staging: lustre: avoid intensive reconnecting for ko2iblnd
authorLiang Zhen <liang.zhen@intel.com>
Wed, 2 Mar 2016 23:53:29 +0000 (18:53 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 3 Mar 2016 00:01:38 +0000 (16:01 -0800)
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/17892
Reviewed-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
drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.h
drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd_cb.c

index 56c221b..135ccf1 100644 (file)
@@ -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);
 
index 6a4c4ac..bfcbdd1 100644 (file)
@@ -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);
 
index 199c105..79051c0 100644 (file)
@@ -1257,6 +1257,7 @@ kiblnd_connect_peer(kib_peer_t *peer)
 
        LASSERT(net);
        LASSERT(peer->ibp_connecting > 0);
+       LASSERT(!peer->ibp_reconnecting);
 
        cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
                                     IB_QPT_RC);
@@ -1312,6 +1313,56 @@ kiblnd_connect_peer(kib_peer_t *peer)
        kiblnd_peer_connect_failed(peer, 1, rc);
 }
 
+bool
+kiblnd_reconnect_peer(kib_peer_t *peer)
+{
+       rwlock_t *glock = &kiblnd_data.kib_global_lock;
+       char *reason = NULL;
+       struct list_head txs;
+       unsigned long flags;
+
+       INIT_LIST_HEAD(&txs);
+
+       write_lock_irqsave(glock, flags);
+       if (!peer->ibp_reconnecting) {
+               if (peer->ibp_accepting)
+                       reason = "accepting";
+               else if (peer->ibp_connecting)
+                       reason = "connecting";
+               else if (!list_empty(&peer->ibp_conns))
+                       reason = "connected";
+               else /* connected then closed */
+                       reason = "closed";
+
+               goto no_reconnect;
+       }
+
+       LASSERT(!peer->ibp_accepting && !peer->ibp_connecting &&
+               list_empty(&peer->ibp_conns));
+       peer->ibp_reconnecting = 0;
+
+       if (!kiblnd_peer_active(peer)) {
+               list_splice_init(&peer->ibp_tx_queue, &txs);
+               reason = "unlinked";
+               goto no_reconnect;
+       }
+
+       peer->ibp_connecting++;
+       peer->ibp_reconnected++;
+       write_unlock_irqrestore(glock, flags);
+
+       kiblnd_connect_peer(peer);
+       return true;
+
+no_reconnect:
+       write_unlock_irqrestore(glock, flags);
+
+       CWARN("Abort reconnection of %s: %s\n",
+             libcfs_nid2str(peer->ibp_nid), reason);
+       kiblnd_txlist_done(peer->ibp_ni, &txs, -ECONNABORTED);
+       return false;
+}
+
 void
 kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
 {
@@ -1357,8 +1408,7 @@ kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
        if (peer) {
                if (list_empty(&peer->ibp_conns)) {
                        /* found a peer, but it's still connecting... */
-                       LASSERT(peer->ibp_connecting ||
-                               peer->ibp_accepting);
+                       LASSERT(kiblnd_peer_connecting(peer));
                        if (tx)
                                list_add_tail(&tx->tx_list,
                                              &peer->ibp_tx_queue);
@@ -1396,8 +1446,7 @@ kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
        if (peer2) {
                if (list_empty(&peer2->ibp_conns)) {
                        /* found a peer, but it's still connecting... */
-                       LASSERT(peer2->ibp_connecting ||
-                               peer2->ibp_accepting);
+                       LASSERT(kiblnd_peer_connecting(peer2));
                        if (tx)
                                list_add_tail(&tx->tx_list,
                                              &peer2->ibp_tx_queue);
@@ -1817,10 +1866,7 @@ kiblnd_peer_notify(kib_peer_t *peer)
 
        read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
 
-       if (list_empty(&peer->ibp_conns) &&
-           !peer->ibp_accepting &&
-           !peer->ibp_connecting &&
-           peer->ibp_error) {
+       if (kiblnd_peer_idle(peer) && peer->ibp_error) {
                error = peer->ibp_error;
                peer->ibp_error = 0;
 
@@ -2020,14 +2066,14 @@ kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
                peer->ibp_accepting--;
        }
 
-       if (peer->ibp_connecting ||
-           peer->ibp_accepting) {
+       if (kiblnd_peer_connecting(peer)) {
                /* another connection attempt under way... */
                write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
                                        flags);
                return;
        }
 
+       peer->ibp_reconnected = 0;
        if (list_empty(&peer->ibp_conns)) {
                /* Take peer's blocked transmits to complete with error */
                list_add(&zombies, &peer->ibp_tx_queue);
@@ -2100,6 +2146,7 @@ kiblnd_connreq_done(kib_conn_t *conn, int status)
         */
        kiblnd_conn_addref(conn);              /* +1 ref for ibc_list */
        list_add(&conn->ibc_list, &peer->ibp_conns);
+       peer->ibp_reconnected = 0;
        if (active)
                peer->ibp_connecting--;
        else
@@ -2355,10 +2402,16 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
                if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
                    peer2->ibp_version     != version) {
                        kiblnd_close_peer_conns_locked(peer2, -ESTALE);
+
+                       if (kiblnd_peer_active(peer2)) {
+                               peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
+                               peer2->ibp_version = version;
+                       }
                        write_unlock_irqrestore(g_lock, flags);
 
-                       CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
-                             libcfs_nid2str(nid), peer2->ibp_version, version);
+                       CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
+                             libcfs_nid2str(nid), peer2->ibp_version, version,
+                             peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
 
                        kiblnd_peer_decref(peer);
                        rej.ibr_why = IBLND_REJECT_CONN_STALE;
@@ -2377,6 +2430,11 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
                        goto failed;
                }
 
+               /**
+                * passive connection is allowed even this peer is waiting for
+                * reconnection.
+                */
+               peer2->ibp_reconnecting = 0;
                peer2->ibp_accepting++;
                kiblnd_peer_addref(peer2);
 
@@ -2478,75 +2536,79 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
 }
 
 static void
-kiblnd_reconnect(kib_conn_t *conn, int version,
-                __u64 incarnation, int why, kib_connparams_t *cp)
+kiblnd_check_reconnect(kib_conn_t *conn, int version,
+                      __u64 incarnation, int why, kib_connparams_t *cp)
 {
+       rwlock_t *glock = &kiblnd_data.kib_global_lock;
        kib_peer_t *peer = conn->ibc_peer;
        char *reason;
-       int retry = 0;
+       int msg_size = IBLND_MSG_SIZE;
+       int frag_num = -1;
+       int queue_dep = -1;
+       bool reconnect;
        unsigned long flags;
 
        LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
        LASSERT(peer->ibp_connecting > 0);     /* 'conn' at least */
+       LASSERT(!peer->ibp_reconnecting);
 
-       write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+       if (cp) {
+               msg_size = cp->ibcp_max_msg_size;
+               frag_num = cp->ibcp_max_frags;
+               queue_dep = cp->ibcp_queue_depth;
+       }
 
-       /*
+       write_lock_irqsave(glock, flags);
+       /**
         * retry connection if it's still needed and no other connection
         * attempts (active or passive) are in progress
         * NB: reconnect is still needed even when ibp_tx_queue is
         * empty if ibp_version != version because reconnect may be
         * initiated by kiblnd_query()
         */
-       if ((!list_empty(&peer->ibp_tx_queue) ||
-            peer->ibp_version != version) &&
-           peer->ibp_connecting == 1 &&
-           !peer->ibp_accepting) {
-               retry = 1;
-               peer->ibp_connecting++;
-
-               peer->ibp_version     = version;
-               peer->ibp_incarnation = incarnation;
+       reconnect = (!list_empty(&peer->ibp_tx_queue) ||
+                    peer->ibp_version != version) &&
+                   peer->ibp_connecting == 1 &&
+                   !peer->ibp_accepting;
+       if (!reconnect) {
+               reason = "no need";
+               goto out;
        }
 
-       write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
-
-       if (!retry)
-               return;
-
        switch (why) {
        default:
                reason = "Unknown";
                break;
 
        case IBLND_REJECT_RDMA_FRAGS:
-               if (!cp)
-                       goto failed;
-
-               if (conn->ibc_max_frags <= cp->ibcp_max_frags) {
-                       CNETERR("Unsupported max frags, peer supports %d\n",
-                               cp->ibcp_max_frags);
-                       goto failed;
-               } else if (!*kiblnd_tunables.kib_map_on_demand) {
-                       CNETERR("map_on_demand must be enabled to support map_on_demand peers\n");
-                       goto failed;
+               if (!cp) {
+                       reason = "can't negotiate max frags";
+                       goto out;
+               }
+               if (!*kiblnd_tunables.kib_map_on_demand) {
+                       reason = "map_on_demand must be enabled";
+                       goto out;
+               }
+               if (conn->ibc_max_frags <= frag_num) {
+                       reason = "unsupported max frags";
+                       goto out;
                }
 
-               peer->ibp_max_frags = cp->ibcp_max_frags;
+               peer->ibp_max_frags = frag_num;
                reason = "rdma fragments";
                break;
 
        case IBLND_REJECT_MSG_QUEUE_SIZE:
-               if (!cp)
-                       goto failed;
-
-               if (conn->ibc_queue_depth <= cp->ibcp_queue_depth) {
-                       CNETERR("Unsupported queue depth, peer supports %d\n",
-                               cp->ibcp_queue_depth);
-                       goto failed;
+               if (!cp) {
+                       reason = "can't negotiate queue depth";
+                       goto out;
+               }
+               if (conn->ibc_queue_depth <= queue_dep) {
+                       reason = "unsupported queue depth";
+                       goto out;
                }
 
-               peer->ibp_queue_depth = cp->ibcp_queue_depth;
+               peer->ibp_queue_depth = queue_dep;
                reason = "queue depth";
                break;
 
@@ -2563,20 +2625,24 @@ kiblnd_reconnect(kib_conn_t *conn, int version,
                break;
        }
 
-       CNETERR("%s: retrying (%s), %x, %x, queue_dep: %d, max_frag: %d, msg_size: %d\n",
-               libcfs_nid2str(peer->ibp_nid),
-               reason, IBLND_MSG_VERSION, version,
-               conn->ibc_queue_depth, conn->ibc_max_frags,
-               cp ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
-
-       kiblnd_connect_peer(peer);
-       return;
-failed:
-       write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
-       peer->ibp_connecting--;
-       write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+       conn->ibc_reconnect = 1;
+       peer->ibp_reconnecting = 1;
+       peer->ibp_version = version;
+       if (incarnation)
+               peer->ibp_incarnation = incarnation;
+out:
+        write_unlock_irqrestore(glock, flags);
 
-       return;
+       CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
+               libcfs_nid2str(peer->ibp_nid),
+               reconnect ? "reconnect" : "don't reconnect",
+               reason, IBLND_MSG_VERSION, version, msg_size,
+               conn->ibc_queue_depth, queue_dep,
+               conn->ibc_max_frags, frag_num);
+        /**
+        * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer
+        * while destroying the zombie
+        */
 }
 
 static void
@@ -2589,8 +2655,8 @@ kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
 
        switch (reason) {
        case IB_CM_REJ_STALE_CONN:
-               kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
-                                IBLND_REJECT_CONN_STALE, NULL);
+               kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
+                                      IBLND_REJECT_CONN_STALE, NULL);
                break;
 
        case IB_CM_REJ_INVALID_SERVICE_ID:
@@ -2674,8 +2740,9 @@ kiblnd_rejected(kib_conn_t *conn, int reason, void *priv, int priv_nob)
                        case IBLND_REJECT_CONN_UNCOMPAT:
                        case IBLND_REJECT_MSG_QUEUE_SIZE:
                        case IBLND_REJECT_RDMA_FRAGS:
-                               kiblnd_reconnect(conn, rej->ibr_version,
-                                                incarnation, rej->ibr_why, cp);
+                               kiblnd_check_reconnect(conn, rej->ibr_version,
+                                                      incarnation,
+                                                      rej->ibr_why, cp);
                                break;
 
                        case IBLND_REJECT_NO_RESOURCES:
@@ -3179,9 +3246,21 @@ kiblnd_disconnect_conn(kib_conn_t *conn)
        kiblnd_peer_notify(conn->ibc_peer);
 }
 
+/**
+ * High-water for reconnection to the same peer, reconnection attempt should
+ * be delayed after trying more than KIB_RECONN_HIGH_RACE.
+ */
+#define KIB_RECONN_HIGH_RACE   10
+/**
+ * Allow connd to take a break and handle other things after consecutive
+ * reconnection attemps.
+ */
+#define KIB_RECONN_BREAK       100
+
 int
 kiblnd_connd(void *arg)
 {
+       spinlock_t *lock= &kiblnd_data.kib_connd_lock;
        wait_queue_t wait;
        unsigned long flags;
        kib_conn_t *conn;
@@ -3196,23 +3275,40 @@ kiblnd_connd(void *arg)
        init_waitqueue_entry(&wait, current);
        kiblnd_data.kib_connd = current;
 
-       spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+       spin_lock_irqsave(lock, flags);
 
        while (!kiblnd_data.kib_shutdown) {
+               int reconn = 0;
+
                dropped_lock = 0;
 
                if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
+                       kib_peer_t *peer = NULL;
+
                        conn = list_entry(kiblnd_data.kib_connd_zombies.next,
                                          kib_conn_t, ibc_list);
                        list_del(&conn->ibc_list);
+                       if (conn->ibc_reconnect) {
+                               peer = conn->ibc_peer;
+                               kiblnd_peer_addref(peer);
+                       }
 
-                       spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
-                                              flags);
+                       spin_unlock_irqrestore(lock, flags);
                        dropped_lock = 1;
 
-                       kiblnd_destroy_conn(conn);
+                       kiblnd_destroy_conn(conn, !peer);
 
-                       spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+                       spin_lock_irqsave(lock, flags);
+                       if (!peer)
+                               continue;
+
+                       conn->ibc_peer = peer;
+                       if (peer->ibp_reconnected < KIB_RECONN_HIGH_RACE)
+                               list_add_tail(&conn->ibc_list,
+                                             &kiblnd_data.kib_reconn_list);
+                       else
+                               list_add_tail(&conn->ibc_list,
+                                             &kiblnd_data.kib_reconn_wait);
                }
 
                if (!list_empty(&kiblnd_data.kib_connd_conns)) {
@@ -3220,14 +3316,38 @@ kiblnd_connd(void *arg)
                                          kib_conn_t, ibc_list);
                        list_del(&conn->ibc_list);
 
-                       spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
-                                              flags);
+                       spin_unlock_irqrestore(lock, flags);
                        dropped_lock = 1;
 
                        kiblnd_disconnect_conn(conn);
                        kiblnd_conn_decref(conn);
 
-                       spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+                       spin_lock_irqsave(lock, flags);
+               }
+
+               while (reconn < KIB_RECONN_BREAK) {
+                       if (kiblnd_data.kib_reconn_sec !=
+                           ktime_get_real_seconds()) {
+                               kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
+                               list_splice_init(&kiblnd_data.kib_reconn_wait,
+                                                &kiblnd_data.kib_reconn_list);
+                       }
+
+                       if (list_empty(&kiblnd_data.kib_reconn_list))
+                               break;
+
+                       conn = list_entry(kiblnd_data.kib_reconn_list.next,
+                                         kib_conn_t, ibc_list);
+                       list_del(&conn->ibc_list);
+
+                       spin_unlock_irqrestore(lock, flags);
+                       dropped_lock = 1;
+
+                       reconn += kiblnd_reconnect_peer(conn->ibc_peer);
+                       kiblnd_peer_decref(conn->ibc_peer);
+                       LIBCFS_FREE(conn, sizeof(*conn));
+
+                       spin_lock_irqsave(lock, flags);
                }
 
                /* careful with the jiffy wrap... */
@@ -3237,7 +3357,7 @@ kiblnd_connd(void *arg)
                        const int p = 1;
                        int chunk = kiblnd_data.kib_peer_hash_size;
 
-                       spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+                       spin_unlock_irqrestore(lock, flags);
                        dropped_lock = 1;
 
                        /*
@@ -3262,7 +3382,7 @@ kiblnd_connd(void *arg)
                        }
 
                        deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
-                       spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+                       spin_lock_irqsave(lock, flags);
                }
 
                if (dropped_lock)
@@ -3271,15 +3391,15 @@ kiblnd_connd(void *arg)
                /* Nothing to do for 'timeout'  */
                set_current_state(TASK_INTERRUPTIBLE);
                add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
-               spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+               spin_unlock_irqrestore(lock, flags);
 
                schedule_timeout(timeout);
 
                remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
-               spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+               spin_lock_irqsave(lock, flags);
        }
 
-       spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+       spin_unlock_irqrestore(lock, flags);
 
        kiblnd_thread_fini();
        return 0;