Whamcloud - gitweb
- fixed using of few deprected functions:
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
index 11431c9..d45f1e4 100644 (file)
 #include <linux/lustre_mgmt.h>
 #include <linux/lustre_dlm.h>
 #include <linux/lustre_net.h>
+#include <linux/lustre_sec.h>
+
+/* @priority: if non-zero, move the selected to the list head
+ * @nocreate: if non-zero, only search in existed connections
+ */
+static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
+                           int priority, int nocreate)
+{
+        struct ptlrpc_connection *ptlrpc_conn;
+        struct obd_import_conn *imp_conn = NULL, *item;
+        int rc = 0;
+        ENTRY;
+
+        LASSERT(!(nocreate && !priority));
+
+        ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
+        if (!ptlrpc_conn) {
+                CERROR("can't find connection %s\n", uuid->uuid);
+                RETURN (-EINVAL);
+        }
+
+        if (!nocreate) {
+                OBD_ALLOC(imp_conn, sizeof(*imp_conn));
+                if (!imp_conn) {
+                        CERROR("fail to alloc memory\n");
+                        GOTO(out_put, rc = -ENOMEM);
+                }
+        }
+
+        spin_lock(&imp->imp_lock);
+        list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
+                if (obd_uuid_equals(uuid, &item->oic_uuid)) {
+                        if (priority) {
+                                list_del(&item->oic_item);
+                                list_add(&item->oic_item, &imp->imp_conn_list);
+                                item->oic_last_attempt = 0;
+                        }
+                        CDEBUG(D_HA, "imp %p@%s: find existed conn %s%s\n",
+                               imp, imp->imp_obd->obd_name, uuid->uuid,
+                               (priority ? ", move to head." : ""));
+                        spin_unlock(&imp->imp_lock);
+                        GOTO(out_free, rc = 0);
+                }
+        }
+        /* not found */
+        if (!nocreate) {
+                imp_conn->oic_conn = ptlrpc_conn;
+                imp_conn->oic_uuid = *uuid;
+                imp_conn->oic_last_attempt = 0;
+                if (priority)
+                        list_add(&imp_conn->oic_item, &imp->imp_conn_list);
+                else
+                        list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
+                CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
+                       imp, imp->imp_obd->obd_name, uuid->uuid,
+                       (priority ? "head" : "tail"));
+        } else
+                rc = -ENOENT;
+
+        spin_unlock(&imp->imp_lock);
+        RETURN(0);
+out_free:
+        if (imp_conn)
+                OBD_FREE(imp_conn, sizeof(*imp_conn));
+out_put:
+        ptlrpc_put_connection(ptlrpc_conn);
+        RETURN(rc);
+}
+
+int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
+{
+        return import_set_conn(imp, uuid, 1, 1);
+}
+
+int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
+                           int priority)
+{
+        return import_set_conn(imp, uuid, priority, 0);
+}
+
+int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
+{
+        struct obd_import_conn *imp_conn;
+        struct obd_export *dlmexp;
+        int rc = -ENOENT;
+        ENTRY;
+
+        spin_lock(&imp->imp_lock);
+        if (list_empty(&imp->imp_conn_list)) {
+                LASSERT(!imp->imp_conn_current);
+                LASSERT(!imp->imp_connection);
+                GOTO(out, rc);
+        }
+
+        list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
+                if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
+                        continue;
+                LASSERT(imp_conn->oic_conn);
+
+                /* is current conn? */
+                if (imp_conn == imp->imp_conn_current) {
+                        LASSERT(imp_conn->oic_conn == imp->imp_connection);
+
+                        if (imp->imp_state != LUSTRE_IMP_CLOSED &&
+                            imp->imp_state != LUSTRE_IMP_DISCON) {
+                                CERROR("can't remove current connection\n");
+                                GOTO(out, rc = -EBUSY);
+                        }
+
+                        ptlrpc_put_connection(imp->imp_connection);
+                        imp->imp_connection = NULL;
+
+                        dlmexp = class_conn2export(&imp->imp_dlm_handle);
+                        if (dlmexp && dlmexp->exp_connection) {
+                                LASSERT(dlmexp->exp_connection ==
+                                        imp_conn->oic_conn);
+                                ptlrpc_put_connection(dlmexp->exp_connection);
+                                dlmexp->exp_connection = NULL;
+                        }
+                }
+
+                list_del(&imp_conn->oic_item);
+                ptlrpc_put_connection(imp_conn->oic_conn);
+                OBD_FREE(imp_conn, sizeof(*imp_conn));
+                CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
+                       imp, imp->imp_obd->obd_name, uuid->uuid);
+                rc = 0;
+                break;
+        }
+out:
+        spin_unlock(&imp->imp_lock);
+        if (rc == -ENOENT)
+                CERROR("connection %s not found\n", uuid->uuid);
+        RETURN(rc);
+}
 
 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 {
-        struct ptlrpc_connection *conn;
         struct lustre_cfg* lcfg = buf;
         struct client_obd *cli = &obddev->u.cli;
         struct obd_import *imp;
@@ -98,19 +232,32 @@ int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 
         cli->cl_dirty = 0;
         cli->cl_avail_grant = 0;
+        /* FIXME: should limit this for the sum of all cl_dirty_max */
         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
+        if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
+                cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
         INIT_LIST_HEAD(&cli->cl_cache_waiters);
         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
         INIT_LIST_HEAD(&cli->cl_loi_write_list);
         INIT_LIST_HEAD(&cli->cl_loi_read_list);
         spin_lock_init(&cli->cl_loi_list_lock);
-        cli->cl_brw_in_flight = 0;
+        cli->cl_r_in_flight = 0;
+        cli->cl_w_in_flight = 0;
         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
-        cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
-        cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
+
+        if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
+                cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
+                cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
+        } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
+                cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
+                cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
+        } else {
+                cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
+                cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
+        }
 
         rc = ldlm_get_ref();
         if (rc) {
@@ -118,19 +265,12 @@ int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
                 GOTO(err, rc);
         }
 
-        conn = ptlrpc_uuid_to_connection(&server_uuid);
-        if (conn == NULL)
-                GOTO(err_ldlm, rc = -ENOENT);
-
         ptlrpc_init_client(rq_portal, rp_portal, name,
                            &obddev->obd_ldlm_client);
 
         imp = class_new_import();
-        if (imp == NULL) {
-                ptlrpc_put_connection(conn);
+        if (imp == NULL) 
                 GOTO(err_ldlm, rc = -ENOENT);
-        }
-        imp->imp_connection = conn;
         imp->imp_client = &obddev->obd_ldlm_client;
         imp->imp_obd = obddev;
         imp->imp_connect_op = connect_op;
@@ -141,6 +281,12 @@ int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
               lcfg->lcfg_inllen1);
         class_import_put(imp);
 
+        rc = client_import_add_conn(imp, &server_uuid, 1);
+        if (rc) {
+                CERROR("can't add initial connection\n");
+                GOTO(err_import, rc);
+        }
+
         cli->cl_import = imp;
         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
@@ -172,14 +318,14 @@ int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
                         GOTO(err_import, rc = -ENOSYS);
                 }
 
-                register_f = inter_module_get("mgmtcli_register_for_events");
+                register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
                 if (!register_f) {
                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
                         GOTO(err_import, rc = -ENOSYS);
                 }
 
                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
-                inter_module_put("mgmtcli_register_for_events");
+                symbol_put("mgmtcli_register_for_events");
 
                 if (!rc)
                         cli->cl_mgmtcli_obd = mgmt_obd;
@@ -199,17 +345,25 @@ err:
 int client_obd_cleanup(struct obd_device *obddev, int flags)
 {
         struct client_obd *cli = &obddev->u.cli;
+        ENTRY;
 
         if (!cli->cl_import)
                 RETURN(-EINVAL);
         if (cli->cl_mgmtcli_obd) {
                 mgmtcli_deregister_for_events_t dereg_f;
 
-                dereg_f = inter_module_get("mgmtcli_deregister_for_events");
+                dereg_f = symbol_get("mgmtcli_deregister_for_events");
                 dereg_f(cli->cl_mgmtcli_obd, obddev);
-                inter_module_put("mgmtcli_deregister_for_events");
+                symbol_put("mgmtcli_deregister_for_events");
         }
+
+        /* Here we try to drop the security structure after destroy import,
+         * to avoid issue of "sleep in spinlock".
+         */
+        class_import_get(cli->cl_import);
         class_destroy_import(cli->cl_import);
+        ptlrpcs_import_drop_sec(cli->cl_import);
+        class_import_put(cli->cl_import);
         cli->cl_import = NULL;
 
         ldlm_put_ref(flags & OBD_OPT_FORCE);
@@ -219,7 +373,9 @@ int client_obd_cleanup(struct obd_device *obddev, int flags)
 
 int client_connect_import(struct lustre_handle *dlm_handle,
                           struct obd_device *obd,
-                          struct obd_uuid *cluuid)
+                          struct obd_uuid *cluuid,
+                          struct obd_connect_data *conn_data,
+                          unsigned long connect_flags)
 {
         struct client_obd *cli = &obd->u.cli;
         struct obd_import *imp = cli->cl_import;
@@ -244,18 +400,25 @@ int client_connect_import(struct lustre_handle *dlm_handle,
         if (obd->obd_namespace == NULL)
                 GOTO(out_disco, rc = -ENOMEM);
 
+        rc = ptlrpcs_import_get_sec(imp);
+        if (rc != 0)
+                GOTO(out_ldlm, rc);
+
         imp->imp_dlm_handle = *dlm_handle;
         rc = ptlrpc_init_import(imp);
         if (rc != 0) 
                 GOTO(out_ldlm, rc);
 
-        exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
+        imp->imp_connect_flags = connect_flags;
+        if (conn_data)
+                memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
+
         rc = ptlrpc_connect_import(imp, NULL);
         if (rc != 0) {
                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
                 GOTO(out_ldlm, rc);
         }
-
+        LASSERT(exp->exp_connection);
         ptlrpc_pinger_add_import(imp);
         EXIT;
 
@@ -274,7 +437,7 @@ out_sem:
         return rc;
 }
 
-int client_disconnect_export(struct obd_export *exp, int failover)
+int client_disconnect_export(struct obd_export *exp, unsigned long flags)
 {
         struct obd_device *obd = class_exp2obd(exp);
         struct client_obd *cli = &obd->u.cli;
@@ -312,7 +475,9 @@ int client_disconnect_export(struct obd_export *exp, int failover)
                 obd->obd_namespace = NULL;
         }
 
-        /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
+        /* 
+         * Yeah, obd_no_recov also (mainly) means "forced shutdown".
+         */
         if (obd->obd_no_recov)
                 ptlrpc_invalidate_import(imp, 0);
         else
@@ -333,19 +498,16 @@ int client_disconnect_export(struct obd_export *exp, int failover)
  * -------------------------------------------------------------------------- */
 
 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
-                            struct obd_uuid *cluuid)
+                            struct obd_uuid *cluuid, int initial_conn)
 {
-        if (exp->exp_connection) {
+        if (exp->exp_connection && !initial_conn) {
                 struct lustre_handle *hdl;
                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
                 /* Might be a re-connect after a partition. */
-#warning "FIXME ASAP"
-                memcpy(&hdl->cookie, &conn->cookie, sizeof(conn->cookie));
-                if (1 || !memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
+                if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
                         CERROR("%s reconnecting\n", cluuid->uuid);
                         conn->cookie = exp->exp_handle.h_cookie;
-                        /*RETURN(EALREADY);*/
-                        RETURN(0);
+                        RETURN(EALREADY);
                 } else {
                         CERROR("%s reconnecting from %s, "
                                "handle mismatch (ours "LPX64", theirs "
@@ -364,8 +526,17 @@ int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
         RETURN(0);
 }
 
-int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
+static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
+{
+        ptl_process_id_t myid;
+
+        PtlGetId(peer->peer_ni->pni_ni_h, &myid);
+        return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
+}
+
+int target_handle_connect(struct ptlrpc_request *req)
 {
+        unsigned long connect_flags = 0, *cfp;
         struct obd_device *target;
         struct obd_export *export = NULL;
         struct obd_import *revimp;
@@ -374,15 +545,21 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
         struct obd_uuid cluuid;
         struct obd_uuid remote_uuid;
         struct list_head *p;
+        struct obd_connect_data *conn_data;
+        int conn_data_size = sizeof(*conn_data);
         char *str, *tmp;
-        int rc = 0, abort_recovery;
+        int rc = 0;
         unsigned long flags;
+        int initial_conn = 0;
+        char peer_str[PTL_NALFMT_SIZE];
+        const int offset = 1;
         ENTRY;
 
         OBD_RACE(OBD_FAIL_TGT_CONN_RACE); 
 
-        LASSERT_REQSWAB (req, 0);
-        str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
+        LASSERT_REQSWAB (req, offset + 0);
+        str = lustre_msg_string(req->rq_reqmsg, offset + 0,
+                                sizeof(tgtuuid) - 1);
         if (str == NULL) {
                 CERROR("bad target UUID for connect\n");
                 GOTO(out, rc = -EINVAL);
@@ -390,18 +567,17 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 
         obd_str2uuid (&tgtuuid, str);
         target = class_uuid2obd(&tgtuuid);
-        if (!target) {
+        if (!target)
                 target = class_name2obd(str);
-        }
         
         if (!target || target->obd_stopping || !target->obd_set_up) {
-                CERROR("UUID '%s' is not available for connect\n", str);
-
+                CERROR("UUID '%s' is not available for connect from %s\n",
+                       str, req->rq_peerstr);
                 GOTO(out, rc = -ENODEV);
         }
 
-        LASSERT_REQSWAB (req, 1);
-        str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
+        LASSERT_REQSWAB (req, offset + 1);
+        str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
         if (str == NULL) {
                 CERROR("bad client UUID for connect\n");
                 GOTO(out, rc = -EINVAL);
@@ -414,32 +590,38 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
                 /* NB the casts only avoid compiler warnings */
         case 8:
                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
-                         "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_nid);
+                         "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
                 break;
         case 4:
                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
-                         "NET_%x_UUID", (__u32)req->rq_peer.peer_nid);
+                         "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
                 break;
         default:
                 LBUG();
         }
 
-        spin_lock_bh(&target->obd_processing_task_lock);
-        abort_recovery = target->obd_abort_recovery;
-        spin_unlock_bh(&target->obd_processing_task_lock);
-        if (abort_recovery)
-                target_abort_recovery(target);
-
-        tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
+        tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof conn);
         if (tmp == NULL)
                 GOTO(out, rc = -EPROTO);
 
         memcpy(&conn, tmp, sizeof conn);
 
-        rc = lustre_pack_reply(req, 0, NULL, NULL);
+        cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
+        LASSERT(cfp != NULL);
+        connect_flags = *cfp;
+
+        conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
+                                       lustre_swab_connect);
+        if (!conn_data)
+                GOTO(out, rc = -EPROTO);
+
+        rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
         if (rc)
                 GOTO(out, rc);
-
+        
+        if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
+                initial_conn = 1;
+        
         /* lctl gets a backstage, all-access pass. */
         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
                 goto dont_check_exports;
@@ -451,7 +633,8 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
                         spin_unlock(&target->obd_dev_lock);
                         LASSERT(export->exp_obd == target);
 
-                        rc = target_handle_reconnect(&conn, export, &cluuid);
+                        rc = target_handle_reconnect(&conn, export, &cluuid,
+                                                     initial_conn);
                         break;
                 }
                 export = NULL;
@@ -459,36 +642,65 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
         /* If we found an export, we already unlocked. */
         if (!export) {
                 spin_unlock(&target->obd_dev_lock);
-        } else if (req->rq_reqmsg->conn_cnt == 1) {
+        } else if (req->rq_export == NULL && 
+                   atomic_read(&export->exp_rpc_count) > 0) {
+                CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
+                      target->obd_name, cluuid.uuid,
+                      ptlrpc_peernid2str(&req->rq_peer, peer_str),
+                      export, atomic_read(&export->exp_refcount));
+                GOTO(out, rc = -EBUSY);
+        } else if (req->rq_export != NULL &&
+                   atomic_read(&export->exp_rpc_count) > 1) {
+                CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
+                      target->obd_name, cluuid.uuid,
+                      ptlrpc_peernid2str(&req->rq_peer, peer_str),
+                      export, atomic_read(&export->exp_rpc_count));
+                GOTO(out, rc = -EBUSY);
+        } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
                        cluuid.uuid);
-#warning "FIXME ASAP"
-                /*GOTO(out, rc = -EALREADY);*/
+                GOTO(out, rc = -EALREADY);
         }
 
         /* Tell the client if we're in recovery. */
         /* If this is the first client, start the recovery timer */
+        CWARN("%s: connection from %s@%s/%lu %s\n", target->obd_name, cluuid.uuid,
+              ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
+              target->obd_recovering ? "(recovering)" : "");
+
         if (target->obd_recovering) {
                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
-                target_start_recovery_timer(target, handler);
+                target_start_recovery_timer(target);
         }
+
 #if 0
         /* Tell the client if we support replayable requests */
         if (target->obd_replayable)
                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 #endif
+
         if (export == NULL) {
                 if (target->obd_recovering) {
-                        CERROR("denying connection for new client %s: "
-                               "%d clients in recovery for %lds\n", cluuid.uuid,
+                        CERROR("%s denying connection for new client %s@%s: "
+                               "%d clients in recovery for %lds\n", target->obd_name, 
+                               cluuid.uuid,
+                               ptlrpc_peernid2str(&req->rq_peer, peer_str),
                                target->obd_recoverable_clients,
                                (target->obd_recovery_timer.expires-jiffies)/HZ);
                         rc = -EBUSY;
                 } else {
  dont_check_exports:
-                        rc = obd_connect(&conn, target, &cluuid);
+                        rc = obd_connect(&conn, target, &cluuid, conn_data,
+                                         connect_flags);
                 }
         }
+
+        /* Return only the parts of obd_connect_data that we understand, so the
+         * client knows that we don't understand the rest. */
+        conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
+        memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
+               sizeof(*conn_data));
+
         /* Tell the client if we support replayable requests */
         if (target->obd_replayable)
                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
@@ -499,11 +711,6 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
         if (rc && rc != EALREADY)
                 GOTO(out, rc);
 
-        /* XXX track this all the time? */
-        if (target->obd_recovering) {
-                target->obd_connected_clients++;
-        }
-
         req->rq_repmsg->handle = conn;
 
         /* If the client and the server are the same node, we will already
@@ -521,14 +728,16 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
         LASSERT(export != NULL);
 
         spin_lock_irqsave(&export->exp_lock, flags);
-#warning "FIXME ASAP"
-        if (0 && export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
-                CERROR("%s: already connected at a higher conn_cnt: %d > %d\n",
-                       cluuid.uuid, export->exp_conn_cnt, 
+        if (initial_conn) {
+                req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
+        } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
+                CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
+                       cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
+                       export->exp_conn_cnt, 
                        req->rq_reqmsg->conn_cnt);
                 spin_unlock_irqrestore(&export->exp_lock, flags);
                 GOTO(out, rc = -EALREADY);
-        }
+        } 
         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
         spin_unlock_irqrestore(&export->exp_lock, flags);
 
@@ -536,6 +745,15 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
                 export->exp_libclient = 1;
 
+        if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
+            ptlrpc_peer_is_local(&req->rq_peer)) {
+                CWARN("%s: exp %p set sync\n", target->obd_name, export);
+                export->exp_sync = 1;
+        } else {
+                CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
+                export->exp_sync = 0;
+        }
+
         if (export->exp_connection != NULL)
                 ptlrpc_put_connection(export->exp_connection);
         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
@@ -547,19 +765,47 @@ int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
                 GOTO(out, rc = 0);
         }
 
-        memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
-               sizeof conn);
+        if (target->obd_recovering)
+                target->obd_connected_clients++;
 
-        if (export->exp_imp_reverse != NULL)
+        memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
+               sizeof(conn));
+
+        if (export->exp_imp_reverse != NULL) {
+                /* same logic as client_obd_cleanup */
+                class_import_get(export->exp_imp_reverse);
                 class_destroy_import(export->exp_imp_reverse);
+                ptlrpcs_import_drop_sec(export->exp_imp_reverse);
+                class_import_put(export->exp_imp_reverse);
+        }
+
+        /* for the rest part, we return -ENOTCONN in case of errors
+         * in order to let client initialize connection again.
+         */
         revimp = export->exp_imp_reverse = class_new_import();
+        if (!revimp) {
+                CERROR("fail to alloc new reverse import.\n");
+                GOTO(out, rc = -ENOTCONN);
+        }
+
         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
         revimp->imp_remote_handle = conn;
         revimp->imp_obd = target;
         revimp->imp_dlm_fake = 1;
         revimp->imp_state = LUSTRE_IMP_FULL;
+
+        rc = ptlrpcs_import_get_sec(revimp);
+        if (rc) {
+                CERROR("reverse import can not get sec: %d\n", rc);
+                class_destroy_import(revimp);
+                export->exp_imp_reverse = NULL;
+                GOTO(out, rc = -ENOTCONN);
+        }
+
         class_import_put(revimp);
+
+        rc = obd_connect_post(export, connect_flags);
 out:
         if (rc)
                 req->rq_status = rc;
@@ -586,8 +832,10 @@ void target_destroy_export(struct obd_export *exp)
 {
         /* exports created from last_rcvd data, and "fake"
            exports created by lctl don't have an import */
-        if (exp->exp_imp_reverse != NULL)
+        if (exp->exp_imp_reverse != NULL) {
+                ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
                 class_destroy_import(exp->exp_imp_reverse);
+        }
 
         /* We cancel locks at disconnect time, but this will catch any locks
          * granted in a race with recovery-induced disconnect. */
@@ -599,21 +847,86 @@ void target_destroy_export(struct obd_export *exp)
  * Recovery functions
  */
 
-static void abort_delayed_replies(struct obd_device *obd)
+struct ptlrpc_request *
+ptlrpc_clone_req( struct ptlrpc_request *orig_req) 
+{
+        struct ptlrpc_request *copy_req;
+        struct lustre_msg *copy_reqmsg;
+
+        OBD_ALLOC(copy_req, sizeof *copy_req);
+        if (!copy_req)
+                return NULL;
+        OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
+        if (!copy_reqmsg){
+                OBD_FREE(copy_req, sizeof *copy_req);
+                return NULL;
+        }
+
+        memcpy(copy_req, orig_req, sizeof *copy_req);
+        memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
+        /* the copied req takes over the reply state and security data */
+        orig_req->rq_reply_state = NULL;
+        orig_req->rq_sec_svcdata = NULL;
+
+        copy_req->rq_reqmsg = copy_reqmsg;
+        class_export_get(copy_req->rq_export);
+        INIT_LIST_HEAD(&copy_req->rq_list);
+
+        return copy_req;
+}
+
+void ptlrpc_free_clone( struct ptlrpc_request *req) 
+{
+        if (req->rq_svcsec)
+                svcsec_cleanup_req(req);
+
+        class_export_put(req->rq_export);
+        list_del(&req->rq_list);
+        OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
+        OBD_FREE(req, sizeof *req);
+}
+
+static void target_release_saved_req(struct ptlrpc_request *req)
+{
+        if (req->rq_svcsec)
+                svcsec_cleanup_req(req);
+
+        class_export_put(req->rq_export);
+        OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
+        OBD_FREE(req, sizeof *req);
+}
+
+static void target_finish_recovery(struct obd_device *obd)
 {
-        struct ptlrpc_request *req;
         struct list_head *tmp, *n;
+        int rc;
+
+        CWARN("%s: sending delayed replies to recovered clients\n",
+              obd->obd_name);
+
+        ldlm_reprocess_all_ns(obd->obd_namespace);
+
+        /* when recovery finished, cleanup orphans on mds and ost */
+        if (OBT(obd) && OBP(obd, postrecov)) {
+                rc = OBP(obd, postrecov)(obd);
+                if (rc >= 0)
+                        CWARN("%s: all clients recovered, %d MDS "
+                              "orphans deleted\n", obd->obd_name, rc);
+                else
+                        CERROR("postrecov failed %d\n", rc);
+        }
+
+
         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
+                struct ptlrpc_request *req;
                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
-                DEBUG_REQ(D_ERROR, req, "aborted:");
-                req->rq_status = -ENOTCONN;
-                req->rq_type = PTL_RPC_MSG_ERR;
-                ptlrpc_reply(req);
-                class_export_put(req->rq_export);
                 list_del(&req->rq_list);
-                OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
-                OBD_FREE(req, sizeof *req);
+                DEBUG_REQ(D_ERROR, req, "delayed:");
+                ptlrpc_reply(req);
+                target_release_saved_req(req);
         }
+        obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
+        return;
 }
 
 static void abort_recovery_queue(struct obd_device *obd)
@@ -624,6 +937,7 @@ static void abort_recovery_queue(struct obd_device *obd)
 
         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
+                list_del(&req->rq_list);
                 DEBUG_REQ(D_ERROR, req, "aborted:");
                 req->rq_status = -ENOTCONN;
                 req->rq_type = PTL_RPC_MSG_ERR;
@@ -634,45 +948,62 @@ static void abort_recovery_queue(struct obd_device *obd)
                         DEBUG_REQ(D_ERROR, req,
                                   "packing failed for abort-reply; skipping");
                 }
-                list_del(&req->rq_list);
-                class_export_put(req->rq_export);
-                OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
-                OBD_FREE(req, sizeof *req);
+                target_release_saved_req(req);
         }
 }
 
-void target_abort_recovery(void *data)
+/* Called from a cleanup function if the device is being cleaned up
+   forcefully.  The exports should all have been disconnected already,
+   the only thing left to do is
+     - clear the recovery flags
+     - cancel the timer
+     - free queued requests and replies, but don't send replies
+   Because the obd_stopping flag is set, no new requests should be received.
+
+*/
+void target_cleanup_recovery(struct obd_device *obd)
 {
-        struct obd_device *obd = data;
-        int rc;
+        struct list_head *tmp, *n;
+        struct ptlrpc_request *req;
 
-        CERROR("disconnecting clients and aborting recovery\n");
         spin_lock_bh(&obd->obd_processing_task_lock);
         if (!obd->obd_recovering) {
                 spin_unlock_bh(&obd->obd_processing_task_lock);
                 EXIT;
                 return;
         }
-
         obd->obd_recovering = obd->obd_abort_recovery = 0;
-
-        wake_up(&obd->obd_next_transno_waitq);
         target_cancel_recovery_timer(obd);
         spin_unlock_bh(&obd->obd_processing_task_lock);
 
-        class_disconnect_exports(obd, 0);
-
-        /* when recovery was aborted, cleanup orphans on mds and ost */
-        if (OBT(obd) && OBP(obd, postrecov)) {
-                rc = OBP(obd, postrecov)(obd);
-                if (rc >= 0)
-                        CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
-                else
-                        CERROR("postrecov failed %d\n", rc);
+        list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
+                req = list_entry(tmp, struct ptlrpc_request, rq_list);
+                list_del(&req->rq_list);
+                LASSERT (req->rq_reply_state);
+                lustre_free_reply_state(req->rq_reply_state);
+                target_release_saved_req(req);
         }
+        list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
+                req = list_entry(tmp, struct ptlrpc_request, rq_list);
+                list_del(&req->rq_list);
+                LASSERT (req->rq_reply_state == 0);
+                target_release_saved_req(req);
+        }
+}
+
+static void target_abort_recovery(void *data)
+{
+        struct obd_device *obd = data;
+
+        LASSERT(!obd->obd_recovering);
+
+        class_disconnect_stale_exports(obd, 0);
+
+        CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
+               obd->obd_name);
 
-        abort_delayed_replies(obd);
         abort_recovery_queue(obd);
+        target_finish_recovery(obd);
         ptlrpc_run_recovery_over_upcall(obd);
 }
 
@@ -681,7 +1012,9 @@ static void target_recovery_expired(unsigned long castmeharder)
         struct obd_device *obd = (struct obd_device *)castmeharder;
         CERROR("recovery timed out, aborting\n");
         spin_lock_bh(&obd->obd_processing_task_lock);
-        obd->obd_abort_recovery = 1;
+        if (obd->obd_recovering)
+                obd->obd_abort_recovery = 1;
+
         wake_up(&obd->obd_next_transno_waitq);
         spin_unlock_bh(&obd->obd_processing_task_lock);
 }
@@ -694,6 +1027,7 @@ void target_cancel_recovery_timer(struct obd_device *obd)
         del_timer(&obd->obd_recovery_timer);
 }
 
+#ifdef __KERNEL__
 static void reset_recovery_timer(struct obd_device *obd)
 {
         spin_lock_bh(&obd->obd_processing_task_lock);
@@ -706,52 +1040,60 @@ static void reset_recovery_timer(struct obd_device *obd)
         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
         spin_unlock_bh(&obd->obd_processing_task_lock);
 }
-
+#endif
 
 /* Only start it the first time called */
-void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
+void target_start_recovery_timer(struct obd_device *obd)
 {
         spin_lock_bh(&obd->obd_processing_task_lock);
-        if (obd->obd_recovery_handler) {
+        if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
                 spin_unlock_bh(&obd->obd_processing_task_lock);
                 return;
         }
         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
                OBD_RECOVERY_TIMEOUT / HZ);
-        obd->obd_recovery_handler = handler;
         obd->obd_recovery_timer.function = target_recovery_expired;
         obd->obd_recovery_timer.data = (unsigned long)obd;
+        mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
         spin_unlock_bh(&obd->obd_processing_task_lock);
-
-        reset_recovery_timer(obd);
 }
 
+#ifdef __KERNEL__
 static int check_for_next_transno(struct obd_device *obd)
 {
-        struct ptlrpc_request *req;
+        struct ptlrpc_request *req = NULL;
         int wake_up = 0, connected, completed, queue_len, max;
         __u64 next_transno, req_transno;
 
         spin_lock_bh(&obd->obd_processing_task_lock);
-        req = list_entry(obd->obd_recovery_queue.next,
-                         struct ptlrpc_request, rq_list);
+        if (!list_empty(&obd->obd_recovery_queue)) {
+                req = list_entry(obd->obd_recovery_queue.next,
+                                 struct ptlrpc_request, rq_list);
+                req_transno = req->rq_reqmsg->transno;
+        } else {
+                req_transno = 0;
+        }
+
         max = obd->obd_max_recoverable_clients;
-        req_transno = req->rq_reqmsg->transno;
         connected = obd->obd_connected_clients;
         completed = max - obd->obd_recoverable_clients;
         queue_len = obd->obd_requests_queued_for_recovery;
         next_transno = obd->obd_next_recovery_transno;
 
+        CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
+               "req_transno: "LPU64", next_transno: "LPU64"\n",
+               max, connected, completed, queue_len, req_transno, next_transno);
         if (obd->obd_abort_recovery) {
                 CDEBUG(D_HA, "waking for aborted recovery\n");
                 wake_up = 1;
-        } else if (!obd->obd_recovering) {
-                CDEBUG(D_HA, "waking for completed recovery (?)\n");
+        } else if (max == completed) {
+                CDEBUG(D_HA, "waking for completed recovery\n");
                 wake_up = 1;
         } else if (req_transno == next_transno) {
                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
                 wake_up = 1;
         } else if (queue_len + completed == max) {
+                LASSERT(req->rq_reqmsg->transno >= next_transno);
                 CDEBUG(D_ERROR,
                        "waking for skipped transno (skip: "LPD64
                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
@@ -760,64 +1102,130 @@ static int check_for_next_transno(struct obd_device *obd)
                 wake_up = 1;
         }
         spin_unlock_bh(&obd->obd_processing_task_lock);
-        LASSERT(req->rq_reqmsg->transno >= next_transno);
+        
         return wake_up;
 }
 
-static void process_recovery_queue(struct obd_device *obd)
+static struct ptlrpc_request *
+target_next_replay_req(struct obd_device *obd)
 {
-        struct ptlrpc_request *req;
-        int abort_recovery = 0;
         struct l_wait_info lwi = { 0 };
-        ENTRY;
+        struct ptlrpc_request *req;
 
-        for (;;) {
-                spin_lock_bh(&obd->obd_processing_task_lock);
-                LASSERT(obd->obd_processing_task == current->pid);
+        CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
+               obd->obd_next_recovery_transno);
+        l_wait_event(obd->obd_next_transno_waitq,
+                     check_for_next_transno(obd), &lwi);
+        
+        spin_lock_bh(&obd->obd_processing_task_lock);
+        if (obd->obd_abort_recovery) {
+                req = NULL;
+        } else if (!list_empty(&obd->obd_recovery_queue)) {
                 req = list_entry(obd->obd_recovery_queue.next,
                                  struct ptlrpc_request, rq_list);
+                list_del_init(&req->rq_list);
+                obd->obd_requests_queued_for_recovery--;
+        } else {
+                req = NULL;
+        }
+        spin_unlock_bh(&obd->obd_processing_task_lock);
+        return req;
+}
 
-                if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
-                        spin_unlock_bh(&obd->obd_processing_task_lock);
-                        CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
-                               LPD64")\n",
-                               obd->obd_next_recovery_transno,
-                               req->rq_reqmsg->transno);
-                        l_wait_event(obd->obd_next_transno_waitq,
-                                     check_for_next_transno(obd), &lwi);
+static int target_recovery_thread(void *arg)
+{
+        struct obd_device *obd = arg;
+        struct ptlrpc_request *req;
+        struct target_recovery_data *trd = &obd->obd_recovery_data;
+        unsigned long flags;
+        ENTRY;
+
+        kportal_daemonize("tgt-recov");
+
+        SIGNAL_MASK_LOCK(current, flags);
+        sigfillset(&current->blocked);
+        RECALC_SIGPENDING;
+        SIGNAL_MASK_UNLOCK(current, flags);
+
+        CERROR("%s: started recovery thread pid %d\n", obd->obd_name, 
+               current->pid);
+        trd->trd_processing_task = current->pid;
+
+        obd->obd_recovering = 1;
+        complete(&trd->trd_starting);
+
+        while (obd->obd_recovering) {
+                LASSERT(trd->trd_processing_task == current->pid);
+                req = target_next_replay_req(obd);
+                if (req != NULL) {
+                        char peer_str[PTL_NALFMT_SIZE];
+                        DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ", 
+                                  req->rq_reqmsg->transno, 
+                                  ptlrpc_peernid2str(&req->rq_peer, peer_str));
+                        (void)trd->trd_recovery_handler(req);
+                        obd->obd_replayed_requests++;
+                        reset_recovery_timer(obd);
+                        /* bug 1580: decide how to properly sync() in recovery*/
+                        //mds_fsync_super(mds->mds_sb);
+                        ptlrpc_free_clone(req);
                         spin_lock_bh(&obd->obd_processing_task_lock);
-                        abort_recovery = obd->obd_abort_recovery;
+                        obd->obd_next_recovery_transno++;
                         spin_unlock_bh(&obd->obd_processing_task_lock);
-                        if (abort_recovery) {
-                                target_abort_recovery(obd);
-                                return;
+                } else {
+                        /* recovery is over */
+                        spin_lock_bh(&obd->obd_processing_task_lock);
+                        obd->obd_recovering = 0;
+                        target_cancel_recovery_timer(obd);
+                        if (obd->obd_abort_recovery) {
+                                obd->obd_abort_recovery = 0;
+                                spin_unlock_bh(&obd->obd_processing_task_lock);
+                                target_abort_recovery(obd); 
+                        } else {
+                                LASSERT(obd->obd_recoverable_clients == 0);
+                                spin_unlock_bh(&obd->obd_processing_task_lock);
+                                target_finish_recovery(obd);
                         }
-                        continue;
                 }
-                list_del_init(&req->rq_list);
-                obd->obd_requests_queued_for_recovery--;
-                spin_unlock_bh(&obd->obd_processing_task_lock);
+        }
 
-                DEBUG_REQ(D_HA, req, "processing: ");
-                (void)obd->obd_recovery_handler(req);
-                obd->obd_replayed_requests++;
-                reset_recovery_timer(obd);
-                /* bug 1580: decide how to properly sync() in recovery */
-                //mds_fsync_super(mds->mds_sb);
-                class_export_put(req->rq_export);
-                OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
-                OBD_FREE(req, sizeof *req);
-                spin_lock_bh(&obd->obd_processing_task_lock);
-                obd->obd_next_recovery_transno++;
-                if (list_empty(&obd->obd_recovery_queue)) {
-                        obd->obd_processing_task = 0;
-                        spin_unlock_bh(&obd->obd_processing_task_lock);
-                        break;
-                }
+        trd->trd_processing_task = 0;
+        complete(&trd->trd_finishing);
+        return 0;
+}
+
+int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
+{
+        int rc = 0;
+        struct target_recovery_data *trd = &obd->obd_recovery_data;
+
+        memset(trd, 0, sizeof(*trd));
+        init_completion(&trd->trd_starting);
+        init_completion(&trd->trd_finishing);
+        trd->trd_recovery_handler = handler;
+
+        if (kernel_thread(target_recovery_thread, obd, 0) == 0) 
+                wait_for_completion(&trd->trd_starting);
+        else
+                rc = -ECHILD;
+
+        return rc;
+}
+
+void target_stop_recovery_thread(struct obd_device *obd)
+{
+        spin_lock_bh(&obd->obd_processing_task_lock);
+        if (obd->obd_recovery_data.trd_processing_task > 0) {
+                struct target_recovery_data *trd = &obd->obd_recovery_data;
+                CERROR("%s: aborting recovery\n", obd->obd_name);
+                obd->obd_abort_recovery = 1;
+                wake_up(&obd->obd_next_transno_waitq);
+                spin_unlock_bh(&obd->obd_processing_task_lock);
+                wait_for_completion(&trd->trd_finishing);
+        } else {
                 spin_unlock_bh(&obd->obd_processing_task_lock);
         }
-        EXIT;
 }
+#endif
 
 int target_queue_recovery_request(struct ptlrpc_request *req,
                                   struct obd_device *obd)
@@ -825,8 +1233,6 @@ int target_queue_recovery_request(struct ptlrpc_request *req,
         struct list_head *tmp;
         int inserted = 0;
         __u64 transno = req->rq_reqmsg->transno;
-        struct ptlrpc_request *saved_req;
-        struct lustre_msg *reqmsg;
 
         /* CAVEAT EMPTOR: The incoming request message has been swabbed
          * (i.e. buflens etc are in my own byte order), but type-dependent
@@ -838,16 +1244,6 @@ int target_queue_recovery_request(struct ptlrpc_request *req,
                 return 1;
         }
 
-        /* XXX If I were a real man, these LBUGs would be sane cleanups. */
-        /* XXX just like the request-dup code in queue_final_reply */
-        OBD_ALLOC(saved_req, sizeof *saved_req);
-        if (!saved_req)
-                LBUG();
-        OBD_ALLOC(reqmsg, req->rq_reqlen);
-        if (!reqmsg)
-                LBUG();
-
-        spin_lock_bh(&obd->obd_processing_task_lock);
 
         /* If we're processing the queue, we want don't want to queue this
          * message.
@@ -855,23 +1251,33 @@ int target_queue_recovery_request(struct ptlrpc_request *req,
          * Also, if this request has a transno less than the one we're waiting
          * for, we should process it now.  It could (and currently always will)
          * be an open request for a descriptor that was opened some time ago.
+         *
+         * Also, a resent, replayed request that has already been
+         * handled will pass through here and be processed immediately.
          */
-        if (obd->obd_processing_task == current->pid ||
+        spin_lock_bh(&obd->obd_processing_task_lock);
+        if (obd->obd_recovery_data.trd_processing_task == current->pid ||
             transno < obd->obd_next_recovery_transno) {
                 /* Processing the queue right now, don't re-add. */
                 LASSERT(list_empty(&req->rq_list));
                 spin_unlock_bh(&obd->obd_processing_task_lock);
-                OBD_FREE(reqmsg, req->rq_reqlen);
-                OBD_FREE(saved_req, sizeof *saved_req);
                 return 1;
         }
+        spin_unlock_bh(&obd->obd_processing_task_lock);
+
+        /* A resent, replayed request that is still on the queue; just drop it.
+           The queued request will handle this. */
+        if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
+            == (MSG_RESENT | MSG_REPLAY)) {
+                DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
+                return 0;
+        }
+
+        req = ptlrpc_clone_req(req);
+        if (req == NULL)
+                return -ENOMEM;
 
-        memcpy(saved_req, req, sizeof *req);
-        memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
-        req = saved_req;
-        req->rq_reqmsg = reqmsg;
-        class_export_get(req->rq_export);
-        INIT_LIST_HEAD(&req->rq_list);
+        spin_lock_bh(&obd->obd_processing_task_lock);
 
         /* XXX O(n^2) */
         list_for_each(tmp, &obd->obd_recovery_queue) {
@@ -885,28 +1291,12 @@ int target_queue_recovery_request(struct ptlrpc_request *req,
                 }
         }
 
-        if (!inserted) {
+        if (!inserted)
                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
-        }
 
         obd->obd_requests_queued_for_recovery++;
-
-        if (obd->obd_processing_task != 0) {
-                /* Someone else is processing this queue, we'll leave it to
-                 * them.
-                 */
-                wake_up(&obd->obd_next_transno_waitq);
-                spin_unlock_bh(&obd->obd_processing_task_lock);
-                return 0;
-        }
-
-        /* Nobody is processing, and we know there's (at least) one to process
-         * now, so we'll do the honours.
-         */
-        obd->obd_processing_task = current->pid;
+        wake_up(&obd->obd_next_transno_waitq);
         spin_unlock_bh(&obd->obd_processing_task_lock);
-
-        process_recovery_queue(obd);
         return 0;
 }
 
@@ -918,10 +1308,6 @@ struct obd_device * target_req2obd(struct ptlrpc_request *req)
 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 {
         struct obd_device *obd = target_req2obd(req);
-        struct ptlrpc_request *saved_req;
-        struct lustre_msg *reqmsg;
-        int recovery_done = 0;
-        int rc2;
 
         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
 
@@ -934,63 +1320,23 @@ int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 
         LASSERT (!req->rq_reply_state->rs_difficult);
         LASSERT(list_empty(&req->rq_list));
-        /* XXX a bit like the request-dup code in queue_recovery_request */
-        OBD_ALLOC(saved_req, sizeof *saved_req);
-        if (!saved_req)
-                LBUG();
-        OBD_ALLOC(reqmsg, req->rq_reqlen);
-        if (!reqmsg)
-                LBUG();
-        memcpy(saved_req, req, sizeof *saved_req);
-        memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
-        /* the copied req takes over the reply state */
-        req->rq_reply_state = NULL;
-        req = saved_req;
-        req->rq_reqmsg = reqmsg;
-        class_export_get(req->rq_export);
-        list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
+        
+        req = ptlrpc_clone_req(req);
 
         spin_lock_bh(&obd->obd_processing_task_lock);
-        --obd->obd_recoverable_clients;
-        recovery_done = (obd->obd_recoverable_clients == 0);
-        spin_unlock_bh(&obd->obd_processing_task_lock);
 
-        if (recovery_done) {
-                struct list_head *tmp, *n;
-                ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
-                CWARN("%s: all clients recovered, sending delayed replies\n",
-                       obd->obd_name);
-                spin_lock_bh(&obd->obd_processing_task_lock);
-                obd->obd_recovering = 0;
-                target_cancel_recovery_timer(obd);
-                spin_unlock_bh(&obd->obd_processing_task_lock);
-
-                /* when recovery finished, cleanup orphans on mds and ost */
-                if (OBT(obd) && OBP(obd, postrecov)) {
-                        rc2 = OBP(obd, postrecov)(obd);
-                        if (rc2 >= 0)
-                                CWARN("%s: all clients recovered, %d MDS "
-                                      "orphans deleted\n", obd->obd_name, rc2);
-                        else
-                                CERROR("postrecov failed %d\n", rc2);
-                }
+        list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 
-                list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
-                        req = list_entry(tmp, struct ptlrpc_request, rq_list);
-                        DEBUG_REQ(D_ERROR, req, "delayed:");
-                        ptlrpc_reply(req);
-                        class_export_put(req->rq_export);
-                        list_del(&req->rq_list);
-                        OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
-                        OBD_FREE(req, sizeof *req);
-                }
-                ptlrpc_run_recovery_over_upcall(obd);
-        } else {
+        /* only count the first "replay over" request from each
+           export */
+        if (req->rq_export->exp_replay_needed) {
+                --obd->obd_recoverable_clients;
+                req->rq_export->exp_replay_needed = 0;
                 CWARN("%s: %d recoverable clients remain\n",
-                       obd->obd_name, obd->obd_recoverable_clients);
-                wake_up(&obd->obd_next_transno_waitq);
+                      obd->obd_name, obd->obd_recoverable_clients);
         }
-
+        wake_up(&obd->obd_next_transno_waitq);
+        spin_unlock_bh(&obd->obd_processing_task_lock);
         return 1;
 }
 
@@ -1016,15 +1362,8 @@ target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
         }
 
         if (rc) {
-                DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
-                if (req->rq_reply_state == NULL) {
-                        rc = lustre_pack_reply (req, 0, NULL, NULL);
-                        if (rc != 0) {
-                                CERROR ("can't allocate reply\n");
-                                return (rc);
-                        }
-                }
-                req->rq_type = PTL_RPC_MSG_ERR;
+                req->rq_status = rc;
+                return (ptlrpc_error(req));
         } else {
                 DEBUG_REQ(D_NET, req, "sending reply");
         }