Whamcloud - gitweb
ORNL-28 recovery: rework extend_recovery_timer()
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
index 44b6130..3b96a4b 100644 (file)
@@ -30,6 +30,9 @@
  * Use is subject to license terms.
  */
 /*
+ * Copyright (c) 2011 Whamcloud, Inc.
+ */
+/*
  * This file is part of Lustre, http://www.lustre.org/
  * Lustre is a trademark of Sun Microsystems, Inc.
  */
@@ -190,6 +193,31 @@ out:
         RETURN(rc);
 }
 
+/**
+ * Find conn uuid by peer nid. @peer is a server nid. This function is used
+ * to find a conn uuid of @imp which can reach @peer.
+ */
+int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
+                            struct obd_uuid *uuid)
+{
+        struct obd_import_conn *conn;
+        int rc = -ENOENT;
+        ENTRY;
+
+        cfs_spin_lock(&imp->imp_lock);
+        cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
+                /* check if conn uuid does have this peer nid */
+                if (class_check_uuid(&conn->oic_uuid, peer)) {
+                        *uuid = conn->oic_uuid;
+                        rc = 0;
+                        break;
+                }
+        }
+        cfs_spin_unlock(&imp->imp_lock);
+        RETURN(rc);
+}
+EXPORT_SYMBOL(client_import_find_conn);
+
 void client_destroy_import(struct obd_import *imp)
 {
         /* drop security policy instance after all rpc finished/aborted
@@ -214,6 +242,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
         struct obd_uuid server_uuid;
         int rq_portal, rp_portal, connect_op;
         char *name = obddev->obd_type->typ_name;
+        ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
         int rc;
         ENTRY;
 
@@ -225,12 +254,16 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
                 connect_op = OST_CONNECT;
                 cli->cl_sp_me = LUSTRE_SP_CLI;
                 cli->cl_sp_to = LUSTRE_SP_OST;
+                ns_type = LDLM_NS_TYPE_OSC;
+
         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
                 rq_portal = MDS_REQUEST_PORTAL;
                 rp_portal = MDC_REPLY_PORTAL;
                 connect_op = MDS_CONNECT;
                 cli->cl_sp_me = LUSTRE_SP_CLI;
                 cli->cl_sp_to = LUSTRE_SP_MDT;
+                ns_type = LDLM_NS_TYPE_MDC;
+
         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
                 rq_portal = MGS_REQUEST_PORTAL;
                 rp_portal = MGC_REPLY_PORTAL;
@@ -238,6 +271,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
                 cli->cl_sp_me = LUSTRE_SP_MGC;
                 cli->cl_sp_to = LUSTRE_SP_MGS;
                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
+                ns_type = LDLM_NS_TYPE_MGC;
+
         } else {
                 CERROR("unknown client OBD type \"%s\", can't setup\n",
                        name);
@@ -366,7 +401,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 
         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
                                                    LDLM_NAMESPACE_CLIENT,
-                                                   LDLM_NAMESPACE_GREEDY);
+                                                   LDLM_NAMESPACE_GREEDY,
+                                                   ns_type);
         if (obddev->obd_namespace == NULL) {
                 CERROR("Unable to create client namespace - %s\n",
                        obddev->obd_name);
@@ -435,7 +471,7 @@ int client_connect_import(const struct lu_env *env,
                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
         }
 
-        rc = ptlrpc_connect_import(imp, NULL);
+        rc = ptlrpc_connect_import(imp);
         if (rc != 0) {
                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
                 GOTO(out_ldlm, rc);
@@ -568,18 +604,15 @@ int server_disconnect_export(struct obd_export *exp)
                                        struct ptlrpc_reply_state, rs_exp_list);
                 struct ptlrpc_service *svc = rs->rs_service;
 
-                cfs_spin_lock(&svc->srv_lock);
+                cfs_spin_lock(&svc->srv_rs_lock);
                 cfs_list_del_init(&rs->rs_exp_list);
                 cfs_spin_lock(&rs->rs_lock);
                 ptlrpc_schedule_difficult_reply(rs);
                 cfs_spin_unlock(&rs->rs_lock);
-                cfs_spin_unlock(&svc->srv_lock);
+                cfs_spin_unlock(&svc->srv_rs_lock);
         }
         cfs_spin_unlock(&exp->exp_lock);
 
-        /* release nid stat refererence */
-        lprocfs_exp_cleanup(exp);
-
         RETURN(rc);
 }
 
@@ -639,10 +672,17 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 }
 EXPORT_SYMBOL(target_client_add_cb);
 
+#ifdef __KERNEL__
 static void
-target_start_and_reset_recovery_timer(struct obd_device *obd,
-                                      struct ptlrpc_request *req,
-                                      int new_client);
+check_and_start_recovery_timer(struct obd_device *obd,
+                               struct ptlrpc_request *req, int new_client);
+#else
+static inline void
+check_and_start_recovery_timer(struct obd_device *obd,
+                               struct ptlrpc_request *req, int new_client)
+{
+}
+#endif
 
 int target_handle_connect(struct ptlrpc_request *req)
 {
@@ -658,6 +698,7 @@ int target_handle_connect(struct ptlrpc_request *req)
         int rc = 0;
         int mds_conn = 0;
         struct obd_connect_data *data, *tmpdata;
+        int size, tmpsize;
         lnet_nid_t *client_nid = NULL;
         ENTRY;
 
@@ -676,7 +717,7 @@ int target_handle_connect(struct ptlrpc_request *req)
 
         if (!target || target->obd_stopping || !target->obd_set_up) {
                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
-                                   " for connect (%s)\n", str,
+                                   "for connect (%s)\n", str,
                                    !target ? "no target" :
                                    (target->obd_stopping ? "stopping" :
                                    "not set up"));
@@ -725,6 +766,8 @@ int target_handle_connect(struct ptlrpc_request *req)
 
         conn = *tmp;
 
+        size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
+                                    RCL_CLIENT);
         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
         if (!data)
                 GOTO(out, rc = -EPROTO);
@@ -865,15 +908,10 @@ no_export:
               export, (long)cfs_time_current_sec(),
               export ? (long)export->exp_last_request_time : 0);
 
-        /* Tell the client if we're in recovery. */
-        if (target->obd_recovering) {
-                lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
-                /* If this is the first time a client connects,
-                   reset the recovery timer */
-                if (rc == 0)
-                        target_start_and_reset_recovery_timer(target, req,
-                                                              !export);
-        }
+        /* If this is the first time a client connects, reset the recovery
+         * timer */
+        if (rc == 0 && target->obd_recovering)
+                check_and_start_recovery_timer(target, req, export == NULL);
 
         /* We want to handle EALREADY but *not* -EALREADY from
          * target_handle_reconnect(), return reconnection state in a flag */
@@ -920,13 +958,21 @@ dont_check_exports:
         }
         if (rc)
                 GOTO(out, rc);
+
+        LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
+        data->ocd_instance = target->u.obt.obt_instance;
+
         /* Return only the parts of obd_connect_data that we understand, so the
          * client knows that we don't understand the rest. */
         if (data) {
-                 tmpdata = req_capsule_server_get(&req->rq_pill,
-                                                  &RMF_CONNECT_DATA);
-                  //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
-                 *tmpdata = *data;
+                tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
+                                               RCL_SERVER);
+                tmpdata = req_capsule_server_get(&req->rq_pill,
+                                                 &RMF_CONNECT_DATA);
+                /* Don't use struct assignment here, because the client reply
+                 * buffer may be smaller/larger than the local struct
+                 * obd_connect_data. */
+                memcpy(tmpdata, data, min(tmpsize, size));
         }
 
         /* If all else goes well, this is our RPC return code. */
@@ -992,28 +1038,43 @@ dont_check_exports:
                              &export->exp_nid_hash);
         }
 
-        cfs_spin_lock(&target->obd_recovery_task_lock);
         if (target->obd_recovering && !export->exp_in_recovery) {
+                int has_transno;
+                __u64 transno = data->ocd_transno;
+
                 cfs_spin_lock(&export->exp_lock);
                 export->exp_in_recovery = 1;
                 export->exp_req_replay_needed = 1;
                 export->exp_lock_replay_needed = 1;
                 cfs_spin_unlock(&export->exp_lock);
-                if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
-                     && (data->ocd_transno == 0))
+
+                has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
+                                 MSG_CONNECT_TRANSNO);
+                if (has_transno && transno == 0)
                         CWARN("Connect with zero transno!\n");
 
-                if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
-                     && data->ocd_transno < target->obd_next_recovery_transno)
-                        target->obd_next_recovery_transno = data->ocd_transno;
-                target->obd_connected_clients++;
+                if (has_transno && transno > 0 &&
+                    transno < target->obd_next_recovery_transno &&
+                    transno > target->obd_last_committed) {
+                        /* another way is to use cmpxchg() so it will be
+                         * lock free */
+                        cfs_spin_lock(&target->obd_recovery_task_lock);
+                        if (transno < target->obd_next_recovery_transno)
+                                target->obd_next_recovery_transno = transno;
+                        cfs_spin_unlock(&target->obd_recovery_task_lock);
+                }
+
                 cfs_atomic_inc(&target->obd_req_replay_clients);
                 cfs_atomic_inc(&target->obd_lock_replay_clients);
-                if (target->obd_connected_clients ==
+                if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
                     target->obd_max_recoverable_clients)
                         cfs_waitq_signal(&target->obd_next_transno_waitq);
         }
-        cfs_spin_unlock(&target->obd_recovery_task_lock);
+
+        /* Tell the client we're in recovery, when client is involved in it. */
+        if (target->obd_recovering)
+                lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
+
         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
         conn = *tmp;
 
@@ -1099,10 +1160,10 @@ void target_destroy_export(struct obd_export *exp)
         if (exp->exp_imp_reverse != NULL)
                 client_destroy_import(exp->exp_imp_reverse);
 
-        LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
-        LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
-        LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
-        LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
+        LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
+        LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
+        LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
+        LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
 }
 
 /*
@@ -1113,25 +1174,20 @@ static void target_request_copy_get(struct ptlrpc_request *req)
         class_export_rpc_get(req->rq_export);
         LASSERT(cfs_list_empty(&req->rq_list));
         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
+
         /* increase refcount to keep request in queue */
-        LASSERT(cfs_atomic_read(&req->rq_refcount));
         cfs_atomic_inc(&req->rq_refcount);
         /** let export know it has replays to be handled */
         cfs_atomic_inc(&req->rq_export->exp_replay_count);
-        /* release service thread while request is queued 
-         * we are moving the request from active processing
-         * to waiting on the replay queue */
-        ptlrpc_server_active_request_dec(req);
 }
 
 static void target_request_copy_put(struct ptlrpc_request *req)
 {
         LASSERT(cfs_list_empty(&req->rq_replay_list));
-        LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
+        LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
+
         cfs_atomic_dec(&req->rq_export->exp_replay_count);
         class_export_rpc_put(req->rq_export);
-        /* ptlrpc_server_drop_request() assumes the request is active */
-        ptlrpc_server_active_request_inc(req);
         ptlrpc_server_drop_request(req);
 }
 
@@ -1181,9 +1237,17 @@ static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
 #ifdef __KERNEL__
 static void target_finish_recovery(struct obd_device *obd)
 {
+        time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
+                                    obd->obd_recovery_start);
         ENTRY;
-        LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
-                      obd->obd_name);
+
+        LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
+                      "%d recovered and %d %s evicted.\n", obd->obd_name,
+                      (int)elapsed_time / 60, (int)elapsed_time % 60,
+                      obd->obd_max_recoverable_clients,
+                      cfs_atomic_read(&obd->obd_connected_clients),
+                      obd->obd_stale_clients,
+                      obd->obd_stale_clients == 1 ? "was" : "were");
 
         ldlm_reprocess_all_ns(obd->obd_namespace);
         cfs_spin_lock(&obd->obd_recovery_task_lock);
@@ -1254,7 +1318,6 @@ static void abort_lock_replay_queue(struct obd_device *obd)
                 target_request_copy_put(req);
         }
 }
-#endif
 
 /* Called from a cleanup function if the device is being cleaned up
    forcefully.  The exports should all have been disconnected already,
@@ -1312,56 +1375,71 @@ void target_cancel_recovery_timer(struct obd_device *obd)
         cfs_timer_disarm(&obd->obd_recovery_timer);
 }
 
-/* extend = 1 means require at least "duration" seconds left in the timer,
-   extend = 0 means set the total duration (start_recovery_timer) */
-static void reset_recovery_timer(struct obd_device *obd, int duration,
-                                 int extend)
+static void target_start_recovery_timer(struct obd_device *obd)
 {
-        cfs_time_t now = cfs_time_current_sec();
-        cfs_duration_t left;
+        if (obd->obd_recovery_start != 0)
+                return;
 
-        cfs_spin_lock(&obd->obd_recovery_task_lock);
+        cfs_spin_lock(&obd->obd_dev_lock);
         if (!obd->obd_recovering || obd->obd_abort_recovery) {
-                cfs_spin_unlock(&obd->obd_recovery_task_lock);
+                cfs_spin_unlock(&obd->obd_dev_lock);
                 return;
         }
 
-        left = cfs_time_sub(obd->obd_recovery_end, now);
+        LASSERT(obd->obd_recovery_timeout != 0);
 
-        if (extend && (duration > left))
-                obd->obd_recovery_timeout += duration - left;
-        else if (!extend && (duration > obd->obd_recovery_timeout))
-                /* Track the client's largest expected replay time */
-                obd->obd_recovery_timeout = duration;
+        if (obd->obd_recovery_start != 0) {
+                cfs_spin_unlock(&obd->obd_dev_lock);
+                return;
+        }
 
-        /* Hard limit of obd_recovery_time_hard which should not happen */
-        if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
-                obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
+        cfs_timer_arm(&obd->obd_recovery_timer,
+                      cfs_time_shift(obd->obd_recovery_timeout));
+        obd->obd_recovery_start = cfs_time_current_sec();
+        cfs_spin_unlock(&obd->obd_dev_lock);
 
-        obd->obd_recovery_end = obd->obd_recovery_start +
-                                obd->obd_recovery_timeout;
-        if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
-            cfs_time_before(now, obd->obd_recovery_end)) {
-                left = cfs_time_sub(obd->obd_recovery_end, now);
-                cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
-        }
-        cfs_spin_unlock(&obd->obd_recovery_task_lock);
-        CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
-               obd->obd_name, (unsigned)left);
+        LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
+                      "or until %d client%s reconnect%s\n",
+                      obd->obd_name,
+                      obd->obd_recovery_timeout / 60,
+                      obd->obd_recovery_timeout % 60,
+                      obd->obd_max_recoverable_clients,
+                      (obd->obd_max_recoverable_clients == 1) ? "" : "s",
+                      (obd->obd_max_recoverable_clients == 1) ? "s": "");
 }
 
-static void check_and_start_recovery_timer(struct obd_device *obd)
+/* extend recovery window to have extra @duration seconds at least. */
+static void extend_recovery_timer(struct obd_device *obd, int drt)
 {
-        cfs_spin_lock(&obd->obd_recovery_task_lock);
-        if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
-                cfs_spin_unlock(&obd->obd_recovery_task_lock);
+        cfs_time_t now;
+        cfs_time_t end;
+        cfs_duration_t left;
+
+        cfs_spin_lock(&obd->obd_dev_lock);
+        if (!obd->obd_recovering || obd->obd_abort_recovery) {
+                cfs_spin_unlock(&obd->obd_dev_lock);
                 return;
         }
-        CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
-        obd->obd_recovery_start = cfs_time_current_sec();
-        cfs_spin_unlock(&obd->obd_recovery_task_lock);
 
-        reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
+        LASSERT(obd->obd_recovery_start != 0);
+
+        now = cfs_time_current_sec();
+        end = obd->obd_recovery_start + obd->obd_recovery_timeout;
+        left = cfs_time_sub(end, now);
+        if (left < 0) {
+                obd->obd_recovery_timeout += drt - left;
+        } else if (left < drt) {
+                drt -= left;
+                obd->obd_recovery_timeout += drt;
+        } else {
+                drt = left;
+        }
+
+        cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(drt));
+        cfs_spin_unlock(&obd->obd_dev_lock);
+
+        CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
+               obd->obd_name, (unsigned)drt);
 }
 
 /* Reset the timer with each new client connection */
@@ -1376,11 +1454,13 @@ static void check_and_start_recovery_timer(struct obd_device *obd)
  */
 
 static void
-target_start_and_reset_recovery_timer(struct obd_device *obd,
-                                      struct ptlrpc_request *req,
-                                      int new_client)
+check_and_start_recovery_timer(struct obd_device *obd,
+                               struct ptlrpc_request *req,
+                               int new_client)
 {
         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
+        struct obd_device_target *obt = &obd->u.obt;
+        struct lustre_sb_info *lsi;
 
         if (!new_client && service_time)
                 /* Teach server about old server's estimates, as first guess
@@ -1388,7 +1468,7 @@ target_start_and_reset_recovery_timer(struct obd_device *obd,
                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
                             service_time);
 
-        check_and_start_recovery_timer(obd);
+        target_start_recovery_timer(obd);
 
         /* convert the service time to rpc timeout,
          * reuse service_time to limit stack usage */
@@ -1397,14 +1477,18 @@ target_start_and_reset_recovery_timer(struct obd_device *obd,
         /* We expect other clients to timeout within service_time, then try
          * to reconnect, then try the failover server.  The max delay between
          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
-        service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
-                             INITIAL_CONNECT_TIMEOUT);
-        if (service_time > obd->obd_recovery_timeout && !new_client)
-                reset_recovery_timer(obd, service_time, 0);
+        service_time += 2 * INITIAL_CONNECT_TIMEOUT;
+
+        LASSERT(obt->obt_magic == OBT_MAGIC);
+        lsi = s2lsi(obt->obt_sb);
+        if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
+                service_time += 2 * (CONNECTION_SWITCH_MAX +
+                                     CONNECTION_SWITCH_INC);
+        service_time -= obd->obd_recovery_timeout;
+        if (service_time > 0)
+                extend_recovery_timer(obd, service_time);
 }
 
-#ifdef __KERNEL__
-
 /** Health checking routines */
 static inline int exp_connect_healthy(struct obd_export *exp)
 {
@@ -1437,12 +1521,13 @@ static inline int exp_finished(struct obd_export *exp)
 /** Checking routines for recovery */
 static int check_for_clients(struct obd_device *obd)
 {
+        unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
+
         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
                 return 1;
-        LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
+        LASSERT(clnts <= obd->obd_max_recoverable_clients);
         if (obd->obd_no_conn == 0 &&
-            obd->obd_connected_clients + obd->obd_stale_clients ==
-            obd->obd_max_recoverable_clients)
+            clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
                 return 1;
         return 0;
 }
@@ -1463,7 +1548,7 @@ static int check_for_next_transno(struct obd_device *obd)
                 req_transno = 0;
         }
 
-        connected = obd->obd_connected_clients;
+        connected = cfs_atomic_read(&obd->obd_connected_clients);
         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
         queue_len = obd->obd_requests_queued_for_recovery;
         next_transno = obd->obd_next_recovery_transno;
@@ -1567,7 +1652,7 @@ repeat:
                  * reset timer, recovery will proceed with versions now,
                  * timeout is set just to handle reconnection delays
                  */
-                reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
+                extend_recovery_timer(obd, RECONNECT_DELAY_MAX);
                 /** Wait for recovery events again, after evicting bad clients */
                 goto repeat;
         }
@@ -1685,13 +1770,15 @@ static int handle_recovery_req(struct ptlrpc_thread *thread,
         lu_context_fini(&req->rq_recov_session);
         /* don't reset timer for final stage */
         if (!exp_finished(req->rq_export)) {
+                int to = obd_timeout;
+
                 /**
                  * Add request timeout to the recovery time so next request from
                  * this client may come in recovery time
                  */
-                 reset_recovery_timer(class_exp2obd(req->rq_export),
-                                      AT_OFF ? obd_timeout :
-                                      lustre_msg_get_timeout(req->rq_reqmsg), 1);
+                if (!AT_OFF)
+                        to = lustre_msg_get_timeout(req->rq_reqmsg);
+                 extend_recovery_timer(class_exp2obd(req->rq_export), to);
         }
 reqcopy_put:
         RETURN(rc);
@@ -1726,7 +1813,7 @@ static int target_recovery_thread(void *arg)
         env.le_ctx.lc_thread = thread;
         thread->t_data = NULL;
 
-        CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
+        CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
                cfs_curproc_pid());
         trd->trd_processing_task = cfs_curproc_pid();
 
@@ -1833,7 +1920,7 @@ static int target_start_recovery_thread(struct lu_target *lut,
         cfs_init_completion(&trd->trd_finishing);
         trd->trd_recovery_handler = handler;
 
-        if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
+        if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
                 cfs_wait_for_completion(&trd->trd_starting);
                 LASSERT(obd->obd_recovering != 0);
         } else
@@ -1873,7 +1960,7 @@ static void target_recovery_expired(unsigned long castmeharder)
                " after %lds (%d clients connected)\n",
                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
                cfs_time_current_sec()- obd->obd_recovery_start,
-               obd->obd_connected_clients);
+               cfs_atomic_read(&obd->obd_connected_clients));
 
         obd->obd_recovery_expired = 1;
         cfs_waitq_signal(&obd->obd_next_transno_waitq);
@@ -1896,11 +1983,6 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
         obd->obd_recovery_start = 0;
         obd->obd_recovery_end = 0;
 
-        /* both values can be get from mount data already */
-        if (obd->obd_recovery_timeout == 0)
-                obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
-        if (obd->obd_recovery_time_hard == 0)
-                obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
         target_start_recovery_thread(lut, handler);
 }
@@ -1915,25 +1997,29 @@ static int target_process_req_flags(struct obd_device *obd,
         LASSERT(exp != NULL);
         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
                 /* client declares he's ready to replay locks */
+                cfs_spin_lock(&exp->exp_lock);
                 if (exp->exp_req_replay_needed) {
-                        LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) >
-                                0);
-                        cfs_spin_lock(&exp->exp_lock);
                         exp->exp_req_replay_needed = 0;
                         cfs_spin_unlock(&exp->exp_lock);
+
+                        LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
                         cfs_atomic_dec(&obd->obd_req_replay_clients);
+                } else {
+                        cfs_spin_unlock(&exp->exp_lock);
                 }
         }
         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
                 /* client declares he's ready to complete recovery
                  * so, we put the request on th final queue */
+                cfs_spin_lock(&exp->exp_lock);
                 if (exp->exp_lock_replay_needed) {
-                        LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) >
-                                0);
-                        cfs_spin_lock(&exp->exp_lock);
                         exp->exp_lock_replay_needed = 0;
                         cfs_spin_unlock(&exp->exp_lock);
+
+                        LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
+                } else {
+                        cfs_spin_unlock(&exp->exp_lock);
                 }
         }
         return 0;
@@ -2185,7 +2271,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
 
         netrc = target_send_reply_msg (req, rc, fail_id);
 
-        cfs_spin_lock(&svc->srv_lock);
+        cfs_spin_lock(&svc->srv_rs_lock);
 
         cfs_atomic_inc(&svc->srv_n_difficult_replies);
 
@@ -2197,7 +2283,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
                  * reply_out_callback leaves alone) */
                 rs->rs_on_net = 0;
                 ptlrpc_rs_addref(rs);
-                cfs_atomic_inc (&svc->srv_outstanding_replies);
         }
 
         cfs_spin_lock(&rs->rs_lock);
@@ -2212,7 +2297,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
         }
         cfs_spin_unlock(&rs->rs_lock);
-        cfs_spin_unlock(&svc->srv_lock);
+        cfs_spin_unlock(&svc->srv_rs_lock);
         EXIT;
 }
 
@@ -2456,3 +2541,107 @@ void ldlm_dump_export_locks(struct obd_export *exp)
         cfs_spin_unlock(&exp->exp_locks_list_guard);
 }
 #endif
+
+static int target_bulk_timeout(void *data)
+{
+        ENTRY;
+        /* We don't fail the connection here, because having the export
+         * killed makes the (vital) call to commitrw very sad.
+         */
+        RETURN(1);
+}
+
+static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
+{
+        return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
+}
+
+int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
+                   struct l_wait_info *lwi)
+{
+        struct ptlrpc_request *req = desc->bd_req;
+        int rc = 0;
+        ENTRY;
+
+        /* Check if there is eviction in progress, and if so, wait for
+         * it to finish */
+        if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
+                *lwi = LWI_INTR(NULL, NULL);
+                rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
+                                  !cfs_atomic_read(&exp->exp_obd->
+                                                   obd_evict_inprogress),
+                                  lwi);
+        }
+
+        /* Check if client was evicted or tried to reconnect already */
+        if (exp->exp_failed || exp->exp_abort_active_req) {
+                rc = -ENOTCONN;
+        } else {
+                if (desc->bd_type == BULK_PUT_SINK)
+                        rc = sptlrpc_svc_wrap_bulk(req, desc);
+                if (rc == 0)
+                        rc = ptlrpc_start_bulk_transfer(desc);
+        }
+
+        if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
+                ptlrpc_abort_bulk(desc);
+        } else if (rc == 0) {
+                time_t start = cfs_time_current_sec();
+                do {
+                        long timeoutl = req->rq_deadline - cfs_time_current_sec();
+                        cfs_duration_t timeout = timeoutl <= 0 ?
+                                CFS_TICK : cfs_time_seconds(timeoutl);
+                        *lwi = LWI_TIMEOUT_INTERVAL(timeout,
+                                                    cfs_time_seconds(1),
+                                                   target_bulk_timeout,
+                                                   desc);
+                        rc = l_wait_event(desc->bd_waitq,
+                                          !ptlrpc_server_bulk_active(desc) ||
+                                          exp->exp_failed ||
+                                          exp->exp_abort_active_req,
+                                          lwi);
+                        LASSERT(rc == 0 || rc == -ETIMEDOUT);
+                        /* Wait again if we changed deadline */
+                } while ((rc == -ETIMEDOUT) &&
+                         (req->rq_deadline > cfs_time_current_sec()));
+
+                if (rc == -ETIMEDOUT) {
+                        DEBUG_REQ(D_ERROR, req,
+                                  "timeout on bulk %s after %ld%+lds",
+                                  bulk2type(desc),
+                                  req->rq_deadline - start,
+                                  cfs_time_current_sec() -
+                                  req->rq_deadline);
+                        ptlrpc_abort_bulk(desc);
+                } else if (exp->exp_failed) {
+                        DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
+                                  bulk2type(desc));
+                        rc = -ENOTCONN;
+                        ptlrpc_abort_bulk(desc);
+                } else if (exp->exp_abort_active_req) {
+                        DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
+                                  bulk2type(desc));
+                        /* we don't reply anyway */
+                        rc = -ETIMEDOUT;
+                        ptlrpc_abort_bulk(desc);
+                } else if (!desc->bd_success ||
+                           desc->bd_nob_transferred != desc->bd_nob) {
+                        DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
+                                  desc->bd_success ?
+                                  "truncated" : "network error on",
+                                  bulk2type(desc),
+                                  desc->bd_nob_transferred,
+                                  desc->bd_nob);
+                        /* XXX should this be a different errno? */
+                        rc = -ETIMEDOUT;
+                } else if (desc->bd_type == BULK_GET_SINK) {
+                        rc = sptlrpc_svc_unwrap_bulk(req, desc);
+                }
+        } else {
+                DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
+                          bulk2type(desc), rc);
+        }
+
+        RETURN(rc);
+}
+EXPORT_SYMBOL(target_bulk_io);