X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd.c;h=a197c99868c7196692babcba5adcfd7b5d8adb67;hp=fcc05fa0e33048e54e36f806a1bd3e230140b72c;hb=4f526f5c20073d7025154d97885ec2fdfe832e1c;hpb=4d381ef9f179b21217c237ad1cc83055a2448550

diff --git a/lnet/klnds/gnilnd/gnilnd.c b/lnet/klnds/gnilnd/gnilnd.c
index fcc05fa..a197c99 100644
--- a/lnet/klnds/gnilnd/gnilnd.c
+++ b/lnet/klnds/gnilnd/gnilnd.c
@@ -1,7 +1,8 @@
 /*
  * Copyright (C) 2012 Cray, Inc.
  *
- *   Author: Igor Gorodetsky <iogordet@cray.com>
+ * Copyright (c) 2013, 2017, Intel Corporation.
+ *
  *   Author: Nic Henke <nic@cray.com>
  *   Author: James Shimek <jshimek@cray.com>
  *
@@ -24,7 +25,7 @@
 #include "gnilnd.h"
 
 /* Primary entry points from LNET.  There are no guarantees against reentrance. */
-lnd_t the_kgnilnd = {
+const struct lnet_lnd the_kgnilnd = {
 	.lnd_type       = GNILND,
 	.lnd_startup    = kgnilnd_startup,
 	.lnd_shutdown   = kgnilnd_shutdown,
@@ -36,7 +37,52 @@ lnd_t the_kgnilnd = {
 };
 
 kgn_data_t      kgnilnd_data;
-kgn_hssops_t	kgnilnd_hssops;
+
+int
+kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
+{
+	struct task_struct *thrd;
+
+	thrd = kthread_run(fn, arg, "%s_%02d", name, id);
+	if (IS_ERR(thrd))
+		return PTR_ERR(thrd);
+
+	atomic_inc(&kgnilnd_data.kgn_nthreads);
+	return 0;
+}
+
+/* bind scheduler threads to cpus */
+int
+kgnilnd_start_sd_threads(void)
+{
+	int cpu;
+	int i = 0;
+	struct task_struct *task;
+
+	for_each_online_cpu(cpu) {
+		/* don't bind to cpu 0 - all interrupts are processed here */
+		if (cpu == 0)
+			continue;
+
+		task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
+				      "%s_%02d", "kgnilnd_sd", i);
+		if (!IS_ERR(task)) {
+			kthread_bind(task, cpu);
+			wake_up_process(task);
+		} else {
+			CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
+				PTR_ERR(task));
+			return PTR_ERR(task);
+		}
+		atomic_inc(&kgnilnd_data.kgn_nthreads);
+
+		if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
+			break;
+		}
+	}
+
+	return 0;
+}
 
 /* needs write_lock on kgn_peer_conn_lock */
 int
@@ -67,8 +113,8 @@ kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
 		    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",
+				" connstamp %llu(%llu)"
+				" peerstamp %llu(%llu)\n",
 				conn, newconn->gnc_my_connstamp,
 				conn->gnc_peer_connstamp,
 				newconn->gnc_peer_connstamp,
@@ -78,25 +124,25 @@ kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
 
 		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 0x%p peerstamp %llu >= "
+				"newconn 0x%p peerstamp %llu\n",
 				conn, conn->gnc_peerstamp,
 				newconn, newconn->gnc_peerstamp);
 
 			CDEBUG(D_NET, "Closing stale conn nid: %s "
-			       " peerstamp:"LPX64"("LPX64")\n",
+			       " peerstamp:%#llx(%#llx)\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 0x%p peer_connstamp %llu >= "
+				"newconn 0x%p peer_connstamp %llu\n",
 				conn, conn->gnc_peer_connstamp,
 				newconn, newconn->gnc_peer_connstamp);
 
 			CDEBUG(D_NET, "Closing stale conn nid: %s"
-			       " connstamp:"LPU64"("LPU64")\n",
+			       " connstamp:%llu(%llu)\n",
 			       libcfs_nid2str(peer->gnp_nid),
 			       conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
 		}
@@ -125,8 +171,8 @@ kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
 	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
+			" lo %d new %llu existing %llu"
+			" new peer %llu existing peer %llu"
 			" new dev %p existing dev %p\n",
 			conn, libcfs_nid2str(peer->gnp_nid),
 			loopback,
@@ -177,9 +223,9 @@ kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
 int
 kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 {
-	kgn_conn_t    *conn;
-	gni_return_t   rrc;
-	int            rc = 0;
+	kgn_conn_t	*conn;
+	gni_return_t	rrc;
+	int		rc = 0;
 
 	LASSERT (!in_interrupt());
 	atomic_inc(&kgnilnd_data.kgn_nconns);
@@ -198,30 +244,35 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 		return -ENOMEM;
 	}
 
-	LIBCFS_ALLOC(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
+	conn->gnc_tx_ref_table =
+		kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
 	if (conn->gnc_tx_ref_table == NULL) {
 		CERROR("Can't allocate conn tx_ref_table\n");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
+	mutex_init(&conn->gnc_smsg_mutex);
+	mutex_init(&conn->gnc_rdma_mutex);
 	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);
+	INIT_LIST_HEAD(&conn->gnc_delaylist);
 	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);
+	BUILD_BUG_ON(GNILND_MAX_MSG_ID >= GNILND_MSGID_CLOSE);
 
 	/* get a new unique CQ id for this conn */
 	write_lock(&kgnilnd_data.kgn_peer_conn_lock);
@@ -231,8 +282,7 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 
 	if (conn->gnc_cqid == 0) {
 		CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
-		rc = -E2BIG;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -E2BIG);
 	}
 
 	CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
@@ -251,10 +301,8 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 	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) {
-		rc = -ENETDOWN;
-		GOTO(failed, rc);
-	}
+	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);
@@ -268,7 +316,8 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
 
 failed:
 	atomic_dec(&kgnilnd_data.kgn_nconns);
-	LIBCFS_FREE(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
+	kgnilnd_vfree(conn->gnc_tx_ref_table,
+		      GNILND_MAX_MSG_ID * sizeof(void *));
 	LIBCFS_FREE(conn, sizeof(*conn));
 	return rc;
 }
@@ -278,7 +327,6 @@ kgn_conn_t *
 kgnilnd_find_conn_locked(kgn_peer_t *peer)
 {
 	kgn_conn_t      *conn = NULL;
-	ENTRY;
 
 	/* if we are in reset, this conn is going to die soon */
 	if (unlikely(kgnilnd_data.kgn_in_reset)) {
@@ -399,17 +447,21 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
 		list_empty(&conn->gnc_list) &&
 		list_empty(&conn->gnc_hashlist) &&
 		list_empty(&conn->gnc_schedlist) &&
-		list_empty(&conn->gnc_mdd_list),
-		"conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p lists %d/%d/%d/%d\n",
+		list_empty(&conn->gnc_mdd_list) &&
+		list_empty(&conn->gnc_delaylist) &&
+		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/%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));
+		list_empty(&conn->gnc_mdd_list),
+		list_empty(&conn->gnc_delaylist));
 
 	/* Tripping these is especially bad, as it means we have items on the
 	 *  lists that didn't keep their refcount on the connection - or
@@ -424,8 +476,16 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
 	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);
 	}
 
@@ -433,8 +493,8 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
 		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 *));
+		kgnilnd_vfree(conn->gnc_tx_ref_table,
+			      GNILND_MAX_MSG_ID * sizeof(void *));
 	}
 
 	LIBCFS_FREE(conn, sizeof(*conn));
@@ -445,11 +505,13 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
 void
 kgnilnd_peer_alive(kgn_peer_t *peer)
 {
-	set_mb(peer->gnp_last_alive, jiffies);
+	time64_t now = ktime_get_seconds();
+
+	set_mb(peer->gnp_last_alive, now);
 }
 
 void
-kgnilnd_peer_notify(kgn_peer_t *peer, int error)
+kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive)
 {
 	int                     tell_lnet = 0;
 	int                     nnets = 0;
@@ -478,10 +540,10 @@ kgnilnd_peer_notify(kgn_peer_t *peer, int error)
 	       peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
 	       kgnilnd_data.kgn_in_reset, error);
 
-	if ((peer->gnp_connecting == GNILND_PEER_IDLE) &&
+	if (((peer->gnp_connecting == GNILND_PEER_IDLE) &&
 	    (conn == NULL) &&
 	    (!kgnilnd_data.kgn_in_reset) &&
-	    (!kgnilnd_conn_clean_errno(error))) {
+	    (!kgnilnd_conn_clean_errno(error))) || alive) {
 		tell_lnet = 1;
 	}
 
@@ -541,12 +603,13 @@ kgnilnd_peer_notify(kgn_peer_t *peer, int error)
 			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",
+			CDEBUG(D_NET, "peer 0x%p->%s last_alive %lld (%llds 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);
+				ktime_get_seconds() - peer->gnp_last_alive);
 
+			lnet_notify(net->gnn_ni, peer_nid, alive,
+				    (alive) ? true : false,
+				    peer->gnp_last_alive);
 
 			kgnilnd_net_decref(net);
 		}
@@ -574,7 +637,8 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
 	}
 
 	/* if we NETERROR, make sure it is rate limited */
-	if (!kgnilnd_conn_clean_errno(error)) {
+	if (!kgnilnd_conn_clean_errno(error) &&
+	    peer->gnp_state != GNILND_PEER_DOWN) {
 		CNETERR("closing conn to %s: error %d\n",
 		       libcfs_nid2str(peer->gnp_nid), error);
 	} else {
@@ -600,6 +664,7 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
 	/* 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 */
@@ -609,6 +674,10 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
 		conn->gnc_state = GNILND_CONN_CLOSING;
 	}
 
+	if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
+		msleep_interruptible(MSEC_PER_SEC);
+	}
+
 	/* 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 */
@@ -628,11 +697,6 @@ kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
 	 * gnd_ready_conns and allows us to find it in quiesce processing */
 	kgnilnd_schedule_conn(conn);
 
-	/* lose peer's ref */
-	kgnilnd_conn_decref(conn);
-	/* -1 for conn table */
-	kgnilnd_conn_decref(conn);
-
 	EXIT;
 }
 
@@ -676,8 +740,24 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 		 kgnilnd_conn_state2str(conn));
 
 	LASSERT(list_empty(&conn->gnc_hashlist));
+	/* We shouldnt be on the delay list, the conn can 
+	 * get added to this list during a retransmit, and retransmits
+	 * only occur within scheduler threads.
+	 */
+	LASSERT(list_empty(&conn->gnc_delaylist));
 
 	/* 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... */
@@ -713,9 +793,13 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 
 	/* nobody should have marked this as needing scheduling after
 	 * we called close - so only ref should be us handling it */
-	LASSERTF(conn->gnc_scheduled == GNILND_CONN_PROCESS,
-		 "conn 0x%p scheduled %d\n", conn, conn->gnc_scheduled);
-
+	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);
@@ -732,19 +816,13 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 	logmsg = (nlive + nrdma + nq_rdma);
 
 	if (logmsg) {
-		if (conn->gnc_peer_error != 0) {
-			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 {
-			CNETERR("Closed conn 0x%p->%s (errno %d): "
-				"canceled %d TX, %d/%d RDMA\n",
-				conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
-				conn->gnc_error,
-				nlive, nq_rdma, nrdma);
-		}
+		int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ?
+				D_NETERROR : D_NET;
+		CDEBUG(level, "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);
@@ -767,6 +845,8 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 	/* 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
@@ -781,8 +861,8 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
 
 	/* 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);
+			    conn->gnc_error == -ECONNRESET ?
+			    conn->gnc_peer_error : conn->gnc_error, 0);
 
 	EXIT;
 }
@@ -795,6 +875,7 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram)
 	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 */
 
@@ -829,7 +910,6 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram)
 			&connreq->gncr_gnparams.gnpr_smsg_attr);
 	if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
 		gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
-		gni_smsg_attr_t *remote = &connreq->gncr_gnparams.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"
@@ -846,8 +926,8 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram)
 
 	/* 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",
+		" local cqid %u SMSG %p->%u hndl %#llx.%#llx"
+		" remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n",
 		conn, libcfs_nid2str(connreq->gncr_srcnid),
 		libcfs_nid2str(connreq->gncr_dstnid),
 		&conn->gnpr_smsg_attr,
@@ -864,6 +944,7 @@ kgnilnd_set_conn_params(kgn_dgram_t *dgram)
 
 	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));
@@ -890,10 +971,13 @@ return_out:
  * 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)
+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;
+	kgn_peer_t	*peer;
+	int		rc;
 
 	LASSERT(nid != LNET_NID_ANY);
 
@@ -922,6 +1006,7 @@ kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net)
 		return -ENOMEM;
 	}
 	peer->gnp_nid = nid;
+	peer->gnp_state = node_state;
 
 	/* translate from nid to nic addr & store */
 	rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
@@ -1028,13 +1113,12 @@ kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
 	CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
 		conn->gnc_device);
 
-	/* add ref for mbox purgatory hold */
-	kgnilnd_peer_addref(peer);
-	kgnilnd_conn_addref(conn);
+	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_nid = peer->gnp_nid;
+	mbox->mbx_prev_purg_nid = peer->gnp_nid;
 	mbox->mbx_add_purgatory = jiffies;
 	kgnilnd_release_mbox(conn, 1);
 
@@ -1085,7 +1169,6 @@ kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
 		 * on the peer's conn_list anymore.
 		 */
 
-		kgnilnd_peer_decref(conn->gnc_peer);
 		list_del_init(&conn->gnc_list);
 
 		/* NB - only unlinking if we set pending in del_peer_locked from admin or
@@ -1138,12 +1221,12 @@ kgnilnd_release_purgatory_list(struct list_head *conn_list)
 		 * 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);
+		kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
 
 		list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
 					 gmp_list) {
 			CDEBUG(D_NET,
-			       "dev %p releasing held mdd "LPX64"."LPX64"\n",
+			       "dev %p releasing held mdd %#llx.%#llx\n",
 			       conn->gnc_device, gmp->gmp_map_key.qword1,
 			       gmp->gmp_map_key.qword2);
 
@@ -1253,9 +1336,6 @@ kgnilnd_get_peer_info(int index,
 		list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
 			peer = list_entry(ptmp, kgn_peer_t, gnp_list);
 
-			if (peer->gnp_nid != *id)
-				continue;
-
 			if (index-- > 0)
 				continue;
 
@@ -1316,11 +1396,14 @@ 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);
@@ -1328,7 +1411,7 @@ kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
 		rc = -ESHUTDOWN;
 		RETURN(rc);
 	}
-	rc = kgnilnd_create_peer_safe(&peer, nid, net);
+	rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
 	if (rc != 0) {
 		up_read(&kgnilnd_data.kgn_net_rw_sem);
 		RETURN(rc);
@@ -1495,9 +1578,6 @@ kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
 
 	write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 
-	/* release all of the souls found held in purgatory */
-	kgnilnd_release_purgatory_list(&souls);
-
 	/* nuke peer TX */
 	kgnilnd_txlist_done(&zombies, error);
 
@@ -1519,7 +1599,8 @@ kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
 	       atomic_read(&kgnilnd_data.kgn_npending_detach)  ||
 	       atomic_read(&kgnilnd_data.kgn_npending_unlink)) {
 
-		cfs_pause(cfs_time_seconds(1));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(cfs_time_seconds(1));
 		i++;
 
 		CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
@@ -1589,8 +1670,8 @@ kgnilnd_get_conn_info(kgn_peer_t *peer,
 
 	*device_id = conn->gnc_device->gnd_host_id;
 	*peerstamp = conn->gnc_peerstamp;
-	*tx_seq = conn->gnc_tx_seq;
-	*rx_seq = conn->gnc_rx_seq;
+	*tx_seq = atomic_read(&conn->gnc_tx_seq);
+	*rx_seq = atomic_read(&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);
@@ -1628,7 +1709,101 @@ kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
 }
 
 int
-kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
+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_PEER_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_state = down;
+
+	if (down == GNILND_PEER_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_PEER_DOWN) {
+		/* using ENETRESET so we don't get messages from
+		 * kgnilnd_tx_done
+		 */
+		kgnilnd_txlist_done(&zombies, -ENETRESET);
+		kgnilnd_peer_notify(peer, -ECONNRESET, 0);
+		LCONSOLE_INFO("Received down event for nid %d\n",
+			      LNET_NIDADDR(nid));
+	}
+
+	return 0;
+}
+
+int
+kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
 {
 	struct libcfs_ioctl_data *data = arg;
 	kgn_net_t                *net = ni->ni_data;
@@ -1748,13 +1923,16 @@ kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
 }
 
 void
-kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
+kgnilnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_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 = LUSTRE_SRV_LNET_PID};
+	struct lnet_process_id       id = {
+		.nid = nid,
+		.pid = LNET_PID_LUSTRE,
+	};
 	ENTRY;
 
 	/* I expect to find him, so only take a read lock */
@@ -1764,7 +1942,7 @@ kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
 		/* 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;
+			*when = ktime_get_seconds();
 			read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
 			GOTO(out, 0);
 		}
@@ -1780,7 +1958,7 @@ kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
 			 * - 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;
+				*when = ktime_get_seconds();
 			}
 			/* we still want to fire a TX and new conn in this case */
 		} else {
@@ -1810,7 +1988,7 @@ kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
 		kgnilnd_launch_tx(tx, net, &id);
 	}
 out:
-	CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lu\n", peer,
+	CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lld\n", peer,
 	       libcfs_nid2str(nid), *when);
 	EXIT;
 }
@@ -1830,12 +2008,11 @@ kgnilnd_dev_init(kgn_device_t *dev)
 	cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
 
 	rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
-				 GNILND_COOKIE, 0,
+				 *kgnilnd_tunables.kgn_pkey, 0,
 				 &dev->gnd_domain);
 	if (rrc != GNI_RC_SUCCESS) {
 		CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
-		rc = -ENODEV;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENODEV);
 	}
 
 	rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
@@ -1843,28 +2020,28 @@ kgnilnd_dev_init(kgn_device_t *dev)
 	if (rrc != GNI_RC_SUCCESS) {
 		CERROR("Can't attach CDM to device %d (%d)\n",
 			dev->gnd_id, rrc);
-		rc = -ENODEV;
-		GOTO(failed, rc);
+		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) {
-		rc = -ENODEV;
-		GOTO(failed, rc);
-	}
+	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,
+		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);
-			rc = -ENODEV;
-			GOTO(failed, rc);
+			GOTO(failed, rc = -ENODEV);
 		}
 
 		rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
@@ -1872,8 +2049,7 @@ kgnilnd_dev_init(kgn_device_t *dev)
 		if (rc != GNI_RC_SUCCESS) {
 			CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
 				dev->gnd_id, rrc);
-			rc = -ENODEV;
-			GOTO(failed, rc);
+			GOTO(failed, rc = -ENODEV);
 		}
 	}
 
@@ -1884,19 +2060,17 @@ kgnilnd_dev_init(kgn_device_t *dev)
 			CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
 				dev->gnd_host_id, rc);
 		}
-		rc = -ESRCH;
-		GOTO(failed, 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,
+	rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
 				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);
-		rc = -EINVAL;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -EINVAL);
 	}
 
 	rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
@@ -1905,8 +2079,7 @@ kgnilnd_dev_init(kgn_device_t *dev)
 	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);
-		rc = -EINVAL;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -EINVAL);
 	}
 
 	/* This one we size differently - overflows are possible and it needs to be
@@ -1918,8 +2091,12 @@ kgnilnd_dev_init(kgn_device_t *dev)
 	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);
-		rc = -EINVAL;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -EINVAL);
+	}
+
+	rrc = kgnilnd_register_smdd_buf(dev);
+	if (rrc != GNI_RC_SUCCESS) {
+		GOTO(failed, rc = -EINVAL);
 	}
 
 	RETURN(0);
@@ -1938,9 +2115,12 @@ kgnilnd_dev_fini(kgn_device_t *dev)
 	/* 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",
+		 list_empty(&dev->gnd_rdmaq) &&
+		 list_empty(&dev->gnd_delay_conns),
+		 "dev 0x%p ready_conns %d@0x%p delay_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_delay_conns), &dev->gnd_delay_conns,
 		 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
 		 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
 
@@ -1950,7 +2130,7 @@ kgnilnd_dev_fini(kgn_device_t *dev)
 		 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",
+		"%d virtual mappings of %llu bytes still mapped\n",
 		 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
 
 	LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
@@ -1972,6 +2152,13 @@ kgnilnd_dev_fini(kgn_device_t *dev)
 		atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
 		"tried to shutdown with threads active\n");
 
+	if (dev->gnd_smdd_hold_buf) {
+		rrc = kgnilnd_deregister_smdd_buf(dev);
+		LASSERTF(rrc == GNI_RC_SUCCESS,
+			"bad rc from deregistion of sMDD buffer: %d\n", rrc);
+		dev->gnd_smdd_hold_buf = NULL;
+	}
+
 	if (dev->gnd_rcv_fma_cqh) {
 		rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
 		LASSERTF(rrc == GNI_RC_SUCCESS,
@@ -2010,7 +2197,6 @@ kgnilnd_dev_fini(kgn_device_t *dev)
 	EXIT;
 }
 
-
 int kgnilnd_base_startup(void)
 {
 	struct timeval       tv;
@@ -2019,6 +2205,15 @@ int kgnilnd_base_startup(void)
 	int                  i;
 	kgn_device_t        *dev;
 	struct task_struct  *thrd;
+
+#if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
+	/* limit how much memory can be allocated for fma blocks in
+	 * instances where many nodes need to reconnects at the same time */
+	struct sysinfo si;
+	si_meminfo(&si);
+	kgnilnd_data.free_pages_limit = si.totalram/4;
+#endif
+
 	ENTRY;
 
 	LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
@@ -2026,7 +2221,7 @@ int kgnilnd_base_startup(void)
 
 	/* zero pointers, flags etc */
 	memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
-	memset(&kgnilnd_hssops, 0, sizeof(kgnilnd_hssops));
+	kgnilnd_check_kgni_version();
 
 	/* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
 	 * a unique (for all time) connstamp so we can uniquely identify
@@ -2046,10 +2241,11 @@ int kgnilnd_base_startup(void)
 
 		dev->gnd_id = i;
 		INIT_LIST_HEAD(&dev->gnd_ready_conns);
+		INIT_LIST_HEAD(&dev->gnd_delay_conns);
 		INIT_LIST_HEAD(&dev->gnd_map_tx);
 		INIT_LIST_HEAD(&dev->gnd_fma_buffs);
 		mutex_init(&dev->gnd_cq_mutex);
-		sema_init(&dev->gnd_fmablk_sem, 1);
+		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);
@@ -2066,24 +2262,28 @@ int kgnilnd_base_startup(void)
 		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) {
-			rc = -ENOMEM;
-			GOTO(failed, rc);
-		}
+		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 */
@@ -2093,24 +2293,27 @@ int kgnilnd_base_startup(void)
 	init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
 	spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
 
-	sema_init(&kgnilnd_data.kgn_quiesce_sem, 1);
+	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;
-	PORTAL_MODULE_USE;
+	if (!try_module_get(THIS_MODULE))
+		GOTO(failed, rc = -ENOENT);
 
 	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) {
-		rc = -ENOMEM;
-		GOTO(failed, rc);
-	}
+	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]);
@@ -2119,10 +2322,8 @@ int kgnilnd_base_startup(void)
 	LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
 		    sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
 
-	if (kgnilnd_data.kgn_conns == NULL) {
-		rc = -ENOMEM;
-		GOTO(failed, rc);
-	}
+	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]);
@@ -2131,68 +2332,49 @@ int kgnilnd_base_startup(void)
 	LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
 		    sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);
 
-	if (kgnilnd_data.kgn_nets == NULL) {
-		rc = -ENOMEM;
-		GOTO(failed, rc);
-	}
+	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 =
-		cfs_mem_cache_create("kgn_mbox_block",
-				     KMALLOC_MAX_SIZE,
-				     0,    /* offset */
-				     SLAB_HWCACHE_ALIGN);   /* flags */
+		kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0,
+				  SLAB_HWCACHE_ALIGN, NULL);
 	if (kgnilnd_data.kgn_mbox_cache == NULL) {
 		CERROR("Can't create slab for physical mbox blocks\n");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
 	kgnilnd_data.kgn_rx_cache =
-		cfs_mem_cache_create("kgn_rx_t",
-				     sizeof(kgn_rx_t),
-				     0,    /* offset */
-				     0);   /* flags */
+		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");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
 	kgnilnd_data.kgn_tx_cache =
-		cfs_mem_cache_create("kgn_tx_t",
-				     sizeof(kgn_tx_t),
-				     0,    /* offset */
-				     0);   /* flags */
+		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");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
 	kgnilnd_data.kgn_tx_phys_cache =
-		cfs_mem_cache_create("kgn_tx_phys",
-				     LNET_MAX_IOV * sizeof(gni_mem_segment_t),
-				     0,    /* offset */
-				     0);   /* flags */
+		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");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
 	kgnilnd_data.kgn_dgram_cache =
-		cfs_mem_cache_create("kgn_dgram_t",
-				     sizeof(kgn_dgram_t),
-				     0,    /* offset */
-				     0);   /* flags */
+		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");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 
 	/* allocate a MAX_IOV array of page pointers for each cpu */
@@ -2200,8 +2382,7 @@ int kgnilnd_base_startup(void)
 						   GFP_KERNEL);
 	if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
 		CERROR("Can't allocate vmap cksum pages\n");
-		rc = -ENOMEM;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENOMEM);
 	}
 	kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
 	memset(kgnilnd_data.kgn_cksum_map_pages, 0,
@@ -2212,8 +2393,7 @@ int kgnilnd_base_startup(void)
 							      GFP_KERNEL);
 		if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
 			CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
-			rc = -ENOMEM;
-			GOTO(failed, rc);
+			GOTO(failed, rc = -ENOMEM);
 		}
 	}
 
@@ -2229,16 +2409,14 @@ int kgnilnd_base_startup(void)
 			kgnilnd_data.kgn_ndevs++;
 
 			rc = kgnilnd_allocate_phys_fmablk(dev);
-			if (rc) {
+			if (rc)
 				GOTO(failed, rc);
-			}
 		}
 	}
 
 	if (kgnilnd_data.kgn_ndevs == 0) {
 		CERROR("Can't initialise any GNI devices\n");
-		rc = -ENODEV;
-		GOTO(failed, rc);
+		GOTO(failed, rc = -ENODEV);
 	}
 
 	rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
@@ -2247,6 +2425,12 @@ int kgnilnd_base_startup(void)
 		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
@@ -2261,13 +2445,20 @@ int kgnilnd_base_startup(void)
 	}
 
 	/* 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);
+	if (*kgnilnd_tunables.kgn_thread_affinity) {
+		rc = kgnilnd_start_sd_threads();
+		if (rc != 0)
 			GOTO(failed, rc);
+	} else {
+		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);
+			}
 		}
 	}
 
@@ -2298,8 +2489,6 @@ int kgnilnd_base_startup(void)
 		}
 	}
 
-
-
 	/* flag everything initialised */
 	kgnilnd_data.kgn_init = GNILND_INIT_ALL;
 	/*****************************************************/
@@ -2316,7 +2505,7 @@ failed:
 void
 kgnilnd_base_shutdown(void)
 {
-	int           i;
+	int			i, j;
 	ENTRY;
 
 	while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
@@ -2326,10 +2515,30 @@ kgnilnd_base_shutdown(void)
 	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));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(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
@@ -2347,7 +2556,8 @@ kgnilnd_base_shutdown(void)
 		i++;
 		CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
 		       "Waiting for ruhroh thread to terminate\n");
-		cfs_pause(cfs_time_seconds(1));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(cfs_time_seconds(1));
 	}
 
        /* Flag threads to terminate */
@@ -2357,7 +2567,7 @@ kgnilnd_base_shutdown(void)
 		kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
 
 		/* should clear all the MDDs */
-		kgnilnd_unmap_phys_fmablk(dev);
+		kgnilnd_unmap_fma_blocks(dev);
 
 		kgnilnd_schedule_device(dev);
 		wake_up_all(&dev->gnd_dgram_waitq);
@@ -2369,6 +2579,9 @@ kgnilnd_base_shutdown(void)
 	wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
 	spin_unlock(&kgnilnd_data.kgn_reaper_lock);
 
+	if (atomic_read(&kgnilnd_data.kgn_nthreads))
+		kgnilnd_wakeup_rca_thread();
+
 	/* Wait for threads to exit */
 	i = 2;
 	while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
@@ -2376,7 +2589,8 @@ kgnilnd_base_shutdown(void)
 		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));
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(cfs_time_seconds(1));
 	}
 
 	LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
@@ -2433,30 +2647,20 @@ kgnilnd_base_shutdown(void)
 		kgnilnd_free_phys_fmablk(dev);
 	}
 
-	if (kgnilnd_data.kgn_mbox_cache != NULL) {
-		i = cfs_mem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
-		LASSERTF(i == 0, "rc %d destroying kgn_mbox_cache\n", i);
-	}
+	if (kgnilnd_data.kgn_mbox_cache != NULL)
+		kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
 
-	if (kgnilnd_data.kgn_rx_cache != NULL) {
-		i = cfs_mem_cache_destroy(kgnilnd_data.kgn_rx_cache);
-		LASSERTF(i == 0, "rc %d destroying kgn_rx_cache\n", i);
-	}
+	if (kgnilnd_data.kgn_rx_cache != NULL)
+		kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
 
-	if (kgnilnd_data.kgn_tx_cache != NULL) {
-		i = cfs_mem_cache_destroy(kgnilnd_data.kgn_tx_cache);
-		LASSERTF(i == 0, "rc %d destroying kgn_tx_cache\n", i);
-	}
+	if (kgnilnd_data.kgn_tx_cache != NULL)
+		kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
 
-	if (kgnilnd_data.kgn_tx_phys_cache != NULL) {
-		i = cfs_mem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
-		LASSERTF(i == 0, "rc %d destroying kgn_tx_phys_cache\n", i);
-	}
+	if (kgnilnd_data.kgn_tx_phys_cache != NULL)
+		kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
 
-	if (kgnilnd_data.kgn_dgram_cache != NULL) {
-		i = cfs_mem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
-		LASSERTF(i == 0, "rc %d destroying kgn_dgram_cache\n", i);
-	}
+	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++) {
@@ -2471,21 +2675,21 @@ kgnilnd_base_shutdown(void)
 	       atomic_read(&libcfs_kmemory));
 
 	kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
-	PORTAL_MODULE_UNUSE;
+	module_put(THIS_MODULE);
 
 	EXIT;
 }
 
 int
-kgnilnd_startup(lnet_ni_t *ni)
+kgnilnd_startup(struct lnet_ni *ni)
 {
 	int               rc, devno;
 	kgn_net_t        *net;
 	ENTRY;
 
-	LASSERTF(ni->ni_lnd == &the_kgnilnd,
+	LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd,
 		"bad LND 0x%p != the_kgnilnd @ 0x%p\n",
-		ni->ni_lnd, &the_kgnilnd);
+		ni->ni_net->net_lnd, &the_kgnilnd);
 
 	if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
 		rc = kgnilnd_base_startup();
@@ -2494,32 +2698,47 @@ kgnilnd_startup(lnet_ni_t *ni)
 	}
 
 	/* Serialize with shutdown. */
-	down(&kgnilnd_data.kgn_quiesce_sem);
+	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");
-		rc = -ENOMEM;
 		/* no need to cleanup the CDM... */
-		GOTO(failed, rc);
+		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 (!ni->ni_net->net_tunables_set) {
+		ni->ni_net->net_tunables.lct_max_tx_credits =
+			*kgnilnd_tunables.kgn_credits;
+		ni->ni_net->net_tunables.lct_peer_tx_credits =
+			*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);
-
-		ni->ni_peertimeout = *kgnilnd_tunables.kgn_timeout + fudge;
+		timeout = *kgnilnd_tunables.kgn_timeout + fudge;
+
+		if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
+			ni->ni_net->net_tunables.lct_peer_timeout =
+				 *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_net->net_tunables.lct_peer_timeout = timeout;
 
 		LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
-			      ni->ni_peertimeout);
+			      ni->ni_net->net_tunables.lct_peer_timeout);
 	}
 
 	atomic_set(&net->gnn_refcount, 1);
@@ -2551,16 +2770,16 @@ kgnilnd_startup(lnet_ni_t *ni)
 
 	/* we need a separate thread to call probe_wait_by_id until
 	 * we get a function callback notifier from kgni */
-	up(&kgnilnd_data.kgn_quiesce_sem);
+	mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
 	RETURN(0);
  failed:
-	up(&kgnilnd_data.kgn_quiesce_sem);
+	mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
 	kgnilnd_shutdown(ni);
 	RETURN(rc);
 }
 
 void
-kgnilnd_shutdown(lnet_ni_t *ni)
+kgnilnd_shutdown(struct lnet_ni *ni)
 {
 	kgn_net_t     *net = ni->ni_data;
 	int           i;
@@ -2573,14 +2792,13 @@ kgnilnd_shutdown(lnet_ni_t *ni)
 		"init %d\n", kgnilnd_data.kgn_init);
 
 	/* Serialize with startup. */
-	down(&kgnilnd_data.kgn_quiesce_sem);
+	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);
-		rc = -EINVAL;
-		GOTO(out, rc);
+		GOTO(out, rc = -EINVAL);
 	}
 
 	LASSERTF(ni == net->gnn_ni,
@@ -2623,7 +2841,8 @@ kgnilnd_shutdown(lnet_ni_t *ni)
 				"Waiting for %d references to clear on net %d\n",
 				atomic_read(&net->gnn_refcount),
 				net->gnn_netnum);
-			cfs_pause(cfs_time_seconds(1));
+			set_current_state(TASK_UNINTERRUPTIBLE);
+			schedule_timeout(cfs_time_seconds(1));
 		}
 
 		/* release ref from kgnilnd_startup */
@@ -2657,22 +2876,18 @@ out:
 	CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
 	       atomic_read(&libcfs_kmemory));
 
-	up(&kgnilnd_data.kgn_quiesce_sem);
+	mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
 	EXIT;
-	return;
 }
 
-void __exit
-kgnilnd_module_fini(void)
+static void __exit kgnilnd_exit(void)
 {
 	lnet_unregister_lnd(&the_kgnilnd);
 	kgnilnd_proc_fini();
 	kgnilnd_remove_sysctl();
-	kgnilnd_tunables_fini();
 }
 
-int __init
-kgnilnd_module_init(void)
+static int __init kgnilnd_init(void)
 {
 	int    rc;
 
@@ -2680,7 +2895,7 @@ kgnilnd_module_init(void)
 	if (rc != 0)
 		return rc;
 
-	printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n");
+	LCONSOLE_INFO("Lustre: kgnilnd build version: "LUSTRE_VERSION_STRING"\n");
 
 	kgnilnd_insert_sysctl();
 	kgnilnd_proc_init();
@@ -2691,8 +2906,9 @@ kgnilnd_module_init(void)
 }
 
 MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
-MODULE_DESCRIPTION("Kernel Gemini LND v"KGNILND_BUILD_REV);
+MODULE_DESCRIPTION("Gemini LNet Network Driver");
+MODULE_VERSION(LUSTRE_VERSION_STRING);
 MODULE_LICENSE("GPL");
 
-module_init(kgnilnd_module_init);
-module_exit(kgnilnd_module_fini);
+module_init(kgnilnd_init);
+module_exit(kgnilnd_exit);