staging: lustre: avoid intensive reconnecting for ko2iblnd

When there is a connection race between two nodes and one side
of the connection is rejected by the other side. o2iblnd will
reconnect immediately, this is going to generate a lot of
trashes if:

 - race winner is slow and can't send out connecting request
   in short time.
 - remote side leaves a cmid in TIMEWAIT state, which will reject
   future connection requests

To resolve this problem, this patch changed the reconnection
behave: reconnection is submitted by connd only if a zombie
connection is being destroyed and there is a pending
reconnection request for the corresponding peer.

Also, after a few rejections, reconnection will have a time
interval between each attempt.
Signed-off-by: Liang Zhen <liang.zhen@intel.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-7569
Reviewed-on: http://review.whamcloud.com/17892Reviewed-by: Doug Oucharek <doug.s.oucharek@intel.com>
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Tested-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.c

@@ -364,9 +364,7 @@ void kiblnd_destroy_peer(kib_peer_t *peer)
 	LASSERT(net);
 	LASSERT(!atomic_read(&peer->ibp_refcount));
 	LASSERT(!kiblnd_peer_active(peer));
-	LASSERT(!peer->ibp_connecting);
-	LASSERT(!peer->ibp_accepting);
-	LASSERT(list_empty(&peer->ibp_conns));
+	LASSERT(kiblnd_peer_idle(peer));
 	LASSERT(list_empty(&peer->ibp_tx_queue));
 
 	LIBCFS_FREE(peer, sizeof(*peer));
@@ -392,10 +390,7 @@ kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid)
 
 	list_for_each(tmp, peer_list) {
 		peer = list_entry(tmp, kib_peer_t, ibp_list);
-
-		LASSERT(peer->ibp_connecting > 0 || /* creating conns */
-			 peer->ibp_accepting > 0 ||
-			 !list_empty(&peer->ibp_conns));  /* active conn */
+		LASSERT(!kiblnd_peer_idle(peer));
 
 		if (peer->ibp_nid != nid)
 			continue;
@@ -432,9 +427,7 @@ static int kiblnd_get_peer_info(lnet_ni_t *ni, int index,
 	for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 		list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 			peer = list_entry(ptmp, kib_peer_t, ibp_list);
-			LASSERT(peer->ibp_connecting > 0 ||
-				peer->ibp_accepting > 0 ||
-				!list_empty(&peer->ibp_conns));
+			LASSERT(!kiblnd_peer_idle(peer));
 
 			if (peer->ibp_ni != ni)
 				continue;
@@ -502,9 +495,7 @@ static int kiblnd_del_peer(lnet_ni_t *ni, lnet_nid_t nid)
 	for (i = lo; i <= hi; i++) {
 		list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 			peer = list_entry(ptmp, kib_peer_t, ibp_list);
-			LASSERT(peer->ibp_connecting > 0 ||
-				peer->ibp_accepting > 0 ||
-				!list_empty(&peer->ibp_conns));
+			LASSERT(!kiblnd_peer_idle(peer));
 
 			if (peer->ibp_ni != ni)
 				continue;
@@ -545,9 +536,7 @@ static kib_conn_t *kiblnd_get_conn_by_idx(lnet_ni_t *ni, int index)
 	for (i = 0; i < kiblnd_data.kib_peer_hash_size; i++) {
 		list_for_each(ptmp, &kiblnd_data.kib_peers[i]) {
 			peer = list_entry(ptmp, kib_peer_t, ibp_list);
-			LASSERT(peer->ibp_connecting > 0 ||
-				peer->ibp_accepting > 0 ||
-				!list_empty(&peer->ibp_conns));
+			LASSERT(!kiblnd_peer_idle(peer));
 
 			if (peer->ibp_ni != ni)
 				continue;
@@ -837,14 +826,14 @@ kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 	return conn;
 
  failed_2:
-	kiblnd_destroy_conn(conn);
+	kiblnd_destroy_conn(conn, true);
  failed_1:
 	LIBCFS_FREE(init_qp_attr, sizeof(*init_qp_attr));
  failed_0:
 	return NULL;
 }
 
-void kiblnd_destroy_conn(kib_conn_t *conn)
+void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn)
 {
 	struct rdma_cm_id *cmid = conn->ibc_cmid;
 	kib_peer_t *peer = conn->ibc_peer;
@@ -984,9 +973,7 @@ static int kiblnd_close_matching_conns(lnet_ni_t *ni, lnet_nid_t nid)
 	for (i = lo; i <= hi; i++) {
 		list_for_each_safe(ptmp, pnxt, &kiblnd_data.kib_peers[i]) {
 			peer = list_entry(ptmp, kib_peer_t, ibp_list);
-			LASSERT(peer->ibp_connecting > 0 ||
-				peer->ibp_accepting > 0 ||
-				!list_empty(&peer->ibp_conns));
+			LASSERT(!kiblnd_peer_idle(peer));
 
 			if (peer->ibp_ni != ni)
 				continue;
@@ -1071,12 +1058,8 @@ static void kiblnd_query(lnet_ni_t *ni, lnet_nid_t nid, unsigned long *when)
 	read_lock_irqsave(glock, flags);
 
 	peer = kiblnd_find_peer_locked(nid);
-	if (peer) {
-		LASSERT(peer->ibp_connecting > 0 || /* creating conns */
-			 peer->ibp_accepting > 0 ||
-			 !list_empty(&peer->ibp_conns));  /* active conn */
+	if (peer)
 		last_alive = peer->ibp_last_alive;
-	}
 
 	read_unlock_irqrestore(glock, flags);
 
@@ -2368,6 +2351,8 @@ static void kiblnd_base_shutdown(void)
 			LASSERT(list_empty(&kiblnd_data.kib_peers[i]));
 		LASSERT(list_empty(&kiblnd_data.kib_connd_zombies));
 		LASSERT(list_empty(&kiblnd_data.kib_connd_conns));
+		LASSERT(list_empty(&kiblnd_data.kib_reconn_list));
+		LASSERT(list_empty(&kiblnd_data.kib_reconn_wait));
 
 		/* flag threads to terminate; wake and wait for them to die */
 		kiblnd_data.kib_shutdown = 1;
@@ -2506,6 +2491,9 @@ static int kiblnd_base_startup(void)
 	spin_lock_init(&kiblnd_data.kib_connd_lock);
 	INIT_LIST_HEAD(&kiblnd_data.kib_connd_conns);
 	INIT_LIST_HEAD(&kiblnd_data.kib_connd_zombies);
+	INIT_LIST_HEAD(&kiblnd_data.kib_reconn_list);
+	INIT_LIST_HEAD(&kiblnd_data.kib_reconn_wait);
+
 	init_waitqueue_head(&kiblnd_data.kib_connd_waitq);
 	init_waitqueue_head(&kiblnd_data.kib_failover_waitq);
 

drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.h

@@ -348,6 +348,16 @@ typedef struct {
 	void *kib_connd; /* the connd task (serialisation assertions) */
 	struct list_head kib_connd_conns;   /* connections to setup/teardown */
 	struct list_head kib_connd_zombies; /* connections with zero refcount */
+	/* connections to reconnect */
+	struct list_head	kib_reconn_list;
+	/* peers wait for reconnection */
+	struct list_head	kib_reconn_wait;
+	/**
+	 * The second that peers are pulled out from \a kib_reconn_wait
+	 * for reconnection.
+	 */
+	time64_t		kib_reconn_sec;
+
 	wait_queue_head_t kib_connd_waitq;  /* connection daemon sleeps here */
 	spinlock_t kib_connd_lock;          /* serialise */
 	struct ib_qp_attr kib_error_qpa;    /* QP->ERROR */
@@ -525,6 +535,8 @@ typedef struct kib_conn {
 	struct list_head ibc_list;             /* stash on peer's conn list */
 	struct list_head      ibc_sched_list;  /* schedule for attention */
 	__u16                 ibc_version;     /* version of connection */
+	/* reconnect later */
+	__u16			ibc_reconnect:1;
 	__u64                 ibc_incarnation; /* which instance of the peer */
 	atomic_t              ibc_refcount;    /* # users */
 	int                   ibc_state;       /* what's happening */
@@ -574,18 +586,25 @@ typedef struct kib_peer {
 	struct list_head ibp_list;        /* stash on global peer list */
 	lnet_nid_t       ibp_nid;         /* who's on the other end(s) */
 	lnet_ni_t        *ibp_ni;         /* LNet interface */
-	atomic_t         ibp_refcount;    /* # users */
 	struct list_head ibp_conns;       /* all active connections */
 	struct list_head ibp_tx_queue;    /* msgs waiting for a conn */
-	__u16            ibp_version;     /* version of peer */
 	__u64            ibp_incarnation; /* incarnation of peer */
-	int              ibp_connecting;  /* current active connection attempts
-					   */
-	int              ibp_accepting;   /* current passive connection attempts
-					   */
-	int              ibp_error;       /* errno on closing this peer */
-	unsigned long    ibp_last_alive;  /* when (in jiffies) I was last alive
-					   */
+	/* when (in jiffies) I was last alive */
+	unsigned long		ibp_last_alive;
+	/* # users */
+	atomic_t		ibp_refcount;
+	/* version of peer */
+	__u16			ibp_version;
+	/* current passive connection attempts */
+	unsigned short		ibp_accepting;
+	/* current active connection attempts */
+	unsigned short		ibp_connecting;
+	/* reconnect this peer later */
+	unsigned short		ibp_reconnecting:1;
+	/* # consecutive reconnection attempts to this peer */
+	unsigned int		ibp_reconnected;
+	/* errno on closing this peer */
+	int              ibp_error;
 	/* max map_on_demand */
 	__u16		 ibp_max_frags;
 	/* max_peer_credits */
@@ -667,6 +686,20 @@ do {							    \
 		kiblnd_destroy_peer(peer);		      \
 } while (0)
 
+static inline bool
+kiblnd_peer_connecting(kib_peer_t *peer)
+{
+	return peer->ibp_connecting ||
+	       peer->ibp_reconnecting ||
+	       peer->ibp_accepting;
+}
+
+static inline bool
+kiblnd_peer_idle(kib_peer_t *peer)
+{
+	return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
+}
+
 static inline struct list_head *
 kiblnd_nid2peerlist(lnet_nid_t nid)
 {
@@ -943,6 +976,7 @@ int  kiblnd_translate_mtu(int value);
 int  kiblnd_dev_failover(kib_dev_t *dev);
 int  kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
 void kiblnd_destroy_peer(kib_peer_t *peer);
+bool kiblnd_reconnect_peer(kib_peer_t *peer);
 void kiblnd_destroy_dev(kib_dev_t *dev);
 void kiblnd_unlink_peer_locked(kib_peer_t *peer);
 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid);
@@ -952,7 +986,7 @@ int  kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why);
 
 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
 			       int state, int version);
-void kiblnd_destroy_conn(kib_conn_t *conn);
+void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
 void kiblnd_close_conn(kib_conn_t *conn, int error);
 void kiblnd_close_conn_locked(kib_conn_t *conn, int error);