X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Ftarget%2Ftgt_lastrcvd.c;h=061f088f8a91967b9888b14c84c028384bf57995;hp=cfcf18c266e0b824b3e50a42d41d2035265a00a8;hb=300858ccfcd00b52663de45e0bb472012242f342;hpb=72057a3af19ee02d9a686bd7e7d074917e381310

diff --git a/lustre/target/tgt_lastrcvd.c b/lustre/target/tgt_lastrcvd.c
index cfcf18c..061f088 100644
--- a/lustre/target/tgt_lastrcvd.c
+++ b/lustre/target/tgt_lastrcvd.c
@@ -23,7 +23,7 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2015, Intel Corporation.
+ * Copyright (c) 2011, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -40,6 +40,8 @@
 
 #include "tgt_internal.h"
 
+/** version recovery epoch */
+#define LR_EPOCH_BITS	32
 
 /* Allocate a bitmap for a chunk of reply data slots */
 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
@@ -146,6 +148,13 @@ static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
 	int chunk;
 	int b;
 
+	if (lut->lut_obd->obd_stopping)
+		/*
+		 * in case of failover keep the bit set in order to
+		 * avoid overwriting slots in reply_data which might
+		 * be required by resent rpcs
+		 */
+		return 0;
 	chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 	b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 
@@ -214,6 +223,9 @@ static int tgt_reply_header_write(const struct lu_env *env,
 		tgt->lut_obd->obd_name, REPLY_DATA,
 		lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 
+	if (tgt->lut_bottom->dd_rdonly)
+		RETURN(0);
+
 	buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
 	buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
 	buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
@@ -319,7 +331,7 @@ static void tgt_free_reply_data(struct lu_target *lut,
 
 	list_del(&trd->trd_list);
 	ted->ted_reply_cnt--;
-	if (lut != NULL)
+	if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
 		tgt_clear_reply_slot(lut, trd->trd_index);
 	OBD_FREE_PTR(trd);
 }
@@ -383,6 +395,8 @@ int tgt_client_alloc(struct obd_export *exp)
 
 	spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
 	INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
+	spin_lock_init(&exp->exp_target_data.ted_fmd_lock);
+	INIT_LIST_HEAD(&exp->exp_target_data.ted_fmd_list);
 
 	OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
 	if (exp->exp_target_data.ted_lcd == NULL)
@@ -406,6 +420,8 @@ void tgt_client_free(struct obd_export *exp)
 
 	LASSERT(exp != exp->exp_obd->obd_self_export);
 
+	tgt_fmd_cleanup(exp);
+
 	/* free reply data */
 	mutex_lock(&ted->ted_lcd_lock);
 	list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
@@ -444,8 +460,22 @@ void tgt_client_free(struct obd_export *exp)
 }
 EXPORT_SYMBOL(tgt_client_free);
 
-int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
-			 struct lsd_client_data *lcd, loff_t *off, int index)
+static inline void tgt_check_lcd(const char *obd_name, int index,
+				 struct lsd_client_data *lcd)
+{
+	size_t uuid_size = sizeof(lcd->lcd_uuid);
+
+	if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
+		lcd->lcd_uuid[uuid_size - 1] = '\0';
+
+		LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
+			       lcd->lcd_uuid, obd_name, index);
+	}
+}
+
+static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
+				struct lsd_client_data *lcd,
+				loff_t *off, int index)
 {
 	struct tgt_thread_info	*tti = tgt_th_info(env);
 	int			 rc;
@@ -453,7 +483,7 @@ int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 	tti_buf_lcd(tti);
 	rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 	if (rc == 0) {
-		check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
+		tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 		lcd_le_to_cpu(&tti->tti_lcd, lcd);
 		lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 		lcd->lcd_last_close_result =
@@ -471,9 +501,10 @@ int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 	return rc;
 }
 
-int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
-			  struct lsd_client_data *lcd, loff_t *off,
-			  struct thandle *th)
+static int tgt_client_data_write(const struct lu_env *env,
+				 struct lu_target *tgt,
+				 struct lsd_client_data *lcd,
+				 loff_t *off, struct thandle *th)
 {
 	struct tgt_thread_info *tti = tgt_th_info(env);
 	struct dt_object	*dto;
@@ -488,6 +519,59 @@ int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
 	return dt_record_write(env, dto, &tti->tti_buf, off, th);
 }
 
+struct tgt_new_client_callback {
+	struct dt_txn_commit_cb	 lncc_cb;
+	struct obd_export	*lncc_exp;
+};
+
+static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
+			      struct dt_txn_commit_cb *cb, int err)
+{
+	struct tgt_new_client_callback *ccb;
+
+	ccb = container_of(cb, struct tgt_new_client_callback, lncc_cb);
+
+	LASSERT(ccb->lncc_exp->exp_obd);
+
+	CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
+	       ccb->lncc_exp->exp_obd->obd_name,
+	       ccb->lncc_exp->exp_client_uuid.uuid);
+
+	spin_lock(&ccb->lncc_exp->exp_lock);
+
+	ccb->lncc_exp->exp_need_sync = 0;
+
+	spin_unlock(&ccb->lncc_exp->exp_lock);
+	class_export_cb_put(ccb->lncc_exp);
+
+	OBD_FREE_PTR(ccb);
+}
+
+int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
+{
+	struct tgt_new_client_callback	*ccb;
+	struct dt_txn_commit_cb		*dcb;
+	int				 rc;
+
+	OBD_ALLOC_PTR(ccb);
+	if (ccb == NULL)
+		return -ENOMEM;
+
+	ccb->lncc_exp = class_export_cb_get(exp);
+
+	dcb = &ccb->lncc_cb;
+	dcb->dcb_func = tgt_cb_new_client;
+	INIT_LIST_HEAD(&dcb->dcb_linkage);
+	strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
+
+	rc = dt_trans_cb_add(th, dcb);
+	if (rc) {
+		class_export_cb_put(exp);
+		OBD_FREE_PTR(ccb);
+	}
+	return rc;
+}
+
 /**
  * Update client data in last_rcvd
  */
@@ -508,12 +592,14 @@ static int tgt_client_data_update(const struct lu_env *env,
 		RETURN(-EINVAL);
 	}
 
+	if (tgt->lut_bottom->dd_rdonly)
+		RETURN(0);
+
 	th = dt_trans_create(env, tgt->lut_bottom);
 	if (IS_ERR(th))
 		RETURN(PTR_ERR(th));
 
 	tti_buf_lcd(tti);
-	mutex_lock(&ted->ted_lcd_lock);
 	rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 				     &tti->tti_buf,
 				     ted->ted_lr_off, th);
@@ -523,6 +609,9 @@ static int tgt_client_data_update(const struct lu_env *env,
 	rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 	if (rc)
 		GOTO(out, rc);
+
+	mutex_lock(&ted->ted_lcd_lock);
+
 	/*
 	 * Until this operations will be committed the sync is needed
 	 * for this export. This should be done _after_ starting the
@@ -541,9 +630,11 @@ static int tgt_client_data_update(const struct lu_env *env,
 
 	tti->tti_off = ted->ted_lr_off;
 	rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
+
+	mutex_unlock(&ted->ted_lcd_lock);
+
 	EXIT;
 out:
-	mutex_unlock(&ted->ted_lcd_lock);
 	dt_trans_stop(env, tgt->lut_bottom, th);
 	CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 	       "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
@@ -552,7 +643,7 @@ out:
 	return rc;
 }
 
-int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
+static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 {
 	struct tgt_thread_info	*tti = tgt_th_info(env);
 	int			 rc;
@@ -570,8 +661,8 @@ int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
         return rc;
 }
 
-int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
-			  struct thandle *th)
+static int tgt_server_data_write(const struct lu_env *env,
+				 struct lu_target *tgt, struct thandle *th)
 {
 	struct tgt_thread_info	*tti = tgt_th_info(env);
 	struct dt_object	*dto;
@@ -615,6 +706,9 @@ int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 	tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 	spin_unlock(&tgt->lut_translock);
 
+	if (tgt->lut_bottom->dd_rdonly)
+		RETURN(0);
+
 	th = dt_trans_create(env, tgt->lut_bottom);
 	if (IS_ERR(th))
 		RETURN(PTR_ERR(th));
@@ -642,8 +736,8 @@ out:
 }
 EXPORT_SYMBOL(tgt_server_data_update);
 
-int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
-			   loff_t size)
+static int tgt_truncate_last_rcvd(const struct lu_env *env,
+				  struct lu_target *tgt, loff_t size)
 {
 	struct dt_object *dt = tgt->lut_last_rcvd;
 	struct thandle	 *th;
@@ -652,6 +746,9 @@ int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
 
 	ENTRY;
 
+	if (tgt->lut_bottom->dd_rdonly)
+		RETURN(0);
+
 	attr.la_size = size;
 	attr.la_valid = LA_SIZE;
 
@@ -700,7 +797,7 @@ void tgt_boot_epoch_update(struct lu_target *tgt)
 	struct lu_env		 env;
 	struct ptlrpc_request	*req;
 	__u32			 start_epoch;
-	struct list_head	 client_list;
+	LIST_HEAD(client_list);
 	int			 rc;
 
 	if (tgt->lut_obd->obd_stopping)
@@ -714,12 +811,11 @@ void tgt_boot_epoch_update(struct lu_target *tgt)
 	}
 
 	spin_lock(&tgt->lut_translock);
-	start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
+	start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
 	tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 	tgt->lut_lsd.lsd_start_epoch = start_epoch;
 	spin_unlock(&tgt->lut_translock);
 
-	INIT_LIST_HEAD(&client_list);
 	/**
 	 * The recovery is not yet finished and final queue can still be updated
 	 * with resend requests. Move final list to separate one for processing
@@ -747,7 +843,7 @@ void tgt_boot_epoch_update(struct lu_target *tgt)
 	 * - there is no client to recover or the recovery was aborted
 	 */
 	if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
-	    (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
+	    (atomic_read(&tgt->lut_obd->obd_max_recoverable_clients) == 0 ||
 	    tgt->lut_obd->obd_abort_recovery))
 		tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
 
@@ -771,7 +867,7 @@ static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 {
 	struct tgt_last_committed_callback *ccb;
 
-	ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
+	ccb = container_of(cb, struct tgt_last_committed_callback, llcc_cb);
 
 	LASSERT(ccb->llcc_exp);
 	LASSERT(ccb->llcc_tgt != NULL);
@@ -800,16 +896,21 @@ static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 	} else {
 		spin_unlock(&ccb->llcc_tgt->lut_translock);
 	}
+
+	CDEBUG(D_HA, "%s: transno %lld is committed\n",
+	       ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
+
 out:
 	class_export_cb_put(ccb->llcc_exp);
-	if (ccb->llcc_transno)
-		CDEBUG(D_HA, "%s: transno %lld is committed\n",
-		       ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 	OBD_FREE_PTR(ccb);
 }
 
-int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
-			   struct obd_export *exp, __u64 transno)
+/**
+ * Add commit callback function, it returns a non-zero value to inform
+ * caller to use sync transaction if necessary.
+ */
+static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
+				  struct obd_export *exp, __u64 transno)
 {
 	struct tgt_last_committed_callback	*ccb;
 	struct dt_txn_commit_cb			*dcb;
@@ -838,60 +939,29 @@ int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 		/* report failure to force synchronous operation */
 		return -EPERM;
 
-	return rc;
+	/* if exp_need_sync is set, return non-zero value to force
+	 * a sync transaction. */
+	return rc ? rc : exp->exp_need_sync;
 }
 
-struct tgt_new_client_callback {
-	struct dt_txn_commit_cb	 lncc_cb;
-	struct obd_export	*lncc_exp;
-};
-
-static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
-			      struct dt_txn_commit_cb *cb, int err)
+static int tgt_is_local_client(const struct lu_env *env,
+				      struct obd_export *exp)
 {
-	struct tgt_new_client_callback *ccb;
-
-	ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
-
-	LASSERT(ccb->lncc_exp->exp_obd);
-
-	CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
-	       ccb->lncc_exp->exp_obd->obd_name,
-	       ccb->lncc_exp->exp_client_uuid.uuid);
-
-	spin_lock(&ccb->lncc_exp->exp_lock);
-
-	ccb->lncc_exp->exp_need_sync = 0;
+	struct lu_target	*tgt = class_exp2tgt(exp);
+	struct tgt_session_info *tsi = tgt_ses_info(env);
+	struct ptlrpc_request	*req = tgt_ses_req(tsi);
 
-	spin_unlock(&ccb->lncc_exp->exp_lock);
-	class_export_cb_put(ccb->lncc_exp);
+	if (exp_connect_flags(exp) & OBD_CONNECT_MDS ||
+	    exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS)
+		return 0;
+	if (tgt->lut_local_recovery)
+		return 0;
+	if (!req)
+		return 0;
+	if (!LNetIsPeerLocal(req->rq_peer.nid))
+		return 0;
 
-	OBD_FREE_PTR(ccb);
-}
-
-int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
-{
-	struct tgt_new_client_callback	*ccb;
-	struct dt_txn_commit_cb		*dcb;
-	int				 rc;
-
-	OBD_ALLOC_PTR(ccb);
-	if (ccb == NULL)
-		return -ENOMEM;
-
-	ccb->lncc_exp = class_export_cb_get(exp);
-
-	dcb = &ccb->lncc_cb;
-	dcb->dcb_func = tgt_cb_new_client;
-	INIT_LIST_HEAD(&dcb->dcb_linkage);
-	strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
-
-	rc = dt_trans_cb_add(th, dcb);
-	if (rc) {
-		class_export_cb_put(exp);
-		OBD_FREE_PTR(ccb);
-	}
-	return rc;
+	return 1;
 }
 
 /**
@@ -915,6 +985,13 @@ int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 	if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 		RETURN(0);
 
+	if (tgt_is_local_client(env, exp)) {
+		LCONSOLE_WARN("%s: local client %s w/o recovery\n",
+			      exp->exp_obd->obd_name, ted->ted_lcd->lcd_uuid);
+		exp->exp_no_recovery = 1;
+		RETURN(0);
+	}
+
 	/* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 	 * there's no need for extra complication here
 	 */
@@ -1042,7 +1119,8 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 	/* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 	if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
 		    (char *)tgt->lut_obd->obd_uuid.uuid) ||
-	    exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
+	    exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT ||
+	    exp->exp_no_recovery)
 		RETURN(0);
 
 	/* Slot may be not yet assigned, use case is race between Client
@@ -1069,6 +1147,9 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 	if (exp->exp_flags & OBD_OPT_FAILOVER)
 		RETURN(0);
 
+	if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
+		RETURN(0);
+
 	/* Make sure the server's last_transno is up to date.
 	 * This should be done before zeroing client slot so last_transno will
 	 * be in server data or in client data in case of failure */
@@ -1091,12 +1172,35 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 }
 EXPORT_SYMBOL(tgt_client_del);
 
-int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
+static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
+{
+	struct tg_export_data	*ted = &exp->exp_target_data;
+	struct lu_target	*lut = class_exp2tgt(exp);
+	struct tg_reply_data	*trd, *tmp;
+
+	if (tag == 0)
+		return;
+
+	list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
+		if (trd->trd_tag != tag)
+			continue;
+
+		LASSERT(ergo(tgt_is_increasing_xid_client(exp),
+			     trd->trd_reply.lrd_xid <= xid));
+
+		ted->ted_release_tag++;
+		tgt_release_reply_data(lut, ted, trd);
+	}
+}
+
+static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
 		       struct tg_export_data *ted, struct tg_reply_data *trd,
+		       struct ptlrpc_request *req,
 		       struct thandle *th, bool update_lrd_file)
 {
 	struct lsd_reply_data	*lrd;
 	int	i;
+	int	rc;
 
 	lrd = &trd->trd_reply;
 	/* update export last transno */
@@ -1105,30 +1209,51 @@ int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
 		ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
 	mutex_unlock(&ted->ted_lcd_lock);
 
-	/* find a empty slot */
-	i = tgt_find_free_reply_slot(tgt);
-	if (unlikely(i < 0)) {
-		CERROR("%s: couldn't find a slot for reply data: "
-		       "rc = %d\n", tgt_name(tgt), i);
-		RETURN(i);
-	}
-	trd->trd_index = i;
+	if (tgt != NULL) {
+		/* find a empty slot */
+		i = tgt_find_free_reply_slot(tgt);
+		if (unlikely(i < 0)) {
+			CERROR("%s: couldn't find a slot for reply data: "
+			       "rc = %d\n", tgt_name(tgt), i);
+			RETURN(i);
+		}
+		trd->trd_index = i;
 
-	if (update_lrd_file) {
-		loff_t	off;
-		int	rc;
+		if (update_lrd_file) {
+			loff_t	off;
 
-		/* write reply data to disk */
-		off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
-		rc = tgt_reply_data_write(env, tgt, lrd, off, th);
-		if (unlikely(rc != 0)) {
-			CERROR("%s: can't update %s file: rc = %d\n",
-			       tgt_name(tgt), REPLY_DATA, rc);
-			RETURN(rc);
+			/* write reply data to disk */
+			off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
+			rc = tgt_reply_data_write(env, tgt, lrd, off, th);
+			if (unlikely(rc != 0)) {
+				CERROR("%s: can't update %s file: rc = %d\n",
+				       tgt_name(tgt), REPLY_DATA, rc);
+				GOTO(free_slot, rc);
+			}
 		}
+	} else {
+		trd->trd_index = TRD_INDEX_MEMORY;
 	}
+
 	/* add reply data to target export's reply list */
 	mutex_lock(&ted->ted_lcd_lock);
+	if (req != NULL) {
+		int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
+			      MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
+
+		if (req->rq_obsolete) {
+			CDEBUG(D_INFO,
+			       "drop reply data update for obsolete req xid=%llu,"
+			       "transno=%llu, tag=%hu\n", req->rq_xid,
+			       lrd->lrd_transno, trd->trd_tag);
+			mutex_unlock(&ted->ted_lcd_lock);
+			GOTO(free_slot, rc = -EBADR);
+		}
+
+		if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
+			tgt_clean_by_tag(req->rq_export, req->rq_xid,
+					 trd->trd_tag);
+	}
 	list_add(&trd->trd_list, &ted->ted_reply_list);
 	ted->ted_reply_cnt++;
 	if (ted->ted_reply_cnt > ted->ted_reply_max)
@@ -1138,10 +1263,76 @@ int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
 	CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
 	       "tag %hu, client gen %u, slot idx %d\n",
 	       trd, lrd->lrd_xid, lrd->lrd_transno,
-	       trd->trd_tag, lrd->lrd_client_gen, i);
+	       trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
+
 	RETURN(0);
+
+free_slot:
+	if (tgt != NULL)
+		tgt_clear_reply_slot(tgt, trd->trd_index);
+	return rc;
+}
+
+int tgt_mk_reply_data(const struct lu_env *env,
+		      struct lu_target *tgt,
+		      struct tg_export_data *ted,
+		      struct ptlrpc_request *req,
+		      __u64 opdata,
+		      struct thandle *th,
+		      bool write_update,
+		      __u64 transno)
+{
+	struct tg_reply_data	*trd;
+	struct lsd_reply_data	*lrd;
+	__u64			*pre_versions = NULL;
+	int			rc;
+
+	OBD_ALLOC_PTR(trd);
+	if (unlikely(trd == NULL))
+		RETURN(-ENOMEM);
+
+	/* fill reply data information */
+	lrd = &trd->trd_reply;
+	lrd->lrd_transno = transno;
+	if (req != NULL) {
+		lrd->lrd_xid = req->rq_xid;
+		trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
+		lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
+		if (write_update) {
+			pre_versions = lustre_msg_get_versions(req->rq_repmsg);
+			lrd->lrd_result = th->th_result;
+		}
+	} else {
+		struct tgt_session_info *tsi;
+
+		LASSERT(env != NULL);
+		tsi = tgt_ses_info(env);
+		LASSERT(tsi->tsi_xid != 0);
+
+		lrd->lrd_xid = tsi->tsi_xid;
+		lrd->lrd_result = tsi->tsi_result;
+		lrd->lrd_client_gen = tsi->tsi_client_gen;
+	}
+
+	lrd->lrd_data = opdata;
+	if (pre_versions) {
+		trd->trd_pre_versions[0] = pre_versions[0];
+		trd->trd_pre_versions[1] = pre_versions[1];
+		trd->trd_pre_versions[2] = pre_versions[2];
+		trd->trd_pre_versions[3] = pre_versions[3];
+	}
+
+	rc = tgt_add_reply_data(env, tgt, ted, trd, req,
+				th, write_update);
+	if (rc < 0) {
+		OBD_FREE_PTR(trd);
+		if (rc == -EBADR)
+			rc = 0;
+	}
+	return rc;
+
 }
-EXPORT_SYMBOL(tgt_add_reply_data);
+EXPORT_SYMBOL(tgt_mk_reply_data);
 
 /*
  * last_rcvd & last_committed update callbacks
@@ -1152,11 +1343,11 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 {
 	struct tgt_thread_info	*tti = tgt_th_info(env);
 	struct tgt_session_info *tsi = tgt_ses_info(env);
-	struct obd_export	*exp = tsi->tsi_exp;
-	struct tg_export_data	*ted;
-	__u64			*transno_p;
-	int			 rc = 0;
-	bool			 lw_client;
+	struct obd_export *exp = tsi->tsi_exp;
+	struct tg_export_data *ted;
+	__u64 *transno_p;
+	bool nolcd = false;
+	int rc = 0;
 
 	ENTRY;
 
@@ -1164,11 +1355,15 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 	LASSERT(exp != NULL);
 	ted = &exp->exp_target_data;
 
-	lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
-	if (ted->ted_lr_idx < 0 && !lw_client)
-		/* ofd connect may cause transaction before export has
-		 * last_rcvd slot */
-		RETURN(0);
+	/* Some clients don't support recovery, and they don't have last_rcvd
+	 * client data:
+	 * 1. lightweight clients.
+	 * 2. local clients on MDS which doesn't enable "localrecov".
+	 * 3. OFD connect may cause transaction before export has last_rcvd
+	 *    slot.
+	 */
+	if (ted->ted_lr_idx < 0)
+		nolcd = true;
 
 	if (req != NULL)
 		tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
@@ -1209,14 +1404,13 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 	/* if can't add callback, do sync write */
 	th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
 
-	if (lw_client) {
-		/* All operations performed by LW clients are synchronous and
-		 * we store the committed transno in the last_rcvd header */
+	if (nolcd) {
+		/* store transno in the last_rcvd header */
 		spin_lock(&tgt->lut_translock);
 		if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
 			tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
 			spin_unlock(&tgt->lut_translock);
-			/* Although lightweight (LW) connections have no slot
+			/* Although current connection doesn't have slot
 			 * in the last_rcvd, we still want to maintain
 			 * the in-memory lsd_client_data structure in order to
 			 * properly handle reply reconstruction. */
@@ -1232,47 +1426,8 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 
 	/* Target that supports multiple reply data */
 	if (tgt_is_multimodrpcs_client(exp)) {
-		struct tg_reply_data	*trd;
-		struct lsd_reply_data	*lrd;
-		__u64			*pre_versions;
-		bool			write_update;
-
-		OBD_ALLOC_PTR(trd);
-		if (unlikely(trd == NULL))
-			RETURN(-ENOMEM);
-
-		/* fill reply data information */
-		lrd = &trd->trd_reply;
-		lrd->lrd_transno = tti->tti_transno;
-		if (req != NULL) {
-			lrd->lrd_xid = req->rq_xid;
-			trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
-			pre_versions = lustre_msg_get_versions(req->rq_repmsg);
-			lrd->lrd_result = th->th_result;
-			lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
-			write_update = true;
-		} else {
-			LASSERT(tsi->tsi_xid != 0);
-			lrd->lrd_xid = tsi->tsi_xid;
-			lrd->lrd_result = tsi->tsi_result;
-			lrd->lrd_client_gen = tsi->tsi_client_gen;
-			trd->trd_tag = 0;
-			pre_versions = NULL;
-			write_update = false;
-		}
-
-		lrd->lrd_data = opdata;
-		if (pre_versions) {
-			trd->trd_pre_versions[0] = pre_versions[0];
-			trd->trd_pre_versions[1] = pre_versions[1];
-			trd->trd_pre_versions[2] = pre_versions[2];
-			trd->trd_pre_versions[3] = pre_versions[3];
-		}
-
-		rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
-		if (rc < 0)
-			OBD_FREE_PTR(trd);
-		return rc;
+		return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
+					 !!(req != NULL), tti->tti_transno);
 	}
 
 	/* Enough for update replay, let's return */
@@ -1328,9 +1483,13 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 		}
 	}
 
-	if (!lw_client) {
+	if (!nolcd) {
 		tti->tti_off = ted->ted_lr_off;
-		rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
+		if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
+			rc = -EIO;
+		else
+			rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
+						   &tti->tti_off, th);
 		if (rc < 0) {
 			mutex_unlock(&ted->ted_lcd_lock);
 			RETURN(rc);
@@ -1401,8 +1560,11 @@ static int tgt_clients_data_init(const struct lu_env *env,
 
 	ENTRY;
 
-	CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
-		 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
+	if (tgt->lut_bottom->dd_rdonly)
+		RETURN(0);
+
+	BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
+		     sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
 
 	OBD_ALLOC_PTR(lcd);
 	if (lcd == NULL)
@@ -1467,7 +1629,7 @@ static int tgt_clients_data_init(const struct lu_env *env,
 		exp->exp_connecting = 0;
 		exp->exp_in_recovery = 0;
 		spin_unlock(&exp->exp_lock);
-		obd->obd_max_recoverable_clients++;
+		atomic_inc(&obd->obd_max_recoverable_clients);
 
 		if (tgt->lut_lsd.lsd_feature_incompat &
 		    OBD_INCOMPAT_MULTI_RPCS &&
@@ -1556,8 +1718,8 @@ int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
 	last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
 
 	/* ensure padding in the struct is the correct size */
-	CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
-		 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
+	BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
+		     sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
 
 	rc = server_name2index(tgt_name(tgt), &index, NULL);
 	if (rc < 0) {
@@ -1600,19 +1762,29 @@ int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
 			RETURN(rc);
 		}
 		if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
-			LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
-					   "using the wrong disk %s. Were the"
-					   " /dev/ assignments rearranged?\n",
-					   tgt->lut_obd->obd_uuid.uuid,
-					   lsd->lsd_uuid);
-			RETURN(-EINVAL);
+			if (tgt->lut_bottom->dd_rdonly) {
+				/* Such difference may be caused by mounting
+				 * up snapshot with new fsname under rd_only
+				 * mode. But even if it was NOT, it will not
+				 * damage the system because of "rd_only". */
+				memcpy(lsd->lsd_uuid,
+				       tgt->lut_obd->obd_uuid.uuid,
+				       sizeof(lsd->lsd_uuid));
+			} else {
+				LCONSOLE_ERROR_MSG(0x157, "Trying to start "
+						   "OBD %s using the wrong "
+						   "disk %s. Were the /dev/ "
+						   "assignments rearranged?\n",
+						   tgt->lut_obd->obd_uuid.uuid,
+						   lsd->lsd_uuid);
+				RETURN(-EINVAL);
+			}
 		}
 
 		if (lsd->lsd_osd_index != index) {
-			LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
-					   "is different with the index %d in"
-					   "config log, It might be disk"
-					   "corruption!\n", tgt_name(tgt),
+			LCONSOLE_ERROR_MSG(0x157,
+					   "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
+					   tgt_name(tgt),
 					   lsd->lsd_osd_index, index);
 			RETURN(-EINVAL);
 		}
@@ -1722,6 +1894,14 @@ int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
 	struct dt_object	*dto;
 	int			 rc;
 
+	/* For readonly case, the caller should have got failure
+	 * when start the transaction. If the logic comes here,
+	 * there must be something wrong. */
+	if (unlikely(tgt->lut_bottom->dd_rdonly)) {
+		dump_stack();
+		LBUG();
+	}
+
 	/* if there is no session, then this transaction is not result of
 	 * request processing but some local operation */
 	if (env->le_ses == NULL)
@@ -1741,13 +1921,10 @@ int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
 		 * because a replay slot has not been assigned.  This should be
 		 * replaced by dmu_tx_hold_append() when available.
 		 */
-		tti->tti_off = atomic_read(&tgt->lut_num_clients) * 8 *
-				sizeof(struct lsd_reply_data);
 		tti->tti_buf.lb_buf = NULL;
 		tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
 		dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
-		rc = dt_declare_record_write(env, dto, &tti->tti_buf,
-					     tti->tti_off, th);
+		rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
 		if (rc)
 			return rc;
 	} else {
@@ -1823,7 +2000,6 @@ int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
 	unsigned long		 reply_data_size;
 	int			 rc;
 	struct lsd_reply_header	*lrh = NULL;
-	struct lsd_client_data  *lcd = NULL;
 	struct tg_reply_data	*trd = NULL;
 	int                      idx;
 	loff_t			 off;
@@ -1872,10 +2048,6 @@ int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
 		if (hash == NULL)
 			GOTO(out, rc = -ENODEV);
 
-		OBD_ALLOC_PTR(lcd);
-		if (lcd == NULL)
-			GOTO(out, rc = -ENOMEM);
-
 		OBD_ALLOC_PTR(trd);
 		if (trd == NULL)
 			GOTO(out, rc = -ENOMEM);
@@ -1927,6 +2099,13 @@ int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
 			/* update export last committed transation */
 			exp->exp_last_committed = max(exp->exp_last_committed,
 						      lrd->lrd_transno);
+			/* Update lcd_last_transno as well for check in
+			 * tgt_release_reply_data() or the latest client
+			 * transno can be lost.
+			 */
+			ted->ted_lcd->lcd_last_transno =
+				max(ted->ted_lcd->lcd_last_transno,
+				    exp->exp_last_committed);
 
 			mutex_unlock(&ted->ted_lcd_lock);
 			class_export_put(exp);
@@ -1958,8 +2137,6 @@ int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
 out:
 	if (hash != NULL)
 		cfs_hash_putref(hash);
-	if (lcd != NULL)
-		OBD_FREE_PTR(lcd);
 	if (trd != NULL)
 		OBD_FREE_PTR(trd);
 	if (lrh != NULL)
@@ -1967,43 +2144,70 @@ out:
 	return rc;
 }
 
-struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
-					      __u64 xid)
+static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
+				struct tg_reply_data *trd)
 {
-	struct tg_reply_data	*found = NULL;
-	struct tg_reply_data	*reply;
+	struct tg_export_data *ted = &req->rq_export->exp_target_data;
+	struct lu_target *lut = class_exp2tgt(req->rq_export);
+	__u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
+	int rc = 0;
+	struct tg_reply_data *reply;
+	bool check_increasing;
+
+	if (tag == 0)
+		return 0;
+
+	check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
+			   !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+	if (!lookup && !check_increasing)
+		return 0;
 
-	mutex_lock(&ted->ted_lcd_lock);
 	list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
-		if (reply->trd_reply.lrd_xid == xid) {
-			found = reply;
+		if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
+			rc = 1;
+			if (trd != NULL)
+				*trd = *reply;
+			break;
+		} else if (check_increasing && reply->trd_tag == tag &&
+			   reply->trd_reply.lrd_xid > req->rq_xid) {
+			rc = -EPROTO;
+			CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
+			       tgt_name(lut), tag, req->rq_xid, trd,
+			       reply->trd_reply.lrd_xid,
+			       reply->trd_reply.lrd_transno,
+			       reply->trd_reply.lrd_client_gen,
+			       reply->trd_index, rc);
 			break;
 		}
 	}
-	mutex_unlock(&ted->ted_lcd_lock);
-	return found;
+
+	return rc;
 }
-EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
 
 /* Look for a reply data matching specified request @req
  * A copy is returned in @trd if the pointer is not NULL
  */
-bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
+int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
 {
-	struct tg_export_data	*ted = &req->rq_export->exp_target_data;
-	struct tg_reply_data	*reply;
-	bool			 found = false;
-
-	reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
-	if (reply != NULL) {
-		found = true;
-		if (trd != NULL)
-			*trd = *reply;
+	struct tg_export_data *ted = &req->rq_export->exp_target_data;
+	int found = 0;
+	bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
+
+	mutex_lock(&ted->ted_lcd_lock);
+	if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
+		/* A check for the last_xid is needed here in case there is
+		 * no reply data is left in the list. It may happen if another
+		 * RPC on another slot increased the last_xid between our
+		 * process_req_last_xid & tgt_lookup_reply calls */
+		found = -EPROTO;
+	} else {
+		found = tgt_check_lookup_req(req, 1, trd);
 	}
+	mutex_unlock(&ted->ted_lcd_lock);
 
-	CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
-	       tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
-	       found ? 1 : 0);
+	CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
+	       tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
+	       req->rq_export->exp_last_xid);
 
 	return found;
 }
@@ -2015,37 +2219,19 @@ int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
 	struct lu_target	*lut = class_exp2tgt(exp);
 	struct tg_reply_data	*trd, *tmp;
 
-	mutex_lock(&ted->ted_lcd_lock);
+
 	list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
 		if (trd->trd_reply.lrd_xid > rcvd_xid)
 			continue;
 		ted->ted_release_xid++;
 		tgt_release_reply_data(lut, ted, trd);
 	}
-	mutex_unlock(&ted->ted_lcd_lock);
 
 	return 0;
 }
 
-int tgt_handle_tag(struct obd_export *exp, __u16 tag)
+int tgt_handle_tag(struct ptlrpc_request *req)
 {
-	struct tg_export_data	*ted = &exp->exp_target_data;
-	struct lu_target	*lut = class_exp2tgt(exp);
-	struct tg_reply_data	*trd, *tmp;
-
-	if (tag == 0)
-		return 0;
-
-	mutex_lock(&ted->ted_lcd_lock);
-	list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
-		if (trd->trd_tag != tag)
-			continue;
-		ted->ted_release_tag++;
-		tgt_release_reply_data(lut, ted, trd);
-		break;
-	}
-	mutex_unlock(&ted->ted_lcd_lock);
-
-	return 0;
+	return tgt_check_lookup_req(req, 0, NULL);
 }