lnet/klnds/gmlnd/gmlnd_comm.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2003 Los Alamos National Laboratory (LANL)
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  */
  38
  39 /*
  40  *      This file contains all gmnal send and receive functions
  41  */
  42
  43 #include "gmlnd.h"
  44
  45 void
  46 gmnal_notify_peer_down(gmnal_tx_t *tx)
  47 {
  48         time_t             then;
  49
  50         then = cfs_time_current_sec() -
  51                 cfs_duration_sec(cfs_time_current() -
  52                                  tx->tx_launchtime);
  53
  54         lnet_notify(tx->tx_gmni->gmni_ni, tx->tx_nid, 0, then);
  55 }
  56
  57 void
  58 gmnal_pack_msg(gmnal_ni_t *gmni, gmnal_msg_t *msg,
  59                lnet_nid_t dstnid, int type)
  60 {
  61         /* CAVEAT EMPTOR! this only sets the common message fields. */
  62         msg->gmm_magic    = GMNAL_MSG_MAGIC;
  63         msg->gmm_version  = GMNAL_MSG_VERSION;
  64         msg->gmm_type     = type;
  65         msg->gmm_srcnid   = gmni->gmni_ni->ni_nid;
  66         msg->gmm_dstnid   = dstnid;
  67 }
  68
  69 int
  70 gmnal_unpack_msg(gmnal_ni_t *gmni, gmnal_rx_t *rx)
  71 {
  72         gmnal_msg_t *msg = GMNAL_NETBUF_MSG(&rx->rx_buf);
  73         const int    hdr_size = offsetof(gmnal_msg_t, gmm_u);
  74         int          buffnob = rx->rx_islarge ? gmni->gmni_large_msgsize :
  75                                                 gmni->gmni_small_msgsize;
  76         int          flip;
  77
  78         /* rc = 0:SUCCESS -ve:failure +ve:version mismatch */
  79
  80         /* GM may not overflow our buffer */
  81         LASSERT (rx->rx_recv_nob <= buffnob);
  82
  83         /* 6 bytes are enough to have received magic + version */
  84         if (rx->rx_recv_nob < 6) {
  85                 CERROR("Short message from gmid %u: %d\n",
  86                        rx->rx_recv_gmid, rx->rx_recv_nob);
  87                 return -EPROTO;
  88         }
  89
  90         if (msg->gmm_magic == GMNAL_MSG_MAGIC) {
  91                 flip = 0;
  92         } else if (msg->gmm_magic == __swab32(GMNAL_MSG_MAGIC)) {
  93                 flip = 1;
  94         } else if (msg->gmm_magic == LNET_PROTO_MAGIC ||
  95                    msg->gmm_magic == __swab32(LNET_PROTO_MAGIC)) {
  96                 return EPROTO;
  97         } else {
  98                 CERROR("Bad magic from gmid %u: %08x\n",
  99                        rx->rx_recv_gmid, msg->gmm_magic);
 100                 return -EPROTO;
 101         }
 102
 103         if (msg->gmm_version !=
 104             (flip ? __swab16(GMNAL_MSG_VERSION) : GMNAL_MSG_VERSION)) {
 105                 return EPROTO;
 106         }
 107
 108         if (rx->rx_recv_nob < hdr_size) {
 109                 CERROR("Short message from %u: %d\n",
 110                        rx->rx_recv_gmid, rx->rx_recv_nob);
 111                 return -EPROTO;
 112         }
 113
 114         if (flip) {
 115                 /* leave magic unflipped as a clue to peer endianness */
 116                 __swab16s(&msg->gmm_version);
 117                 __swab16s(&msg->gmm_type);
 118                 __swab64s(&msg->gmm_srcnid);
 119                 __swab64s(&msg->gmm_dstnid);
 120         }
 121
 122         if (msg->gmm_srcnid == LNET_NID_ANY) {
 123                 CERROR("Bad src nid from %u: %s\n",
 124                        rx->rx_recv_gmid, libcfs_nid2str(msg->gmm_srcnid));
 125                 return -EPROTO;
 126         }
 127
 128         if (gmni->gmni_ni->ni_nid != msg->gmm_dstnid) {
 129                 CERROR("Bad dst nid from %u: %s\n",
 130                        rx->rx_recv_gmid, libcfs_nid2str(msg->gmm_dstnid));
 131                 return -EPROTO;
 132         }
 133
 134         switch (msg->gmm_type) {
 135         default:
 136                 CERROR("Unknown message type from %u: %x\n",
 137                        rx->rx_recv_gmid, msg->gmm_type);
 138                 return -EPROTO;
 139
 140         case GMNAL_MSG_IMMEDIATE:
 141                 if (rx->rx_recv_nob <
 142                     offsetof(gmnal_msg_t, gmm_u.immediate.gmim_payload[0])) {
 143                         CERROR("Short IMMEDIATE from %u: %d(%lu)\n",
 144                                rx->rx_recv_gmid, rx->rx_recv_nob,
 145                                (long)offsetof(gmnal_msg_t,
 146                                               gmm_u.immediate.gmim_payload[0]));
 147                         return -EPROTO;
 148                 }
 149                 break;
 150         }
 151         return 0;
 152 }
 153
 154 gmnal_tx_t *
 155 gmnal_get_tx(gmnal_ni_t *gmni)
 156 {
 157         gmnal_tx_t *tx = NULL;
 158
 159         spin_lock(&gmni->gmni_tx_lock);
 160
 161         if (gmni->gmni_shutdown ||
 162             list_empty(&gmni->gmni_idle_txs)) {
 163                 spin_unlock(&gmni->gmni_tx_lock);
 164                 return NULL;
 165         }
 166
 167         tx = list_entry(gmni->gmni_idle_txs.next, gmnal_tx_t, tx_list);
 168         list_del(&tx->tx_list);
 169
 170         spin_unlock(&gmni->gmni_tx_lock);
 171
 172         LASSERT (tx->tx_lntmsg == NULL);
 173         LASSERT (tx->tx_ltxb == NULL);
 174         LASSERT (!tx->tx_credit);
 175
 176         return tx;
 177 }
 178
 179 void
 180 gmnal_tx_done(gmnal_tx_t *tx, int rc)
 181 {
 182         gmnal_ni_t *gmni = tx->tx_gmni;
 183         int         wake_sched = 0;
 184         lnet_msg_t *lnetmsg = tx->tx_lntmsg;
 185
 186         tx->tx_lntmsg = NULL;
 187
 188         spin_lock(&gmni->gmni_tx_lock);
 189
 190         if (tx->tx_ltxb != NULL) {
 191                 wake_sched = 1;
 192                 list_add_tail(&tx->tx_ltxb->txb_list, &gmni->gmni_idle_ltxbs);
 193                 tx->tx_ltxb = NULL;
 194         }
 195
 196         if (tx->tx_credit) {
 197                 wake_sched = 1;
 198                 gmni->gmni_tx_credits++;
 199                 tx->tx_credit = 0;
 200         }
 201
 202         list_add_tail(&tx->tx_list, &gmni->gmni_idle_txs);
 203
 204         if (wake_sched)
 205                 gmnal_check_txqueues_locked(gmni);
 206
 207         spin_unlock(&gmni->gmni_tx_lock);
 208
 209         /* Delay finalize until tx is free */
 210         if (lnetmsg != NULL)
 211                 lnet_finalize(gmni->gmni_ni, lnetmsg, rc);
 212 }
 213
 214 void
 215 gmnal_drop_sends_callback(struct gm_port *gm_port, void *context,
 216                           gm_status_t status)
 217 {
 218         gmnal_tx_t *tx = (gmnal_tx_t*)context;
 219
 220         LASSERT(!in_interrupt());
 221
 222         CDEBUG(D_NET, "status for tx [%p] is [%d][%s], nid %s\n",
 223                tx, status, gmnal_gmstatus2str(status),
 224                libcfs_nid2str(tx->tx_nid));
 225
 226         gmnal_tx_done(tx, -EIO);
 227 }
 228
 229 void
 230 gmnal_tx_callback(gm_port_t *gm_port, void *context, gm_status_t status)
 231 {
 232         gmnal_tx_t *tx = (gmnal_tx_t*)context;
 233         gmnal_ni_t *gmni = tx->tx_gmni;
 234
 235         LASSERT(!in_interrupt());
 236
 237         switch(status) {
 238         case GM_SUCCESS:
 239                 gmnal_tx_done(tx, 0);
 240                 return;
 241
 242         case GM_SEND_DROPPED:
 243                 CDEBUG(D_NETERROR, "Dropped tx %p to %s\n",
 244                        tx, libcfs_nid2str(tx->tx_nid));
 245                 /* Another tx failed and called gm_drop_sends() which made this
 246                  * one complete immediately */
 247                 gmnal_tx_done(tx, -EIO);
 248                 return;
 249
 250         default:
 251                 /* Some error; NB don't complete tx yet; we need its credit for
 252                  * gm_drop_sends() */
 253                 CDEBUG(D_NETERROR, "tx %p error %d(%s), nid %s\n",
 254                        tx, status, gmnal_gmstatus2str(status),
 255                        libcfs_nid2str(tx->tx_nid));
 256
 257                 gmnal_notify_peer_down(tx);
 258
 259                 spin_lock(&gmni->gmni_gm_lock);
 260                 gm_drop_sends(gmni->gmni_port,
 261                               tx->tx_ltxb != NULL ?
 262                               GMNAL_LARGE_PRIORITY : GMNAL_SMALL_PRIORITY,
 263                               tx->tx_gmlid, *gmnal_tunables.gm_port,
 264                               gmnal_drop_sends_callback, tx);
 265                 spin_unlock(&gmni->gmni_gm_lock);
 266                 return;
 267         }
 268
 269         /* not reached */
 270         LBUG();
 271 }
 272
 273 void
 274 gmnal_check_txqueues_locked (gmnal_ni_t *gmni)
 275 {
 276         gmnal_tx_t    *tx;
 277         gmnal_txbuf_t *ltxb;
 278         int            gmsize;
 279         int            pri;
 280         void          *netaddr;
 281
 282         tx = list_empty(&gmni->gmni_buf_txq) ? NULL :
 283              list_entry(gmni->gmni_buf_txq.next, gmnal_tx_t, tx_list);
 284
 285         if (tx != NULL &&
 286             (tx->tx_large_nob == 0 ||
 287              !list_empty(&gmni->gmni_idle_ltxbs))) {
 288
 289                 /* consume tx */
 290                 list_del(&tx->tx_list);
 291
 292                 LASSERT (tx->tx_ltxb == NULL);
 293
 294                 if (tx->tx_large_nob != 0) {
 295                         ltxb = list_entry(gmni->gmni_idle_ltxbs.next,
 296                                           gmnal_txbuf_t, txb_list);
 297
 298                         /* consume large buffer */
 299                         list_del(&ltxb->txb_list);
 300
 301                         spin_unlock(&gmni->gmni_tx_lock);
 302
 303                         /* Unlocking here allows sends to get re-ordered,
 304                          * but we want to allow other CPUs to progress... */
 305
 306                         tx->tx_ltxb = ltxb;
 307
 308                         /* marshall message in tx_ltxb...
 309                          * 1. Copy what was marshalled so far (in tx_buf) */
 310                         memcpy(GMNAL_NETBUF_MSG(&ltxb->txb_buf),
 311                                GMNAL_NETBUF_MSG(&tx->tx_buf), tx->tx_msgnob);
 312
 313                         /* 2. Copy the payload */
 314                         if (tx->tx_large_iskiov)
 315                                 lnet_copy_kiov2kiov(
 316                                         gmni->gmni_large_pages,
 317                                         ltxb->txb_buf.nb_kiov,
 318                                         tx->tx_msgnob,
 319                                         tx->tx_large_niov,
 320                                         tx->tx_large_frags.kiov,
 321                                         tx->tx_large_offset,
 322                                         tx->tx_large_nob);
 323                         else
 324                                 lnet_copy_iov2kiov(
 325                                         gmni->gmni_large_pages,
 326                                         ltxb->txb_buf.nb_kiov,
 327                                         tx->tx_msgnob,
 328                                         tx->tx_large_niov,
 329                                         tx->tx_large_frags.iov,
 330                                         tx->tx_large_offset,
 331                                         tx->tx_large_nob);
 332
 333                         tx->tx_msgnob += tx->tx_large_nob;
 334
 335                         spin_lock(&gmni->gmni_tx_lock);
 336                 }
 337
 338                 list_add_tail(&tx->tx_list, &gmni->gmni_cred_txq);
 339         }
 340
 341         if (!list_empty(&gmni->gmni_cred_txq) &&
 342             gmni->gmni_tx_credits != 0) {
 343
 344                 tx = list_entry(gmni->gmni_cred_txq.next, gmnal_tx_t, tx_list);
 345
 346                 /* consume tx and 1 credit */
 347                 list_del(&tx->tx_list);
 348                 gmni->gmni_tx_credits--;
 349
 350                 spin_unlock(&gmni->gmni_tx_lock);
 351
 352                 /* Unlocking here allows sends to get re-ordered, but we want
 353                  * to allow other CPUs to progress... */
 354
 355                 LASSERT(!tx->tx_credit);
 356                 tx->tx_credit = 1;
 357
 358                 tx->tx_launchtime = cfs_time_current();
 359
 360                 if (tx->tx_msgnob <= gmni->gmni_small_msgsize) {
 361                         LASSERT (tx->tx_ltxb == NULL);
 362                         netaddr = GMNAL_NETBUF_LOCAL_NETADDR(&tx->tx_buf);
 363                         gmsize = gmni->gmni_small_gmsize;
 364                         pri = GMNAL_SMALL_PRIORITY;
 365                 } else {
 366                         LASSERT (tx->tx_ltxb != NULL);
 367                         netaddr = GMNAL_NETBUF_LOCAL_NETADDR(&tx->tx_ltxb->txb_buf);
 368                         gmsize = gmni->gmni_large_gmsize;
 369                         pri = GMNAL_LARGE_PRIORITY;
 370                 }
 371
 372                 spin_lock(&gmni->gmni_gm_lock);
 373
 374                 gm_send_to_peer_with_callback(gmni->gmni_port,
 375                                               netaddr, gmsize,
 376                                               tx->tx_msgnob,
 377                                               pri,
 378                                               tx->tx_gmlid,
 379                                               gmnal_tx_callback,
 380                                               (void*)tx);
 381
 382                 spin_unlock(&gmni->gmni_gm_lock);
 383                 spin_lock(&gmni->gmni_tx_lock);
 384         }
 385 }
 386
 387 void
 388 gmnal_post_rx(gmnal_ni_t *gmni, gmnal_rx_t *rx)
 389 {
 390         int   gmsize = rx->rx_islarge ? gmni->gmni_large_gmsize :
 391                                         gmni->gmni_small_gmsize;
 392         int   pri    = rx->rx_islarge ? GMNAL_LARGE_PRIORITY :
 393                                         GMNAL_SMALL_PRIORITY;
 394         void *buffer = GMNAL_NETBUF_LOCAL_NETADDR(&rx->rx_buf);
 395
 396         CDEBUG(D_NET, "posting rx %p buf %p\n", rx, buffer);
 397
 398         spin_lock(&gmni->gmni_gm_lock);
 399         gm_provide_receive_buffer_with_tag(gmni->gmni_port,
 400                                            buffer, gmsize, pri, 0);
 401         spin_unlock(&gmni->gmni_gm_lock);
 402 }
 403
 404 void
 405 gmnal_version_reply (gmnal_ni_t *gmni, gmnal_rx_t *rx)
 406 {
 407         /* Future protocol version compatibility support!
 408          * The next gmlnd-specific protocol rev will first send a message to
 409          * check version; I reply with a stub message containing my current
 410          * magic+version... */
 411         gmnal_msg_t *msg;
 412         gmnal_tx_t  *tx = gmnal_get_tx(gmni);
 413
 414         if (tx == NULL) {
 415                 CERROR("Can't allocate tx to send version info to %u\n",
 416                        rx->rx_recv_gmid);
 417                 return;
 418         }
 419
 420         LASSERT (tx->tx_lntmsg == NULL);        /* no finalize */
 421
 422         tx->tx_nid = LNET_NID_ANY;
 423         tx->tx_gmlid = rx->rx_recv_gmid;
 424
 425         msg = GMNAL_NETBUF_MSG(&tx->tx_buf);
 426         msg->gmm_magic   = GMNAL_MSG_MAGIC;
 427         msg->gmm_version = GMNAL_MSG_VERSION;
 428
 429         /* just send magic + version */
 430         tx->tx_msgnob = offsetof(gmnal_msg_t, gmm_type);
 431         tx->tx_large_nob = 0;
 432
 433         spin_lock(&gmni->gmni_tx_lock);
 434
 435         list_add_tail(&tx->tx_list, &gmni->gmni_buf_txq);
 436         gmnal_check_txqueues_locked(gmni);
 437
 438         spin_unlock(&gmni->gmni_tx_lock);
 439 }
 440
 441 int
 442 gmnal_rx_thread(void *arg)
 443 {
 444         gmnal_ni_t      *gmni = arg;
 445         gm_recv_event_t *rxevent = NULL;
 446         gm_recv_t       *recv = NULL;
 447         gmnal_rx_t      *rx;
 448         int              rc;
 449
 450         cfs_daemonize("gmnal_rxd");
 451
 452         while (!gmni->gmni_shutdown) {
 453                 rc = down_interruptible(&gmni->gmni_rx_mutex);
 454                 LASSERT (rc == 0 || rc == -EINTR);
 455                 if (rc != 0)
 456                         continue;
 457
 458                 spin_lock(&gmni->gmni_gm_lock);
 459                 rxevent = gm_blocking_receive_no_spin(gmni->gmni_port);
 460                 spin_unlock(&gmni->gmni_gm_lock);
 461
 462                 switch (GM_RECV_EVENT_TYPE(rxevent)) {
 463                 default:
 464                         gm_unknown(gmni->gmni_port, rxevent);
 465                         up(&gmni->gmni_rx_mutex);
 466                         continue;
 467
 468                 case GM_FAST_RECV_EVENT:
 469                 case GM_FAST_PEER_RECV_EVENT:
 470                 case GM_PEER_RECV_EVENT:
 471                 case GM_FAST_HIGH_RECV_EVENT:
 472                 case GM_FAST_HIGH_PEER_RECV_EVENT:
 473                 case GM_HIGH_PEER_RECV_EVENT:
 474                 case GM_RECV_EVENT:
 475                 case GM_HIGH_RECV_EVENT:
 476                         break;
 477                 }
 478
 479                 recv = &rxevent->recv;
 480                 rx = gm_hash_find(gmni->gmni_rx_hash,
 481                                   gm_ntohp(recv->buffer));
 482                 LASSERT (rx != NULL);
 483
 484                 rx->rx_recv_nob  = gm_ntoh_u32(recv->length);
 485                 rx->rx_recv_gmid = gm_ntoh_u16(recv->sender_node_id);
 486                 rx->rx_recv_port = gm_ntoh_u8(recv->sender_port_id);
 487                 rx->rx_recv_type = gm_ntoh_u8(recv->type);
 488
 489                 switch (GM_RECV_EVENT_TYPE(rxevent)) {
 490                 case GM_FAST_RECV_EVENT:
 491                 case GM_FAST_PEER_RECV_EVENT:
 492                 case GM_FAST_HIGH_RECV_EVENT:
 493                 case GM_FAST_HIGH_PEER_RECV_EVENT:
 494                         LASSERT (rx->rx_recv_nob <= PAGE_SIZE);
 495
 496                         memcpy(GMNAL_NETBUF_MSG(&rx->rx_buf),
 497                                gm_ntohp(recv->message), rx->rx_recv_nob);
 498                         break;
 499                 }
 500
 501                 up(&gmni->gmni_rx_mutex);
 502
 503                 CDEBUG (D_NET, "rx %p: buf %p(%p) nob %d\n", rx,
 504                         GMNAL_NETBUF_LOCAL_NETADDR(&rx->rx_buf),
 505                         gm_ntohp(recv->buffer), rx->rx_recv_nob);
 506
 507                 /* We're connectionless: simply drop packets with
 508                  * errors */
 509                 rc = gmnal_unpack_msg(gmni, rx);
 510
 511                 if (rc == 0) {
 512                         gmnal_msg_t *msg = GMNAL_NETBUF_MSG(&rx->rx_buf);
 513
 514                         LASSERT (msg->gmm_type == GMNAL_MSG_IMMEDIATE);
 515                         rc = lnet_parse(gmni->gmni_ni,
 516                                         &msg->gmm_u.immediate.gmim_hdr,
 517                                         msg->gmm_srcnid, rx, 0);
 518                 } else if (rc > 0) {
 519                         gmnal_version_reply(gmni, rx);
 520                         rc = -EPROTO;           /* repost rx */
 521                 }
 522
 523                 if (rc < 0)                     /* parse failure */
 524                         gmnal_post_rx(gmni, rx);
 525         }
 526
 527         CDEBUG(D_NET, "exiting\n");
 528         atomic_dec(&gmni->gmni_nthreads);
 529         return 0;
 530 }
 531
 532 void
 533 gmnal_stop_threads(gmnal_ni_t *gmni)
 534 {
 535         int count = 2;
 536
 537         gmni->gmni_shutdown = 1;
 538         mb();
 539
 540         /* wake rxthread owning gmni_rx_mutex with an alarm. */
 541         spin_lock(&gmni->gmni_gm_lock);
 542         gm_set_alarm(gmni->gmni_port, &gmni->gmni_alarm, 0, NULL, NULL);
 543         spin_unlock(&gmni->gmni_gm_lock);
 544
 545         while (atomic_read(&gmni->gmni_nthreads) != 0) {
 546                 count++;
 547                 if ((count & (count - 1)) == 0)
 548                         CWARN("Waiting for %d threads to stop\n",
 549                               atomic_read(&gmni->gmni_nthreads));
 550                 gmnal_yield(1);
 551         }
 552 }
 553
 554 int
 555 gmnal_start_threads(gmnal_ni_t *gmni)
 556 {
 557         int     i;
 558         int     pid;
 559
 560         LASSERT (!gmni->gmni_shutdown);
 561         LASSERT (atomic_read(&gmni->gmni_nthreads) == 0);
 562
 563         gm_initialize_alarm(&gmni->gmni_alarm);
 564
 565         for (i = 0; i < num_online_cpus(); i++) {
 566
 567                 pid = kernel_thread(gmnal_rx_thread, (void*)gmni, 0);
 568                 if (pid < 0) {
 569                         CERROR("rx thread failed to start: %d\n", pid);
 570                         gmnal_stop_threads(gmni);
 571                         return pid;
 572                 }
 573
 574                 atomic_inc(&gmni->gmni_nthreads);
 575         }
 576
 577         return 0;
 578 }