Whamcloud - gitweb
LU-5569 ptlrpc: change reverse import life cycle
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
index 448eb9b..1d4d6b7 100644 (file)
@@ -260,17 +260,18 @@ static int osc_on_mdt(char *obdname)
  */
 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 {
-        struct client_obd *cli = &obddev->u.cli;
-        struct obd_import *imp;
-        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;
+       struct client_obd *cli = &obddev->u.cli;
+       struct obd_import *imp;
+       struct obd_uuid server_uuid;
+       int rq_portal, rp_portal, connect_op;
+       char *name = obddev->obd_type->typ_name;
+       enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
+       char *cli_name = lustre_cfg_buf(lcfg, 0);
+       int rc;
+       ENTRY;
 
-        /* In a more perfect world, we would hang a ptlrpc_client off of
-         * obd_type and just use the values from there. */
+       /* In a more perfect world, we would hang a ptlrpc_client off of
+        * obd_type and just use the values from there. */
        if (!strcmp(name, LUSTRE_OSC_NAME)) {
                rq_portal = OST_REQUEST_PORTAL;
                rp_portal = OSC_REPLY_PORTAL;
@@ -283,7 +284,12 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
                rq_portal = MDS_REQUEST_PORTAL;
                rp_portal = MDC_REPLY_PORTAL;
                connect_op = MDS_CONNECT;
-               cli->cl_sp_me = LUSTRE_SP_CLI;
+               if (is_lwp_on_ost(cli_name))
+                       cli->cl_sp_me = LUSTRE_SP_OST;
+               else if (is_lwp_on_mdt(cli_name))
+                       cli->cl_sp_me = LUSTRE_SP_MDT;
+               else
+                       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_OSP_NAME)) {
@@ -301,7 +307,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
                        rq_portal = OST_REQUEST_PORTAL;
                }
                rp_portal = OSC_REPLY_PORTAL;
-               cli->cl_sp_me = LUSTRE_SP_CLI;
+               cli->cl_sp_me = LUSTRE_SP_MDT;
         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
                 rq_portal = MGS_REQUEST_PORTAL;
                 rp_portal = MGC_REPLY_PORTAL;
@@ -682,8 +688,7 @@ int server_disconnect_export(struct obd_export *exp)
        if (exp->exp_imp_reverse)
                ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 
-       if (exp->exp_obd->obd_namespace != NULL)
-               ldlm_cancel_locks_for_export(exp);
+       ldlm_bl_thread_wakeup();
 
         /* complete all outstanding replies */
        spin_lock(&exp->exp_lock);
@@ -716,13 +721,13 @@ static int target_handle_reconnect(struct lustre_handle *conn,
                                    struct obd_export *exp,
                                    struct obd_uuid *cluuid)
 {
-        ENTRY;
+       struct lustre_handle *hdl;
+       ENTRY;
 
-        if (exp->exp_connection && exp->exp_imp_reverse) {
-                struct lustre_handle *hdl;
+       hdl = &exp->exp_imp_reverse->imp_remote_handle;
+       if (exp->exp_connection && lustre_handle_is_used(hdl)) {
                 struct obd_device *target;
 
-                hdl = &exp->exp_imp_reverse->imp_remote_handle;
                 target = exp->exp_obd;
 
                 /* Might be a re-connect after a partition. */
@@ -792,17 +797,124 @@ static void
 check_and_start_recovery_timer(struct obd_device *obd,
                                struct ptlrpc_request *req, int new_client);
 
+/**
+ * update flags for import during reconnect process
+ */
+static int rev_import_flags_update(struct obd_import *revimp,
+                                  struct ptlrpc_request *req)
+{
+       int rc;
+       struct obd_connect_data *data;
+
+       data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
+
+       if (data->ocd_connect_flags & OBD_CONNECT_AT)
+               revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
+       else
+               revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
+
+       revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
+
+       rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
+       if (rc) {
+               CERROR("%s: cannot get reverse import %s security: rc = %d\n",
+                       revimp->imp_client->cli_name,
+                       libcfs_id2str(req->rq_peer), rc);
+               return rc;
+       }
+
+       return 0;
+}
+
+/**
+ * Allocate a new reverse import for an export.
+ *
+ * \retval -errno in case error hit
+ * \retval 0 if reverse import correctly init
+ **/
+int rev_import_init(struct obd_export *export)
+{
+       struct obd_device *obd = export->exp_obd;
+       struct obd_import *revimp;
+
+       LASSERT(export->exp_imp_reverse == NULL);
+
+       revimp = class_new_import(obd);
+       if (revimp == NULL)
+               return -ENOMEM;
+
+       revimp->imp_remote_handle.cookie = 0ULL;
+       revimp->imp_client = &obd->obd_ldlm_client;
+       revimp->imp_dlm_fake = 1;
+
+       /* it is safe to connect import in new state as no sends possible */
+       spin_lock(&export->exp_lock);
+       export->exp_imp_reverse = revimp;
+       spin_unlock(&export->exp_lock);
+       class_import_put(revimp);
+
+       return 0;
+}
+EXPORT_SYMBOL(rev_import_init);
+
+/**
+ * Handle reconnect for an export.
+ *
+ * \param exp export to handle reconnect process
+ * \param req client reconnect request
+ *
+ * \retval -rc in case securitfy flavor can't be changed
+ * \retval 0 in case none problems
+ */
+static int rev_import_reconnect(struct obd_export *exp,
+                               struct ptlrpc_request *req)
+{
+       struct obd_import *revimp = exp->exp_imp_reverse;
+       struct lustre_handle *lh;
+       int rc;
+
+       /* avoid sending a request until import flags are changed */
+       ptlrpc_import_enter_resend(revimp);
+
+       if (revimp->imp_connection != NULL)
+               ptlrpc_connection_put(revimp->imp_connection);
+
+       /*
+        * client from recovery don't have a handle so we need to take from
+        * request. it may produce situation when wrong client connected
+        * to recovery as we trust a client uuid
+        */
+       lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
+       revimp->imp_remote_handle = *lh;
+
+       /* unknown versions will be caught in
+        * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
+       revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
+
+       revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
+
+       rc = rev_import_flags_update(revimp, req);
+       if (rc != 0) {
+               /* it is safe to still be in RECOVERY phase as we are not able
+                * to setup correct security flavor so requests are not able to
+                * be delivered correctly */
+               return rc;
+       }
+
+       /* resend all rpc's via new connection */
+       return ptlrpc_import_recovery_state_machine(revimp);
+}
+
 int target_handle_connect(struct ptlrpc_request *req)
 {
-       struct obd_device *target = NULL, *targref = NULL;
-        struct obd_export *export = NULL;
-        struct obd_import *revimp;
-       struct obd_import *tmp_imp = NULL;
-        struct lustre_handle conn;
-        struct lustre_handle *tmp;
+       struct obd_device *target = NULL;
+       struct obd_export *export = NULL;
+       /* connect handle - filled from target_handle_reconnect in
+        * reconnect case */
+       struct lustre_handle conn;
+       struct lustre_handle *tmp;
         struct obd_uuid tgtuuid;
         struct obd_uuid cluuid;
-        struct obd_uuid remote_uuid;
         char *str;
         int rc = 0;
         char *target_start;
@@ -839,6 +951,11 @@ int target_handle_connect(struct ptlrpc_request *req)
        }
 
        spin_lock(&target->obd_dev_lock);
+       /* Make sure the target isn't cleaned up while we're here. Yes,
+        * there's still a race between the above check and our incref here.
+        * Really, class_uuid2obd should take the ref. */
+       class_incref(target, __func__, current);
+
        if (target->obd_stopping || !target->obd_set_up) {
                spin_unlock(&target->obd_dev_lock);
 
@@ -860,11 +977,6 @@ int target_handle_connect(struct ptlrpc_request *req)
                GOTO(out, rc = -EAGAIN);
        }
 
-       /* Make sure the target isn't cleaned up while we're here. Yes,
-        * there's still a race between the above check and our incref here.
-        * Really, class_uuid2obd should take the ref. */
-       targref = class_incref(target, __FUNCTION__, current);
-
        target->obd_conn_inprogress++;
        spin_unlock(&target->obd_dev_lock);
 
@@ -876,21 +988,6 @@ int target_handle_connect(struct ptlrpc_request *req)
 
         obd_str2uuid(&cluuid, str);
 
-       /* XXX Extract a nettype and format accordingly. */
-       switch (sizeof(lnet_nid_t)) {
-       /* NB the casts only avoid compiler warnings. */
-        case 8:
-                snprintf(remote_uuid.uuid, sizeof remote_uuid,
-                         "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
-                break;
-        case 4:
-                snprintf(remote_uuid.uuid, sizeof remote_uuid,
-                         "NET_%x_UUID", (__u32)req->rq_peer.nid);
-                break;
-        default:
-                LBUG();
-        }
-
         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
         if (tmp == NULL)
                 GOTO(out, rc = -EPROTO);
@@ -1139,8 +1236,10 @@ dont_check_exports:
                        if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
                                lustre_msg_add_op_flags(req->rq_repmsg,
                                                        MSG_CONNECT_RECOVERING);
-                       if (rc == 0)
+                       if (rc == 0) {
                                conn.cookie = export->exp_handle.h_cookie;
+                               rc = rev_import_init(export);
+                       }
 
                        if (mds_mds_conn)
                                new_mds_mds_conn = true;
@@ -1168,11 +1267,6 @@ dont_check_exports:
                 memcpy(tmpdata, data, min(tmpsize, size));
         }
 
-        /* If all else goes well, this is our RPC return code. */
-        req->rq_status = 0;
-
-        lustre_msg_set_handle(req->rq_repmsg, &conn);
-
         /* If the client and the server are the same node, we will already
          * have an export that really points to the client's DLM export,
          * because we have a shared handles table.
@@ -1219,14 +1313,19 @@ dont_check_exports:
                 ptlrpc_connection_put(export->exp_connection);
         }
 
-        export->exp_connection = ptlrpc_connection_get(req->rq_peer,
-                                                       req->rq_self,
-                                                       &remote_uuid);
-       if (hlist_unhashed(&export->exp_nid_hash)) {
-                cfs_hash_add(export->exp_obd->obd_nid_hash,
-                             &export->exp_connection->c_peer.nid,
-                             &export->exp_nid_hash);
-        }
+       export->exp_connection = ptlrpc_connection_get(req->rq_peer,
+                                                      req->rq_self,
+                                                      &cluuid);
+       if (hlist_unhashed(&export->exp_nid_hash))
+               cfs_hash_add(export->exp_obd->obd_nid_hash,
+                            &export->exp_connection->c_peer.nid,
+                            &export->exp_nid_hash);
+
+       lustre_msg_set_handle(req->rq_repmsg, &conn);
+
+       rc = rev_import_reconnect(export, req);
+       if (rc != 0)
+               GOTO(out, rc);
 
        if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
                 int has_transno;
@@ -1278,54 +1377,7 @@ dont_check_exports:
        if (target->obd_recovering && !lw_client)
                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 
-        tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
-        conn = *tmp;
-
-       /* Return -ENOTCONN in case of errors to let client reconnect. */
-       revimp = class_new_import(target);
-       if (revimp == NULL) {
-               CERROR("fail to alloc new reverse import.\n");
-               GOTO(out, rc = -ENOTCONN);
-       }
-
-       spin_lock(&export->exp_lock);
-       if (export->exp_imp_reverse != NULL)
-               /* destroyed import can be still referenced in ctxt */
-               tmp_imp = export->exp_imp_reverse;
-       export->exp_imp_reverse = revimp;
-       spin_unlock(&export->exp_lock);
-
-        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_dlm_fake = 1;
-        revimp->imp_state = LUSTRE_IMP_FULL;
-
-       /* Unknown versions will be caught in
-        * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
-        revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
-
-       if (data->ocd_connect_flags & OBD_CONNECT_AT)
-               revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
-       else
-               revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
-
-       revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
-
-       rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
-       if (rc) {
-               CERROR("Failed to get sec for reverse import: %d\n", rc);
-               spin_lock(&export->exp_lock);
-               export->exp_imp_reverse = NULL;
-               spin_unlock(&export->exp_lock);
-               class_destroy_import(revimp);
-       }
-
-       class_import_put(revimp);
-
 out:
-       if (tmp_imp != NULL)
-               client_destroy_import(tmp_imp);
        if (export) {
                spin_lock(&export->exp_lock);
                export->exp_connecting = 0;
@@ -1333,15 +1385,14 @@ out:
 
                class_export_put(export);
        }
-       if (targref) {
+       if (target != NULL) {
                spin_lock(&target->obd_dev_lock);
                target->obd_conn_inprogress--;
                spin_unlock(&target->obd_dev_lock);
 
-               class_decref(targref, __func__, current);
+               class_decref(target, __func__, current);
        }
-       if (rc)
-               req->rq_status = rc;
+       req->rq_status = rc;
        RETURN(rc);
 }
 
@@ -1411,6 +1462,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
         struct obd_export     *exp = req->rq_export;
         struct ptlrpc_request *reqiter;
+       struct ptlrpc_request *dup_req = NULL;
         int                    dup = 0;
 
         LASSERT(exp);
@@ -1419,6 +1471,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
        list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
                                 rq_replay_list) {
                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
+                       dup_req = reqiter;
                         dup = 1;
                         break;
                 }
@@ -1430,6 +1483,16 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
                         CERROR("invalid flags %x of resent replay\n",
                                lustre_msg_get_flags(req->rq_reqmsg));
+
+               if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+                       __u32 new_conn;
+
+                       new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
+                       if (new_conn >
+                           lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
+                               lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
+                                                       new_conn);
+               }
         } else {
                list_add_tail(&req->rq_replay_list,
                                   &exp->exp_req_replay_queue);
@@ -1485,8 +1548,11 @@ static void target_finish_recovery(struct lu_target *lut)
        spin_unlock(&obd->obd_recovery_task_lock);
 
        if (lut->lut_tdtd != NULL &&
-           !list_empty(&lut->lut_tdtd->tdtd_replay_list))
+           (!list_empty(&lut->lut_tdtd->tdtd_replay_list) ||
+           !list_empty(&lut->lut_tdtd->tdtd_replay_finish_list))) {
                dtrq_list_dump(lut->lut_tdtd, D_ERROR);
+               dtrq_list_destroy(lut->lut_tdtd);
+       }
 
         obd->obd_recovery_end = cfs_time_current_sec();
 
@@ -1554,7 +1620,6 @@ void target_cleanup_recovery(struct obd_device *obd)
 {
         struct ptlrpc_request *req, *n;
        struct list_head clean_list;
-        ENTRY;
 
        INIT_LIST_HEAD(&clean_list);
        spin_lock(&obd->obd_dev_lock);
@@ -1754,6 +1819,7 @@ static int check_for_next_transno(struct lu_target *lut)
 {
        struct ptlrpc_request *req = NULL;
        struct obd_device *obd = lut->lut_obd;
+       struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
        int wake_up = 0, connected, completed, queue_len;
        __u64 req_transno = 0;
        __u64 update_transno = 0;
@@ -1767,12 +1833,8 @@ static int check_for_next_transno(struct lu_target *lut)
                req_transno = lustre_msg_get_transno(req->rq_reqmsg);
        }
 
-       if (lut->lut_tdtd != NULL) {
-               struct target_distribute_txn_data *tdtd;
-
-               tdtd = lut->lut_tdtd;
-               update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
-       }
+       if (tdtd != NULL)
+               update_transno = distribute_txn_get_next_transno(tdtd);
 
        connected = atomic_read(&obd->obd_connected_clients);
        completed = connected - atomic_read(&obd->obd_req_replay_clients);
@@ -1790,6 +1852,13 @@ static int check_for_next_transno(struct lu_target *lut)
        } else if (obd->obd_recovery_expired) {
                CDEBUG(D_HA, "waking for expired recovery\n");
                wake_up = 1;
+       } else if (tdtd != NULL && req != NULL &&
+                  is_req_replayed_by_update(req)) {
+               LASSERTF(req_transno < next_transno, "req_transno "LPU64
+                        "next_transno"LPU64"\n", req_transno, next_transno);
+               CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
+                      req_transno);
+               wake_up = 1;
        } else if (req_transno == next_transno ||
                   (update_transno != 0 && update_transno <= next_transno)) {
                CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
@@ -2118,100 +2187,6 @@ static void drop_duplicate_replay_req(struct lu_env *env,
        obd->obd_replayed_requests++;
 }
 
-/**
- * Update last_rcvd of the update
- *
- * Because update recovery might update the last_rcvd by updates, i.e.
- * it will not update the last_rcvd information in memory, so we need
- * refresh these information in memory after update recovery.
- *
- * \param[in] obd      obd_device under recoverying.
- * \param[in] dtrq     the update replay requests being replayed.
- */
-static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
-                             struct distribute_txn_replay_req *dtrq)
-{
-       struct obd_device       *obd = lut->lut_obd;
-       struct obd_export       *export;
-       struct tg_export_data   *ted;
-       struct distribute_txn_replay_req_sub    *dtrqs;
-       struct seq_server_site *site;
-       struct update_records   *ur;
-       const struct lu_fid     *fid;
-       struct update_ops       *ops;
-       struct update_params    *params;
-       struct update_op        *op;
-       __u32                   mdt_index;
-       unsigned int            i;
-       struct lsd_client_data  *lcd = NULL;
-
-       /* if Updates has been executed(committed) on the recovery target,
-        * i.e. the updates is not being executed on the target, so we do
-        * not need update it in memory */
-       site = lu_site2seq(obd->obd_lu_dev->ld_site);
-       mdt_index = site->ss_node_id;
-       dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
-       if (dtrqs != NULL)
-               return;
-
-       if (dtrq->dtrq_lur == NULL)
-               return;
-
-       /* Find the update last_rcvd record */
-       fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
-       ur = &dtrq->dtrq_lur->lur_update_rec;
-       ops = &ur->ur_ops;
-       params = update_records_get_params(ur);
-       for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
-            i++, op = update_op_next_op(op)) {
-               __u64 pos;
-               __u16 size;
-               void *buf;
-
-               if (!lu_fid_eq(&op->uop_fid, fid))
-                       continue;
-
-               if (op->uop_type != OUT_WRITE)
-                       continue;
-
-               buf = update_params_get_param_buf(params, op->uop_params_off[1],
-                                                 ur->ur_param_count, NULL);
-               if (buf == NULL)
-                       continue;
-
-               pos = le64_to_cpu(*(__u64 *)buf);
-               if (pos == 0)
-                       continue;
-
-               buf = update_params_get_param_buf(params, op->uop_params_off[0],
-                                                 ur->ur_param_count, &size);
-               if (buf == NULL)
-                       continue;
-
-               if (size != sizeof(*lcd))
-                       continue;
-               lcd = buf;
-       }
-
-       if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
-               return;
-
-       /* locate the export then update the exp_target_data if needed */
-       export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
-       if (export == NULL)
-               return;
-
-       ted = &export->exp_target_data;
-       if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
-               CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
-                      lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
-                      lcd->lcd_last_xid);
-               ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
-               ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
-       }
-       class_export_put(export);
-}
-
 static void replay_request_or_update(struct lu_env *env,
                                     struct lu_target *lut,
                                     struct target_recovery_data *trd,
@@ -2247,23 +2222,35 @@ static void replay_request_or_update(struct lu_env *env,
 
                spin_lock(&obd->obd_recovery_task_lock);
                transno = get_next_transno(lut, &type);
-               if (type == REQUEST_RECOVERY && tdtd != NULL &&
-                   transno == tdtd->tdtd_last_update_transno) {
+               if (type == REQUEST_RECOVERY && transno != 0) {
                        /* Drop replay request from client side, if the
                         * replay has been executed by update with the
                         * same transno */
                        req = list_entry(obd->obd_req_replay_queue.next,
                                        struct ptlrpc_request, rq_list);
+
                        list_del_init(&req->rq_list);
                        obd->obd_requests_queued_for_recovery--;
                        spin_unlock(&obd->obd_recovery_task_lock);
-                       drop_duplicate_replay_req(env, obd, req);
-               } else if (type == REQUEST_RECOVERY && transno != 0) {
-                       req = list_entry(obd->obd_req_replay_queue.next,
-                                            struct ptlrpc_request, rq_list);
-                       list_del_init(&req->rq_list);
-                       obd->obd_requests_queued_for_recovery--;
-                       spin_unlock(&obd->obd_recovery_task_lock);
+
+                       /* Let's check if the request has been redone by
+                        * update replay */
+                       if (is_req_replayed_by_update(req)) {
+                               struct distribute_txn_replay_req *dtrq;
+
+                               dtrq = distribute_txn_lookup_finish_list(tdtd,
+                                                                  req->rq_xid);
+                               LASSERT(dtrq != NULL);
+                               spin_lock(&tdtd->tdtd_replay_list_lock);
+                               list_del_init(&dtrq->dtrq_list);
+                               spin_unlock(&tdtd->tdtd_replay_list_lock);
+                               dtrq_destroy(dtrq);
+
+                               drop_duplicate_replay_req(env, obd, req);
+
+                               continue;
+                       }
+
                        LASSERT(trd->trd_processing_task == current_pid());
                        DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
                                  lustre_msg_get_transno(req->rq_reqmsg),
@@ -2292,14 +2279,27 @@ static void replay_request_or_update(struct lu_env *env,
                        tdtd->tdtd_replay_handler(env, tdtd, dtrq);
                        lu_context_exit(&thread->t_env->le_ctx);
                        extend_recovery_timer(obd, obd_timeout, true);
-                       LASSERT(tdtd->tdtd_last_update_transno <= transno);
-                       tdtd->tdtd_last_update_transno = transno;
+
+                       /* Add it to the replay finish list */
+                       spin_lock(&tdtd->tdtd_replay_list_lock);
+                       if (dtrq->dtrq_xid != 0) {
+                               CDEBUG(D_HA, "Move x"LPU64" t"LPU64
+                                      " to finish list\n", dtrq->dtrq_xid,
+                                      dtrq->dtrq_master_transno);
+                               list_add(&dtrq->dtrq_list,
+                                        &tdtd->tdtd_replay_finish_list);
+                       } else {
+                               dtrq_destroy(dtrq);
+                       }
+                       spin_unlock(&tdtd->tdtd_replay_list_lock);
+
                        spin_lock(&obd->obd_recovery_task_lock);
-                       if (transno > obd->obd_next_recovery_transno)
-                               obd->obd_next_recovery_transno = transno;
+                       if (transno == obd->obd_next_recovery_transno)
+                               obd->obd_next_recovery_transno++;
+                       else if (transno > obd->obd_next_recovery_transno)
+                               obd->obd_next_recovery_transno = transno + 1;
                        spin_unlock(&obd->obd_recovery_task_lock);
-                       target_update_lcd(env, lut, dtrq);
-                       dtrq_destroy(dtrq);
+
                } else {
                        spin_unlock(&obd->obd_recovery_task_lock);
                        LASSERT(list_empty(&obd->obd_req_replay_queue));
@@ -2630,8 +2630,13 @@ int target_queue_recovery_request(struct ptlrpc_request *req,
         CDEBUG(D_HA, "Next recovery transno: "LPU64
                ", current: "LPU64", replaying\n",
                obd->obd_next_recovery_transno, transno);
+
+       /* If the request has been replayed by update replay, then sends this
+        * request to the recovery thread (replay_request_or_update()), where
+        * it will be handled */
        spin_lock(&obd->obd_recovery_task_lock);
-       if (transno < obd->obd_next_recovery_transno) {
+       if (transno < obd->obd_next_recovery_transno &&
+           !is_req_replayed_by_update(req)) {
                /* Processing the queue right now, don't re-add. */
                LASSERT(list_empty(&req->rq_list));
                spin_unlock(&obd->obd_recovery_task_lock);
@@ -2740,20 +2745,23 @@ int target_pack_pool_reply(struct ptlrpc_request *req)
 static int target_send_reply_msg(struct ptlrpc_request *req,
                                 int rc, int fail_id)
 {
-        if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
-                DEBUG_REQ(D_ERROR, req, "dropping reply");
-                return (-ECOMM);
-        }
+       if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
+               DEBUG_REQ(D_ERROR, req, "dropping reply");
+               return -ECOMM;
+       }
+       if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
+                    OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
+               return -ECOMM;
 
-        if (unlikely(rc)) {
-                DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
-                req->rq_status = rc;
-                return (ptlrpc_send_error(req, 1));
-        } else {
-                DEBUG_REQ(D_NET, req, "sending reply");
-        }
+       if (unlikely(rc)) {
+               DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
+               req->rq_status = rc;
+               return ptlrpc_send_error(req, 1);
+       } else {
+               DEBUG_REQ(D_NET, req, "sending reply");
+       }
 
-        return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
+       return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
 }
 
 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
@@ -2848,7 +2856,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
        EXIT;
 }
 
-ldlm_mode_t lck_compat_array[] = {
+enum ldlm_mode lck_compat_array[] = {
        [LCK_EX]    = LCK_COMPAT_EX,
        [LCK_PW]    = LCK_COMPAT_PW,
        [LCK_PR]    = LCK_COMPAT_PR,
@@ -2863,12 +2871,12 @@ ldlm_mode_t lck_compat_array[] = {
  * Rather arbitrary mapping from LDLM error codes to errno values. This should
  * not escape to the user level.
  */
-int ldlm_error2errno(ldlm_error_t error)
+int ldlm_error2errno(enum ldlm_error error)
 {
-        int result;
+       int result;
 
-        switch (error) {
-        case ELDLM_OK:
+       switch (error) {
+       case ELDLM_OK:
        case ELDLM_LOCK_MATCHED:
                 result = 0;
                 break;
@@ -2890,22 +2898,22 @@ int ldlm_error2errno(ldlm_error_t error)
         case ELDLM_BAD_NAMESPACE:
                 result = -EBADF;
                 break;
-        default:
-                if (((int)error) < 0)  /* cast to signed type */
-                        result = error; /* as ldlm_error_t can be unsigned */
-                else {
-                        CERROR("Invalid DLM result code: %d\n", error);
-                        result = -EPROTO;
-                }
-        }
-        return result;
+       default:
+               if (((int)error) < 0) { /* cast to signed type */
+                       result = error; /* as ldlm_error can be unsigned */
+               } else {
+                       CERROR("Invalid DLM result code: %d\n", error);
+                       result = -EPROTO;
+               }
+       }
+       return result;
 }
 EXPORT_SYMBOL(ldlm_error2errno);
 
 /**
- * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
+ * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
  */
-ldlm_error_t ldlm_errno2error(int err_no)
+enum ldlm_error ldlm_errno2error(int err_no)
 {
         int error;