627                 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2628                         conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2629                                           struct ksock_conn, ksnc_list);
2630                         list_del(&conn->ksnc_list);
2631
2632                         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2633
2634                         ksocknal_destroy_conn(conn);
2635
2636                         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2637                         continue;
2638                 }
2639
2640                 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2641                                  &enomem_conns);
2642
2643                 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2644
2645                 /* reschedule all the connections that stalled with ENOMEM... */
2646                 nenomem_conns = 0;
2647                 while (!list_empty(&enomem_conns)) {
2648                         conn = list_entry(enomem_conns.next,
2649                                           struct ksock_conn, ksnc_tx_list);
2650                         list_del(&conn->ksnc_tx_list);
2651
2652                         sched = conn->ksnc_scheduler;
2653
2654                         spin_lock_bh(&sched->kss_lock);
2655
2656                         LASSERT(conn->ksnc_tx_scheduled);
2657                         conn->ksnc_tx_ready = 1;
2658                         list_add_tail(&conn->ksnc_tx_list,
2659                                           &sched->kss_tx_conns);
2660                         wake_up(&sched->kss_waitq);
2661
2662                         spin_unlock_bh(&sched->kss_lock);
2663                         nenomem_conns++;
2664                 }
2665
2666                 /* careful with the jiffy wrap... */
2667                 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2668                         const int n = 4;
2669                         const int p = 1;
2670                         int  chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2671                         unsigned int lnd_timeout;
2672
2673                         /* Time to check for timeouts on a few more peers: I
2674                          * do checks every 'p' seconds on a proportion of the
2675                          * peer_ni table and I need to check every connection
2676                          * 'n' times within a timeout interval, to ensure I
2677                          * detect a timeout on any connection within (n+1)/n
2678                          * times the timeout interval.
2679                          */
2680
2681                         lnd_timeout = lnet_get_lnd_timeout();
2682                         if (lnd_timeout > n * p)
2683                                 chunk = (chunk * n * p) / lnd_timeout;
2684                         if (chunk == 0)
2685                                 chunk = 1;
2686
2687                         for (i = 0; i < chunk; i++) {
2688                                 ksocknal_check_peer_timeouts(peer_index);
2689                                 peer_index = (peer_index + 1) %
2690                                         HASH_SIZE(ksocknal_data.ksnd_peers);
2691                         }
2692
2693                         deadline += p;
2694                 }
2695
2696                 if (nenomem_conns != 0) {
2697                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2698                          * This also prevents me getting woken immediately
2699                          * if any go back on my enomem list. */
2700                         timeout = SOCKNAL_ENOMEM_RETRY;
2701                 }
2702                 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2703                                                      timeout;
2704
2705                 set_current_state(TASK_INTERRUPTIBLE);
2706                 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2707
2708                 if (!ksocknal_data.ksnd_shuttingdown &&
2709                     list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2710                     list_empty(&ksocknal_data.ksnd_zombie_conns))
2711                         schedule_timeout(cfs_time_seconds(timeout));
2712
2713                 set_current_state(TASK_RUNNING);
2714                 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2715
2716                 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2717         }
2718
2719         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2720
2721         ksocknal_thread_fini();
2722         return 0;
2723 }