lnet/klnds/gnilnd/gnilnd.c

   1 /*
   2  * Copyright (C) 2012 Cray, Inc.
   3  *
   4  * Copyright (c) 2013, 2017, Intel Corporation.
   5  *
   6  *   Author: Nic Henke <nic@cray.com>
   7  *   Author: James Shimek <jshimek@cray.com>
   8  *
   9  *   This file is part of Lustre, http://www.lustre.org.
  10  *
  11  *   Lustre is free software; you can redistribute it and/or
  12  *   modify it under the terms of version 2 of the GNU General Public
  13  *   License as published by the Free Software Foundation.
  14  *
  15  *   Lustre is distributed in the hope that it will be useful,
  16  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  *   GNU General Public License for more details.
  19  *
  20  *   You should have received a copy of the GNU General Public License
  21  *   along with Lustre; if not, write to the Free Software
  22  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  *
  24  */
  25 #include "gnilnd.h"
  26
  27 /* Primary entry points from LNET.  There are no guarantees against reentrance. */
  28 const struct lnet_lnd the_kgnilnd = {
  29         .lnd_type       = GNILND,
  30         .lnd_startup    = kgnilnd_startup,
  31         .lnd_shutdown   = kgnilnd_shutdown,
  32         .lnd_ctl        = kgnilnd_ctl,
  33         .lnd_send       = kgnilnd_send,
  34         .lnd_recv       = kgnilnd_recv,
  35         .lnd_eager_recv = kgnilnd_eager_recv,
  36 };
  37
  38 kgn_data_t      kgnilnd_data;
  39
  40 int
  41 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
  42 {
  43         struct task_struct *thrd;
  44
  45         thrd = kthread_run(fn, arg, "%s_%02d", name, id);
  46         if (IS_ERR(thrd))
  47                 return PTR_ERR(thrd);
  48
  49         atomic_inc(&kgnilnd_data.kgn_nthreads);
  50         return 0;
  51 }
  52
  53 /* bind scheduler threads to cpus */
  54 int
  55 kgnilnd_start_sd_threads(void)
  56 {
  57         int cpu;
  58         int i = 0;
  59         struct task_struct *task;
  60
  61         for_each_online_cpu(cpu) {
  62                 /* don't bind to cpu 0 - all interrupts are processed here */
  63                 if (cpu == 0)
  64                         continue;
  65
  66                 task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
  67                                       "%s_%02d", "kgnilnd_sd", i);
  68                 if (!IS_ERR(task)) {
  69                         kthread_bind(task, cpu);
  70                         wake_up_process(task);
  71                 } else {
  72                         CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
  73                                 PTR_ERR(task));
  74                         return PTR_ERR(task);
  75                 }
  76                 atomic_inc(&kgnilnd_data.kgn_nthreads);
  77
  78                 if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
  79                         break;
  80                 }
  81         }
  82
  83         return 0;
  84 }
  85
  86 /* needs write_lock on kgn_peer_conn_lock */
  87 int
  88 kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
  89 {
  90         kgn_conn_t         *conn;
  91         struct list_head   *ctmp, *cnxt;
  92         int                 loopback;
  93         int                 count = 0;
  94
  95         loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
  96
  97         list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
  98                 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
  99
 100                 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
 101                         continue;
 102
 103                 if (conn == newconn)
 104                         continue;
 105
 106                 if (conn->gnc_device != newconn->gnc_device)
 107                         continue;
 108
 109                 /* This is a two connection loopback - one talking to the other */
 110                 if (loopback &&
 111                     newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
 112                     newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
 113                         CDEBUG(D_NET, "skipping prune of %p, "
 114                                 "loopback and matching stamps"
 115                                 " connstamp %llu(%llu)"
 116                                 " peerstamp %llu(%llu)\n",
 117                                 conn, newconn->gnc_my_connstamp,
 118                                 conn->gnc_peer_connstamp,
 119                                 newconn->gnc_peer_connstamp,
 120                                 conn->gnc_my_connstamp);
 121                         continue;
 122                 }
 123
 124                 if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
 125                         LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
 126                                 "conn 0x%p peerstamp %llu >= "
 127                                 "newconn 0x%p peerstamp %llu\n",
 128                                 conn, conn->gnc_peerstamp,
 129                                 newconn, newconn->gnc_peerstamp);
 130
 131                         CDEBUG(D_NET, "Closing stale conn nid: %s "
 132                                " peerstamp:%#llx(%#llx)\n",
 133                                libcfs_nid2str(peer->gnp_nid),
 134                                conn->gnc_peerstamp, newconn->gnc_peerstamp);
 135                 } else {
 136
 137                         LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
 138                                 "conn 0x%p peer_connstamp %llu >= "
 139                                 "newconn 0x%p peer_connstamp %llu\n",
 140                                 conn, conn->gnc_peer_connstamp,
 141                                 newconn, newconn->gnc_peer_connstamp);
 142
 143                         CDEBUG(D_NET, "Closing stale conn nid: %s"
 144                                " connstamp:%llu(%llu)\n",
 145                                libcfs_nid2str(peer->gnp_nid),
 146                                conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
 147                 }
 148
 149                 count++;
 150                 kgnilnd_close_conn_locked(conn, -ESTALE);
 151         }
 152
 153         if (count != 0) {
 154                 CWARN("Closed %d stale conns to %s\n", count, libcfs_nid2str(peer->gnp_nid));
 155         }
 156
 157         RETURN(count);
 158 }
 159
 160 int
 161 kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
 162 {
 163         kgn_conn_t       *conn;
 164         struct list_head *tmp;
 165         int               loopback;
 166         ENTRY;
 167
 168         loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
 169
 170         list_for_each(tmp, &peer->gnp_conns) {
 171                 conn = list_entry(tmp, kgn_conn_t, gnc_list);
 172                 CDEBUG(D_NET, "checking conn 0x%p for peer %s"
 173                         " lo %d new %llu existing %llu"
 174                         " new peer %llu existing peer %llu"
 175                         " new dev %p existing dev %p\n",
 176                         conn, libcfs_nid2str(peer->gnp_nid),
 177                         loopback,
 178                         newconn->gnc_peerstamp, conn->gnc_peerstamp,
 179                         newconn->gnc_peer_connstamp, conn->gnc_peer_connstamp,
 180                         newconn->gnc_device, conn->gnc_device);
 181
 182                 /* conn is in the process of closing */
 183                 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
 184                         continue;
 185
 186                 /* 'newconn' is from an earlier version of 'peer'!!! */
 187                 if (newconn->gnc_peerstamp < conn->gnc_peerstamp)
 188                         RETURN(1);
 189
 190                 /* 'conn' is from an earlier version of 'peer': it will be
 191                  * removed when we cull stale conns later on... */
 192                 if (newconn->gnc_peerstamp > conn->gnc_peerstamp)
 193                         continue;
 194
 195                 /* Different devices are OK */
 196                 if (conn->gnc_device != newconn->gnc_device)
 197                         continue;
 198
 199                 /* It's me connecting to myself */
 200                 if (loopback &&
 201                     newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
 202                     newconn->gnc_peer_connstamp == conn->gnc_my_connstamp)
 203                         continue;
 204
 205                 /* 'newconn' is an earlier connection from 'peer'!!! */
 206                 if (newconn->gnc_peer_connstamp < conn->gnc_peer_connstamp)
 207                         RETURN(2);
 208
 209                 /* 'conn' is an earlier connection from 'peer': it will be
 210                  * removed when we cull stale conns later on... */
 211                 if (newconn->gnc_peer_connstamp > conn->gnc_peer_connstamp)
 212                         continue;
 213
 214                 /* 'newconn' has the SAME connection stamp; 'peer' isn't
 215                  * playing the game... */
 216                 RETURN(3);
 217         }
 218
 219         RETURN(0);
 220 }
 221
 222 int
 223 kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 224 {
 225         kgn_conn_t      *conn;
 226         gni_return_t    rrc;
 227         int             rc = 0;
 228
 229         LASSERT (!in_interrupt());
 230         atomic_inc(&kgnilnd_data.kgn_nconns);
 231
 232         /* divide by 2 to allow for complete reset and immediate reconnect */
 233         if (atomic_read(&kgnilnd_data.kgn_nconns) >= GNILND_MAX_CQID/2) {
 234                 CERROR("Too many conn are live: %d > %d\n",
 235                         atomic_read(&kgnilnd_data.kgn_nconns), GNILND_MAX_CQID/2);
 236                 atomic_dec(&kgnilnd_data.kgn_nconns);
 237                 return -E2BIG;
 238         }
 239
 240         LIBCFS_ALLOC(conn, sizeof(*conn));
 241         if (conn == NULL) {
 242                 atomic_dec(&kgnilnd_data.kgn_nconns);
 243                 return -ENOMEM;
 244         }
 245
 246         conn->gnc_tx_ref_table =
 247                 kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
 248         if (conn->gnc_tx_ref_table == NULL) {
 249                 CERROR("Can't allocate conn tx_ref_table\n");
 250                 GOTO(failed, rc = -ENOMEM);
 251         }
 252
 253         mutex_init(&conn->gnc_smsg_mutex);
 254         mutex_init(&conn->gnc_rdma_mutex);
 255         atomic_set(&conn->gnc_refcount, 1);
 256         atomic_set(&conn->gnc_reaper_noop, 0);
 257         atomic_set(&conn->gnc_sched_noop, 0);
 258         atomic_set(&conn->gnc_tx_in_use, 0);
 259         INIT_LIST_HEAD(&conn->gnc_list);
 260         INIT_LIST_HEAD(&conn->gnc_hashlist);
 261         INIT_LIST_HEAD(&conn->gnc_schedlist);
 262         INIT_LIST_HEAD(&conn->gnc_fmaq);
 263         INIT_LIST_HEAD(&conn->gnc_mdd_list);
 264         INIT_LIST_HEAD(&conn->gnc_delaylist);
 265         spin_lock_init(&conn->gnc_list_lock);
 266         spin_lock_init(&conn->gnc_tx_lock);
 267         conn->gnc_magic = GNILND_CONN_MAGIC;
 268
 269         /* set tx id to nearly the end to make sure we find wrapping
 270          * issues soon */
 271         conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;
 272
 273         /* if this fails, we have conflicts and MAX_TX is too large */
 274         BUILD_BUG_ON(GNILND_MAX_MSG_ID >= GNILND_MSGID_CLOSE);
 275
 276         /* get a new unique CQ id for this conn */
 277         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
 278         conn->gnc_my_connstamp = kgnilnd_data.kgn_connstamp++;
 279         conn->gnc_cqid = kgnilnd_get_cqid_locked();
 280         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 281
 282         if (conn->gnc_cqid == 0) {
 283                 CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
 284                 GOTO(failed, rc = -E2BIG);
 285         }
 286
 287         CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
 288                 conn->gnc_cqid, conn);
 289
 290         /* need to be set before gnc_ephandle to allow kgnilnd_destroy_conn_ep to
 291          * check context */
 292         conn->gnc_device = dev;
 293
 294         conn->gnc_timeout = max(*kgnilnd_tunables.kgn_timeout,
 295                                  GNILND_MIN_TIMEOUT);
 296         kgnilnd_update_reaper_timeout(conn->gnc_timeout);
 297
 298         /* this is the ep_handle for doing SMSG & BTE */
 299         mutex_lock(&dev->gnd_cq_mutex);
 300         rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
 301                                 &conn->gnc_ephandle);
 302         mutex_unlock(&dev->gnd_cq_mutex);
 303         if (rrc != GNI_RC_SUCCESS)
 304                 GOTO(failed, rc = -ENETDOWN);
 305
 306         CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
 307                conn, conn->gnc_ephandle);
 308
 309         /* add ref for EP canceling */
 310         kgnilnd_conn_addref(conn);
 311         atomic_inc(&dev->gnd_neps);
 312
 313         *connp = conn;
 314         return 0;
 315
 316 failed:
 317         atomic_dec(&kgnilnd_data.kgn_nconns);
 318         kgnilnd_vfree(conn->gnc_tx_ref_table,
 319                       GNILND_MAX_MSG_ID * sizeof(void *));
 320         LIBCFS_FREE(conn, sizeof(*conn));
 321         return rc;
 322 }
 323
 324 /* needs to be called with kgn_peer_conn_lock held (read or write) */
 325 kgn_conn_t *
 326 kgnilnd_find_conn_locked(kgn_peer_t *peer)
 327 {
 328         kgn_conn_t      *conn = NULL;
 329
 330         /* if we are in reset, this conn is going to die soon */
 331         if (unlikely(kgnilnd_data.kgn_in_reset)) {
 332                 RETURN(NULL);
 333         }
 334
 335         /* just return the first ESTABLISHED connection */
 336         list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
 337                 /* kgnilnd_finish_connect doesn't put connections on the
 338                  * peer list until they are actually established */
 339                 LASSERTF(conn->gnc_state >= GNILND_CONN_ESTABLISHED,
 340                         "found conn %p state %s on peer %p (%s)\n",
 341                         conn, kgnilnd_conn_state2str(conn), peer,
 342                         libcfs_nid2str(peer->gnp_nid));
 343                 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
 344                         continue;
 345
 346                 RETURN(conn);
 347         }
 348         RETURN(NULL);
 349 }
 350
 351 /* needs write_lock on kgn_peer_conn_lock held */
 352 kgn_conn_t *
 353 kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer) {
 354
 355         kgn_device_t    *dev = peer->gnp_net->gnn_dev;
 356         kgn_conn_t      *conn;
 357
 358         conn = kgnilnd_find_conn_locked(peer);
 359
 360         if (conn != NULL) {
 361                 return conn;
 362         }
 363
 364         /* if the peer was previously connecting, check if we should
 365          * trigger another connection attempt yet. */
 366         if (time_before(jiffies, peer->gnp_reconnect_time)) {
 367                 return NULL;
 368         }
 369
 370         /* This check prevents us from creating a new connection to a peer while we are
 371          * still in the process of closing an existing connection to the peer.
 372          */
 373         list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
 374                 if (conn->gnc_ephandle != NULL) {
 375                         CDEBUG(D_NET, "Not connecting non-null ephandle found peer 0x%p->%s\n", peer,
 376                                 libcfs_nid2str(peer->gnp_nid));
 377                         return NULL;
 378                 }
 379         }
 380
 381         if (peer->gnp_connecting != GNILND_PEER_IDLE) {
 382                 /* if we are not connecting, fire up a new connection */
 383                 /* or if we are anything but IDLE DONT start a new connection */
 384                return NULL;
 385         }
 386
 387         CDEBUG(D_NET, "starting connect to %s\n",
 388                 libcfs_nid2str(peer->gnp_nid));
 389         peer->gnp_connecting = GNILND_PEER_CONNECT;
 390         kgnilnd_peer_addref(peer); /* extra ref for connd */
 391
 392         spin_lock(&dev->gnd_connd_lock);
 393         list_add_tail(&peer->gnp_connd_list, &dev->gnd_connd_peers);
 394         spin_unlock(&dev->gnd_connd_lock);
 395
 396         kgnilnd_schedule_dgram(dev);
 397         CDEBUG(D_NETTRACE, "scheduling new connect\n");
 398
 399         return NULL;
 400 }
 401
 402 /* Caller is responsible for deciding if/when to call this */
 403 void
 404 kgnilnd_destroy_conn_ep(kgn_conn_t *conn)
 405 {
 406         gni_return_t    rrc;
 407         gni_ep_handle_t tmp_ep;
 408
 409         /* only if we actually initialized it,
 410          *  then set NULL to tell kgnilnd_destroy_conn to leave it alone */
 411
 412         tmp_ep = xchg(&conn->gnc_ephandle, NULL);
 413         if (tmp_ep != NULL) {
 414                 /* we never re-use the EP, so unbind is not needed */
 415                 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
 416                 rrc = kgnilnd_ep_destroy(tmp_ep);
 417
 418                 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
 419
 420                 /* if this fails, it could hork up kgni smsg retransmit and others
 421                  * since we could free the SMSG mbox memory, etc. */
 422                 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d conn 0x%p ep 0x%p\n",
 423                          rrc, conn, conn->gnc_ephandle);
 424
 425                 atomic_dec(&conn->gnc_device->gnd_neps);
 426
 427                 /* clear out count added in kgnilnd_close_conn_locked
 428                  * conn will have a peer once it hits finish_connect, where it
 429                  * is the first spot we'll mark it ESTABLISHED as well */
 430                 if (conn->gnc_peer) {
 431                         kgnilnd_admin_decref(conn->gnc_peer->gnp_dirty_eps);
 432                 }
 433
 434                 /* drop ref for EP */
 435                 kgnilnd_conn_decref(conn);
 436         }
 437 }
 438
 439 void
 440 kgnilnd_destroy_conn(kgn_conn_t *conn)
 441 {
 442         LASSERTF(!in_interrupt() &&
 443                 !conn->gnc_scheduled &&
 444                 !conn->gnc_in_purgatory &&
 445                 conn->gnc_ephandle == NULL &&
 446                 list_empty(&conn->gnc_list) &&
 447                 list_empty(&conn->gnc_hashlist) &&
 448                 list_empty(&conn->gnc_schedlist) &&
 449                 list_empty(&conn->gnc_mdd_list) &&
 450                 list_empty(&conn->gnc_delaylist) &&
 451                 conn->gnc_magic == GNILND_CONN_MAGIC,
 452                 "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n",
 453                 conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
 454                                      : "<?>",
 455                 !!in_interrupt(), conn->gnc_scheduled,
 456                 conn->gnc_in_purgatory,
 457                 conn->gnc_ephandle,
 458                 conn->gnc_magic,
 459                 list_empty(&conn->gnc_list),
 460                 list_empty(&conn->gnc_hashlist),
 461                 list_empty(&conn->gnc_schedlist),
 462                 list_empty(&conn->gnc_mdd_list),
 463                 list_empty(&conn->gnc_delaylist));
 464
 465         /* Tripping these is especially bad, as it means we have items on the
 466          *  lists that didn't keep their refcount on the connection - or
 467          *  somebody evil released their own */
 468         LASSERTF(list_empty(&conn->gnc_fmaq) &&
 469                  atomic_read(&conn->gnc_nlive_fma) == 0 &&
 470                  atomic_read(&conn->gnc_nlive_rdma) == 0,
 471                  "conn 0x%p fmaq %d@0x%p nfma %d nrdma %d\n",
 472                  conn, kgnilnd_count_list(&conn->gnc_fmaq), &conn->gnc_fmaq,
 473                  atomic_read(&conn->gnc_nlive_fma), atomic_read(&conn->gnc_nlive_rdma));
 474
 475         CDEBUG(D_NET, "destroying conn %p ephandle %p error %d\n",
 476                 conn, conn->gnc_ephandle, conn->gnc_error);
 477
 478         /* We are freeing this memory remove the magic value from the connection */
 479         conn->gnc_magic = 0;
 480
 481         /* if there is an FMA blk left here, we'll tear it down */
 482         if (conn->gnc_fma_blk) {
 483                 if (conn->gnc_peer) {
 484                         kgn_mbox_info_t *mbox;
 485                         mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
 486                         mbox->mbx_prev_nid = conn->gnc_peer->gnp_nid;
 487                 }
 488                 kgnilnd_release_mbox(conn, 0);
 489         }
 490
 491         if (conn->gnc_peer != NULL)
 492                 kgnilnd_peer_decref(conn->gnc_peer);
 493
 494         if (conn->gnc_tx_ref_table != NULL) {
 495                 kgnilnd_vfree(conn->gnc_tx_ref_table,
 496                               GNILND_MAX_MSG_ID * sizeof(void *));
 497         }
 498
 499         LIBCFS_FREE(conn, sizeof(*conn));
 500         atomic_dec(&kgnilnd_data.kgn_nconns);
 501 }
 502
 503 /* peer_alive and peer_notify done in the style of the o2iblnd */
 504 void
 505 kgnilnd_peer_alive(kgn_peer_t *peer)
 506 {
 507         time64_t now = ktime_get_seconds();
 508
 509         set_mb(peer->gnp_last_alive, now);
 510 }
 511
 512 void
 513 kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive)
 514 {
 515         int                     tell_lnet = 0;
 516         int                     nnets = 0;
 517         int                     rc;
 518         int                     i, j;
 519         kgn_conn_t             *conn;
 520         kgn_net_t             **nets;
 521         kgn_net_t              *net;
 522
 523
 524         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DONT_NOTIFY))
 525                 return;
 526
 527         /* Tell LNet we are giving ups on this peer - but only
 528          * if it isn't already reconnected or trying to reconnect */
 529         read_lock(&kgnilnd_data.kgn_peer_conn_lock);
 530
 531         /* use kgnilnd_find_conn_locked to avoid any conns in the process of being nuked
 532          *
 533          * don't tell LNet if we are in reset - we assume that everyone will be able to
 534          * reconnect just fine
 535          */
 536         conn = kgnilnd_find_conn_locked(peer);
 537
 538         CDEBUG(D_NETTRACE, "peer 0x%p->%s ting %d conn 0x%p, rst %d error %d\n",
 539                peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
 540                kgnilnd_data.kgn_in_reset, error);
 541
 542         if (((peer->gnp_connecting == GNILND_PEER_IDLE) &&
 543             (conn == NULL) &&
 544             (!kgnilnd_data.kgn_in_reset) &&
 545             (!kgnilnd_conn_clean_errno(error))) || alive) {
 546                 tell_lnet = 1;
 547         }
 548
 549         read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 550
 551         if (!tell_lnet) {
 552                 /* short circuit if we dont need to notify Lnet */
 553                 return;
 554         }
 555
 556         rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
 557
 558         if (rc) {
 559             /* dont do this if this fails since LNET is in shutdown or something else
 560              */
 561
 562                 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
 563                         list_for_each_entry(net , &kgnilnd_data.kgn_nets[i], gnn_list) {
 564                                 /* if gnn_shutdown set for any net shutdown is in progress just return */
 565                                 if (net->gnn_shutdown) {
 566                                         up_read(&kgnilnd_data.kgn_net_rw_sem);
 567                                         return;
 568                                 }
 569                                 nnets++;
 570                         }
 571                 }
 572
 573                 if (nnets == 0) {
 574                         /* shutdown in progress most likely */
 575                         up_read(&kgnilnd_data.kgn_net_rw_sem);
 576                         return;
 577                 }
 578
 579                 CFS_ALLOC_PTR_ARRAY(nets, nnets);
 580
 581                 if (nets == NULL) {
 582                         up_read(&kgnilnd_data.kgn_net_rw_sem);
 583                         CERROR("Failed to allocate nets[%d]\n", nnets);
 584                         return;
 585                 }
 586
 587                 j = 0;
 588                 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
 589                         list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
 590                                 nets[j] = net;
 591                                 kgnilnd_net_addref(net);
 592                                 j++;
 593                         }
 594                 }
 595                 up_read(&kgnilnd_data.kgn_net_rw_sem);
 596
 597                 for (i = 0; i < nnets; i++) {
 598                         lnet_nid_t peer_nid;
 599
 600                         net = nets[i];
 601
 602                         peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid,
 603                                                                  peer->gnp_nid);
 604
 605                         CDEBUG(D_NET, "peer 0x%p->%s last_alive %lld (%llds ago)\n",
 606                                 peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
 607                                 ktime_get_seconds() - peer->gnp_last_alive);
 608
 609                         lnet_notify(net->gnn_ni, peer_nid, alive, true,
 610                                     peer->gnp_last_alive);
 611
 612                         kgnilnd_net_decref(net);
 613                 }
 614
 615                 CFS_FREE_PTR_ARRAY(nets, nnets);
 616         }
 617 }
 618
 619 /* need write_lock on kgn_peer_conn_lock */
 620 void
 621 kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
 622 {
 623         kgn_peer_t        *peer = conn->gnc_peer;
 624         ENTRY;
 625
 626         LASSERT(!in_interrupt());
 627
 628         /* store error for tx completion */
 629         conn->gnc_error = error;
 630         peer->gnp_last_errno = error;
 631
 632         /* use real error from peer if possible */
 633         if (error == -ECONNRESET) {
 634                 error = conn->gnc_peer_error;
 635         }
 636
 637         /* if we NETERROR, make sure it is rate limited */
 638         if (!kgnilnd_conn_clean_errno(error) &&
 639             peer->gnp_state != GNILND_PEER_DOWN) {
 640                 CNETERR("closing conn to %s: error %d\n",
 641                        libcfs_nid2str(peer->gnp_nid), error);
 642         } else {
 643                 CDEBUG(D_NET, "closing conn to %s: error %d\n",
 644                        libcfs_nid2str(peer->gnp_nid), error);
 645         }
 646
 647         LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
 648                 "conn %p to %s with bogus state %s\n", conn,
 649                 libcfs_nid2str(conn->gnc_peer->gnp_nid),
 650                 kgnilnd_conn_state2str(conn));
 651         LASSERT(!list_empty(&conn->gnc_hashlist));
 652         LASSERT(!list_empty(&conn->gnc_list));
 653
 654
 655         /* mark peer count here so any place the EP gets destroyed will
 656          * open up the peer count so that a new ESTABLISHED conn is then free
 657          * to send new messages -- sending before the previous EPs are destroyed
 658          * could end up with messages on the network for the old conn _after_
 659          * the new conn and break the mbox safety protocol */
 660         kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
 661
 662         /* Remove from conn hash table: no new callbacks */
 663         list_del_init(&conn->gnc_hashlist);
 664         kgnilnd_data.kgn_conn_version++;
 665         kgnilnd_conn_decref(conn);
 666
 667         /* if we are in reset, go right to CLOSED as there is no scheduler
 668          * thread to move from CLOSING to CLOSED */
 669         if (unlikely(kgnilnd_data.kgn_in_reset)) {
 670                 conn->gnc_state = GNILND_CONN_CLOSED;
 671         } else {
 672                 conn->gnc_state = GNILND_CONN_CLOSING;
 673         }
 674
 675         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
 676                 msleep_interruptible(MSEC_PER_SEC);
 677         }
 678
 679         /* leave on peer->gnp_conns to make sure we don't let the reaper
 680          * or others try to unlink this peer until the conn is fully
 681          * processed for closing */
 682
 683         if (kgnilnd_check_purgatory_conn(conn)) {
 684                 kgnilnd_add_purgatory_locked(conn, conn->gnc_peer);
 685         }
 686
 687         /* Reset RX timeout to ensure we wait for an incoming CLOSE
 688          * for the full timeout.  If we get a CLOSE we know the
 689          * peer has stopped all RDMA.  Otherwise if we wait for
 690          * the full timeout we can also be sure all RDMA has stopped. */
 691         conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
 692         mb();
 693
 694         /* schedule sending CLOSE - if we are in quiesce, this adds to
 695          * gnd_ready_conns and allows us to find it in quiesce processing */
 696         kgnilnd_schedule_conn(conn);
 697
 698         EXIT;
 699 }
 700
 701 void
 702 kgnilnd_close_conn(kgn_conn_t *conn, int error)
 703 {
 704         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
 705         /* need to check the state here - this call is racy and we don't
 706          * know the state until after the lock is grabbed */
 707         if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
 708                 kgnilnd_close_conn_locked(conn, error);
 709         }
 710         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 711 }
 712
 713 void
 714 kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 715 {
 716         LIST_HEAD               (sinners);
 717         kgn_tx_t               *tx, *txn;
 718         int                     nlive = 0;
 719         int                     nrdma = 0;
 720         int                     nq_rdma = 0;
 721         int                     logmsg;
 722         ENTRY;
 723
 724         /* Dump log  on cksum error - wait until complete phase to let
 725          * RX of error happen */
 726         if (*kgnilnd_tunables.kgn_checksum_dump &&
 727             (conn != NULL && conn->gnc_peer_error == -ENOKEY)) {
 728                 libcfs_debug_dumplog();
 729         }
 730
 731         /* _CLOSED set in kgnilnd_process_fmaq once we decide to
 732          * send the CLOSE or not */
 733         LASSERTF(conn->gnc_state == GNILND_CONN_CLOSED,
 734                  "conn 0x%p->%s with bad state %s\n",
 735                  conn, conn->gnc_peer ?
 736                         libcfs_nid2str(conn->gnc_peer->gnp_nid) :
 737                         "<?>",
 738                  kgnilnd_conn_state2str(conn));
 739
 740         LASSERT(list_empty(&conn->gnc_hashlist));
 741         /* We shouldnt be on the delay list, the conn can
 742          * get added to this list during a retransmit, and retransmits
 743          * only occur within scheduler threads.
 744          */
 745         LASSERT(list_empty(&conn->gnc_delaylist));
 746
 747         /* we've sent the close, start nuking */
 748         if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
 749                 kgnilnd_schedule_conn(conn);
 750
 751         if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
 752                 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
 753                                 "done, Attempting to recover conn 0x%p "
 754                                 "scheduled %d function: %s line: %d\n", conn,
 755                                 conn->gnc_scheduled, conn->gnc_sched_caller,
 756                                 conn->gnc_sched_line);
 757                 RETURN_EXIT;
 758         }
 759
 760         /* we don't use lists to track things that we can get out of the
 761          * tx_ref table... */
 762
 763         /* need to hold locks for tx_list_state, sampling it is too racy:
 764          * - the lock actually protects tx != NULL, but we can't take the proper
 765          *   lock until we check tx_list_state, which would be too late and
 766          *   we could have the TX change under us.
 767          * gnd_rdmaq_lock and gnd_lock and not used together, so taking both
 768          * should be fine */
 769         spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
 770         spin_lock(&conn->gnc_device->gnd_lock);
 771
 772         for (nrdma = 0; nrdma < GNILND_MAX_MSG_ID; nrdma++) {
 773                 tx = conn->gnc_tx_ref_table[nrdma];
 774
 775                 if (tx != NULL) {
 776                         /* only print the first error and if not CLOSE, we often don't see
 777                          * CQ events for that by the time we get here... and really don't care */
 778                         if (nlive || tx->tx_msg.gnm_type == GNILND_MSG_CLOSE)
 779                                 tx->tx_state |= GNILND_TX_QUIET_ERROR;
 780                         nlive++;
 781                         GNIDBG_TX(D_NET, tx, "cleaning up on close, nlive %d", nlive);
 782
 783                         /* don't worry about gnc_lock here as nobody else should be
 784                          * touching this conn */
 785                         kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
 786                         list_add_tail(&tx->tx_list, &sinners);
 787                 }
 788         }
 789         spin_unlock(&conn->gnc_device->gnd_lock);
 790         spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
 791
 792         /* nobody should have marked this as needing scheduling after
 793          * we called close - so only ref should be us handling it */
 794         if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
 795                 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
 796                                 "done, Attempting to recover conn 0x%p "
 797                                 "scheduled %d function %s line: %d\n", conn,
 798                                 conn->gnc_scheduled, conn->gnc_sched_caller,
 799                                 conn->gnc_sched_line);
 800         }
 801         /* now reset a few to actual counters... */
 802         nrdma = atomic_read(&conn->gnc_nlive_rdma);
 803         nq_rdma = atomic_read(&conn->gnc_nq_rdma);
 804
 805         if (!list_empty(&sinners)) {
 806                 list_for_each_entry_safe(tx, txn, &sinners, tx_list) {
 807                         /* clear tx_list to make tx_add_list_locked happy */
 808                         list_del_init(&tx->tx_list);
 809                         /* The error codes determine if we hold onto the MDD */
 810                         kgnilnd_tx_done(tx, conn->gnc_error);
 811                 }
 812         }
 813
 814         logmsg = (nlive + nrdma + nq_rdma);
 815
 816         if (logmsg) {
 817                 int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ?
 818                                 D_NETERROR : D_NET;
 819                 CDEBUG(level, "Closed conn 0x%p->%s (errno %d,"
 820                         " peer errno %d): canceled %d TX, %d/%d RDMA\n",
 821                         conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
 822                         conn->gnc_error, conn->gnc_peer_error,
 823                         nlive, nq_rdma, nrdma);
 824         }
 825
 826         kgnilnd_destroy_conn_ep(conn);
 827
 828         /* Bug 765042 - race this with completing a new conn to same peer - we need
 829          * finish_connect to detach purgatory before we can do it ourselves here */
 830         CFS_RACE(CFS_FAIL_GNI_FINISH_PURG);
 831
 832         /* now it is safe to remove from peer list - anyone looking at
 833          * gnp_conns now is free to unlink if not on purgatory */
 834         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
 835
 836         conn->gnc_state = GNILND_CONN_DONE;
 837
 838         /* Decrement counter if we are marked by del_conn_or_peers for closing
 839          */
 840         if (conn->gnc_needs_closing)
 841                 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_conns);
 842
 843         /* Remove from peer's list of valid connections if its not in purgatory */
 844         if (!conn->gnc_in_purgatory) {
 845                 list_del_init(&conn->gnc_list);
 846                 /* Lose peers reference on the conn */
 847                 kgnilnd_conn_decref(conn);
 848         }
 849
 850         /* NB - only unlinking if we set pending in del_peer_locked from admin or
 851          * shutdown */
 852         if (kgnilnd_peer_active(conn->gnc_peer) &&
 853             conn->gnc_peer->gnp_pending_unlink &&
 854             kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
 855                 kgnilnd_unlink_peer_locked(conn->gnc_peer);
 856         }
 857
 858         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 859
 860         /* I'm telling Mommy! - use peer_error if they initiated close */
 861         kgnilnd_peer_notify(conn->gnc_peer,
 862                             conn->gnc_error == -ECONNRESET ?
 863                             conn->gnc_peer_error : conn->gnc_error, 0);
 864
 865         EXIT;
 866 }
 867
 868 int
 869 kgnilnd_set_conn_params(kgn_dgram_t *dgram)
 870 {
 871         kgn_conn_t             *conn = dgram->gndg_conn;
 872         kgn_connreq_t          *connreq = &dgram->gndg_conn_in;
 873         kgn_gniparams_t        *rem_param = &connreq->gncr_gnparams;
 874         gni_return_t            rrc;
 875         int                     rc = 0;
 876         gni_smsg_attr_t        *remote = &connreq->gncr_gnparams.gnpr_smsg_attr;
 877
 878         /* set timeout vals in conn early so we can use them for the NAK */
 879
 880         /* use max of the requested and our timeout, peer will do the same */
 881         conn->gnc_timeout = max(conn->gnc_timeout, connreq->gncr_timeout);
 882
 883         /* only ep_bind really mucks around with the CQ */
 884         /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
 885          * is necessary as you can only bind an ep once and we must make sure we dont bind when already bound.
 886          */
 887         if (connreq->gncr_dstnid != LNET_NID_ANY && dgram->gndg_conn_out.gncr_dstnid != connreq->gncr_srcnid) {
 888                 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
 889                 rrc = kgnilnd_ep_bind(conn->gnc_ephandle,
 890                         connreq->gncr_gnparams.gnpr_host_id,
 891                         conn->gnc_cqid);
 892                 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
 893                 if (rrc != GNI_RC_SUCCESS) {
 894                         rc = -ECONNABORTED;
 895                         goto return_out;
 896                 }
 897         }
 898
 899         rrc = kgnilnd_ep_set_eventdata(conn->gnc_ephandle, conn->gnc_cqid,
 900                          connreq->gncr_gnparams.gnpr_cqid);
 901         if (rrc != GNI_RC_SUCCESS) {
 902                 rc = -ECONNABORTED;
 903                 goto cleanup_out;
 904         }
 905
 906         /* Initialize SMSG */
 907         rrc = kgnilnd_smsg_init(conn->gnc_ephandle, &conn->gnpr_smsg_attr,
 908                         &connreq->gncr_gnparams.gnpr_smsg_attr);
 909         if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
 910                 gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
 911                 /* help folks figure out if there is a tunable off, etc. */
 912                 LCONSOLE_ERROR("SMSG attribute mismatch. Data from local/remote:"
 913                                " type %d/%d msg_maxsize %u/%u"
 914                                " mbox_maxcredit %u/%u. Please check kgni"
 915                                " logs for further data\n",
 916                                local->msg_type, remote->msg_type,
 917                                local->msg_maxsize, remote->msg_maxsize,
 918                                local->mbox_maxcredit, remote->mbox_maxcredit);
 919         }
 920         if (rrc != GNI_RC_SUCCESS) {
 921                 rc = -ECONNABORTED;
 922                 goto cleanup_out;
 923         }
 924
 925         /* log this for help in debuggin SMSG buffer re-use */
 926         CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
 927                 " local cqid %u SMSG %p->%u hndl %#llx.%#llx"
 928                 " remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n",
 929                 conn, libcfs_nid2str(connreq->gncr_srcnid),
 930                 libcfs_nid2str(connreq->gncr_dstnid),
 931                 &conn->gnpr_smsg_attr,
 932                 conn->gnc_cqid,
 933                 conn->gnpr_smsg_attr.msg_buffer,
 934                 conn->gnpr_smsg_attr.mbox_offset,
 935                 conn->gnpr_smsg_attr.mem_hndl.qword1,
 936                 conn->gnpr_smsg_attr.mem_hndl.qword2,
 937                 rem_param->gnpr_cqid,
 938                 rem_param->gnpr_smsg_attr.msg_buffer,
 939                 rem_param->gnpr_smsg_attr.mbox_offset,
 940                 rem_param->gnpr_smsg_attr.mem_hndl.qword1,
 941                 rem_param->gnpr_smsg_attr.mem_hndl.qword2);
 942
 943         conn->gnc_peerstamp = connreq->gncr_peerstamp;
 944         conn->gnc_peer_connstamp = connreq->gncr_connstamp;
 945         conn->remote_mbox_addr = (void *)((char *)remote->msg_buffer + remote->mbox_offset);
 946
 947         /* We update the reaper timeout once we have a valid conn and timeout */
 948         kgnilnd_update_reaper_timeout(GNILND_TO2KA(conn->gnc_timeout));
 949
 950         return 0;
 951
 952 cleanup_out:
 953         rrc = kgnilnd_ep_unbind(conn->gnc_ephandle);
 954         /* not sure I can just let this fly */
 955         LASSERTF(rrc == GNI_RC_SUCCESS,
 956                 "bad rc from gni_ep_unbind trying to cleanup: %d\n", rrc);
 957
 958 return_out:
 959         LASSERTF(rc != 0, "SOFTWARE BUG: rc == 0\n");
 960         CERROR("Error setting connection params from %s: %d\n",
 961                libcfs_nid2str(connreq->gncr_srcnid), rc);
 962         return rc;
 963 }
 964
 965 /* needs down_read on kgn_net_rw_sem held from before this call until
 966  * after the write_lock on kgn_peer_conn_lock - this ensures we stay sane
 967  * with kgnilnd_shutdown - it'll get the sem and set shutdown, then get the
 968  * kgn_peer_conn_lock to start del_peer'ing. If we hold the sem until after
 969  * kgn_peer_conn_lock is held, we guarantee that nobody calls
 970  * kgnilnd_add_peer_locked without checking gnn_shutdown */
 971 int
 972 kgnilnd_create_peer_safe(kgn_peer_t **peerp,
 973                          lnet_nid_t nid,
 974                          kgn_net_t *net,
 975                          int node_state)
 976 {
 977         kgn_peer_t      *peer;
 978         int             rc;
 979
 980         LASSERT(nid != LNET_NID_ANY);
 981
 982         /* We dont pass the net around in the dgram anymore so here is where we find it
 983          * this will work unless its in shutdown or the nid has a net that is invalid.
 984          * Either way error code needs to be returned in that case.
 985          *
 986          * If the net passed in is not NULL then we can use it, this alleviates looking it
 987          * when the calling function has access to the data.
 988          */
 989         if (net == NULL) {
 990                 rc = kgnilnd_find_net(nid, &net);
 991                 if (rc < 0)
 992                         return rc;
 993         } else {
 994                 /* find net adds a reference on the net if we are not using
 995                  * it we must do it manually so the net references are
 996                  * correct when tearing down the net
 997                  */
 998                 kgnilnd_net_addref(net);
 999         }
1000
1001         LIBCFS_ALLOC(peer, sizeof(*peer));
1002         if (peer == NULL) {
1003                 kgnilnd_net_decref(net);
1004                 return -ENOMEM;
1005         }
1006         peer->gnp_nid = nid;
1007         peer->gnp_state = node_state;
1008
1009         /* translate from nid to nic addr & store */
1010         rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
1011         if (rc <= 0) {
1012                 kgnilnd_net_decref(net);
1013                 LIBCFS_FREE(peer, sizeof(*peer));
1014                 return -ESRCH;
1015         }
1016         CDEBUG(D_NET, "peer 0x%p->%s -> NIC 0x%x\n", peer,
1017                 libcfs_nid2str(nid), peer->gnp_host_id);
1018
1019         atomic_set(&peer->gnp_refcount, 1);     /* 1 ref for caller */
1020         atomic_set(&peer->gnp_dirty_eps, 0);
1021
1022         INIT_LIST_HEAD(&peer->gnp_list);
1023         INIT_LIST_HEAD(&peer->gnp_connd_list);
1024         INIT_LIST_HEAD(&peer->gnp_conns);
1025         INIT_LIST_HEAD(&peer->gnp_tx_queue);
1026
1027         /* the first reconnect should happen immediately, so we leave
1028          * gnp_reconnect_interval set to 0 */
1029
1030         LASSERTF(net != NULL, "peer 0x%p->%s with NULL net\n",
1031                  peer, libcfs_nid2str(nid));
1032
1033         /* must have kgn_net_rw_sem held for this...  */
1034         if (net->gnn_shutdown) {
1035                 /* shutdown has started already */
1036                 kgnilnd_net_decref(net);
1037                 LIBCFS_FREE(peer, sizeof(*peer));
1038                 return -ESHUTDOWN;
1039         }
1040
1041         peer->gnp_net = net;
1042
1043         atomic_inc(&kgnilnd_data.kgn_npeers);
1044
1045         *peerp = peer;
1046         return 0;
1047 }
1048
1049 void
1050 kgnilnd_destroy_peer(kgn_peer_t *peer)
1051 {
1052         CDEBUG(D_NET, "peer %s %p deleted\n",
1053                libcfs_nid2str(peer->gnp_nid), peer);
1054         LASSERTF(atomic_read(&peer->gnp_refcount) == 0,
1055                  "peer 0x%p->%s refs %d\n",
1056                  peer, libcfs_nid2str(peer->gnp_nid),
1057                  atomic_read(&peer->gnp_refcount));
1058         LASSERTF(atomic_read(&peer->gnp_dirty_eps) == 0,
1059                  "peer 0x%p->%s dirty eps %d\n",
1060                  peer, libcfs_nid2str(peer->gnp_nid),
1061                  atomic_read(&peer->gnp_dirty_eps));
1062         LASSERTF(peer->gnp_net != NULL, "peer %p (%s) with NULL net\n",
1063                  peer, libcfs_nid2str(peer->gnp_nid));
1064         LASSERTF(!kgnilnd_peer_active(peer),
1065                  "peer 0x%p->%s\n",
1066                 peer, libcfs_nid2str(peer->gnp_nid));
1067         LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE || peer->gnp_connecting == GNILND_PEER_KILL,
1068                  "peer 0x%p->%s, connecting %d\n",
1069                 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1070         LASSERTF(list_empty(&peer->gnp_conns),
1071                  "peer 0x%p->%s\n",
1072                 peer, libcfs_nid2str(peer->gnp_nid));
1073         LASSERTF(list_empty(&peer->gnp_tx_queue),
1074                  "peer 0x%p->%s\n",
1075                 peer, libcfs_nid2str(peer->gnp_nid));
1076         LASSERTF(list_empty(&peer->gnp_connd_list),
1077                  "peer 0x%p->%s\n",
1078                 peer, libcfs_nid2str(peer->gnp_nid));
1079
1080         /* NB a peer's connections keep a reference on their peer until
1081          * they are destroyed, so we can be assured that _all_ state to do
1082          * with this peer has been cleaned up when its refcount drops to
1083          * zero. */
1084
1085         atomic_dec(&kgnilnd_data.kgn_npeers);
1086         kgnilnd_net_decref(peer->gnp_net);
1087
1088         LIBCFS_FREE(peer, sizeof(*peer));
1089 }
1090
1091 /* the conn might not have made it all the way through to a connected
1092  * state - but we need to purgatory any conn that a remote peer might
1093  * have seen through a posted dgram as well */
1094 void
1095 kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
1096 {
1097         kgn_mbox_info_t *mbox = NULL;
1098         ENTRY;
1099
1100         /* NB - the caller should own conn by removing him from the
1101          * scheduler thread when finishing the close */
1102
1103         LASSERTF(peer != NULL, "conn %p with NULL peer\n", conn);
1104
1105         /* If this is still true, need to add the calls to unlink back in and
1106          * figure out how to close the hole on loopback conns */
1107         LASSERTF(kgnilnd_peer_active(peer), "can't use inactive peer %s (%p)"
1108                 " we'll never recover the resources\n",
1109                 libcfs_nid2str(peer->gnp_nid), peer);
1110
1111         CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
1112                 conn->gnc_device);
1113
1114         LASSERTF(conn->gnc_in_purgatory == 0,
1115                 "Conn already in purgatory\n");
1116         conn->gnc_in_purgatory = 1;
1117
1118         mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1119         mbox->mbx_prev_purg_nid = peer->gnp_nid;
1120         mbox->mbx_add_purgatory = jiffies;
1121         kgnilnd_release_mbox(conn, 1);
1122
1123         LASSERTF(list_empty(&conn->gnc_mdd_list),
1124                 "conn 0x%p->%s with active purgatory hold MDD %d\n",
1125                 conn, libcfs_nid2str(peer->gnp_nid),
1126                 kgnilnd_count_list(&conn->gnc_mdd_list));
1127
1128         EXIT;
1129 }
1130
1131 /* Instead of detaching everything from purgatory here we just mark the conn as needing
1132  * detach, when the reaper checks the conn the next time it will detach it.
1133  * Calling function requires write_lock held on kgn_peer_conn_lock
1134  */
1135 void
1136 kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer) {
1137         kgn_conn_t       *conn;
1138
1139         list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
1140                 if (conn->gnc_in_purgatory && !conn->gnc_needs_detach) {
1141                         conn->gnc_needs_detach = 1;
1142                         kgnilnd_admin_addref(kgnilnd_data.kgn_npending_detach);
1143                 }
1144         }
1145 }
1146
1147 /* Calling function needs a write_lock held on kgn_peer_conn_lock */
1148 void
1149 kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
1150 {
1151         kgn_mbox_info_t *mbox = NULL;
1152
1153         /* if needed, add the conn purgatory data to the list passed in */
1154         if (conn->gnc_in_purgatory) {
1155                 CDEBUG(D_NET, "peer %p->%s purg_conn %p@%s mdd_list #tx %d\n",
1156                         conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1157                         conn, kgnilnd_conn_state2str(conn),
1158                         kgnilnd_count_list(&conn->gnc_mdd_list));
1159
1160                 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1161                 mbox->mbx_detach_of_purgatory = jiffies;
1162
1163                 /* conn->gnc_list is the entry point on peer->gnp_conns, so detaching it
1164                  * here removes it from the list of 'valid' peer connections.
1165                  * We put the current conn onto a list of conns to call kgnilnd_release_purgatory_locked()
1166                  * and as such the caller of kgnilnd_detach_purgatory_locked() now owns that conn, since its not
1167                  * on the peer's conn_list anymore.
1168                  */
1169
1170                 list_del_init(&conn->gnc_list);
1171
1172                 /* NB - only unlinking if we set pending in del_peer_locked from admin or
1173                  * shutdown */
1174                 if (kgnilnd_peer_active(conn->gnc_peer) &&
1175                     conn->gnc_peer->gnp_pending_unlink &&
1176                     kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
1177                         kgnilnd_unlink_peer_locked(conn->gnc_peer);
1178                 }
1179                 /* The reaper will not call detach unless the conn is fully through kgnilnd_complete_closed_conn.
1180                  * If the conn is not in a DONE state somehow we are attempting to detach even though
1181                  * the conn has not been fully cleaned up. If we detach while the conn is still closing
1182                  * we will end up with an orphaned connection that has valid ep_handle, that is not on a
1183                  * peer.
1184                  */
1185
1186                 LASSERTF(conn->gnc_state == GNILND_CONN_DONE, "Conn in invalid state  %p@%s \n",
1187                                 conn, kgnilnd_conn_state2str(conn));
1188
1189                 /* move from peer to the delayed release list */
1190                 list_add_tail(&conn->gnc_list, conn_list);
1191         }
1192 }
1193
1194 void
1195 kgnilnd_release_purgatory_list(struct list_head *conn_list)
1196 {
1197         kgn_device_t            *dev;
1198         kgn_conn_t              *conn, *connN;
1199         kgn_mdd_purgatory_t     *gmp, *gmpN;
1200
1201         list_for_each_entry_safe(conn, connN, conn_list, gnc_list) {
1202                 dev = conn->gnc_device;
1203
1204                 kgnilnd_release_mbox(conn, -1);
1205                 conn->gnc_in_purgatory = 0;
1206
1207                 list_del_init(&conn->gnc_list);
1208
1209                 /* gnc_needs_detach is set in kgnilnd_del_conn_or_peer. It is used to keep track
1210                  * of conns that have been marked for detach by kgnilnd_del_conn_or_peer.
1211                  * The function uses kgn_npending_detach to verify the conn has
1212                  * actually been detached.
1213                  */
1214
1215                 if (conn->gnc_needs_detach)
1216                         kgnilnd_admin_decref(kgnilnd_data.kgn_npending_detach);
1217
1218                 /* if this guy is really dead (we are doing release from reaper),
1219                  * make sure we tell LNet - if this is from other context,
1220                  * the checks in the function will prevent an errant
1221                  * notification */
1222                 kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
1223
1224                 list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
1225                                          gmp_list) {
1226                         CDEBUG(D_NET,
1227                                "dev %p releasing held mdd %#llx.%#llx\n",
1228                                conn->gnc_device, gmp->gmp_map_key.qword1,
1229                                gmp->gmp_map_key.qword2);
1230
1231                         atomic_dec(&dev->gnd_n_mdd_held);
1232                         kgnilnd_mem_mdd_release(conn->gnc_device->gnd_handle,
1233                                                 &gmp->gmp_map_key);
1234                         /* ignoring the return code - if kgni/ghal can't find it
1235                          * it must be released already */
1236
1237                         list_del_init(&gmp->gmp_list);
1238                         LIBCFS_FREE(gmp, sizeof(*gmp));
1239                 }
1240                 /* lose conn ref for purgatory */
1241                 kgnilnd_conn_decref(conn);
1242         }
1243 }
1244
1245 /* needs write_lock on kgnilnd_data.kgn_peer_conn_lock held */
1246 void
1247 kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer)
1248 {
1249         int current_to;
1250
1251         current_to = peer->gnp_reconnect_interval;
1252
1253         /* we'll try to reconnect fast the first time, then back-off */
1254         if (current_to == 0) {
1255                 peer->gnp_reconnect_time = jiffies - 1;
1256                 current_to = *kgnilnd_tunables.kgn_min_reconnect_interval;
1257         } else {
1258                 peer->gnp_reconnect_time = jiffies + cfs_time_seconds(current_to);
1259                 /* add 50% of min timeout & retry */
1260                 current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
1261         }
1262
1263         current_to = min(current_to,
1264                          *kgnilnd_tunables.kgn_max_reconnect_interval);
1265
1266         peer->gnp_reconnect_interval = current_to;
1267         CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
1268                libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_time,
1269                peer->gnp_reconnect_interval);
1270 }
1271
1272 /* needs kgnilnd_data.kgn_peer_conn_lock held */
1273 kgn_peer_t *
1274 kgnilnd_find_peer_locked(lnet_nid_t nid)
1275 {
1276         struct list_head *peer_list = kgnilnd_nid2peerlist(nid);
1277         kgn_peer_t       *peer;
1278
1279         /* Chopping nid down to only NIDADDR using LNET_NIDADDR so we only
1280          * have a single peer per device instead of a peer per nid/net combo.
1281          */
1282
1283         list_for_each_entry(peer, peer_list, gnp_list) {
1284                 if (LNET_NIDADDR(nid) != LNET_NIDADDR(peer->gnp_nid))
1285                         continue;
1286
1287                 CDEBUG(D_NET, "got peer [%p] -> %s c %d (%d)\n",
1288                        peer, libcfs_nid2str(nid),
1289                        peer->gnp_connecting,
1290                        atomic_read(&peer->gnp_refcount));
1291                 return peer;
1292         }
1293         return NULL;
1294 }
1295
1296 /* need write_lock on kgn_peer_conn_lock */
1297 void
1298 kgnilnd_unlink_peer_locked(kgn_peer_t *peer)
1299 {
1300         LASSERTF(list_empty(&peer->gnp_conns),
1301                 "peer 0x%p->%s\n",
1302                  peer, libcfs_nid2str(peer->gnp_nid));
1303         LASSERTF(list_empty(&peer->gnp_tx_queue),
1304                 "peer 0x%p->%s\n",
1305                  peer, libcfs_nid2str(peer->gnp_nid));
1306         LASSERTF(kgnilnd_peer_active(peer),
1307                 "peer 0x%p->%s\n",
1308                  peer, libcfs_nid2str(peer->gnp_nid));
1309         CDEBUG(D_NET, "unlinking peer 0x%p->%s\n",
1310                 peer, libcfs_nid2str(peer->gnp_nid));
1311
1312         list_del_init(&peer->gnp_list);
1313         kgnilnd_data.kgn_peer_version++;
1314         kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1315         /* lose peerlist's ref */
1316         kgnilnd_peer_decref(peer);
1317 }
1318
1319 int
1320 kgnilnd_get_peer_info(int index,
1321                       kgn_peer_t **found_peer,
1322                       lnet_nid_t *id, __u32 *nic_addr,
1323                       int *refcount, int *connecting)
1324 {
1325         struct list_head  *ptmp;
1326         kgn_peer_t        *peer;
1327         int               i;
1328         int               rc = -ENOENT;
1329
1330         read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1331
1332         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1333
1334                 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1335                         peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1336
1337                         if (index-- > 0)
1338                                 continue;
1339
1340                         CDEBUG(D_NET, "found peer %p (%s) at index %d\n",
1341                                peer, libcfs_nid2str(peer->gnp_nid), index);
1342
1343                         *found_peer  = peer;
1344                         *id          = peer->gnp_nid;
1345                         *nic_addr    = peer->gnp_host_id;
1346                         *refcount    = atomic_read(&peer->gnp_refcount);
1347                         *connecting  = peer->gnp_connecting;
1348
1349                         rc = 0;
1350                         goto out;
1351                 }
1352         }
1353 out:
1354         read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1355         if (rc)
1356                 CDEBUG(D_NET, "no gni peer at index %d\n", index);
1357         return rc;
1358 }
1359
1360 /* requires write_lock on kgn_peer_conn_lock held */
1361 void
1362 kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp)
1363 {
1364         kgn_peer_t        *peer, *peer2;
1365
1366         LASSERTF(new_stub_peer != NULL, "bad stub peer for nid %s\n",
1367                  libcfs_nid2str(nid));
1368
1369         peer2 = kgnilnd_find_peer_locked(nid);
1370         if (peer2 != NULL) {
1371                 /* A peer was created during the lock transition, so drop
1372                  * the new one we created */
1373                 kgnilnd_peer_decref(new_stub_peer);
1374                 peer = peer2;
1375         } else {
1376                 peer = new_stub_peer;
1377                 /* peer table takes existing ref on peer */
1378
1379                 LASSERTF(!kgnilnd_peer_active(peer),
1380                         "peer 0x%p->%s already in peer table\n",
1381                         peer, libcfs_nid2str(peer->gnp_nid));
1382                 list_add_tail(&peer->gnp_list,
1383                               kgnilnd_nid2peerlist(nid));
1384                 kgnilnd_data.kgn_peer_version++;
1385         }
1386
1387         LASSERTF(peer->gnp_net != NULL, "peer 0x%p->%s with NULL net\n",
1388                  peer, libcfs_nid2str(peer->gnp_nid));
1389         *peerp = peer;
1390 }
1391
1392 int
1393 kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
1394 {
1395         kgn_peer_t        *peer;
1396         int                rc;
1397         int                node_state;
1398         ENTRY;
1399
1400         if (nid == LNET_NID_ANY)
1401                 return -EINVAL;
1402
1403         node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));
1404
1405         /* NB - this will not block during normal operations -
1406          * the only writer of this is in the startup/shutdown path. */
1407         rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1408         if (!rc) {
1409                 rc = -ESHUTDOWN;
1410                 RETURN(rc);
1411         }
1412         rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
1413         if (rc != 0) {
1414                 up_read(&kgnilnd_data.kgn_net_rw_sem);
1415                 RETURN(rc);
1416         }
1417
1418         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1419         up_read(&kgnilnd_data.kgn_net_rw_sem);
1420
1421         kgnilnd_add_peer_locked(nid, peer, peerp);
1422
1423         CDEBUG(D_NET, "peer 0x%p->%s connecting %d\n",
1424                peerp, libcfs_nid2str((*peerp)->gnp_nid),
1425                (*peerp)->gnp_connecting);
1426
1427         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1428         RETURN(0);
1429 }
1430
1431 /* needs write_lock on kgn_peer_conn_lock */
1432 void
1433 kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies)
1434 {
1435         kgn_tx_t        *tx, *txn;
1436
1437         /* we do care about state of gnp_connecting - we could be between
1438          * reconnect attempts, so try to find the dgram and cancel the TX
1439          * anyways. If we are in the process of posting DONT do anything;
1440          * once it fails or succeeds we can nuke the connect attempt.
1441          * We have no idea where in kgnilnd_post_dgram we are so we cant
1442          * attempt to cancel until the function is done.
1443          */
1444
1445         /* make sure peer isn't in process of connecting or waiting for connect*/
1446         spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1447         if (!(list_empty(&peer->gnp_connd_list))) {
1448                 list_del_init(&peer->gnp_connd_list);
1449                 /* remove connd ref */
1450                 kgnilnd_peer_decref(peer);
1451         }
1452         spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1453
1454         if (peer->gnp_connecting == GNILND_PEER_POSTING || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1455                 peer->gnp_connecting = GNILND_PEER_NEEDS_DEATH;
1456                 /* We are in process of posting right now the xchg set it up for us to
1457                  * cancel the connect so we are finished for now */
1458         } else {
1459                 /* no need for exchange we have the peer lock and its ready for us to nuke */
1460                 LASSERTF(peer->gnp_connecting != GNILND_PEER_POSTING,
1461                         "Peer in invalid state 0x%p->%s, connecting %d\n",
1462                         peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1463                 peer->gnp_connecting = GNILND_PEER_IDLE;
1464                 set_mb(peer->gnp_last_dgram_errno, -ETIMEDOUT);
1465                 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
1466                                                       peer->gnp_nid);
1467         }
1468
1469         /* The least we can do is nuke the tx's no matter what.... */
1470         list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1471                 kgnilnd_tx_del_state_locked(tx, peer, NULL,
1472                                            GNILND_TX_ALLOCD);
1473                 list_add_tail(&tx->tx_list, zombies);
1474         }
1475 }
1476
1477 /* needs write_lock on kgn_peer_conn_lock */
1478 void
1479 kgnilnd_del_peer_locked(kgn_peer_t *peer, int error)
1480 {
1481         /* this peer could be passive and only held for purgatory,
1482          * take a ref to ensure it doesn't disappear in this function */
1483         kgnilnd_peer_addref(peer);
1484
1485         CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1486
1487         /* if purgatory release cleared it out, don't try again */
1488         if (kgnilnd_peer_active(peer)) {
1489                 /* always do this to allow kgnilnd_start_connect and
1490                  * kgnilnd_finish_connect to catch this before they
1491                  * wrap up their operations */
1492                 if (kgnilnd_can_unlink_peer_locked(peer)) {
1493                         /* already released purgatory, so only active
1494                          * conns hold it */
1495                         kgnilnd_unlink_peer_locked(peer);
1496                 } else {
1497                         kgnilnd_close_peer_conns_locked(peer, error);
1498                         /* peer unlinks itself when last conn is closed */
1499                 }
1500         }
1501
1502         /* we are done, release back to the wild */
1503         kgnilnd_peer_decref(peer);
1504 }
1505
1506 int
1507 kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
1508                           int error)
1509 {
1510         LIST_HEAD               (souls);
1511         LIST_HEAD               (zombies);
1512         struct list_head        *ptmp, *pnxt;
1513         kgn_peer_t              *peer;
1514         int                     lo;
1515         int                     hi;
1516         int                     i;
1517         int                     rc = -ENOENT;
1518
1519         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1520
1521         if (nid != LNET_NID_ANY)
1522                 lo = hi = kgnilnd_nid2peerlist(nid) - kgnilnd_data.kgn_peers;
1523         else {
1524                 lo = 0;
1525                 hi = *kgnilnd_tunables.kgn_peer_hash_size - 1;
1526                 /* wildcards always succeed */
1527                 rc = 0;
1528         }
1529
1530         for (i = lo; i <= hi; i++) {
1531                 list_for_each_safe(ptmp, pnxt, &kgnilnd_data.kgn_peers[i]) {
1532                         peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1533
1534                         LASSERTF(peer->gnp_net != NULL,
1535                                 "peer %p (%s) with NULL net\n",
1536                                  peer, libcfs_nid2str(peer->gnp_nid));
1537
1538                         if (net != NULL && peer->gnp_net != net)
1539                                 continue;
1540
1541                         if (!(nid == LNET_NID_ANY || LNET_NIDADDR(peer->gnp_nid) == LNET_NIDADDR(nid)))
1542                                 continue;
1543
1544                         /* In both cases, we want to stop any in-flight
1545                          * connect attempts */
1546                         kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1547
1548                         switch (command) {
1549                         case GNILND_DEL_CONN:
1550                                 kgnilnd_close_peer_conns_locked(peer, error);
1551                                 break;
1552                         case GNILND_DEL_PEER:
1553                                 peer->gnp_pending_unlink = 1;
1554                                 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_unlink);
1555                                 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1556                                 kgnilnd_del_peer_locked(peer, error);
1557                                 break;
1558                         case GNILND_CLEAR_PURGATORY:
1559                                 /* Mark everything ready for detach reaper will cleanup
1560                                  * once we release the kgn_peer_conn_lock
1561                                  */
1562                                 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1563                                 peer->gnp_last_errno = -EISCONN;
1564                                 /* clear reconnect so he can reconnect soon */
1565                                 peer->gnp_reconnect_time = 0;
1566                                 peer->gnp_reconnect_interval = 0;
1567                                 break;
1568                         default:
1569                                 CERROR("bad command %d\n", command);
1570                                 LBUG();
1571                         }
1572                         /* we matched something */
1573                         rc = 0;
1574                 }
1575         }
1576
1577         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1578
1579         /* nuke peer TX */
1580         kgnilnd_txlist_done(&zombies, error);
1581
1582         /* This function does not return until the commands it initiated have completed,
1583          * since they have to work there way through the other threads. In the case of shutdown
1584          * threads are not woken up until after this call is initiated so we cannot wait, we just
1585          * need to return. The same applies for stack reset we shouldnt wait as the reset thread
1586          * handles closing.
1587          */
1588
1589         CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
1590
1591         if (error == -ENOTRECOVERABLE || error == -ESHUTDOWN) {
1592                 return rc;
1593         }
1594
1595         wait_var_event_warning(&kgnilnd_data,
1596                                !atomic_read(&kgnilnd_data.kgn_npending_conns) &&
1597                                !atomic_read(&kgnilnd_data.kgn_npending_detach) &&
1598                                !atomic_read(&kgnilnd_data.kgn_npending_unlink),
1599                                "Waiting on %d peers %d closes %d detaches\n",
1600                                 atomic_read(&kgnilnd_data.kgn_npending_unlink),
1601                                 atomic_read(&kgnilnd_data.kgn_npending_conns),
1602                                 atomic_read(&kgnilnd_data.kgn_npending_detach));
1603
1604         return rc;
1605 }
1606
1607 kgn_conn_t *
1608 kgnilnd_get_conn_by_idx(int index)
1609 {
1610         kgn_peer_t        *peer;
1611         struct list_head  *ptmp;
1612         kgn_conn_t        *conn;
1613         struct list_head  *ctmp;
1614         int                i;
1615
1616
1617         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1618                 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1619                 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1620
1621                         peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1622
1623                         list_for_each(ctmp, &peer->gnp_conns) {
1624                                 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1625
1626                                 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1627                                         continue;
1628
1629                                 if (index-- > 0)
1630                                         continue;
1631
1632                                 CDEBUG(D_NET, "++conn[%p] -> %s (%d)\n", conn,
1633                                        libcfs_nid2str(conn->gnc_peer->gnp_nid),
1634                                        atomic_read(&conn->gnc_refcount));
1635                                 kgnilnd_conn_addref(conn);
1636                                 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1637                                 return conn;
1638                         }
1639                 }
1640                 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1641         }
1642
1643         return NULL;
1644 }
1645
1646 int
1647 kgnilnd_get_conn_info(kgn_peer_t *peer,
1648                       int *device_id, __u64 *peerstamp,
1649                       int *tx_seq, int *rx_seq,
1650                       int *fmaq_len, int *nfma, int *nrdma)
1651 {
1652         kgn_conn_t        *conn;
1653         int               rc = 0;
1654
1655         read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1656
1657         conn = kgnilnd_find_conn_locked(peer);
1658         if (conn == NULL) {
1659                 rc = -ENOENT;
1660                 goto out;
1661         }
1662
1663         *device_id = conn->gnc_device->gnd_host_id;
1664         *peerstamp = conn->gnc_peerstamp;
1665         *tx_seq = atomic_read(&conn->gnc_tx_seq);
1666         *rx_seq = atomic_read(&conn->gnc_rx_seq);
1667         *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
1668         *nfma = atomic_read(&conn->gnc_nlive_fma);
1669         *nrdma = atomic_read(&conn->gnc_nlive_rdma);
1670 out:
1671         read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1672         return rc;
1673 }
1674
1675 /* needs write_lock on kgn_peer_conn_lock */
1676 int
1677 kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
1678 {
1679         kgn_conn_t         *conn;
1680         struct list_head   *ctmp, *cnxt;
1681         int                 count = 0;
1682
1683         list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
1684                 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1685
1686                 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1687                         continue;
1688
1689                 count++;
1690                 /* we mark gnc_needs closing and increment kgn_npending_conns so that
1691                  * kgnilnd_del_conn_or_peer can wait on the other threads closing
1692                  * and cleaning up the connection.
1693                  */
1694                 if (!conn->gnc_needs_closing) {
1695                         conn->gnc_needs_closing = 1;
1696                         kgnilnd_admin_addref(kgnilnd_data.kgn_npending_conns);
1697                 }
1698                 kgnilnd_close_conn_locked(conn, why);
1699         }
1700         return count;
1701 }
1702
1703 int
1704 kgnilnd_report_node_state(lnet_nid_t nid, int down)
1705 {
1706         int         rc;
1707         kgn_peer_t  *peer, *new_peer;
1708         LIST_HEAD(zombies);
1709
1710         write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1711         peer = kgnilnd_find_peer_locked(nid);
1712
1713         if (peer == NULL) {
1714                 int       i;
1715                 int       found_net = 0;
1716                 kgn_net_t *net;
1717
1718                 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1719
1720                 /* Don't add a peer for node up events */
1721                 if (down == GNILND_PEER_UP)
1722                         return 0;
1723
1724                 /* find any valid net - we don't care which one... */
1725                 down_read(&kgnilnd_data.kgn_net_rw_sem);
1726                 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
1727                         list_for_each_entry(net, &kgnilnd_data.kgn_nets[i],
1728                                             gnn_list) {
1729                                 found_net = 1;
1730                                 break;
1731                         }
1732
1733                         if (found_net) {
1734                                 break;
1735                         }
1736                 }
1737                 up_read(&kgnilnd_data.kgn_net_rw_sem);
1738
1739                 if (!found_net) {
1740                         CNETERR("Could not find a net for nid %lld\n", nid);
1741                         return 1;
1742                 }
1743
1744                 /* The nid passed in does not yet contain the net portion.
1745                  * Let's build it up now
1746                  */
1747                 nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
1748                 rc = kgnilnd_add_peer(net, nid, &new_peer);
1749
1750                 if (rc) {
1751                         CNETERR("Could not add peer for nid %lld, rc %d\n",
1752                                 nid, rc);
1753                         return 1;
1754                 }
1755
1756                 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1757                 peer = kgnilnd_find_peer_locked(nid);
1758
1759                 if (peer == NULL) {
1760                         CNETERR("Could not find peer for nid %lld\n", nid);
1761                         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1762                         return 1;
1763                 }
1764         }
1765
1766         peer->gnp_state = down;
1767
1768         if (down == GNILND_PEER_DOWN) {
1769                 kgn_conn_t *conn;
1770
1771                 peer->gnp_down_event_time = jiffies;
1772                 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1773                 conn = kgnilnd_find_conn_locked(peer);
1774
1775                 if (conn != NULL) {
1776                         kgnilnd_close_conn_locked(conn, -ENETRESET);
1777                 }
1778         } else {
1779                 peer->gnp_up_event_time = jiffies;
1780         }
1781
1782         write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1783
1784         if (down == GNILND_PEER_DOWN) {
1785                 /* using ENETRESET so we don't get messages from
1786                  * kgnilnd_tx_done
1787                  */
1788                 kgnilnd_txlist_done(&zombies, -ENETRESET);
1789                 kgnilnd_peer_notify(peer, -ECONNRESET, 0);
1790                 LCONSOLE_INFO("Received down event for nid %d\n",
1791                               LNET_NIDADDR(nid));
1792         }
1793
1794         return 0;
1795 }
1796
1797 int
1798 kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1799 {
1800         struct libcfs_ioctl_data *data = arg;
1801         kgn_net_t                *net = ni->ni_data;
1802         int                       rc = -EINVAL;
1803
1804         LASSERT(ni == net->gnn_ni);
1805
1806         switch (cmd) {
1807         case IOC_LIBCFS_GET_PEER: {
1808                 lnet_nid_t   nid = 0;
1809                 kgn_peer_t  *peer = NULL;
1810                 __u32 nic_addr = 0;
1811                 __u64 peerstamp = 0;
1812                 int peer_refcount = 0, peer_connecting = 0;
1813                 int device_id = 0;
1814                 int tx_seq = 0, rx_seq = 0;
1815                 int fmaq_len = 0, nfma = 0, nrdma = 0;
1816
1817                 rc = kgnilnd_get_peer_info(data->ioc_count, &peer,
1818                                            &nid, &nic_addr, &peer_refcount,
1819                                            &peer_connecting);
1820                 if (rc)
1821                         break;
1822
1823                 /* Barf */
1824                 /* LNET_MKNID is used to mask from lnet the multiplexing/demultiplexing of connections and peers
1825                  * LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
1826                  * wants to see instead of the underlying network that is being used to send the data
1827                  */
1828                 data->ioc_nid    = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
1829                 data->ioc_flags  = peer_connecting;
1830                 data->ioc_count  = peer_refcount;
1831
1832                 rc = kgnilnd_get_conn_info(peer, &device_id, &peerstamp,
1833                                            &tx_seq, &rx_seq, &fmaq_len,
1834                                            &nfma, &nrdma);
1835
1836                 /* This is allowable - a persistent peer could not
1837                  * have a connection */
1838                 if (rc) {
1839                         /* flag to indicate we are not connected -
1840                          * need to print as such */
1841                         data->ioc_flags |= (1<<16);
1842                         rc = 0;
1843                 } else {
1844                         /* still barf */
1845                         data->ioc_net = device_id;
1846                         data->ioc_u64[0] = peerstamp;
1847                         data->ioc_u32[0] = fmaq_len;
1848                         data->ioc_u32[1] = nfma;
1849                         data->ioc_u32[2] = tx_seq;
1850                         data->ioc_u32[3] = rx_seq;
1851                         data->ioc_u32[4] = nrdma;
1852                 }
1853                 break;
1854         }
1855         case IOC_LIBCFS_ADD_PEER: {
1856                 /* just dummy value to allow using common interface */
1857                 kgn_peer_t      *peer;
1858                 rc = kgnilnd_add_peer(net, data->ioc_nid, &peer);
1859                 break;
1860         }
1861         case IOC_LIBCFS_DEL_PEER: {
1862                 /* NULL is passed in so it affects all peers in existence without regard to network
1863                  * as the peer may not exist on the network LNET believes it to be on.
1864                  */
1865                 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1866                                               GNILND_DEL_PEER, -EUCLEAN);
1867                 break;
1868         }
1869         case IOC_LIBCFS_GET_CONN: {
1870                 kgn_conn_t *conn = kgnilnd_get_conn_by_idx(data->ioc_count);
1871
1872                 if (conn == NULL)
1873                         rc = -ENOENT;
1874                 else {
1875                         rc = 0;
1876                         /* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
1877                          * the generic connection that is used to send the data
1878                          */
1879                         data->ioc_nid    = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
1880                         data->ioc_u32[0] = conn->gnc_device->gnd_id;
1881                         kgnilnd_conn_decref(conn);
1882                 }
1883                 break;
1884         }
1885         case IOC_LIBCFS_CLOSE_CONNECTION: {
1886                 /* use error = -ENETRESET to indicate it was lctl disconnect */
1887                 /* NULL is passed in so it affects all the nets as the connection is virtual
1888                  * and may not exist on the network LNET believes it to be on.
1889                  */
1890                 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1891                                               GNILND_DEL_CONN, -ENETRESET);
1892                 break;
1893         }
1894         case IOC_LIBCFS_PUSH_CONNECTION: {
1895                 /* we use this to flush purgatory */
1896                 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1897                                               GNILND_CLEAR_PURGATORY, -EUCLEAN);
1898                 break;
1899         }
1900         case IOC_LIBCFS_REGISTER_MYNID: {
1901                 /* Ignore if this is a noop */
1902                 if (data->ioc_nid == ni->ni_nid) {
1903                         rc = 0;
1904                 } else {
1905                         CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1906                                libcfs_nid2str(data->ioc_nid),
1907                                libcfs_nid2str(ni->ni_nid));
1908                         rc = -EINVAL;
1909                 }
1910                 break;
1911         }
1912         }
1913
1914         return rc;
1915 }
1916
1917 int
1918 kgnilnd_dev_init(kgn_device_t *dev)
1919 {
1920         gni_return_t      rrc;
1921         int               rc = 0;
1922         unsigned int      cq_size;
1923         ENTRY;
1924
1925         /* size of these CQs should be able to accommodate the outgoing
1926          * RDMA and SMSG transactions.  Since we really don't know what we
1927          * really need here, we'll take credits * 2 * 3 to allow a bunch.
1928          * We need to dig into this more with the performance work. */
1929         cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
1930
1931         rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
1932                                  *kgnilnd_tunables.kgn_pkey, 0,
1933                                  &dev->gnd_domain);
1934         if (rrc != GNI_RC_SUCCESS) {
1935                 CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
1936                 GOTO(failed, rc = -ENODEV);
1937         }
1938
1939         rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
1940                                  &dev->gnd_host_id, &dev->gnd_handle);
1941         if (rrc != GNI_RC_SUCCESS) {
1942                 CERROR("Can't attach CDM to device %d (%d)\n",
1943                         dev->gnd_id, rrc);
1944                 GOTO(failed, rc = -ENODEV);
1945         }
1946
1947         /* a bit gross, but not much we can do - Aries Sim doesn't have
1948          * hardcoded NIC/NID that we can use */
1949         rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
1950         if (rc != 0)
1951                 GOTO(failed, rc = -ENODEV);
1952
1953         /* only dev 0 gets the errors - no need to reset the stack twice
1954          * - this works because we have a single PTAG, if we had more
1955          * then we'd need to have multiple handlers */
1956         if (dev->gnd_id == 0) {
1957                 rrc = kgnilnd_subscribe_errors(dev->gnd_handle,
1958                                                 GNI_ERRMASK_CRITICAL |
1959                                                 GNI_ERRMASK_UNKNOWN_TRANSACTION,
1960                                               0, NULL, kgnilnd_critical_error,
1961                                               &dev->gnd_err_handle);
1962                 if (rrc != GNI_RC_SUCCESS) {
1963                         CERROR("Can't subscribe for errors on device %d: rc %d\n",
1964                                 dev->gnd_id, rrc);
1965                         GOTO(failed, rc = -ENODEV);
1966                 }
1967
1968                 rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
1969                                                   kgnilnd_quiesce_end_callback);
1970                 if (rc != GNI_RC_SUCCESS) {
1971                         CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
1972                                 dev->gnd_id, rrc);
1973                         GOTO(failed, rc = -ENODEV);
1974                 }
1975         }
1976
1977         rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
1978         if (rc < 0) {
1979                 /* log messages during startup */
1980                 if (kgnilnd_data.kgn_init < GNILND_INIT_ALL) {
1981                         CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
1982                                 dev->gnd_host_id, rc);
1983                 }
1984                 GOTO(failed, rc = -ESRCH);
1985         }
1986         CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
1987
1988         rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
1989                                 0, kgnilnd_device_callback,
1990                                 dev->gnd_id, &dev->gnd_snd_rdma_cqh);
1991         if (rrc != GNI_RC_SUCCESS) {
1992                 CERROR("Can't create rdma send cq size %u for device "
1993                        "%d (%d)\n", cq_size, dev->gnd_id, rrc);
1994                 GOTO(failed, rc = -EINVAL);
1995         }
1996
1997         rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
1998                         0, kgnilnd_device_callback, dev->gnd_id,
1999                         &dev->gnd_snd_fma_cqh);
2000         if (rrc != GNI_RC_SUCCESS) {
2001                 CERROR("Can't create fma send cq size %u for device %d (%d)\n",
2002                        cq_size, dev->gnd_id, rrc);
2003                 GOTO(failed, rc = -EINVAL);
2004         }
2005
2006         /* This one we size differently - overflows are possible and it needs to be
2007          * sized based on machine size */
2008         rrc = kgnilnd_cq_create(dev->gnd_handle,
2009                         *kgnilnd_tunables.kgn_fma_cq_size,
2010                         0, kgnilnd_device_callback, dev->gnd_id,
2011                         &dev->gnd_rcv_fma_cqh);
2012         if (rrc != GNI_RC_SUCCESS) {
2013                 CERROR("Can't create fma cq size %d for device %d (%d)\n",
2014                        *kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
2015                 GOTO(failed, rc = -EINVAL);
2016         }
2017
2018         rrc = kgnilnd_register_smdd_buf(dev);
2019         if (rrc != GNI_RC_SUCCESS) {
2020                 GOTO(failed, rc = -EINVAL);
2021         }
2022
2023         RETURN(0);
2024
2025 failed:
2026         kgnilnd_dev_fini(dev);
2027         RETURN(rc);
2028 }
2029
2030 void
2031 kgnilnd_dev_fini(kgn_device_t *dev)
2032 {
2033         gni_return_t rrc;
2034         ENTRY;
2035
2036         /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
2037         LASSERTF(list_empty(&dev->gnd_ready_conns) &&
2038                  list_empty(&dev->gnd_map_tx) &&
2039                  list_empty(&dev->gnd_rdmaq) &&
2040                  list_empty(&dev->gnd_delay_conns),
2041                  "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p"
2042                  "map_tx %d@0x%p rdmaq %d@0x%p\n",
2043                  dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
2044                  kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns,
2045                  kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
2046                  kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
2047
2048         /* These should follow from tearing down all connections */
2049         LASSERTF(dev->gnd_map_nphys == 0 && dev->gnd_map_physnop == 0,
2050                 "%d physical mappings of %d pages still mapped\n",
2051                  dev->gnd_map_nphys, dev->gnd_map_physnop);
2052
2053         LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
2054                  atomic_read(&dev->gnd_n_mdd_held) == 0 &&
2055                  atomic64_read(&dev->gnd_nbytes_map) == 0,
2056                  "%d SMSG mappings of %lld bytes still mapped or held %d\n",
2057                  atomic_read(&dev->gnd_n_mdd),
2058                  (u64)atomic64_read(&dev->gnd_nbytes_map),
2059                  atomic_read(&dev->gnd_n_mdd_held));
2060
2061         LASSERT(list_empty(&dev->gnd_map_list));
2062
2063         /* What other assertions needed to ensure all connections torn down ? */
2064
2065         /* check all counters == 0 (EP, MDD, etc) */
2066
2067         /* if we are resetting due to quiese (stack reset), don't check
2068          * thread states */
2069         LASSERTF(kgnilnd_data.kgn_quiesce_trigger ||
2070                 atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
2071                 "tried to shutdown with threads active\n");
2072
2073         if (dev->gnd_smdd_hold_buf) {
2074                 rrc = kgnilnd_deregister_smdd_buf(dev);
2075                 LASSERTF(rrc == GNI_RC_SUCCESS,
2076                         "bad rc from deregistion of sMDD buffer: %d\n", rrc);
2077                 dev->gnd_smdd_hold_buf = NULL;
2078         }
2079
2080         if (dev->gnd_rcv_fma_cqh) {
2081                 rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
2082                 LASSERTF(rrc == GNI_RC_SUCCESS,
2083                         "bad rc from gni_cq_destroy on rcv_fma_cqh: %d\n", rrc);
2084                 dev->gnd_rcv_fma_cqh = NULL;
2085         }
2086
2087         if (dev->gnd_snd_rdma_cqh) {
2088                 rrc = kgnilnd_cq_destroy(dev->gnd_snd_rdma_cqh);
2089                 LASSERTF(rrc == GNI_RC_SUCCESS,
2090                         "bad rc from gni_cq_destroy on send_rdma_cqh: %d\n", rrc);
2091                 dev->gnd_snd_rdma_cqh = NULL;
2092         }
2093
2094         if (dev->gnd_snd_fma_cqh) {
2095                 rrc = kgnilnd_cq_destroy(dev->gnd_snd_fma_cqh);
2096                 LASSERTF(rrc == GNI_RC_SUCCESS,
2097                         "bad rc from gni_cq_destroy on snd_fma_cqh: %d\n", rrc);
2098                 dev->gnd_snd_fma_cqh = NULL;
2099         }
2100
2101         if (dev->gnd_err_handle) {
2102                 rrc = kgnilnd_release_errors(dev->gnd_err_handle);
2103                 LASSERTF(rrc == GNI_RC_SUCCESS,
2104                         "bad rc from gni_release_errors: %d\n", rrc);
2105                 dev->gnd_err_handle = NULL;
2106         }
2107
2108         if (dev->gnd_domain) {
2109                 rrc = kgnilnd_cdm_destroy(dev->gnd_domain);
2110                 LASSERTF(rrc == GNI_RC_SUCCESS,
2111                         "bad rc from gni_cdm_destroy: %d\n", rrc);
2112                 dev->gnd_domain = NULL;
2113         }
2114
2115         EXIT;
2116 }
2117
2118 int kgnilnd_base_startup(void)
2119 {
2120         struct timespec64    ts;
2121         long long            pkmem = libcfs_kmem_read();
2122         int                  rc;
2123         int                  i;
2124         kgn_device_t        *dev;
2125         struct task_struct  *thrd;
2126
2127 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
2128         /* limit how much memory can be allocated for fma blocks in
2129          * instances where many nodes need to reconnects at the same time */
2130         struct sysinfo si;
2131         si_meminfo(&si);
2132         kgnilnd_data.free_pages_limit = si.totalram/4;
2133 #endif
2134
2135         ENTRY;
2136
2137         LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
2138                 "init %d\n", kgnilnd_data.kgn_init);
2139
2140         /* zero pointers, flags etc */
2141         memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
2142         kgnilnd_check_kgni_version();
2143
2144         /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
2145          * a unique (for all time) connstamp so we can uniquely identify
2146          * the sender.  The connstamp is an incrementing counter
2147          * initialised with seconds + microseconds at startup time.  So we
2148          * rely on NOT creating connections more frequently on average than
2149          * 1MHz to ensure we don't use old connstamps when we reboot. */
2150         ktime_get_ts64(&ts);
2151         kgnilnd_data.kgn_connstamp =
2152                  kgnilnd_data.kgn_peerstamp =
2153                         (ts.tv_sec * 1000000) + (ts.tv_nsec / 100);
2154
2155         init_rwsem(&kgnilnd_data.kgn_net_rw_sem);
2156
2157         for (i = 0; i < GNILND_MAXDEVS; i++) {
2158                 kgn_device_t  *dev = &kgnilnd_data.kgn_devices[i];
2159
2160                 dev->gnd_id = i;
2161                 INIT_LIST_HEAD(&dev->gnd_ready_conns);
2162                 INIT_LIST_HEAD(&dev->gnd_delay_conns);
2163                 INIT_LIST_HEAD(&dev->gnd_map_tx);
2164                 INIT_LIST_HEAD(&dev->gnd_fma_buffs);
2165                 mutex_init(&dev->gnd_cq_mutex);
2166                 mutex_init(&dev->gnd_fmablk_mutex);
2167                 spin_lock_init(&dev->gnd_fmablk_lock);
2168                 init_waitqueue_head(&dev->gnd_waitq);
2169                 init_waitqueue_head(&dev->gnd_dgram_waitq);
2170                 init_waitqueue_head(&dev->gnd_dgping_waitq);
2171                 spin_lock_init(&dev->gnd_lock);
2172                 INIT_LIST_HEAD(&dev->gnd_map_list);
2173                 spin_lock_init(&dev->gnd_map_lock);
2174                 atomic_set(&dev->gnd_nfmablk, 0);
2175                 atomic_set(&dev->gnd_fmablk_vers, 1);
2176                 atomic_set(&dev->gnd_neps, 0);
2177                 atomic_set(&dev->gnd_canceled_dgrams, 0);
2178                 INIT_LIST_HEAD(&dev->gnd_connd_peers);
2179                 spin_lock_init(&dev->gnd_connd_lock);
2180                 spin_lock_init(&dev->gnd_dgram_lock);
2181                 spin_lock_init(&dev->gnd_rdmaq_lock);
2182                 INIT_LIST_HEAD(&dev->gnd_rdmaq);
2183                 init_rwsem(&dev->gnd_conn_sem);
2184
2185                 /* alloc & setup nid based dgram table */
2186                 CFS_ALLOC_PTR_ARRAY(dev->gnd_dgrams,
2187                                     *kgnilnd_tunables.kgn_peer_hash_size);
2188
2189                 if (dev->gnd_dgrams == NULL)
2190                         GOTO(failed, rc = -ENOMEM);
2191
2192                 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2193                         INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
2194                 }
2195                 atomic_set(&dev->gnd_ndgrams, 0);
2196                 atomic_set(&dev->gnd_nwcdgrams, 0);
2197                 /* setup timer for RDMAQ processing */
2198                 cfs_timer_setup(&dev->gnd_rdmaq_timer,
2199                                 kgnilnd_schedule_device_timer,
2200                                 (unsigned long)dev, 0);
2201
2202                 /* setup timer for mapping processing */
2203                 cfs_timer_setup(&dev->gnd_map_timer,
2204                                 kgnilnd_schedule_device_timer,
2205                                 (unsigned long)dev, 0);
2206
2207         }
2208
2209         /* CQID 0 isn't allowed, set to MAX_MSG_ID - 1 to check for conflicts early */
2210         kgnilnd_data.kgn_next_cqid = GNILND_MAX_MSG_ID - 1;
2211         kgnilnd_data.kgn_new_min_timeout = *kgnilnd_tunables.kgn_timeout;
2212         init_waitqueue_head(&kgnilnd_data.kgn_reaper_waitq);
2213         init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
2214         spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
2215
2216         mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
2217         atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
2218         atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
2219         atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
2220         atomic_set(&kgnilnd_data.kgn_npending_detach, 0);
2221         atomic_set(&kgnilnd_data.kgn_rev_offset, 0);
2222         atomic_set(&kgnilnd_data.kgn_rev_length, 0);
2223         atomic_set(&kgnilnd_data.kgn_rev_copy_buff, 0);
2224
2225         /* OK to call kgnilnd_api_shutdown() to cleanup now */
2226         kgnilnd_data.kgn_init = GNILND_INIT_DATA;
2227         if (!try_module_get(THIS_MODULE))
2228                 GOTO(failed, rc = -ENOENT);
2229
2230         rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
2231
2232         CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_peers,
2233                             *kgnilnd_tunables.kgn_peer_hash_size);
2234
2235         if (kgnilnd_data.kgn_peers == NULL)
2236                 GOTO(failed, rc = -ENOMEM);
2237
2238         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2239                 INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
2240         }
2241
2242         CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_conns,
2243                             *kgnilnd_tunables.kgn_peer_hash_size);
2244
2245         if (kgnilnd_data.kgn_conns == NULL)
2246                 GOTO(failed, rc = -ENOMEM);
2247
2248         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2249                 INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
2250         }
2251
2252         CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_nets,
2253                             *kgnilnd_tunables.kgn_net_hash_size);
2254
2255         if (kgnilnd_data.kgn_nets == NULL)
2256                 GOTO(failed, rc = -ENOMEM);
2257
2258         for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2259                 INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
2260         }
2261
2262         kgnilnd_data.kgn_mbox_cache =
2263                 kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0,
2264                                   SLAB_HWCACHE_ALIGN, NULL);
2265         if (kgnilnd_data.kgn_mbox_cache == NULL) {
2266                 CERROR("Can't create slab for physical mbox blocks\n");
2267                 GOTO(failed, rc = -ENOMEM);
2268         }
2269
2270         kgnilnd_data.kgn_rx_cache =
2271                 kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
2272         if (kgnilnd_data.kgn_rx_cache == NULL) {
2273                 CERROR("Can't create slab for kgn_rx_t descriptors\n");
2274                 GOTO(failed, rc = -ENOMEM);
2275         }
2276
2277         kgnilnd_data.kgn_tx_cache =
2278                 kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
2279         if (kgnilnd_data.kgn_tx_cache == NULL) {
2280                 CERROR("Can't create slab for kgn_tx_t\n");
2281                 GOTO(failed, rc = -ENOMEM);
2282         }
2283
2284         kgnilnd_data.kgn_tx_phys_cache =
2285                 kmem_cache_create("kgn_tx_phys",
2286                                    LNET_MAX_IOV * sizeof(gni_mem_segment_t),
2287                                    0, 0, NULL);
2288         if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
2289                 CERROR("Can't create slab for kgn_tx_phys\n");
2290                 GOTO(failed, rc = -ENOMEM);
2291         }
2292
2293         kgnilnd_data.kgn_dgram_cache =
2294                 kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
2295         if (kgnilnd_data.kgn_dgram_cache == NULL) {
2296                 CERROR("Can't create slab for outgoing datagrams\n");
2297                 GOTO(failed, rc = -ENOMEM);
2298         }
2299
2300         /* allocate a MAX_IOV array of page pointers for each cpu */
2301         kgnilnd_data.kgn_cksum_map_pages = kmalloc(num_possible_cpus() * sizeof (struct page *),
2302                                                    GFP_KERNEL);
2303         if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
2304                 CERROR("Can't allocate vmap cksum pages\n");
2305                 GOTO(failed, rc = -ENOMEM);
2306         }
2307         kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
2308         memset(kgnilnd_data.kgn_cksum_map_pages, 0,
2309                 kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));
2310
2311         for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2312                 kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
2313                                                               GFP_KERNEL);
2314                 if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
2315                         CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
2316                         GOTO(failed, rc = -ENOMEM);
2317                 }
2318         }
2319
2320         LASSERT(kgnilnd_data.kgn_ndevs == 0);
2321
2322         /* Use all available GNI devices */
2323         for (i = 0; i < GNILND_MAXDEVS; i++) {
2324                 dev = &kgnilnd_data.kgn_devices[kgnilnd_data.kgn_ndevs];
2325
2326                 rc = kgnilnd_dev_init(dev);
2327                 if (rc == 0) {
2328                         /* Increment here so base_shutdown cleans it up */
2329                         kgnilnd_data.kgn_ndevs++;
2330
2331                         rc = kgnilnd_allocate_phys_fmablk(dev);
2332                         if (rc)
2333                                 GOTO(failed, rc);
2334                 }
2335         }
2336
2337         if (kgnilnd_data.kgn_ndevs == 0) {
2338                 CERROR("Can't initialise any GNI devices\n");
2339                 GOTO(failed, rc = -ENODEV);
2340         }
2341
2342         rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
2343         if (rc != 0) {
2344                 CERROR("Can't spawn gnilnd reaper: %d\n", rc);
2345                 GOTO(failed, rc);
2346         }
2347
2348         rc = kgnilnd_start_rca_thread();
2349         if (rc != 0) {
2350                 CERROR("Can't spawn gnilnd rca: %d\n", rc);
2351                 GOTO(failed, rc);
2352         }
2353
2354         /*
2355          * Start ruhroh thread.  We can't use kgnilnd_thread_start() because
2356          * we don't want this thread included in kgnilnd_data.kgn_nthreads
2357          * count.  This thread controls quiesce, so it mustn't
2358          * quiesce itself.
2359          */
2360         thrd = kthread_run(kgnilnd_ruhroh_thread, NULL, "%s_%02d", "kgnilnd_rr", 0);
2361         if (IS_ERR(thrd)) {
2362                 rc = PTR_ERR(thrd);
2363                 CERROR("Can't spawn gnilnd ruhroh thread: %d\n", rc);
2364                 GOTO(failed, rc);
2365         }
2366
2367         /* threads will load balance across devs as they are available */
2368         if (*kgnilnd_tunables.kgn_thread_affinity) {
2369                 rc = kgnilnd_start_sd_threads();
2370                 if (rc != 0)
2371                         GOTO(failed, rc);
2372         } else {
2373                 for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
2374                         rc = kgnilnd_thread_start(kgnilnd_scheduler,
2375                                                   (void *)((long)i),
2376                                                   "kgnilnd_sd", i);
2377                         if (rc != 0) {
2378                                 CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
2379                                        i, rc);
2380                                 GOTO(failed, rc);
2381                         }
2382                 }
2383         }
2384
2385         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2386                 dev = &kgnilnd_data.kgn_devices[i];
2387                 rc = kgnilnd_thread_start(kgnilnd_dgram_mover, dev,
2388                                           "kgnilnd_dg", dev->gnd_id);
2389                 if (rc != 0) {
2390                         CERROR("Can't spawn gnilnd dgram_mover[%d]: %d\n",
2391                                dev->gnd_id, rc);
2392                         GOTO(failed, rc);
2393                 }
2394
2395                 rc = kgnilnd_thread_start(kgnilnd_dgram_waitq, dev,
2396                                           "kgnilnd_dgn", dev->gnd_id);
2397                 if (rc != 0) {
2398                         CERROR("Can't spawn gnilnd dgram_waitq[%d]: %d\n",
2399                                 dev->gnd_id, rc);
2400                         GOTO(failed, rc);
2401                 }
2402
2403                 rc = kgnilnd_setup_wildcard_dgram(dev);
2404
2405                 if (rc != 0) {
2406                         CERROR("Can't create wildcard dgrams[%d]: %d\n",
2407                                 dev->gnd_id, rc);
2408                         GOTO(failed, rc);
2409                 }
2410         }
2411
2412         /* flag everything initialised */
2413         kgnilnd_data.kgn_init = GNILND_INIT_ALL;
2414         /*****************************************************/
2415
2416         CDEBUG(D_MALLOC, "initial kmem %lld\n", pkmem);
2417         RETURN(0);
2418
2419 failed:
2420         kgnilnd_base_shutdown();
2421         kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2422         RETURN(rc);
2423 }
2424
2425 void
2426 kgnilnd_base_shutdown(void)
2427 {
2428         int                     i, j;
2429         ENTRY;
2430
2431         while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
2432
2433         kgnilnd_data.kgn_wc_kill = 1;
2434
2435         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2436                 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2437                 kgnilnd_cancel_wc_dgrams(dev);
2438                 kgnilnd_cancel_dgrams(dev);
2439                 kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2440                 kgnilnd_wait_for_canceled_dgrams(dev);
2441         }
2442
2443         /* We need to verify there are no conns left before we let the threads
2444          * shut down otherwise we could clean up the peers but still have
2445          * some outstanding conns due to orphaned datagram conns that are
2446          * being cleaned up.
2447          */
2448         i = 2;
2449         while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
2450                 i++;
2451
2452                 for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
2453                         kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
2454                         kgnilnd_schedule_device(dev);
2455                 }
2456
2457                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2458                         "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
2459                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2460         }
2461         /* Peer state all cleaned up BEFORE setting shutdown, so threads don't
2462          * have to worry about shutdown races.  NB connections may be created
2463          * while there are still active connds, but these will be temporary
2464          * since peer creation always fails after the listener has started to
2465          * shut down.
2466          * all peers should have been cleared out on the nets */
2467         LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2468                 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2469
2470         /* Wait for the ruhroh thread to shut down. */
2471         kgnilnd_data.kgn_ruhroh_shutdown = 1;
2472         wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
2473         i = 2;
2474         while (kgnilnd_data.kgn_ruhroh_running != 0) {
2475                 i++;
2476                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2477                        "Waiting for ruhroh thread to terminate\n");
2478                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2479         }
2480
2481        /* Flag threads to terminate */
2482         kgnilnd_data.kgn_shutdown = 1;
2483
2484         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2485                 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2486
2487                 /* should clear all the MDDs */
2488                 kgnilnd_unmap_fma_blocks(dev);
2489
2490                 kgnilnd_schedule_device(dev);
2491                 wake_up(&dev->gnd_dgram_waitq);
2492                 wake_up(&dev->gnd_dgping_waitq);
2493                 LASSERT(list_empty(&dev->gnd_connd_peers));
2494         }
2495
2496         spin_lock(&kgnilnd_data.kgn_reaper_lock);
2497         wake_up(&kgnilnd_data.kgn_reaper_waitq);
2498         spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2499
2500         if (atomic_read(&kgnilnd_data.kgn_nthreads))
2501                 kgnilnd_wakeup_rca_thread();
2502
2503         /* Wait for threads to exit */
2504         i = 2;
2505         while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
2506                 i++;
2507                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2508                        "Waiting for %d threads to terminate\n",
2509                        atomic_read(&kgnilnd_data.kgn_nthreads));
2510                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2511         }
2512
2513         LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2514                 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2515
2516         if (kgnilnd_data.kgn_peers != NULL) {
2517                 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2518                         LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));
2519
2520                 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_peers,
2521                                    *kgnilnd_tunables.kgn_peer_hash_size);
2522         }
2523
2524         down_write(&kgnilnd_data.kgn_net_rw_sem);
2525         if (kgnilnd_data.kgn_nets != NULL) {
2526                 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
2527                         LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));
2528
2529                 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_nets,
2530                                    *kgnilnd_tunables.kgn_net_hash_size);
2531         }
2532         up_write(&kgnilnd_data.kgn_net_rw_sem);
2533
2534         LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
2535                 "conns left %d\n", atomic_read(&kgnilnd_data.kgn_nconns));
2536
2537         if (kgnilnd_data.kgn_conns != NULL) {
2538                 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2539                         LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));
2540
2541                 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_conns,
2542                                    *kgnilnd_tunables.kgn_peer_hash_size);
2543         }
2544
2545         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2546                 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2547                 kgnilnd_dev_fini(dev);
2548
2549                 LASSERTF(atomic_read(&dev->gnd_ndgrams) == 0,
2550                         "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));
2551
2552                 if (dev->gnd_dgrams != NULL) {
2553                         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size;
2554                              i++)
2555                                 LASSERT(list_empty(&dev->gnd_dgrams[i]));
2556
2557                         CFS_FREE_PTR_ARRAY(dev->gnd_dgrams,
2558                                            *kgnilnd_tunables.kgn_peer_hash_size);
2559                 }
2560
2561                 kgnilnd_free_phys_fmablk(dev);
2562         }
2563
2564         if (kgnilnd_data.kgn_mbox_cache != NULL)
2565                 kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
2566
2567         if (kgnilnd_data.kgn_rx_cache != NULL)
2568                 kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
2569
2570         if (kgnilnd_data.kgn_tx_cache != NULL)
2571                 kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
2572
2573         if (kgnilnd_data.kgn_tx_phys_cache != NULL)
2574                 kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
2575
2576         if (kgnilnd_data.kgn_dgram_cache != NULL)
2577                 kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
2578
2579         if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
2580                 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2581                         if (kgnilnd_data.kgn_cksum_map_pages[i] != NULL) {
2582                                 kfree(kgnilnd_data.kgn_cksum_map_pages[i]);
2583                         }
2584                 }
2585                 kfree(kgnilnd_data.kgn_cksum_map_pages);
2586         }
2587
2588         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2589                libcfs_kmem_read());
2590
2591         kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2592         module_put(THIS_MODULE);
2593
2594         EXIT;
2595 }
2596
2597 int
2598 kgnilnd_startup(struct lnet_ni *ni)
2599 {
2600         int               rc, devno;
2601         kgn_net_t        *net;
2602         ENTRY;
2603
2604         LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd,
2605                 "bad LND 0x%p != the_kgnilnd @ 0x%p\n",
2606                 ni->ni_net->net_lnd, &the_kgnilnd);
2607
2608         if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
2609                 rc = kgnilnd_base_startup();
2610                 if (rc != 0)
2611                         RETURN(rc);
2612         }
2613
2614         /* Serialize with shutdown. */
2615         mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2616
2617         LIBCFS_ALLOC(net, sizeof(*net));
2618         if (net == NULL) {
2619                 CERROR("could not allocate net for new interface instance\n");
2620                 /* no need to cleanup the CDM... */
2621                 GOTO(failed, rc = -ENOMEM);
2622         }
2623         INIT_LIST_HEAD(&net->gnn_list);
2624         ni->ni_data = net;
2625         net->gnn_ni = ni;
2626         if (!ni->ni_net->net_tunables_set) {
2627                 ni->ni_net->net_tunables.lct_max_tx_credits =
2628                         *kgnilnd_tunables.kgn_credits;
2629                 ni->ni_net->net_tunables.lct_peer_tx_credits =
2630                         *kgnilnd_tunables.kgn_peer_credits;
2631         }
2632
2633         if (*kgnilnd_tunables.kgn_peer_health) {
2634                 int     fudge;
2635                 int     timeout;
2636                 /* give this a bit of leeway - we don't have a hard timeout
2637                  * as we only check timeouts periodically - see comment in kgnilnd_reaper */
2638                 fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
2639                 timeout = *kgnilnd_tunables.kgn_timeout + fudge;
2640
2641                 if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
2642                         ni->ni_net->net_tunables.lct_peer_timeout =
2643                                  *kgnilnd_tunables.kgn_peer_timeout;
2644                 } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
2645                         LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
2646                                         *kgnilnd_tunables.kgn_peer_timeout,
2647                                         timeout);
2648                         ni->ni_data = NULL;
2649                         LIBCFS_FREE(net, sizeof(*net));
2650                         GOTO(failed, rc = -EINVAL);
2651                 } else
2652                         ni->ni_net->net_tunables.lct_peer_timeout = timeout;
2653
2654                 LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
2655                               ni->ni_net->net_tunables.lct_peer_timeout);
2656         }
2657
2658         atomic_set(&net->gnn_refcount, 1);
2659
2660         /* if we have multiple devices, spread the nets around */
2661         net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));
2662
2663         devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
2664         net->gnn_dev = &kgnilnd_data.kgn_devices[devno];
2665
2666         /* allocate a 'dummy' cdm for datagram use. We can only have a single
2667          * datagram between a nid:inst_id and nid2:inst_id. The fake cdm
2668          * give us additional inst_id to use, allowing the datagrams to flow
2669          * like rivers of honey and beer */
2670
2671         /* the instance id for the cdm is the NETNUM offset by MAXDEVS -
2672          * ensuring we'll have a unique id */
2673
2674
2675         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
2676         CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
2677                 net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
2678         /* until the gnn_list is set, we need to cleanup ourselves as
2679          * kgnilnd_shutdown is just gonna get confused */
2680
2681         down_write(&kgnilnd_data.kgn_net_rw_sem);
2682         list_add_tail(&net->gnn_list, kgnilnd_netnum2netlist(net->gnn_netnum));
2683         up_write(&kgnilnd_data.kgn_net_rw_sem);
2684
2685         /* we need a separate thread to call probe_wait_by_id until
2686          * we get a function callback notifier from kgni */
2687         mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2688         RETURN(0);
2689  failed:
2690         mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2691         kgnilnd_shutdown(ni);
2692         RETURN(rc);
2693 }
2694
2695 void
2696 kgnilnd_shutdown(struct lnet_ni *ni)
2697 {
2698         kgn_net_t     *net = ni->ni_data;
2699         int           i;
2700         int           rc;
2701         ENTRY;
2702
2703         CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
2704
2705         LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_ALL,
2706                 "init %d\n", kgnilnd_data.kgn_init);
2707
2708         /* Serialize with startup. */
2709         mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2710         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2711                libcfs_kmem_read());
2712
2713         if (net == NULL) {
2714                 CERROR("got NULL net for ni %p\n", ni);
2715                 GOTO(out, rc = -EINVAL);
2716         }
2717
2718         LASSERTF(ni == net->gnn_ni,
2719                 "ni %p gnn_ni %p\n", net, net->gnn_ni);
2720
2721         ni->ni_data = NULL;
2722
2723         LASSERT(!net->gnn_shutdown);
2724         LASSERTF(atomic_read(&net->gnn_refcount) != 0,
2725                 "net %p refcount %d\n",
2726                  net, atomic_read(&net->gnn_refcount));
2727
2728         if (!list_empty(&net->gnn_list)) {
2729                 /* serialize with peer creation */
2730                 down_write(&kgnilnd_data.kgn_net_rw_sem);
2731                 net->gnn_shutdown = 1;
2732                 up_write(&kgnilnd_data.kgn_net_rw_sem);
2733
2734                 kgnilnd_cancel_net_dgrams(net);
2735
2736                 kgnilnd_del_conn_or_peer(net, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2737
2738                 /* if we are quiesced, need to wake up - we need those threads
2739                  * alive to release peers, etc */
2740                 if (GNILND_IS_QUIESCED) {
2741                         set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
2742                         kgnilnd_quiesce_wait("shutdown");
2743                 }
2744
2745                 kgnilnd_wait_for_canceled_dgrams(net->gnn_dev);
2746
2747                 /* We wait until the nets ref's are 1, we will release final ref which is ours
2748                  * this allows us to make sure everything else is done before we free the
2749                  * net.
2750                  */
2751                 i = 4;
2752                 while (atomic_read(&net->gnn_refcount) != 1) {
2753                         i++;
2754                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2755                                 "Waiting for %d references to clear on net %d\n",
2756                                 atomic_read(&net->gnn_refcount),
2757                                 net->gnn_netnum);
2758                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2759                 }
2760
2761                 /* release ref from kgnilnd_startup */
2762                 kgnilnd_net_decref(net);
2763                 /* serialize with reaper and conn_task looping */
2764                 down_write(&kgnilnd_data.kgn_net_rw_sem);
2765                 list_del_init(&net->gnn_list);
2766                 up_write(&kgnilnd_data.kgn_net_rw_sem);
2767
2768         }
2769
2770         /* not locking, this can't race with writers */
2771         LASSERTF(atomic_read(&net->gnn_refcount) == 0,
2772                 "net %p refcount %d\n",
2773                  net, atomic_read(&net->gnn_refcount));
2774         LIBCFS_FREE(net, sizeof(*net));
2775
2776 out:
2777         down_read(&kgnilnd_data.kgn_net_rw_sem);
2778         for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2779                 if (!list_empty(&kgnilnd_data.kgn_nets[i])) {
2780                         up_read(&kgnilnd_data.kgn_net_rw_sem);
2781                         break;
2782                 }
2783
2784                 if (i == *kgnilnd_tunables.kgn_net_hash_size - 1) {
2785                         up_read(&kgnilnd_data.kgn_net_rw_sem);
2786                         kgnilnd_base_shutdown();
2787                 }
2788         }
2789         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2790                libcfs_kmem_read());
2791
2792         mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2793         EXIT;
2794 }
2795
2796 static void __exit kgnilnd_exit(void)
2797 {
2798         lnet_unregister_lnd(&the_kgnilnd);
2799         kgnilnd_proc_fini();
2800         kgnilnd_remove_sysctl();
2801 }
2802
2803 static int __init kgnilnd_init(void)
2804 {
2805         int    rc;
2806
2807         rc = kgnilnd_tunables_init();
2808         if (rc != 0)
2809                 return rc;
2810
2811         LCONSOLE_INFO("Lustre: kgnilnd build version: "LUSTRE_VERSION_STRING"\n");
2812
2813         kgnilnd_insert_sysctl();
2814         kgnilnd_proc_init();
2815
2816         lnet_register_lnd(&the_kgnilnd);
2817
2818         return 0;
2819 }
2820
2821 MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
2822 MODULE_DESCRIPTION("Gemini LNet Network Driver");
2823 MODULE_VERSION(LUSTRE_VERSION_STRING);
2824 MODULE_LICENSE("GPL");
2825
2826 module_init(kgnilnd_init);
2827 module_exit(kgnilnd_exit);