[fs/lustre-release.git] / lnet / klnds / gnilnd / gnilnd.c

/*
 * Copyright (C) 2012 Cray, Inc.
 *
 * Copyright (c) 2013, 2014, Intel Corporation.
 *
 *   Author: Nic Henke <nic@cray.com>
 *   Author: James Shimek <jshimek@cray.com>
 *
 *   This file is part of Lustre, http://www.lustre.org.
 *
 *   Lustre is free software; you can redistribute it and/or
 *   modify it under the terms of version 2 of the GNU General Public
 *   License as published by the Free Software Foundation.
 *
 *   Lustre is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *   GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with Lustre; if not, write to the Free Software
 *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
 *
 */
#include "gnilnd.h"

/* Primary entry points from LNET.  There are no guarantees against reentrance. */
lnd_t the_kgnilnd = {
#ifdef CONFIG_CRAY_XT
        .lnd_type       = GNILND,
#else
        .lnd_type       = GNIIPLND,
#endif
        .lnd_startup    = kgnilnd_startup,
        .lnd_shutdown   = kgnilnd_shutdown,
        .lnd_ctl        = kgnilnd_ctl,
        .lnd_send       = kgnilnd_send,
        .lnd_recv       = kgnilnd_recv,
        .lnd_eager_recv = kgnilnd_eager_recv,
        .lnd_query      = kgnilnd_query,
};

kgn_data_t      kgnilnd_data;

/* needs write_lock on kgn_peer_conn_lock */
int
kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
{
        kgn_conn_t         *conn;
        struct list_head   *ctmp, *cnxt;
        int                 loopback;
        int                 count = 0;

        loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;

        list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
                conn = list_entry(ctmp, kgn_conn_t, gnc_list);

                if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
                        continue;

                if (conn == newconn)
                        continue;

                if (conn->gnc_device != newconn->gnc_device)
                        continue;

                /* This is a two connection loopback - one talking to the other */
                if (loopback &&
                    newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
                    newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
                        CDEBUG(D_NET, "skipping prune of %p, "
                                "loopback and matching stamps"
                                " connstamp "LPU64"("LPU64")"
                                " peerstamp "LPU64"("LPU64")\n",
                                conn, newconn->gnc_my_connstamp,
                                conn->gnc_peer_connstamp,
                                newconn->gnc_peer_connstamp,
                                conn->gnc_my_connstamp);
                        continue;
                }

                if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
                        LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
                                "conn 0x%p peerstamp "LPU64" >= "
                                "newconn 0x%p peerstamp "LPU64"\n",
                                conn, conn->gnc_peerstamp,
                                newconn, newconn->gnc_peerstamp);

                        CDEBUG(D_NET, "Closing stale conn nid: %s "
                               " peerstamp:"LPX64"("LPX64")\n",
                               libcfs_nid2str(peer->gnp_nid),
                               conn->gnc_peerstamp, newconn->gnc_peerstamp);
                } else {

                        LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
                                "conn 0x%p peer_connstamp "LPU64" >= "
                                "newconn 0x%p peer_connstamp "LPU64"\n",
                                conn, conn->gnc_peer_connstamp,
                                newconn, newconn->gnc_peer_connstamp);

                        CDEBUG(D_NET, "Closing stale conn nid: %s"
                               " connstamp:"LPU64"("LPU64")\n",
                               libcfs_nid2str(peer->gnp_nid),
                               conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
                }

                count++;
                kgnilnd_close_conn_locked(conn, -ESTALE);
        }

        if (count != 0) {
                CWARN("Closed %d stale conns to %s\n", count, libcfs_nid2str(peer->gnp_nid));
        }

        RETURN(count);
}

int
kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
{
        kgn_conn_t       *conn;
        struct list_head *tmp;
        int               loopback;
        ENTRY;

        loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;

        list_for_each(tmp, &peer->gnp_conns) {
                conn = list_entry(tmp, kgn_conn_t, gnc_list);
                CDEBUG(D_NET, "checking conn 0x%p for peer %s"
                        " lo %d new "LPU64" existing "LPU64
                        " new peer "LPU64" existing peer "LPU64
                        " new dev %p existing dev %p\n",
                        conn, libcfs_nid2str(peer->gnp_nid),
                        loopback,
                        newconn->gnc_peerstamp, conn->gnc_peerstamp,
                        newconn->gnc_peer_connstamp, conn->gnc_peer_connstamp,
                        newconn->gnc_device, conn->gnc_device);

                /* conn is in the process of closing */
                if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
                        continue;

                /* 'newconn' is from an earlier version of 'peer'!!! */
                if (newconn->gnc_peerstamp < conn->gnc_peerstamp)
                        RETURN(1);

                /* 'conn' is from an earlier version of 'peer': it will be
                 * removed when we cull stale conns later on... */
                if (newconn->gnc_peerstamp > conn->gnc_peerstamp)
                        continue;

                /* Different devices are OK */
                if (conn->gnc_device != newconn->gnc_device)
                        continue;

                /* It's me connecting to myself */
                if (loopback &&
                    newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
                    newconn->gnc_peer_connstamp == conn->gnc_my_connstamp)
                        continue;

                /* 'newconn' is an earlier connection from 'peer'!!! */
                if (newconn->gnc_peer_connstamp < conn->gnc_peer_connstamp)
                        RETURN(2);

                /* 'conn' is an earlier connection from 'peer': it will be
                 * removed when we cull stale conns later on... */
                if (newconn->gnc_peer_connstamp > conn->gnc_peer_connstamp)
                        continue;

                /* 'newconn' has the SAME connection stamp; 'peer' isn't
                 * playing the game... */
                RETURN(3);
        }

        RETURN(0);
}

int
kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
{
        kgn_conn_t      *conn;
        gni_return_t    rrc;
        int             rc = 0;

        LASSERT (!in_interrupt());
        atomic_inc(&kgnilnd_data.kgn_nconns);

        /* divide by 2 to allow for complete reset and immediate reconnect */
        if (atomic_read(&kgnilnd_data.kgn_nconns) >= GNILND_MAX_CQID/2) {
                CERROR("Too many conn are live: %d > %d\n",
                        atomic_read(&kgnilnd_data.kgn_nconns), GNILND_MAX_CQID/2);
                atomic_dec(&kgnilnd_data.kgn_nconns);
                return -E2BIG;
        }

        LIBCFS_ALLOC(conn, sizeof(*conn));
        if (conn == NULL) {
                atomic_dec(&kgnilnd_data.kgn_nconns);
                return -ENOMEM;
        }

        LIBCFS_ALLOC(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
        if (conn->gnc_tx_ref_table == NULL) {
                CERROR("Can't allocate conn tx_ref_table\n");
                GOTO(failed, rc = -ENOMEM);
        }

        atomic_set(&conn->gnc_refcount, 1);
        atomic_set(&conn->gnc_reaper_noop, 0);
        atomic_set(&conn->gnc_sched_noop, 0);
        atomic_set(&conn->gnc_tx_in_use, 0);
        INIT_LIST_HEAD(&conn->gnc_list);
        INIT_LIST_HEAD(&conn->gnc_hashlist);
        INIT_LIST_HEAD(&conn->gnc_schedlist);
        INIT_LIST_HEAD(&conn->gnc_fmaq);
        INIT_LIST_HEAD(&conn->gnc_mdd_list);
        spin_lock_init(&conn->gnc_list_lock);
        spin_lock_init(&conn->gnc_tx_lock);
        conn->gnc_magic = GNILND_CONN_MAGIC;

        /* set tx id to nearly the end to make sure we find wrapping
         * issues soon */
        conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;

        /* if this fails, we have conflicts and MAX_TX is too large */
        CLASSERT(GNILND_MAX_MSG_ID < GNILND_MSGID_CLOSE);

        /* get a new unique CQ id for this conn */
        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        conn->gnc_my_connstamp = kgnilnd_data.kgn_connstamp++;
        conn->gnc_cqid = kgnilnd_get_cqid_locked();
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        if (conn->gnc_cqid == 0) {
                CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
                GOTO(failed, rc = -E2BIG);
        }

        CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
                conn->gnc_cqid, conn);

        /* need to be set before gnc_ephandle to allow kgnilnd_destroy_conn_ep to
         * check context */
        conn->gnc_device = dev;

        conn->gnc_timeout = MAX(*kgnilnd_tunables.kgn_timeout,
                                GNILND_MIN_TIMEOUT);
        kgnilnd_update_reaper_timeout(conn->gnc_timeout);

        /* this is the ep_handle for doing SMSG & BTE */
        mutex_lock(&dev->gnd_cq_mutex);
        rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
                                &conn->gnc_ephandle);
        mutex_unlock(&dev->gnd_cq_mutex);
        if (rrc != GNI_RC_SUCCESS)
                GOTO(failed, rc = -ENETDOWN);

        CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
               conn, conn->gnc_ephandle);

        /* add ref for EP canceling */
        kgnilnd_conn_addref(conn);
        atomic_inc(&dev->gnd_neps);

        *connp = conn;
        return 0;

failed:
        atomic_dec(&kgnilnd_data.kgn_nconns);
        LIBCFS_FREE(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
        LIBCFS_FREE(conn, sizeof(*conn));
        return rc;
}

/* needs to be called with kgn_peer_conn_lock held (read or write) */
kgn_conn_t *
kgnilnd_find_conn_locked(kgn_peer_t *peer)
{
        kgn_conn_t      *conn = NULL;

        /* if we are in reset, this conn is going to die soon */
        if (unlikely(kgnilnd_data.kgn_in_reset)) {
                RETURN(NULL);
        }

        /* just return the first ESTABLISHED connection */
        list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
                /* kgnilnd_finish_connect doesn't put connections on the
                 * peer list until they are actually established */
                LASSERTF(conn->gnc_state >= GNILND_CONN_ESTABLISHED,
                        "found conn %p state %s on peer %p (%s)\n",
                        conn, kgnilnd_conn_state2str(conn), peer,
                        libcfs_nid2str(peer->gnp_nid));
                if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
                        continue;

                RETURN(conn);
        }
        RETURN(NULL);
}

/* needs write_lock on kgn_peer_conn_lock held */
kgn_conn_t *
kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer) {

        kgn_device_t    *dev = peer->gnp_net->gnn_dev;
        kgn_conn_t      *conn;

        conn = kgnilnd_find_conn_locked(peer);

        if (conn != NULL) {
                return conn;
        }

        /* if the peer was previously connecting, check if we should
         * trigger another connection attempt yet. */
        if (time_before(jiffies, peer->gnp_reconnect_time)) {
                return NULL;
        }

        /* This check prevents us from creating a new connection to a peer while we are
         * still in the process of closing an existing connection to the peer.
         */
        list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
                if (conn->gnc_ephandle != NULL) {
                        CDEBUG(D_NET, "Not connecting non-null ephandle found peer 0x%p->%s\n", peer,
                                libcfs_nid2str(peer->gnp_nid));
                        return NULL;
                }
        }

        if (peer->gnp_connecting != GNILND_PEER_IDLE) {
                /* if we are not connecting, fire up a new connection */
                /* or if we are anything but IDLE DONT start a new connection */
               return NULL;
        }

        CDEBUG(D_NET, "starting connect to %s\n",
                libcfs_nid2str(peer->gnp_nid));
        peer->gnp_connecting = GNILND_PEER_CONNECT;
        kgnilnd_peer_addref(peer); /* extra ref for connd */

        spin_lock(&dev->gnd_connd_lock);
        list_add_tail(&peer->gnp_connd_list, &dev->gnd_connd_peers);
        spin_unlock(&dev->gnd_connd_lock);

        kgnilnd_schedule_dgram(dev);
        CDEBUG(D_NETTRACE, "scheduling new connect\n");

        return NULL;
}

/* Caller is responsible for deciding if/when to call this */
void
kgnilnd_destroy_conn_ep(kgn_conn_t *conn)
{
        gni_return_t    rrc;
        gni_ep_handle_t tmp_ep;

        /* only if we actually initialized it,
         *  then set NULL to tell kgnilnd_destroy_conn to leave it alone */

        tmp_ep = xchg(&conn->gnc_ephandle, NULL);
        if (tmp_ep != NULL) {
                /* we never re-use the EP, so unbind is not needed */
                mutex_lock(&conn->gnc_device->gnd_cq_mutex);
                rrc = kgnilnd_ep_destroy(tmp_ep);

                mutex_unlock(&conn->gnc_device->gnd_cq_mutex);

                /* if this fails, it could hork up kgni smsg retransmit and others
                 * since we could free the SMSG mbox memory, etc. */
                LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d conn 0x%p ep 0x%p\n",
                         rrc, conn, conn->gnc_ephandle);

                atomic_dec(&conn->gnc_device->gnd_neps);

                /* clear out count added in kgnilnd_close_conn_locked
                 * conn will have a peer once it hits finish_connect, where it
                 * is the first spot we'll mark it ESTABLISHED as well */
                if (conn->gnc_peer) {
                        kgnilnd_admin_decref(conn->gnc_peer->gnp_dirty_eps);
                }

                /* drop ref for EP */
                kgnilnd_conn_decref(conn);
        }
}

void
kgnilnd_destroy_conn(kgn_conn_t *conn)
{
        LASSERTF(!in_interrupt() &&
                !conn->gnc_scheduled &&
                !conn->gnc_in_purgatory &&
                conn->gnc_ephandle == NULL &&
                list_empty(&conn->gnc_list) &&
                list_empty(&conn->gnc_hashlist) &&
                list_empty(&conn->gnc_schedlist) &&
                list_empty(&conn->gnc_mdd_list) &&
                conn->gnc_magic == GNILND_CONN_MAGIC,
                "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d\n",
                conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
                                     : "<?>",
                !!in_interrupt(), conn->gnc_scheduled,
                conn->gnc_in_purgatory,
                conn->gnc_ephandle,
                conn->gnc_magic,
                list_empty(&conn->gnc_list),
                list_empty(&conn->gnc_hashlist),
                list_empty(&conn->gnc_schedlist),
                list_empty(&conn->gnc_mdd_list));

        /* Tripping these is especially bad, as it means we have items on the
         *  lists that didn't keep their refcount on the connection - or
         *  somebody evil released their own */
        LASSERTF(list_empty(&conn->gnc_fmaq) &&
                 atomic_read(&conn->gnc_nlive_fma) == 0 &&
                 atomic_read(&conn->gnc_nlive_rdma) == 0,
                 "conn 0x%p fmaq %d@0x%p nfma %d nrdma %d\n",
                 conn, kgnilnd_count_list(&conn->gnc_fmaq), &conn->gnc_fmaq,
                 atomic_read(&conn->gnc_nlive_fma), atomic_read(&conn->gnc_nlive_rdma));

        CDEBUG(D_NET, "destroying conn %p ephandle %p error %d\n",
                conn, conn->gnc_ephandle, conn->gnc_error);

        /* We are freeing this memory remove the magic value from the connection */
        conn->gnc_magic = 0;

        /* if there is an FMA blk left here, we'll tear it down */
        if (conn->gnc_fma_blk) {
                if (conn->gnc_peer) {
                        kgn_mbox_info_t *mbox;
                        mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
                        mbox->mbx_prev_nid = conn->gnc_peer->gnp_nid;
                }
                kgnilnd_release_mbox(conn, 0);
        }

        if (conn->gnc_peer != NULL)
                kgnilnd_peer_decref(conn->gnc_peer);

        if (conn->gnc_tx_ref_table != NULL) {
                LIBCFS_FREE(conn->gnc_tx_ref_table,
                            GNILND_MAX_MSG_ID * sizeof(void *));
        }

        LIBCFS_FREE(conn, sizeof(*conn));
        atomic_dec(&kgnilnd_data.kgn_nconns);
}

/* peer_alive and peer_notify done in the style of the o2iblnd */
void
kgnilnd_peer_alive(kgn_peer_t *peer)
{
        set_mb(peer->gnp_last_alive, jiffies);
}

void
kgnilnd_peer_notify(kgn_peer_t *peer, int error)
{
        int                     tell_lnet = 0;
        int                     nnets = 0;
        int                     rc;
        int                     i, j;
        kgn_conn_t             *conn;
        kgn_net_t             **nets;
        kgn_net_t              *net;


        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DONT_NOTIFY))
                return;

        /* Tell LNet we are giving ups on this peer - but only
         * if it isn't already reconnected or trying to reconnect */
        read_lock(&kgnilnd_data.kgn_peer_conn_lock);

        /* use kgnilnd_find_conn_locked to avoid any conns in the process of being nuked
         *
         * don't tell LNet if we are in reset - we assume that everyone will be able to
         * reconnect just fine
         */
        conn = kgnilnd_find_conn_locked(peer);

        CDEBUG(D_NETTRACE, "peer 0x%p->%s ting %d conn 0x%p, rst %d error %d\n",
               peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
               kgnilnd_data.kgn_in_reset, error);

        if ((peer->gnp_connecting == GNILND_PEER_IDLE) &&
            (conn == NULL) &&
            (!kgnilnd_data.kgn_in_reset) &&
            (!kgnilnd_conn_clean_errno(error))) {
                tell_lnet = 1;
        }

        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        if (!tell_lnet) {
                /* short circuit if we dont need to notify Lnet */
                return;
        }

        rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);

        if (rc) {
            /* dont do this if this fails since LNET is in shutdown or something else
             */

                for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
                        list_for_each_entry(net , &kgnilnd_data.kgn_nets[i], gnn_list) {
                                /* if gnn_shutdown set for any net shutdown is in progress just return */
                                if (net->gnn_shutdown) {
                                        up_read(&kgnilnd_data.kgn_net_rw_sem);
                                        return;
                                }
                                nnets++;
                        }
                }

                if (nnets == 0) {
                        /* shutdown in progress most likely */
                        up_read(&kgnilnd_data.kgn_net_rw_sem);
                        return;
                }

                LIBCFS_ALLOC(nets, nnets * sizeof(*nets));

                if (nets == NULL) {
                        up_read(&kgnilnd_data.kgn_net_rw_sem);
                        CERROR("Failed to allocate nets[%d]\n", nnets);
                        return;
                }

                j = 0;
                for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
                        list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
                                nets[j] = net;
                                kgnilnd_net_addref(net);
                                j++;
                        }
                }
                up_read(&kgnilnd_data.kgn_net_rw_sem);

                for (i = 0; i < nnets; i++) {
                        lnet_nid_t peer_nid;

                        net = nets[i];

                        peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid,
                                                                 peer->gnp_nid);

                        CDEBUG(D_NET, "peer 0x%p->%s last_alive %lu (%lus ago)\n",
                                peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
                                cfs_duration_sec(jiffies - peer->gnp_last_alive));

                        lnet_notify(net->gnn_ni, peer_nid, 0, peer->gnp_last_alive);


                        kgnilnd_net_decref(net);
                }

                LIBCFS_FREE(nets, nnets * sizeof(*nets));
        }
}

/* need write_lock on kgn_peer_conn_lock */
void
kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
{
        kgn_peer_t        *peer = conn->gnc_peer;
        ENTRY;

        LASSERT(!in_interrupt());

        /* store error for tx completion */
        conn->gnc_error = error;
        peer->gnp_last_errno = error;

        /* use real error from peer if possible */
        if (error == -ECONNRESET) {
                error = conn->gnc_peer_error;
        }

        /* if we NETERROR, make sure it is rate limited */
        if (!kgnilnd_conn_clean_errno(error) &&
            peer->gnp_down == GNILND_RCA_NODE_UP) {
                CNETERR("closing conn to %s: error %d\n",
                       libcfs_nid2str(peer->gnp_nid), error);
        } else {
                CDEBUG(D_NET, "closing conn to %s: error %d\n",
                       libcfs_nid2str(peer->gnp_nid), error);
        }

        LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
                "conn %p to %s with bogus state %s\n", conn,
                libcfs_nid2str(conn->gnc_peer->gnp_nid),
                kgnilnd_conn_state2str(conn));
        LASSERT(!list_empty(&conn->gnc_hashlist));
        LASSERT(!list_empty(&conn->gnc_list));


        /* mark peer count here so any place the EP gets destroyed will
         * open up the peer count so that a new ESTABLISHED conn is then free
         * to send new messages -- sending before the previous EPs are destroyed
         * could end up with messages on the network for the old conn _after_
         * the new conn and break the mbox safety protocol */
        kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);

        /* Remove from conn hash table: no new callbacks */
        list_del_init(&conn->gnc_hashlist);
        kgnilnd_data.kgn_conn_version++;
        kgnilnd_conn_decref(conn);

        /* if we are in reset, go right to CLOSED as there is no scheduler
         * thread to move from CLOSING to CLOSED */
        if (unlikely(kgnilnd_data.kgn_in_reset)) {
                conn->gnc_state = GNILND_CONN_CLOSED;
        } else {
                conn->gnc_state = GNILND_CONN_CLOSING;
        }

        /* leave on peer->gnp_conns to make sure we don't let the reaper
         * or others try to unlink this peer until the conn is fully
         * processed for closing */

        if (kgnilnd_check_purgatory_conn(conn)) {
                kgnilnd_add_purgatory_locked(conn, conn->gnc_peer);
        }

        /* Reset RX timeout to ensure we wait for an incoming CLOSE
         * for the full timeout.  If we get a CLOSE we know the
         * peer has stopped all RDMA.  Otherwise if we wait for
         * the full timeout we can also be sure all RDMA has stopped. */
        conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
        mb();

        /* schedule sending CLOSE - if we are in quiesce, this adds to
         * gnd_ready_conns and allows us to find it in quiesce processing */
        kgnilnd_schedule_conn(conn);

        EXIT;
}

void
kgnilnd_close_conn(kgn_conn_t *conn, int error)
{
        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        /* need to check the state here - this call is racy and we don't
         * know the state until after the lock is grabbed */
        if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
                kgnilnd_close_conn_locked(conn, error);
        }
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
}

void
kgnilnd_complete_closed_conn(kgn_conn_t *conn)
{
        LIST_HEAD               (sinners);
        kgn_tx_t               *tx, *txn;
        int                     nlive = 0;
        int                     nrdma = 0;
        int                     nq_rdma = 0;
        int                     logmsg;
        ENTRY;

        /* Dump log  on cksum error - wait until complete phase to let
         * RX of error happen */
        if (*kgnilnd_tunables.kgn_checksum_dump &&
            (conn != NULL && conn->gnc_peer_error == -ENOKEY)) {
                libcfs_debug_dumplog();
        }

        /* _CLOSED set in kgnilnd_process_fmaq once we decide to
         * send the CLOSE or not */
        LASSERTF(conn->gnc_state == GNILND_CONN_CLOSED,
                 "conn 0x%p->%s with bad state %s\n",
                 conn, conn->gnc_peer ?
                        libcfs_nid2str(conn->gnc_peer->gnp_nid) :
                        "<?>",
                 kgnilnd_conn_state2str(conn));

        LASSERT(list_empty(&conn->gnc_hashlist));

        /* we've sent the close, start nuking */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
                kgnilnd_schedule_conn(conn);

        if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
                CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
                                "done, Attempting to recover conn 0x%p "
                                "scheduled %d function: %s line: %d\n", conn,
                                conn->gnc_scheduled, conn->gnc_sched_caller,
                                conn->gnc_sched_line);
                RETURN_EXIT;
        }

        /* we don't use lists to track things that we can get out of the
         * tx_ref table... */

        /* need to hold locks for tx_list_state, sampling it is too racy:
         * - the lock actually protects tx != NULL, but we can't take the proper
         *   lock until we check tx_list_state, which would be too late and
         *   we could have the TX change under us.
         * gnd_rdmaq_lock and gnd_lock and not used together, so taking both
         * should be fine */
        spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
        spin_lock(&conn->gnc_device->gnd_lock);

        for (nrdma = 0; nrdma < GNILND_MAX_MSG_ID; nrdma++) {
                tx = conn->gnc_tx_ref_table[nrdma];

                if (tx != NULL) {
                        /* only print the first error and if not CLOSE, we often don't see
                         * CQ events for that by the time we get here... and really don't care */
                        if (nlive || tx->tx_msg.gnm_type == GNILND_MSG_CLOSE)
                                tx->tx_state |= GNILND_TX_QUIET_ERROR;
                        nlive++;
                        GNIDBG_TX(D_NET, tx, "cleaning up on close, nlive %d", nlive);

                        /* don't worry about gnc_lock here as nobody else should be
                         * touching this conn */
                        kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
                        list_add_tail(&tx->tx_list, &sinners);
                }
        }
        spin_unlock(&conn->gnc_device->gnd_lock);
        spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);

        /* nobody should have marked this as needing scheduling after
         * we called close - so only ref should be us handling it */
        if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
                CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
                                "done, Attempting to recover conn 0x%p "
                                "scheduled %d function %s line: %d\n", conn,
                                conn->gnc_scheduled, conn->gnc_sched_caller,
                                conn->gnc_sched_line);
        }
        /* now reset a few to actual counters... */
        nrdma = atomic_read(&conn->gnc_nlive_rdma);
        nq_rdma = atomic_read(&conn->gnc_nq_rdma);

        if (!list_empty(&sinners)) {
                list_for_each_entry_safe(tx, txn, &sinners, tx_list) {
                        /* clear tx_list to make tx_add_list_locked happy */
                        list_del_init(&tx->tx_list);
                        /* The error codes determine if we hold onto the MDD */
                        kgnilnd_tx_done(tx, conn->gnc_error);
                }
        }

        logmsg = (nlive + nrdma + nq_rdma);

        if (logmsg) {
                if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_UP) {
                        CNETERR("Closed conn 0x%p->%s (errno %d, peer errno %d): "
                                "canceled %d TX, %d/%d RDMA\n",
                                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
                                conn->gnc_error, conn->gnc_peer_error,
                                nlive, nq_rdma, nrdma);
                } else {
                        CDEBUG(D_NET, "Closed conn 0x%p->%s (errno %d,"
                                " peer errno %d): canceled %d TX, %d/%d RDMA\n",
                                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
                                conn->gnc_error, conn->gnc_peer_error,
                                nlive, nq_rdma, nrdma);
                }
        }

        kgnilnd_destroy_conn_ep(conn);

        /* Bug 765042 - race this with completing a new conn to same peer - we need
         * finish_connect to detach purgatory before we can do it ourselves here */
        CFS_RACE(CFS_FAIL_GNI_FINISH_PURG);

        /* now it is safe to remove from peer list - anyone looking at
         * gnp_conns now is free to unlink if not on purgatory */
        write_lock(&kgnilnd_data.kgn_peer_conn_lock);

        conn->gnc_state = GNILND_CONN_DONE;

        /* Decrement counter if we are marked by del_conn_or_peers for closing
         */
        if (conn->gnc_needs_closing)
                kgnilnd_admin_decref(kgnilnd_data.kgn_npending_conns);

        /* Remove from peer's list of valid connections if its not in purgatory */
        if (!conn->gnc_in_purgatory) {
                list_del_init(&conn->gnc_list);
                /* Lose peers reference on the conn */
                kgnilnd_conn_decref(conn);
        }

        /* NB - only unlinking if we set pending in del_peer_locked from admin or
         * shutdown */
        if (kgnilnd_peer_active(conn->gnc_peer) &&
            conn->gnc_peer->gnp_pending_unlink &&
            kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
                kgnilnd_unlink_peer_locked(conn->gnc_peer);
        }

        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        /* I'm telling Mommy! - use peer_error if they initiated close */
        kgnilnd_peer_notify(conn->gnc_peer,
                            conn->gnc_error == -ECONNRESET ? conn->gnc_peer_error
                                                           : conn->gnc_error);

        EXIT;
}

int
kgnilnd_set_conn_params(kgn_dgram_t *dgram)
{
        kgn_conn_t             *conn = dgram->gndg_conn;
        kgn_connreq_t          *connreq = &dgram->gndg_conn_in;
        kgn_gniparams_t        *rem_param = &connreq->gncr_gnparams;
        gni_return_t            rrc;
        int                     rc = 0;
        gni_smsg_attr_t        *remote = &connreq->gncr_gnparams.gnpr_smsg_attr;

        /* set timeout vals in conn early so we can use them for the NAK */

        /* use max of the requested and our timeout, peer will do the same */
        conn->gnc_timeout = MAX(conn->gnc_timeout, connreq->gncr_timeout);

        /* only ep_bind really mucks around with the CQ */
        /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
         * is necessary as you can only bind an ep once and we must make sure we dont bind when already bound.
         */
        if (connreq->gncr_dstnid != LNET_NID_ANY && dgram->gndg_conn_out.gncr_dstnid != connreq->gncr_srcnid) {
                mutex_lock(&conn->gnc_device->gnd_cq_mutex);
                rrc = kgnilnd_ep_bind(conn->gnc_ephandle,
                        connreq->gncr_gnparams.gnpr_host_id,
                        conn->gnc_cqid);
                mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                if (rrc != GNI_RC_SUCCESS) {
                        rc = -ECONNABORTED;
                        goto return_out;
                }
        }

        rrc = kgnilnd_ep_set_eventdata(conn->gnc_ephandle, conn->gnc_cqid,
                         connreq->gncr_gnparams.gnpr_cqid);
        if (rrc != GNI_RC_SUCCESS) {
                rc = -ECONNABORTED;
                goto cleanup_out;
        }

        /* Initialize SMSG */
        rrc = kgnilnd_smsg_init(conn->gnc_ephandle, &conn->gnpr_smsg_attr,
                        &connreq->gncr_gnparams.gnpr_smsg_attr);
        if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
                gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
                /* help folks figure out if there is a tunable off, etc. */
                LCONSOLE_ERROR("SMSG attribute mismatch. Data from local/remote:"
                               " type %d/%d msg_maxsize %u/%u"
                               " mbox_maxcredit %u/%u. Please check kgni"
                               " logs for further data\n",
                               local->msg_type, remote->msg_type,
                               local->msg_maxsize, remote->msg_maxsize,
                               local->mbox_maxcredit, remote->mbox_maxcredit);
        }
        if (rrc != GNI_RC_SUCCESS) {
                rc = -ECONNABORTED;
                goto cleanup_out;
        }

        /* log this for help in debuggin SMSG buffer re-use */
        CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
                " local cqid %u SMSG %p->%u hndl "LPX64"."LPX64
                " remote cqid %u SMSG %p->%u hndl "LPX64"."LPX64"\n",
                conn, libcfs_nid2str(connreq->gncr_srcnid),
                libcfs_nid2str(connreq->gncr_dstnid),
                &conn->gnpr_smsg_attr,
                conn->gnc_cqid,
                conn->gnpr_smsg_attr.msg_buffer,
                conn->gnpr_smsg_attr.mbox_offset,
                conn->gnpr_smsg_attr.mem_hndl.qword1,
                conn->gnpr_smsg_attr.mem_hndl.qword2,
                rem_param->gnpr_cqid,
                rem_param->gnpr_smsg_attr.msg_buffer,
                rem_param->gnpr_smsg_attr.mbox_offset,
                rem_param->gnpr_smsg_attr.mem_hndl.qword1,
                rem_param->gnpr_smsg_attr.mem_hndl.qword2);

        conn->gnc_peerstamp = connreq->gncr_peerstamp;
        conn->gnc_peer_connstamp = connreq->gncr_connstamp;
        conn->remote_mbox_addr = (void *)((char *)remote->msg_buffer + remote->mbox_offset);

        /* We update the reaper timeout once we have a valid conn and timeout */
        kgnilnd_update_reaper_timeout(GNILND_TO2KA(conn->gnc_timeout));

        return 0;

cleanup_out:
        rrc = kgnilnd_ep_unbind(conn->gnc_ephandle);
        /* not sure I can just let this fly */
        LASSERTF(rrc == GNI_RC_SUCCESS,
                "bad rc from gni_ep_unbind trying to cleanup: %d\n", rrc);

return_out:
        LASSERTF(rc != 0, "SOFTWARE BUG: rc == 0\n");
        CERROR("Error setting connection params from %s: %d\n",
               libcfs_nid2str(connreq->gncr_srcnid), rc);
        return rc;
}

/* needs down_read on kgn_net_rw_sem held from before this call until
 * after the write_lock on kgn_peer_conn_lock - this ensures we stay sane
 * with kgnilnd_shutdown - it'll get the sem and set shutdown, then get the
 * kgn_peer_conn_lock to start del_peer'ing. If we hold the sem until after
 * kgn_peer_conn_lock is held, we guarantee that nobody calls
 * kgnilnd_add_peer_locked without checking gnn_shutdown */
int
kgnilnd_create_peer_safe(kgn_peer_t **peerp,
                         lnet_nid_t nid,
                         kgn_net_t *net,
                         int node_state)
{
        kgn_peer_t      *peer;
        int             rc;

        LASSERT(nid != LNET_NID_ANY);

        /* We dont pass the net around in the dgram anymore so here is where we find it
         * this will work unless its in shutdown or the nid has a net that is invalid.
         * Either way error code needs to be returned in that case.
         *
         * If the net passed in is not NULL then we can use it, this alleviates looking it
         * when the calling function has access to the data.
         */
        if (net == NULL) {
                rc = kgnilnd_find_net(nid, &net);
                if (rc < 0)
                        return rc;
        } else {
                /* find net adds a reference on the net if we are not using
                 * it we must do it manually so the net references are
                 * correct when tearing down the net
                 */
                kgnilnd_net_addref(net);
        }

        LIBCFS_ALLOC(peer, sizeof(*peer));
        if (peer == NULL) {
                kgnilnd_net_decref(net);
                return -ENOMEM;
        }
        peer->gnp_nid = nid;
        peer->gnp_down = node_state;

        /* translate from nid to nic addr & store */
        rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
        if (rc <= 0) {
                kgnilnd_net_decref(net);
                LIBCFS_FREE(peer, sizeof(*peer));
                return -ESRCH;
        }
        CDEBUG(D_NET, "peer 0x%p->%s -> NIC 0x%x\n", peer,
                libcfs_nid2str(nid), peer->gnp_host_id);

        atomic_set(&peer->gnp_refcount, 1);     /* 1 ref for caller */
        atomic_set(&peer->gnp_dirty_eps, 0);

        INIT_LIST_HEAD(&peer->gnp_list);
        INIT_LIST_HEAD(&peer->gnp_connd_list);
        INIT_LIST_HEAD(&peer->gnp_conns);
        INIT_LIST_HEAD(&peer->gnp_tx_queue);

        /* the first reconnect should happen immediately, so we leave
         * gnp_reconnect_interval set to 0 */

        LASSERTF(net != NULL, "peer 0x%p->%s with NULL net\n",
                 peer, libcfs_nid2str(nid));

        /* must have kgn_net_rw_sem held for this...  */
        if (net->gnn_shutdown) {
                /* shutdown has started already */
                kgnilnd_net_decref(net);
                LIBCFS_FREE(peer, sizeof(*peer));
                return -ESHUTDOWN;
        }

        peer->gnp_net = net;

        atomic_inc(&kgnilnd_data.kgn_npeers);

        *peerp = peer;
        return 0;
}

void
kgnilnd_destroy_peer(kgn_peer_t *peer)
{
        CDEBUG(D_NET, "peer %s %p deleted\n",
               libcfs_nid2str(peer->gnp_nid), peer);
        LASSERTF(atomic_read(&peer->gnp_refcount) == 0,
                 "peer 0x%p->%s refs %d\n",
                 peer, libcfs_nid2str(peer->gnp_nid),
                 atomic_read(&peer->gnp_refcount));
        LASSERTF(atomic_read(&peer->gnp_dirty_eps) == 0,
                 "peer 0x%p->%s dirty eps %d\n",
                 peer, libcfs_nid2str(peer->gnp_nid),
                 atomic_read(&peer->gnp_dirty_eps));
        LASSERTF(peer->gnp_net != NULL, "peer %p (%s) with NULL net\n",
                 peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(!kgnilnd_peer_active(peer),
                 "peer 0x%p->%s\n",
                peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE || peer->gnp_connecting == GNILND_PEER_KILL,
                 "peer 0x%p->%s, connecting %d\n",
                peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
        LASSERTF(list_empty(&peer->gnp_conns),
                 "peer 0x%p->%s\n",
                peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(list_empty(&peer->gnp_tx_queue),
                 "peer 0x%p->%s\n",
                peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(list_empty(&peer->gnp_connd_list),
                 "peer 0x%p->%s\n",
                peer, libcfs_nid2str(peer->gnp_nid));

        /* NB a peer's connections keep a reference on their peer until
         * they are destroyed, so we can be assured that _all_ state to do
         * with this peer has been cleaned up when its refcount drops to
         * zero. */

        atomic_dec(&kgnilnd_data.kgn_npeers);
        kgnilnd_net_decref(peer->gnp_net);

        LIBCFS_FREE(peer, sizeof(*peer));
}

/* the conn might not have made it all the way through to a connected
 * state - but we need to purgatory any conn that a remote peer might
 * have seen through a posted dgram as well */
void
kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
{
        kgn_mbox_info_t *mbox = NULL;
        ENTRY;

        /* NB - the caller should own conn by removing him from the
         * scheduler thread when finishing the close */

        LASSERTF(peer != NULL, "conn %p with NULL peer\n", conn);

        /* If this is still true, need to add the calls to unlink back in and
         * figure out how to close the hole on loopback conns */
        LASSERTF(kgnilnd_peer_active(peer), "can't use inactive peer %s (%p)"
                " we'll never recover the resources\n",
                libcfs_nid2str(peer->gnp_nid), peer);

        CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
                conn->gnc_device);

        LASSERTF(conn->gnc_in_purgatory == 0,
                "Conn already in purgatory\n");
        conn->gnc_in_purgatory = 1;

        mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
        mbox->mbx_prev_purg_nid = peer->gnp_nid;
        mbox->mbx_add_purgatory = jiffies;
        kgnilnd_release_mbox(conn, 1);

        LASSERTF(list_empty(&conn->gnc_mdd_list),
                "conn 0x%p->%s with active purgatory hold MDD %d\n",
                conn, libcfs_nid2str(peer->gnp_nid),
                kgnilnd_count_list(&conn->gnc_mdd_list));

        EXIT;
}

/* Instead of detaching everything from purgatory here we just mark the conn as needing
 * detach, when the reaper checks the conn the next time it will detach it.
 * Calling function requires write_lock held on kgn_peer_conn_lock
 */
void
kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer) {
        kgn_conn_t       *conn;

        list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
                if (conn->gnc_in_purgatory && !conn->gnc_needs_detach) {
                        conn->gnc_needs_detach = 1;
                        kgnilnd_admin_addref(kgnilnd_data.kgn_npending_detach);
                }
        }
}

/* Calling function needs a write_lock held on kgn_peer_conn_lock */
void
kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
{
        kgn_mbox_info_t *mbox = NULL;

        /* if needed, add the conn purgatory data to the list passed in */
        if (conn->gnc_in_purgatory) {
                CDEBUG(D_NET, "peer %p->%s purg_conn %p@%s mdd_list #tx %d\n",
                        conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
                        conn, kgnilnd_conn_state2str(conn),
                        kgnilnd_count_list(&conn->gnc_mdd_list));

                mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
                mbox->mbx_detach_of_purgatory = jiffies;

                /* conn->gnc_list is the entry point on peer->gnp_conns, so detaching it
                 * here removes it from the list of 'valid' peer connections.
                 * We put the current conn onto a list of conns to call kgnilnd_release_purgatory_locked()
                 * and as such the caller of kgnilnd_detach_purgatory_locked() now owns that conn, since its not
                 * on the peer's conn_list anymore.
                 */

                list_del_init(&conn->gnc_list);

                /* NB - only unlinking if we set pending in del_peer_locked from admin or
                 * shutdown */
                if (kgnilnd_peer_active(conn->gnc_peer) &&
                    conn->gnc_peer->gnp_pending_unlink &&
                    kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
                        kgnilnd_unlink_peer_locked(conn->gnc_peer);
                }
                /* The reaper will not call detach unless the conn is fully through kgnilnd_complete_closed_conn.
                 * If the conn is not in a DONE state somehow we are attempting to detach even though
                 * the conn has not been fully cleaned up. If we detach while the conn is still closing
                 * we will end up with an orphaned connection that has valid ep_handle, that is not on a
                 * peer.
                 */

                LASSERTF(conn->gnc_state == GNILND_CONN_DONE, "Conn in invalid state  %p@%s \n",
                                conn, kgnilnd_conn_state2str(conn));

                /* move from peer to the delayed release list */
                list_add_tail(&conn->gnc_list, conn_list);
        }
}

void
kgnilnd_release_purgatory_list(struct list_head *conn_list)
{
        kgn_device_t            *dev;
        kgn_conn_t              *conn, *connN;
        kgn_mdd_purgatory_t     *gmp, *gmpN;

        list_for_each_entry_safe(conn, connN, conn_list, gnc_list) {
                dev = conn->gnc_device;

                kgnilnd_release_mbox(conn, -1);
                conn->gnc_in_purgatory = 0;

                list_del_init(&conn->gnc_list);

                /* gnc_needs_detach is set in kgnilnd_del_conn_or_peer. It is used to keep track
                 * of conns that have been marked for detach by kgnilnd_del_conn_or_peer.
                 * The function uses kgn_npending_detach to verify the conn has
                 * actually been detached.
                 */

                if (conn->gnc_needs_detach)
                        kgnilnd_admin_decref(kgnilnd_data.kgn_npending_detach);

                /* if this guy is really dead (we are doing release from reaper),
                 * make sure we tell LNet - if this is from other context,
                 * the checks in the function will prevent an errant
                 * notification */
                kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error);

                list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
                                         gmp_list) {
                        CDEBUG(D_NET,
                               "dev %p releasing held mdd "LPX64"."LPX64"\n",
                               conn->gnc_device, gmp->gmp_map_key.qword1,
                               gmp->gmp_map_key.qword2);

                        atomic_dec(&dev->gnd_n_mdd_held);
                        kgnilnd_mem_mdd_release(conn->gnc_device->gnd_handle,
                                                &gmp->gmp_map_key);
                        /* ignoring the return code - if kgni/ghal can't find it
                         * it must be released already */

                        list_del_init(&gmp->gmp_list);
                        LIBCFS_FREE(gmp, sizeof(*gmp));
                }
                /* lose conn ref for purgatory */
                kgnilnd_conn_decref(conn);
        }
}

/* needs write_lock on kgnilnd_data.kgn_peer_conn_lock held */
void
kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer)
{
        int current_to;

        current_to = peer->gnp_reconnect_interval;

        /* we'll try to reconnect fast the first time, then back-off */
        if (current_to == 0) {
                peer->gnp_reconnect_time = jiffies - 1;
                current_to = *kgnilnd_tunables.kgn_min_reconnect_interval;
        } else {
                peer->gnp_reconnect_time = jiffies + cfs_time_seconds(current_to);
                /* add 50% of min timeout & retry */
                current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
        }

        current_to = MIN(current_to,
                                *kgnilnd_tunables.kgn_max_reconnect_interval);

        peer->gnp_reconnect_interval = current_to;
        CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
               libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_time,
               peer->gnp_reconnect_interval);
}

/* needs kgnilnd_data.kgn_peer_conn_lock held */
kgn_peer_t *
kgnilnd_find_peer_locked(lnet_nid_t nid)
{
        struct list_head *peer_list = kgnilnd_nid2peerlist(nid);
        kgn_peer_t       *peer;

        /* Chopping nid down to only NIDADDR using LNET_NIDADDR so we only
         * have a single peer per device instead of a peer per nid/net combo.
         */

        list_for_each_entry(peer, peer_list, gnp_list) {
                if (LNET_NIDADDR(nid) != LNET_NIDADDR(peer->gnp_nid))
                        continue;

                CDEBUG(D_NET, "got peer [%p] -> %s c %d (%d)\n",
                       peer, libcfs_nid2str(nid),
                       peer->gnp_connecting,
                       atomic_read(&peer->gnp_refcount));
                return peer;
        }
        return NULL;
}

/* need write_lock on kgn_peer_conn_lock */
void
kgnilnd_unlink_peer_locked(kgn_peer_t *peer)
{
        LASSERTF(list_empty(&peer->gnp_conns),
                "peer 0x%p->%s\n",
                 peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(list_empty(&peer->gnp_tx_queue),
                "peer 0x%p->%s\n",
                 peer, libcfs_nid2str(peer->gnp_nid));
        LASSERTF(kgnilnd_peer_active(peer),
                "peer 0x%p->%s\n",
                 peer, libcfs_nid2str(peer->gnp_nid));
        CDEBUG(D_NET, "unlinking peer 0x%p->%s\n",
                peer, libcfs_nid2str(peer->gnp_nid));

        list_del_init(&peer->gnp_list);
        kgnilnd_data.kgn_peer_version++;
        kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
        /* lose peerlist's ref */
        kgnilnd_peer_decref(peer);
}

int
kgnilnd_get_peer_info(int index,
                      kgn_peer_t **found_peer,
                      lnet_nid_t *id, __u32 *nic_addr,
                      int *refcount, int *connecting)
{
        struct list_head  *ptmp;
        kgn_peer_t        *peer;
        int               i;
        int               rc = -ENOENT;

        read_lock(&kgnilnd_data.kgn_peer_conn_lock);

        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {

                list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
                        peer = list_entry(ptmp, kgn_peer_t, gnp_list);

                        if (index-- > 0)
                                continue;

                        CDEBUG(D_NET, "found peer %p (%s) at index %d\n",
                               peer, libcfs_nid2str(peer->gnp_nid), index);

                        *found_peer  = peer;
                        *id          = peer->gnp_nid;
                        *nic_addr    = peer->gnp_host_id;
                        *refcount    = atomic_read(&peer->gnp_refcount);
                        *connecting  = peer->gnp_connecting;

                        rc = 0;
                        goto out;
                }
        }
out:
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        if (rc)
                CDEBUG(D_NET, "no gni peer at index %d\n", index);
        return rc;
}

/* requires write_lock on kgn_peer_conn_lock held */
void
kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp)
{
        kgn_peer_t        *peer, *peer2;

        LASSERTF(new_stub_peer != NULL, "bad stub peer for nid %s\n",
                 libcfs_nid2str(nid));

        peer2 = kgnilnd_find_peer_locked(nid);
        if (peer2 != NULL) {
                /* A peer was created during the lock transition, so drop
                 * the new one we created */
                kgnilnd_peer_decref(new_stub_peer);
                peer = peer2;
        } else {
                peer = new_stub_peer;
                /* peer table takes existing ref on peer */

                LASSERTF(!kgnilnd_peer_active(peer),
                        "peer 0x%p->%s already in peer table\n",
                        peer, libcfs_nid2str(peer->gnp_nid));
                list_add_tail(&peer->gnp_list,
                              kgnilnd_nid2peerlist(nid));
                kgnilnd_data.kgn_peer_version++;
        }

        LASSERTF(peer->gnp_net != NULL, "peer 0x%p->%s with NULL net\n",
                 peer, libcfs_nid2str(peer->gnp_nid));
        *peerp = peer;
}

int
kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
{
        kgn_peer_t        *peer;
        int                rc;
        int                node_state;
        ENTRY;

        if (nid == LNET_NID_ANY)
                return -EINVAL;

        node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));

        /* NB - this will not block during normal operations -
         * the only writer of this is in the startup/shutdown path. */
        rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
        if (!rc) {
                rc = -ESHUTDOWN;
                RETURN(rc);
        }
        rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
        if (rc != 0) {
                up_read(&kgnilnd_data.kgn_net_rw_sem);
                RETURN(rc);
        }

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        up_read(&kgnilnd_data.kgn_net_rw_sem);

        kgnilnd_add_peer_locked(nid, peer, peerp);

        CDEBUG(D_NET, "peer 0x%p->%s connecting %d\n",
               peerp, libcfs_nid2str((*peerp)->gnp_nid),
               (*peerp)->gnp_connecting);

        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        RETURN(0);
}

/* needs write_lock on kgn_peer_conn_lock */
void
kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies)
{
        kgn_tx_t        *tx, *txn;

        /* we do care about state of gnp_connecting - we could be between
         * reconnect attempts, so try to find the dgram and cancel the TX
         * anyways. If we are in the process of posting DONT do anything;
         * once it fails or succeeds we can nuke the connect attempt.
         * We have no idea where in kgnilnd_post_dgram we are so we cant
         * attempt to cancel until the function is done.
         */

        /* make sure peer isn't in process of connecting or waiting for connect*/
        spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
        if (!(list_empty(&peer->gnp_connd_list))) {
                list_del_init(&peer->gnp_connd_list);
                /* remove connd ref */
                kgnilnd_peer_decref(peer);
        }
        spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);

        if (peer->gnp_connecting == GNILND_PEER_POSTING || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
                peer->gnp_connecting = GNILND_PEER_NEEDS_DEATH;
                /* We are in process of posting right now the xchg set it up for us to
                 * cancel the connect so we are finished for now */
        } else {
                /* no need for exchange we have the peer lock and its ready for us to nuke */
                LASSERTF(peer->gnp_connecting != GNILND_PEER_POSTING,
                        "Peer in invalid state 0x%p->%s, connecting %d\n",
                        peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
                peer->gnp_connecting = GNILND_PEER_IDLE;
                set_mb(peer->gnp_last_dgram_errno, -ETIMEDOUT);
                kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
                                                      peer->gnp_nid);
        }

        /* The least we can do is nuke the tx's no matter what.... */
        list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
                kgnilnd_tx_del_state_locked(tx, peer, NULL,
                                           GNILND_TX_ALLOCD);
                list_add_tail(&tx->tx_list, zombies);
        }
}

/* needs write_lock on kgn_peer_conn_lock */
void
kgnilnd_del_peer_locked(kgn_peer_t *peer, int error)
{
        /* this peer could be passive and only held for purgatory,
         * take a ref to ensure it doesn't disappear in this function */
        kgnilnd_peer_addref(peer);

        CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);

        /* if purgatory release cleared it out, don't try again */
        if (kgnilnd_peer_active(peer)) {
                /* always do this to allow kgnilnd_start_connect and
                 * kgnilnd_finish_connect to catch this before they
                 * wrap up their operations */
                if (kgnilnd_can_unlink_peer_locked(peer)) {
                        /* already released purgatory, so only active
                         * conns hold it */
                        kgnilnd_unlink_peer_locked(peer);
                } else {
                        kgnilnd_close_peer_conns_locked(peer, error);
                        /* peer unlinks itself when last conn is closed */
                }
        }

        /* we are done, release back to the wild */
        kgnilnd_peer_decref(peer);
}

int
kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
                          int error)
{
        LIST_HEAD               (souls);
        LIST_HEAD               (zombies);
        struct list_head        *ptmp, *pnxt;
        kgn_peer_t              *peer;
        int                     lo;
        int                     hi;
        int                     i;
        int                     rc = -ENOENT;

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);

        if (nid != LNET_NID_ANY)
                lo = hi = kgnilnd_nid2peerlist(nid) - kgnilnd_data.kgn_peers;
        else {
                lo = 0;
                hi = *kgnilnd_tunables.kgn_peer_hash_size - 1;
                /* wildcards always succeed */
                rc = 0;
        }

        for (i = lo; i <= hi; i++) {
                list_for_each_safe(ptmp, pnxt, &kgnilnd_data.kgn_peers[i]) {
                        peer = list_entry(ptmp, kgn_peer_t, gnp_list);

                        LASSERTF(peer->gnp_net != NULL,
                                "peer %p (%s) with NULL net\n",
                                 peer, libcfs_nid2str(peer->gnp_nid));

                        if (net != NULL && peer->gnp_net != net)
                                continue;

                        if (!(nid == LNET_NID_ANY || LNET_NIDADDR(peer->gnp_nid) == LNET_NIDADDR(nid)))
                                continue;

                        /* In both cases, we want to stop any in-flight
                         * connect attempts */
                        kgnilnd_cancel_peer_connect_locked(peer, &zombies);

                        switch (command) {
                        case GNILND_DEL_CONN:
                                kgnilnd_close_peer_conns_locked(peer, error);
                                break;
                        case GNILND_DEL_PEER:
                                peer->gnp_pending_unlink = 1;
                                kgnilnd_admin_addref(kgnilnd_data.kgn_npending_unlink);
                                kgnilnd_mark_for_detach_purgatory_all_locked(peer);
                                kgnilnd_del_peer_locked(peer, error);
                                break;
                        case GNILND_CLEAR_PURGATORY:
                                /* Mark everything ready for detach reaper will cleanup
                                 * once we release the kgn_peer_conn_lock
                                 */
                                kgnilnd_mark_for_detach_purgatory_all_locked(peer);
                                peer->gnp_last_errno = -EISCONN;
                                /* clear reconnect so he can reconnect soon */
                                peer->gnp_reconnect_time = 0;
                                peer->gnp_reconnect_interval = 0;
                                break;
                        default:
                                CERROR("bad command %d\n", command);
                                LBUG();
                        }
                        /* we matched something */
                        rc = 0;
                }
        }

        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        /* nuke peer TX */
        kgnilnd_txlist_done(&zombies, error);

        /* This function does not return until the commands it initiated have completed,
         * since they have to work there way through the other threads. In the case of shutdown
         * threads are not woken up until after this call is initiated so we cannot wait, we just
         * need to return. The same applies for stack reset we shouldnt wait as the reset thread
         * handles closing.
         */

        CFS_RACE(CFS_FAIL_GNI_RACE_RESET);

        if (error == -ENOTRECOVERABLE || error == -ESHUTDOWN) {
                return rc;
        }

        i = 4;
        while (atomic_read(&kgnilnd_data.kgn_npending_conns)   ||
               atomic_read(&kgnilnd_data.kgn_npending_detach)  ||
               atomic_read(&kgnilnd_data.kgn_npending_unlink)) {

                cfs_pause(cfs_time_seconds(1));
                i++;

                CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
                                atomic_read(&kgnilnd_data.kgn_npending_unlink),
                                atomic_read(&kgnilnd_data.kgn_npending_conns),
                                atomic_read(&kgnilnd_data.kgn_npending_detach));
        }

        return rc;
}

kgn_conn_t *
kgnilnd_get_conn_by_idx(int index)
{
        kgn_peer_t        *peer;
        struct list_head  *ptmp;
        kgn_conn_t        *conn;
        struct list_head  *ctmp;
        int                i;


        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
                read_lock(&kgnilnd_data.kgn_peer_conn_lock);
                list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {

                        peer = list_entry(ptmp, kgn_peer_t, gnp_list);

                        list_for_each(ctmp, &peer->gnp_conns) {
                                conn = list_entry(ctmp, kgn_conn_t, gnc_list);

                                if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
                                        continue;

                                if (index-- > 0)
                                        continue;

                                CDEBUG(D_NET, "++conn[%p] -> %s (%d)\n", conn,
                                       libcfs_nid2str(conn->gnc_peer->gnp_nid),
                                       atomic_read(&conn->gnc_refcount));
                                kgnilnd_conn_addref(conn);
                                read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                                return conn;
                        }
                }
                read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        }

        return NULL;
}

int
kgnilnd_get_conn_info(kgn_peer_t *peer,
                      int *device_id, __u64 *peerstamp,
                      int *tx_seq, int *rx_seq,
                      int *fmaq_len, int *nfma, int *nrdma)
{
        kgn_conn_t        *conn;
        int               rc = 0;

        read_lock(&kgnilnd_data.kgn_peer_conn_lock);

        conn = kgnilnd_find_conn_locked(peer);
        if (conn == NULL) {
                rc = -ENOENT;
                goto out;
        }

        *device_id = conn->gnc_device->gnd_host_id;
        *peerstamp = conn->gnc_peerstamp;
        *tx_seq = conn->gnc_tx_seq;
        *rx_seq = conn->gnc_rx_seq;
        *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
        *nfma = atomic_read(&conn->gnc_nlive_fma);
        *nrdma = atomic_read(&conn->gnc_nlive_rdma);
out:
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        return rc;
}

/* needs write_lock on kgn_peer_conn_lock */
int
kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
{
        kgn_conn_t         *conn;
        struct list_head   *ctmp, *cnxt;
        int                 count = 0;

        list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
                conn = list_entry(ctmp, kgn_conn_t, gnc_list);

                if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
                        continue;

                count++;
                /* we mark gnc_needs closing and increment kgn_npending_conns so that
                 * kgnilnd_del_conn_or_peer can wait on the other threads closing
                 * and cleaning up the connection.
                 */
                if (!conn->gnc_needs_closing) {
                        conn->gnc_needs_closing = 1;
                        kgnilnd_admin_addref(kgnilnd_data.kgn_npending_conns);
                }
                kgnilnd_close_conn_locked(conn, why);
        }
        return count;
}

int
kgnilnd_report_node_state(lnet_nid_t nid, int down)
{
        int         rc;
        kgn_peer_t  *peer, *new_peer;
        LIST_HEAD(zombies);

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        peer = kgnilnd_find_peer_locked(nid);

        if (peer == NULL) {
                int       i;
                int       found_net = 0;
                kgn_net_t *net;

                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

                /* Don't add a peer for node up events */
                if (down == GNILND_RCA_NODE_UP) {
                        return 0;
                }

                /* find any valid net - we don't care which one... */
                down_read(&kgnilnd_data.kgn_net_rw_sem);
                for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
                        list_for_each_entry(net, &kgnilnd_data.kgn_nets[i],
                                            gnn_list) {
                                found_net = 1;
                                break;
                        }

                        if (found_net) {
                                break;
                        }
                }
                up_read(&kgnilnd_data.kgn_net_rw_sem);

                if (!found_net) {
                        CNETERR("Could not find a net for nid %lld\n", nid);
                        return 1;
                }

                /* The nid passed in does not yet contain the net portion.
                 * Let's build it up now
                 */
                nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
                rc = kgnilnd_add_peer(net, nid, &new_peer);

                if (rc) {
                        CNETERR("Could not add peer for nid %lld, rc %d\n",
                                nid, rc);
                        return 1;
                }

                write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                peer = kgnilnd_find_peer_locked(nid);

                if (peer == NULL) {
                        CNETERR("Could not find peer for nid %lld\n", nid);
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        return 1;
                }
        }

        peer->gnp_down = down;

        if (down == GNILND_RCA_NODE_DOWN) {
                kgn_conn_t *conn;

                peer->gnp_down_event_time = jiffies;
                kgnilnd_cancel_peer_connect_locked(peer, &zombies);
                conn = kgnilnd_find_conn_locked(peer);

                if (conn != NULL) {
                        kgnilnd_close_conn_locked(conn, -ENETRESET);
                }
        } else {
                peer->gnp_up_event_time = jiffies;
        }

        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        if (down == GNILND_RCA_NODE_DOWN) {
                /* using ENETRESET so we don't get messages from
                 * kgnilnd_tx_done
                 */
                kgnilnd_txlist_done(&zombies, -ENETRESET);

                if (*kgnilnd_tunables.kgn_peer_health) {
                        kgnilnd_peer_notify(peer, -ECONNRESET);
                }
        }

        CDEBUG(D_INFO, "marking nid %lld %s\n", nid, down ? "down" : "up");
        return 0;
}

int
kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
{
        struct libcfs_ioctl_data *data = arg;
        kgn_net_t                *net = ni->ni_data;
        int                       rc = -EINVAL;

        LASSERT(ni == net->gnn_ni);

        switch (cmd) {
        case IOC_LIBCFS_GET_PEER: {
                lnet_nid_t   nid = 0;
                kgn_peer_t  *peer = NULL;
                __u32 nic_addr = 0;
                __u64 peerstamp = 0;
                int peer_refcount = 0, peer_connecting = 0;
                int device_id = 0;
                int tx_seq = 0, rx_seq = 0;
                int fmaq_len = 0, nfma = 0, nrdma = 0;

                rc = kgnilnd_get_peer_info(data->ioc_count, &peer,
                                           &nid, &nic_addr, &peer_refcount,
                                           &peer_connecting);
                if (rc)
                        break;

                /* Barf */
                /* LNET_MKNID is used to mask from lnet the multiplexing/demultiplexing of connections and peers
                 * LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
                 * wants to see instead of the underlying network that is being used to send the data
                 */
                data->ioc_nid    = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
                data->ioc_flags  = peer_connecting;
                data->ioc_count  = peer_refcount;

                rc = kgnilnd_get_conn_info(peer, &device_id, &peerstamp,
                                           &tx_seq, &rx_seq, &fmaq_len,
                                           &nfma, &nrdma);

                /* This is allowable - a persistent peer could not
                 * have a connection */
                if (rc) {
                        /* flag to indicate we are not connected -
                         * need to print as such */
                        data->ioc_flags |= (1<<16);
                        rc = 0;
                } else {
                        /* still barf */
                        data->ioc_net = device_id;
                        data->ioc_u64[0] = peerstamp;
                        data->ioc_u32[0] = fmaq_len;
                        data->ioc_u32[1] = nfma;
                        data->ioc_u32[2] = tx_seq;
                        data->ioc_u32[3] = rx_seq;
                        data->ioc_u32[4] = nrdma;
                }
                break;
        }
        case IOC_LIBCFS_ADD_PEER: {
                /* just dummy value to allow using common interface */
                kgn_peer_t      *peer;
                rc = kgnilnd_add_peer(net, data->ioc_nid, &peer);
                break;
        }
        case IOC_LIBCFS_DEL_PEER: {
                /* NULL is passed in so it affects all peers in existence without regard to network
                 * as the peer may not exist on the network LNET believes it to be on.
                 */
                rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
                                              GNILND_DEL_PEER, -EUCLEAN);
                break;
        }
        case IOC_LIBCFS_GET_CONN: {
                kgn_conn_t *conn = kgnilnd_get_conn_by_idx(data->ioc_count);

                if (conn == NULL)
                        rc = -ENOENT;
                else {
                        rc = 0;
                        /* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
                         * the generic connection that is used to send the data
                         */
                        data->ioc_nid    = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
                        data->ioc_u32[0] = conn->gnc_device->gnd_id;
                        kgnilnd_conn_decref(conn);
                }
                break;
        }
        case IOC_LIBCFS_CLOSE_CONNECTION: {
                /* use error = -ENETRESET to indicate it was lctl disconnect */
                /* NULL is passed in so it affects all the nets as the connection is virtual
                 * and may not exist on the network LNET believes it to be on.
                 */
                rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
                                              GNILND_DEL_CONN, -ENETRESET);
                break;
        }
        case IOC_LIBCFS_PUSH_CONNECTION: {
                /* we use this to flush purgatory */
                rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
                                              GNILND_CLEAR_PURGATORY, -EUCLEAN);
                break;
        }
        case IOC_LIBCFS_REGISTER_MYNID: {
                /* Ignore if this is a noop */
                if (data->ioc_nid == ni->ni_nid) {
                        rc = 0;
                } else {
                        CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
                               libcfs_nid2str(data->ioc_nid),
                               libcfs_nid2str(ni->ni_nid));
                        rc = -EINVAL;
                }
                break;
        }
        }

        return rc;
}

void
kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
{
        kgn_net_t               *net = ni->ni_data;
        kgn_tx_t                *tx;
        kgn_peer_t              *peer = NULL;
        kgn_conn_t              *conn = NULL;
        lnet_process_id_t       id = {
                .nid = nid,
                .pid = LNET_PID_LUSTRE,
        };
        ENTRY;

        /* I expect to find him, so only take a read lock */
        read_lock(&kgnilnd_data.kgn_peer_conn_lock);
        peer = kgnilnd_find_peer_locked(nid);
        if (peer != NULL) {
                /* LIE if in a quiesce - we will update the timeouts after,
                 * but we don't want sends failing during it */
                if (kgnilnd_data.kgn_quiesce_trigger) {
                        *when = jiffies;
                        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        GOTO(out, 0);
                }

                /* Update to best guess, might refine on later checks */
                *when = peer->gnp_last_alive;

                /* we have a peer, how about a conn? */
                conn = kgnilnd_find_conn_locked(peer);

                if (conn == NULL)  {
                        /* if there is no conn, check peer last errno to see if clean disconnect
                         * - if it was, we lie to LNet because we believe a TX would complete
                         * on reconnect */
                        if (kgnilnd_conn_clean_errno(peer->gnp_last_errno)) {
                                *when = jiffies;
                        }
                        /* we still want to fire a TX and new conn in this case */
                } else {
                        /* gnp_last_alive is valid, run for the hills */
                        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        GOTO(out, 0);
                }
        }
        /* if we get here, either we have no peer or no conn for him, so fire off
         * new TX to trigger conn setup */
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        /* if we couldn't find him, we'll fire up a TX and get connected -
         * if we don't do this, after ni_peer_timeout, LNet will declare him dead.
         * So really we treat kgnilnd_query as a bit of a 'connect now' type
         * event because it'll only do this when it wants to send
         *
         * Use a real TX for this to get the proper gnp_tx_queue behavior, etc
         * normally we'd use kgnilnd_send_ctlmsg for this, but we don't really
         * care that this goes out quickly since we already know we need a new conn
         * formed */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
                return;

        tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, ni->ni_nid);
        if (tx != NULL) {
                kgnilnd_launch_tx(tx, net, &id);
        }
out:
        CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lu\n", peer,
               libcfs_nid2str(nid), *when);
        EXIT;
}

int
kgnilnd_dev_init(kgn_device_t *dev)
{
        gni_return_t      rrc;
        int               rc = 0;
        unsigned int      cq_size;
        ENTRY;

        /* size of these CQs should be able to accommodate the outgoing
         * RDMA and SMSG transactions.  Since we really don't know what we
         * really need here, we'll take credits * 2 * 3 to allow a bunch.
         * We need to dig into this more with the performance work. */
        cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;

        rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
                                 GNILND_COOKIE, 0,
                                 &dev->gnd_domain);
        if (rrc != GNI_RC_SUCCESS) {
                CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
                GOTO(failed, rc = -ENODEV);
        }

        rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
                                 &dev->gnd_host_id, &dev->gnd_handle);
        if (rrc != GNI_RC_SUCCESS) {
                CERROR("Can't attach CDM to device %d (%d)\n",
                        dev->gnd_id, rrc);
                GOTO(failed, rc = -ENODEV);
        }

        /* a bit gross, but not much we can do - Aries Sim doesn't have
         * hardcoded NIC/NID that we can use */
        rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
        if (rc != 0)
                GOTO(failed, rc = -ENODEV);

        /* only dev 0 gets the errors - no need to reset the stack twice
         * - this works because we have a single PTAG, if we had more
         * then we'd need to have multiple handlers */
        if (dev->gnd_id == 0) {
                rrc = kgnilnd_subscribe_errors(dev->gnd_handle,
                                                GNI_ERRMASK_CRITICAL |
                                                GNI_ERRMASK_UNKNOWN_TRANSACTION,
                                              0, NULL, kgnilnd_critical_error,
                                              &dev->gnd_err_handle);
                if (rrc != GNI_RC_SUCCESS) {
                        CERROR("Can't subscribe for errors on device %d: rc %d\n",
                                dev->gnd_id, rrc);
                        GOTO(failed, rc = -ENODEV);
                }

                rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
                                                  kgnilnd_quiesce_end_callback);
                if (rc != GNI_RC_SUCCESS) {
                        CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
                                dev->gnd_id, rrc);
                        GOTO(failed, rc = -ENODEV);
                }
        }

        rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
        if (rrc < 0) {
                CERROR("sock_create returned %d\n", rrc);
                GOTO(failed, rrc);
        }

        rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
        if (rc < 0) {
                /* log messages during startup */
                if (kgnilnd_data.kgn_init < GNILND_INIT_ALL) {
                        CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
                                dev->gnd_host_id, rc);
                }
                GOTO(failed, rc = -ESRCH);
        }
        CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);

        rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
                                0, kgnilnd_device_callback,
                                dev->gnd_id, &dev->gnd_snd_rdma_cqh);
        if (rrc != GNI_RC_SUCCESS) {
                CERROR("Can't create rdma send cq size %u for device "
                       "%d (%d)\n", cq_size, dev->gnd_id, rrc);
                GOTO(failed, rc = -EINVAL);
        }

        rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
                        0, kgnilnd_device_callback, dev->gnd_id,
                        &dev->gnd_snd_fma_cqh);
        if (rrc != GNI_RC_SUCCESS) {
                CERROR("Can't create fma send cq size %u for device %d (%d)\n",
                       cq_size, dev->gnd_id, rrc);
                GOTO(failed, rc = -EINVAL);
        }

        /* This one we size differently - overflows are possible and it needs to be
         * sized based on machine size */
        rrc = kgnilnd_cq_create(dev->gnd_handle,
                        *kgnilnd_tunables.kgn_fma_cq_size,
                        0, kgnilnd_device_callback, dev->gnd_id,
                        &dev->gnd_rcv_fma_cqh);
        if (rrc != GNI_RC_SUCCESS) {
                CERROR("Can't create fma cq size %d for device %d (%d)\n",
                       *kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
                GOTO(failed, rc = -EINVAL);
        }

        RETURN(0);

failed:
        kgnilnd_dev_fini(dev);
        RETURN(rc);
}

void
kgnilnd_dev_fini(kgn_device_t *dev)
{
        gni_return_t rrc;
        ENTRY;

        /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
        LASSERTF(list_empty(&dev->gnd_ready_conns) &&
                 list_empty(&dev->gnd_map_tx) &&
                 list_empty(&dev->gnd_rdmaq),
                 "dev 0x%p ready_conns %d@0x%p map_tx %d@0x%p rdmaq %d@0x%p\n",
                 dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
                 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
                 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);

        /* These should follow from tearing down all connections */
        LASSERTF(dev->gnd_map_nphys == 0 && dev->gnd_map_physnop == 0,
                "%d physical mappings of %d pages still mapped\n",
                 dev->gnd_map_nphys, dev->gnd_map_physnop);

        LASSERTF(dev->gnd_map_nvirt == 0 && dev->gnd_map_virtnob == 0,
                "%d virtual mappings of "LPU64" bytes still mapped\n",
                 dev->gnd_map_nvirt, dev->gnd_map_virtnob);

        LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
                 atomic_read(&dev->gnd_n_mdd_held) == 0 &&
                 atomic64_read(&dev->gnd_nbytes_map) == 0,
                "%d SMSG mappings of %ld bytes still mapped or held %d\n",
                 atomic_read(&dev->gnd_n_mdd),
                 atomic64_read(&dev->gnd_nbytes_map), atomic_read(&dev->gnd_n_mdd_held));

        LASSERT(list_empty(&dev->gnd_map_list));

        /* What other assertions needed to ensure all connections torn down ? */

        /* check all counters == 0 (EP, MDD, etc) */

        /* if we are resetting due to quiese (stack reset), don't check
         * thread states */
        LASSERTF(kgnilnd_data.kgn_quiesce_trigger ||
                atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
                "tried to shutdown with threads active\n");

        if (dev->gnd_rcv_fma_cqh) {
                rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
                LASSERTF(rrc == GNI_RC_SUCCESS,
                        "bad rc from gni_cq_destroy on rcv_fma_cqh: %d\n", rrc);
                dev->gnd_rcv_fma_cqh = NULL;
        }

        if (dev->gnd_snd_rdma_cqh) {
                rrc = kgnilnd_cq_destroy(dev->gnd_snd_rdma_cqh);
                LASSERTF(rrc == GNI_RC_SUCCESS,
                        "bad rc from gni_cq_destroy on send_rdma_cqh: %d\n", rrc);
                dev->gnd_snd_rdma_cqh = NULL;
        }

        if (dev->gnd_snd_fma_cqh) {
                rrc = kgnilnd_cq_destroy(dev->gnd_snd_fma_cqh);
                LASSERTF(rrc == GNI_RC_SUCCESS,
                        "bad rc from gni_cq_destroy on snd_fma_cqh: %d\n", rrc);
                dev->gnd_snd_fma_cqh = NULL;
        }

        if (dev->gnd_err_handle) {
                rrc = kgnilnd_release_errors(dev->gnd_err_handle);
                LASSERTF(rrc == GNI_RC_SUCCESS,
                        "bad rc from gni_release_errors: %d\n", rrc);
                dev->gnd_err_handle = NULL;
        }

        if (dev->gnd_domain) {
                rrc = kgnilnd_cdm_destroy(dev->gnd_domain);
                LASSERTF(rrc == GNI_RC_SUCCESS,
                        "bad rc from gni_cdm_destroy: %d\n", rrc);
                dev->gnd_domain = NULL;
        }

        sock_release(kgnilnd_data.kgn_sock);

        EXIT;
}


int kgnilnd_base_startup(void)
{
        struct timeval       tv;
        int                  pkmem = atomic_read(&libcfs_kmemory);
        int                  rc;
        int                  i;
        kgn_device_t        *dev;
        struct task_struct  *thrd;
        ENTRY;

        LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
                "init %d\n", kgnilnd_data.kgn_init);

        /* zero pointers, flags etc */
        memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));

        /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
         * a unique (for all time) connstamp so we can uniquely identify
         * the sender.  The connstamp is an incrementing counter
         * initialised with seconds + microseconds at startup time.  So we
         * rely on NOT creating connections more frequently on average than
         * 1MHz to ensure we don't use old connstamps when we reboot. */
        do_gettimeofday(&tv);
        kgnilnd_data.kgn_connstamp =
                 kgnilnd_data.kgn_peerstamp =
                        (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;

        init_rwsem(&kgnilnd_data.kgn_net_rw_sem);

        for (i = 0; i < GNILND_MAXDEVS; i++) {
                kgn_device_t  *dev = &kgnilnd_data.kgn_devices[i];

                dev->gnd_id = i;
                INIT_LIST_HEAD(&dev->gnd_ready_conns);
                INIT_LIST_HEAD(&dev->gnd_map_tx);
                INIT_LIST_HEAD(&dev->gnd_fma_buffs);
                mutex_init(&dev->gnd_cq_mutex);
                mutex_init(&dev->gnd_fmablk_mutex);
                spin_lock_init(&dev->gnd_fmablk_lock);
                init_waitqueue_head(&dev->gnd_waitq);
                init_waitqueue_head(&dev->gnd_dgram_waitq);
                init_waitqueue_head(&dev->gnd_dgping_waitq);
                spin_lock_init(&dev->gnd_lock);
                INIT_LIST_HEAD(&dev->gnd_map_list);
                spin_lock_init(&dev->gnd_map_lock);
                atomic_set(&dev->gnd_nfmablk, 0);
                atomic_set(&dev->gnd_fmablk_vers, 1);
                atomic_set(&dev->gnd_neps, 0);
                atomic_set(&dev->gnd_canceled_dgrams, 0);
                INIT_LIST_HEAD(&dev->gnd_connd_peers);
                spin_lock_init(&dev->gnd_connd_lock);
                spin_lock_init(&dev->gnd_dgram_lock);
                spin_lock_init(&dev->gnd_rdmaq_lock);
                INIT_LIST_HEAD(&dev->gnd_rdmaq);
                init_rwsem(&dev->gnd_conn_sem);

                /* alloc & setup nid based dgram table */
                LIBCFS_ALLOC(dev->gnd_dgrams,
                            sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);

                if (dev->gnd_dgrams == NULL)
                        GOTO(failed, rc = -ENOMEM);

                for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
                        INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
                }
                atomic_set(&dev->gnd_ndgrams, 0);
                atomic_set(&dev->gnd_nwcdgrams, 0);
                /* setup timer for RDMAQ processing */
                setup_timer(&dev->gnd_rdmaq_timer, kgnilnd_schedule_device_timer,
                            (unsigned long)dev);

                /* setup timer for mapping processing */
                setup_timer(&dev->gnd_map_timer, kgnilnd_schedule_device_timer,
                            (unsigned long)dev);

        }

        /* CQID 0 isn't allowed, set to MAX_MSG_ID - 1 to check for conflicts early */
        kgnilnd_data.kgn_next_cqid = GNILND_MAX_MSG_ID - 1;
        kgnilnd_data.kgn_new_min_timeout = *kgnilnd_tunables.kgn_timeout;
        init_waitqueue_head(&kgnilnd_data.kgn_reaper_waitq);
        init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
        spin_lock_init(&kgnilnd_data.kgn_reaper_lock);

        mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
        atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
        atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
        atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
        atomic_set(&kgnilnd_data.kgn_npending_detach, 0);
        atomic_set(&kgnilnd_data.kgn_rev_offset, 0);
        atomic_set(&kgnilnd_data.kgn_rev_length, 0);
        atomic_set(&kgnilnd_data.kgn_rev_copy_buff, 0);

        /* OK to call kgnilnd_api_shutdown() to cleanup now */
        kgnilnd_data.kgn_init = GNILND_INIT_DATA;
        try_module_get(THIS_MODULE);

        rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);

        LIBCFS_ALLOC(kgnilnd_data.kgn_peers,
                    sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);

        if (kgnilnd_data.kgn_peers == NULL)
                GOTO(failed, rc = -ENOMEM);

        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
                INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
        }

        LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
                    sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);

        if (kgnilnd_data.kgn_conns == NULL)
                GOTO(failed, rc = -ENOMEM);

        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
                INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
        }

        LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
                    sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);

        if (kgnilnd_data.kgn_nets == NULL)
                GOTO(failed, rc = -ENOMEM);

        for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
                INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
        }

        kgnilnd_data.kgn_mbox_cache =
                kmem_cache_create("kgn_mbox_block", KMALLOC_MAX_SIZE, 0,
                                  SLAB_HWCACHE_ALIGN, NULL);
        if (kgnilnd_data.kgn_mbox_cache == NULL) {
                CERROR("Can't create slab for physical mbox blocks\n");
                GOTO(failed, rc = -ENOMEM);
        }

        kgnilnd_data.kgn_rx_cache =
                kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
        if (kgnilnd_data.kgn_rx_cache == NULL) {
                CERROR("Can't create slab for kgn_rx_t descriptors\n");
                GOTO(failed, rc = -ENOMEM);
        }

        kgnilnd_data.kgn_tx_cache =
                kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
        if (kgnilnd_data.kgn_tx_cache == NULL) {
                CERROR("Can't create slab for kgn_tx_t\n");
                GOTO(failed, rc = -ENOMEM);
        }

        kgnilnd_data.kgn_tx_phys_cache =
                kmem_cache_create("kgn_tx_phys",
                                   LNET_MAX_IOV * sizeof(gni_mem_segment_t),
                                   0, 0, NULL);
        if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
                CERROR("Can't create slab for kgn_tx_phys\n");
                GOTO(failed, rc = -ENOMEM);
        }

        kgnilnd_data.kgn_dgram_cache =
                kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
        if (kgnilnd_data.kgn_dgram_cache == NULL) {
                CERROR("Can't create slab for outgoing datagrams\n");
                GOTO(failed, rc = -ENOMEM);
        }

        /* allocate a MAX_IOV array of page pointers for each cpu */
        kgnilnd_data.kgn_cksum_map_pages = kmalloc(num_possible_cpus() * sizeof (struct page *),
                                                   GFP_KERNEL);
        if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
                CERROR("Can't allocate vmap cksum pages\n");
                GOTO(failed, rc = -ENOMEM);
        }
        kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
        memset(kgnilnd_data.kgn_cksum_map_pages, 0,
                kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));

        for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
                kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
                                                              GFP_KERNEL);
                if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
                        CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
                        GOTO(failed, rc = -ENOMEM);
                }
        }

        LASSERT(kgnilnd_data.kgn_ndevs == 0);

        /* Use all available GNI devices */
        for (i = 0; i < GNILND_MAXDEVS; i++) {
                dev = &kgnilnd_data.kgn_devices[kgnilnd_data.kgn_ndevs];

                rc = kgnilnd_dev_init(dev);
                if (rc == 0) {
                        /* Increment here so base_shutdown cleans it up */
                        kgnilnd_data.kgn_ndevs++;

                        rc = kgnilnd_allocate_phys_fmablk(dev);
                        if (rc)
                                GOTO(failed, rc);
                }
        }

        if (kgnilnd_data.kgn_ndevs == 0) {
                CERROR("Can't initialise any GNI devices\n");
                GOTO(failed, rc = -ENODEV);
        }

        rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
        if (rc != 0) {
                CERROR("Can't spawn gnilnd reaper: %d\n", rc);
                GOTO(failed, rc);
        }

        rc = kgnilnd_start_rca_thread();
        if (rc != 0) {
                CERROR("Can't spawn gnilnd rca: %d\n", rc);
                GOTO(failed, rc);
        }

        /*
         * Start ruhroh thread.  We can't use kgnilnd_thread_start() because
         * we don't want this thread included in kgnilnd_data.kgn_nthreads
         * count.  This thread controls quiesce, so it mustn't
         * quiesce itself.
         */
        thrd = kthread_run(kgnilnd_ruhroh_thread, NULL, "%s_%02d", "kgnilnd_rr", 0);
        if (IS_ERR(thrd)) {
                rc = PTR_ERR(thrd);
                CERROR("Can't spawn gnilnd ruhroh thread: %d\n", rc);
                GOTO(failed, rc);
        }

        /* threads will load balance across devs as they are available */
        for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
                rc = kgnilnd_thread_start(kgnilnd_scheduler, (void *)((long)i),
                                          "kgnilnd_sd", i);
                if (rc != 0) {
                        CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
                               i, rc);
                        GOTO(failed, rc);
                }
        }

        for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
                dev = &kgnilnd_data.kgn_devices[i];
                rc = kgnilnd_thread_start(kgnilnd_dgram_mover, dev,
                                          "kgnilnd_dg", dev->gnd_id);
                if (rc != 0) {
                        CERROR("Can't spawn gnilnd dgram_mover[%d]: %d\n",
                               dev->gnd_id, rc);
                        GOTO(failed, rc);
                }

                rc = kgnilnd_thread_start(kgnilnd_dgram_waitq, dev,
                                          "kgnilnd_dgn", dev->gnd_id);
                if (rc != 0) {
                        CERROR("Can't spawn gnilnd dgram_waitq[%d]: %d\n",
                                dev->gnd_id, rc);
                        GOTO(failed, rc);
                }

                rc = kgnilnd_setup_wildcard_dgram(dev);

                if (rc != 0) {
                        CERROR("Can't create wildcard dgrams[%d]: %d\n",
                                dev->gnd_id, rc);
                        GOTO(failed, rc);
                }
        }



        /* flag everything initialised */
        kgnilnd_data.kgn_init = GNILND_INIT_ALL;
        /*****************************************************/

        CDEBUG(D_MALLOC, "initial kmem %d\n", pkmem);
        RETURN(0);

failed:
        kgnilnd_base_shutdown();
        kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
        RETURN(rc);
}

void
kgnilnd_base_shutdown(void)
{
        int                     i, j;
        ENTRY;

        while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};

        kgnilnd_data.kgn_wc_kill = 1;

        for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
                kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
                kgnilnd_cancel_wc_dgrams(dev);
                kgnilnd_cancel_dgrams(dev);
                kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
                kgnilnd_wait_for_canceled_dgrams(dev);
        }

        /* We need to verify there are no conns left before we let the threads
         * shut down otherwise we could clean up the peers but still have
         * some outstanding conns due to orphaned datagram conns that are
         * being cleaned up.
         */
        i = 2;
        while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
                i++;

                for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
                        kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
                        kgnilnd_schedule_device(dev);
                }

                CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
                        "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
                cfs_pause(cfs_time_seconds(1));
        }
        /* Peer state all cleaned up BEFORE setting shutdown, so threads don't
         * have to worry about shutdown races.  NB connections may be created
         * while there are still active connds, but these will be temporary
         * since peer creation always fails after the listener has started to
         * shut down.
         * all peers should have been cleared out on the nets */
        LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
                "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));

        /* Wait for the ruhroh thread to shut down. */
        kgnilnd_data.kgn_ruhroh_shutdown = 1;
        wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
        i = 2;
        while (kgnilnd_data.kgn_ruhroh_running != 0) {
                i++;
                CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
                       "Waiting for ruhroh thread to terminate\n");
                cfs_pause(cfs_time_seconds(1));
        }

       /* Flag threads to terminate */
        kgnilnd_data.kgn_shutdown = 1;

        for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
                kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];

                /* should clear all the MDDs */
                kgnilnd_unmap_fma_blocks(dev);

                kgnilnd_schedule_device(dev);
                wake_up_all(&dev->gnd_dgram_waitq);
                wake_up_all(&dev->gnd_dgping_waitq);
                LASSERT(list_empty(&dev->gnd_connd_peers));
        }

        spin_lock(&kgnilnd_data.kgn_reaper_lock);
        wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
        spin_unlock(&kgnilnd_data.kgn_reaper_lock);

        kgnilnd_wakeup_rca_thread();

        /* Wait for threads to exit */
        i = 2;
        while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
                i++;
                CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
                       "Waiting for %d threads to terminate\n",
                       atomic_read(&kgnilnd_data.kgn_nthreads));
                cfs_pause(cfs_time_seconds(1));
        }

        LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
                "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));

        if (kgnilnd_data.kgn_peers != NULL) {
                for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
                        LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));

                LIBCFS_FREE(kgnilnd_data.kgn_peers,
                            sizeof (struct list_head) *
                            *kgnilnd_tunables.kgn_peer_hash_size);
        }

        down_write(&kgnilnd_data.kgn_net_rw_sem);
        if (kgnilnd_data.kgn_nets != NULL) {
                for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
                        LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));

                LIBCFS_FREE(kgnilnd_data.kgn_nets,
                            sizeof (struct list_head) *
                            *kgnilnd_tunables.kgn_net_hash_size);
        }
        up_write(&kgnilnd_data.kgn_net_rw_sem);

        LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
                "conns left %d\n", atomic_read(&kgnilnd_data.kgn_nconns));

        if (kgnilnd_data.kgn_conns != NULL) {
                for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
                        LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));

                LIBCFS_FREE(kgnilnd_data.kgn_conns,
                            sizeof (struct list_head) *
                            *kgnilnd_tunables.kgn_peer_hash_size);
        }

        for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
                kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
                kgnilnd_dev_fini(dev);

                LASSERTF(atomic_read(&dev->gnd_ndgrams) == 0,
                        "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));

                if (dev->gnd_dgrams != NULL) {
                        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
                                LASSERT(list_empty(&dev->gnd_dgrams[i]));

                        LIBCFS_FREE(dev->gnd_dgrams,
                                    sizeof (struct list_head) *
                                    *kgnilnd_tunables.kgn_peer_hash_size);
                }

                kgnilnd_free_phys_fmablk(dev);
        }

        if (kgnilnd_data.kgn_mbox_cache != NULL)
                kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);

        if (kgnilnd_data.kgn_rx_cache != NULL)
                kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);

        if (kgnilnd_data.kgn_tx_cache != NULL)
                kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);

        if (kgnilnd_data.kgn_tx_phys_cache != NULL)
                kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);

        if (kgnilnd_data.kgn_dgram_cache != NULL)
                kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);

        if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
                for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
                        if (kgnilnd_data.kgn_cksum_map_pages[i] != NULL) {
                                kfree(kgnilnd_data.kgn_cksum_map_pages[i]);
                        }
                }
                kfree(kgnilnd_data.kgn_cksum_map_pages);
        }

        CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
               atomic_read(&libcfs_kmemory));

        kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
        module_put(THIS_MODULE);

        EXIT;
}

int
kgnilnd_startup(lnet_ni_t *ni)
{
        int               rc, devno;
        kgn_net_t        *net;
        ENTRY;

        LASSERTF(ni->ni_lnd == &the_kgnilnd,
                "bad LND 0x%p != the_kgnilnd @ 0x%p\n",
                ni->ni_lnd, &the_kgnilnd);

        if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
                rc = kgnilnd_base_startup();
                if (rc != 0)
                        RETURN(rc);
        }

        /* Serialize with shutdown. */
        mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);

        LIBCFS_ALLOC(net, sizeof(*net));
        if (net == NULL) {
                CERROR("could not allocate net for new interface instance\n");
                /* no need to cleanup the CDM... */
                GOTO(failed, rc = -ENOMEM);
        }
        INIT_LIST_HEAD(&net->gnn_list);
        ni->ni_data = net;
        net->gnn_ni = ni;
        ni->ni_maxtxcredits = *kgnilnd_tunables.kgn_credits;
        ni->ni_peertxcredits = *kgnilnd_tunables.kgn_peer_credits;

        if (*kgnilnd_tunables.kgn_peer_health) {
                int     fudge;
                int     timeout;
                /* give this a bit of leeway - we don't have a hard timeout
                 * as we only check timeouts periodically - see comment in kgnilnd_reaper */
                fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
                timeout = *kgnilnd_tunables.kgn_timeout + fudge;

                if (*kgnilnd_tunables.kgn_peer_timeout >= timeout)
                        ni->ni_peertimeout = *kgnilnd_tunables.kgn_peer_timeout;
                else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
                        LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
                                        *kgnilnd_tunables.kgn_peer_timeout,
                                        timeout);
                        ni->ni_data = NULL;
                        LIBCFS_FREE(net, sizeof(*net));
                        GOTO(failed, rc = -EINVAL);
                } else
                        ni->ni_peertimeout = timeout;

                LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
                              ni->ni_peertimeout);
        }

        atomic_set(&net->gnn_refcount, 1);

        /* if we have multiple devices, spread the nets around */
        net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));

        devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
        net->gnn_dev = &kgnilnd_data.kgn_devices[devno];

        /* allocate a 'dummy' cdm for datagram use. We can only have a single
         * datagram between a nid:inst_id and nid2:inst_id. The fake cdm
         * give us additional inst_id to use, allowing the datagrams to flow
         * like rivers of honey and beer */

        /* the instance id for the cdm is the NETNUM offset by MAXDEVS -
         * ensuring we'll have a unique id */


        ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
        CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
                net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
        /* until the gnn_list is set, we need to cleanup ourselves as
         * kgnilnd_shutdown is just gonna get confused */

        down_write(&kgnilnd_data.kgn_net_rw_sem);
        list_add_tail(&net->gnn_list, kgnilnd_netnum2netlist(net->gnn_netnum));
        up_write(&kgnilnd_data.kgn_net_rw_sem);

        /* we need a separate thread to call probe_wait_by_id until
         * we get a function callback notifier from kgni */
        mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
        RETURN(0);
 failed:
        mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
        kgnilnd_shutdown(ni);
        RETURN(rc);
}

void
kgnilnd_shutdown(lnet_ni_t *ni)
{
        kgn_net_t     *net = ni->ni_data;
        int           i;
        int           rc;
        ENTRY;

        CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);

        LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_ALL,
                "init %d\n", kgnilnd_data.kgn_init);

        /* Serialize with startup. */
        mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
        CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
               atomic_read(&libcfs_kmemory));

        if (net == NULL) {
                CERROR("got NULL net for ni %p\n", ni);
                GOTO(out, rc = -EINVAL);
        }

        LASSERTF(ni == net->gnn_ni,
                "ni %p gnn_ni %p\n", net, net->gnn_ni);

        ni->ni_data = NULL;

        LASSERT(!net->gnn_shutdown);
        LASSERTF(atomic_read(&net->gnn_refcount) != 0,
                "net %p refcount %d\n",
                 net, atomic_read(&net->gnn_refcount));

        if (!list_empty(&net->gnn_list)) {
                /* serialize with peer creation */
                down_write(&kgnilnd_data.kgn_net_rw_sem);
                net->gnn_shutdown = 1;
                up_write(&kgnilnd_data.kgn_net_rw_sem);

                kgnilnd_cancel_net_dgrams(net);

                kgnilnd_del_conn_or_peer(net, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);

                /* if we are quiesced, need to wake up - we need those threads
                 * alive to release peers, etc */
                if (GNILND_IS_QUIESCED) {
                        set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
                        kgnilnd_quiesce_wait("shutdown");
                }

                kgnilnd_wait_for_canceled_dgrams(net->gnn_dev);

                /* We wait until the nets ref's are 1, we will release final ref which is ours
                 * this allows us to make sure everything else is done before we free the
                 * net.
                 */
                i = 4;
                while (atomic_read(&net->gnn_refcount) != 1) {
                        i++;
                        CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
                                "Waiting for %d references to clear on net %d\n",
                                atomic_read(&net->gnn_refcount),
                                net->gnn_netnum);
                        cfs_pause(cfs_time_seconds(1));
                }

                /* release ref from kgnilnd_startup */
                kgnilnd_net_decref(net);
                /* serialize with reaper and conn_task looping */
                down_write(&kgnilnd_data.kgn_net_rw_sem);
                list_del_init(&net->gnn_list);
                up_write(&kgnilnd_data.kgn_net_rw_sem);

        }

        /* not locking, this can't race with writers */
        LASSERTF(atomic_read(&net->gnn_refcount) == 0,
                "net %p refcount %d\n",
                 net, atomic_read(&net->gnn_refcount));
        LIBCFS_FREE(net, sizeof(*net));

out:
        down_read(&kgnilnd_data.kgn_net_rw_sem);
        for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
                if (!list_empty(&kgnilnd_data.kgn_nets[i])) {
                        up_read(&kgnilnd_data.kgn_net_rw_sem);
                        break;
                }

                if (i == *kgnilnd_tunables.kgn_net_hash_size - 1) {
                        up_read(&kgnilnd_data.kgn_net_rw_sem);
                        kgnilnd_base_shutdown();
                }
        }
        CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
               atomic_read(&libcfs_kmemory));

        mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
        EXIT;
}

void __exit
kgnilnd_module_fini(void)
{
        lnet_unregister_lnd(&the_kgnilnd);
        kgnilnd_proc_fini();
        kgnilnd_remove_sysctl();
        kgnilnd_tunables_fini();
}

int __init
kgnilnd_module_init(void)
{
        int    rc;

        rc = kgnilnd_tunables_init();
        if (rc != 0)
                return rc;

        printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n");

        kgnilnd_insert_sysctl();
        kgnilnd_proc_init();

        lnet_register_lnd(&the_kgnilnd);

        return 0;
}

MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
MODULE_DESCRIPTION("Kernel Gemini LND v"KGNILND_BUILD_REV);
MODULE_LICENSE("GPL");

module_init(kgnilnd_module_init);
module_exit(kgnilnd_module_fini);