X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Ftarget%2Ftgt_lastrcvd.c;h=99a39c3cb5309119079ccc7dbe1abf09cd33c435;hb=8bbc27d31389cabe992d68dd9b0900bb26971e66;hp=ac6841a4cbb3d9494e3f99bc981eacc3eddf213e;hpb=f1803e2b950a64ab70a61fd9572b6c3d95d8bc77;p=fs%2Flustre-release.git

diff --git a/lustre/target/tgt_lastrcvd.c b/lustre/target/tgt_lastrcvd.c
index ac6841a..99a39c3 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);
@@ -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_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;
+
+	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;
 
@@ -714,7 +811,7 @@ 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);
@@ -747,7 +844,7 @@ 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;
 
@@ -800,16 +897,21 @@ 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 +940,9 @@ int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 		/* report failure to force synchronous operation */
 		return -EPERM;
 
-	return rc;
-}
-
-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_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;
-
-	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;
+	/* if exp_need_sync is set, return non-zero value to force
+	 * a sync transaction. */
+	return rc ? rc : exp->exp_need_sync;
 }
 
 /**
@@ -1069,6 +1120,9 @@ 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 */
@@ -1091,8 +1145,30 @@ int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 }
 EXPORT_SYMBOL(tgt_client_del);
 
+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);
+	}
+}
+
 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;
@@ -1129,6 +1205,19 @@ int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
 	}
 	/* 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) {
+			mutex_unlock(&ted->ted_lcd_lock);
+			RETURN(-EALREADY);
+		}
+
+		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)
@@ -1138,7 +1227,8 @@ 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);
 }
 EXPORT_SYMBOL(tgt_add_reply_data);
@@ -1269,7 +1359,8 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 			trd->trd_pre_versions[3] = pre_versions[3];
 		}
 
-		rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
+		rc = tgt_add_reply_data(env, tgt, ted, trd, req,
+					th, write_update);
 		if (rc < 0)
 			OBD_FREE_PTR(trd);
 		return rc;
@@ -1330,7 +1421,11 @@ static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 
 	if (!lw_client) {
 		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 +1496,11 @@ 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 +1565,7 @@ 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 +1654,8 @@ 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) {
@@ -1600,12 +1698,23 @@ 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) {
@@ -1722,6 +1831,14 @@ 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)
@@ -1964,43 +2081,70 @@ 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;
 }
@@ -2012,37 +2156,19 @@ 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);
 }