X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosp%2Fosp_trans.c;h=b6e5cf4ff1d9d29eafc2aacdd599d85984ab2044;hb=4f53536d002c13886210b672b657795baa067144;hp=39a42bcceb3dcf3341343693d425beb9664d2296;hpb=a0da0ad1438f694dcff9ee9826d388927ea19d4e;p=fs%2Flustre-release.git

diff --git a/lustre/osp/osp_trans.c b/lustre/osp/osp_trans.c
index 39a42bc..b6e5cf4 100644
--- a/lustre/osp/osp_trans.c
+++ b/lustre/osp/osp_trans.c
@@ -41,7 +41,7 @@
  * OUT RPC.
  *
  * For the asynchronous idempotent operations, such as get_(x)attr, related
- * RPCs will be inserted into a osp_device based shared asynchronous request
+ * RPCs will be inserted into an osp_device based shared asynchronous request
  * queue - osp_device::opd_async_requests. When the queue is full, all the
  * requests in the queue will be packaged into a single OUT RPC and given to
  * the ptlrpcd daemon (for sending), then the queue is purged and other new
@@ -53,6 +53,13 @@
  * will be called one by one to handle each own result.
  *
  *
+ * There are three kinds of transactions
+ *
+ * 1. Local transaction, all of updates of the transaction are in the local MDT.
+ * 2. Remote transaction, all of updates of the transaction are in one remote
+ * MDT, which only happens in LFSCK now.
+ * 3. Distribute transaction, updates for the transaction are in mulitple MDTs.
+ *
  * Author: Di Wang <di.wang@intel.com>
  * Author: Fan, Yong <fan.yong@intel.com>
  */
@@ -61,27 +68,135 @@
 
 #include "osp_internal.h"
 
-struct osp_async_update_args {
+/**
+ * The argument for the interpreter callback of osp request.
+ */
+struct osp_update_args {
 	struct dt_update_request *oaua_update;
+	atomic_t		 *oaua_count;
+	wait_queue_head_t	 *oaua_waitq;
 	bool			  oaua_flow_control;
 };
 
-struct osp_async_request {
+/**
+ * Call back for each update request.
+ */
+struct osp_update_callback {
 	/* list in the dt_update_request::dur_cb_items */
-	struct list_head		 oar_list;
+	struct list_head		 ouc_list;
 
 	/* The target of the async update request. */
-	struct osp_object		*oar_obj;
+	struct osp_object		*ouc_obj;
 
-	/* The data used by oar_interpreter. */
-	void				*oar_data;
+	/* The data used by or_interpreter. */
+	void				*ouc_data;
 
 	/* The interpreter function called after the async request handled. */
-	osp_async_request_interpreter_t	 oar_interpreter;
+	osp_update_interpreter_t	ouc_interpreter;
 };
 
+static struct object_update_request *object_update_request_alloc(size_t size)
+{
+	struct object_update_request *ourq;
+
+	OBD_ALLOC_LARGE(ourq, size);
+	if (ourq == NULL)
+		return ERR_PTR(-ENOMEM);
+
+	ourq->ourq_magic = UPDATE_REQUEST_MAGIC;
+	ourq->ourq_count = 0;
+
+	return ourq;
+}
+
+static void object_update_request_free(struct object_update_request *ourq,
+				       size_t ourq_size)
+{
+	if (ourq != NULL)
+		OBD_FREE_LARGE(ourq, ourq_size);
+}
+
+/**
+ * Allocate and initialize dt_update_request
+ *
+ * dt_update_request is being used to track updates being executed on
+ * this dt_device(OSD or OSP). The update buffer will be 4k initially,
+ * and increased if needed.
+ *
+ * \param [in] dt	dt device
+ *
+ * \retval		dt_update_request being allocated if succeed
+ * \retval		ERR_PTR(errno) if failed
+ */
+struct dt_update_request *dt_update_request_create(struct dt_device *dt)
+{
+	struct dt_update_request *dt_update;
+	struct object_update_request *ourq;
+
+	OBD_ALLOC_PTR(dt_update);
+	if (dt_update == NULL)
+		return ERR_PTR(-ENOMEM);
+
+	ourq = object_update_request_alloc(OUT_UPDATE_INIT_BUFFER_SIZE);
+	if (IS_ERR(ourq)) {
+		OBD_FREE_PTR(dt_update);
+		return ERR_CAST(ourq);
+	}
+
+	dt_update->dur_buf.ub_req = ourq;
+	dt_update->dur_buf.ub_req_size = OUT_UPDATE_INIT_BUFFER_SIZE;
+
+	dt_update->dur_dt = dt;
+	dt_update->dur_batchid = 0;
+	INIT_LIST_HEAD(&dt_update->dur_cb_items);
+
+	return dt_update;
+}
+
+/**
+ * Destroy dt_update_request
+ *
+ * \param [in] dt_update	dt_update_request being destroyed
+ */
+void dt_update_request_destroy(struct dt_update_request *dt_update)
+{
+	if (dt_update == NULL)
+		return;
+
+	object_update_request_free(dt_update->dur_buf.ub_req,
+				   dt_update->dur_buf.ub_req_size);
+	OBD_FREE_PTR(dt_update);
+}
+
+static void
+object_update_request_dump(const struct object_update_request *ourq,
+			   unsigned int mask)
+{
+	unsigned int i;
+	size_t total_size = 0;
+
+	for (i = 0; i < ourq->ourq_count; i++) {
+		struct object_update	*update;
+		size_t			size = 0;
+
+		update = object_update_request_get(ourq, i, &size);
+		LASSERT(update != NULL);
+		CDEBUG(mask, "i = %u fid = "DFID" op = %s master = %u"
+		       "params = %d batchid = "LPU64" size = %zu\n",
+		       i, PFID(&update->ou_fid),
+		       update_op_str(update->ou_type),
+		       update->ou_master_index, update->ou_params_count,
+		       update->ou_batchid, size);
+
+		total_size += size;
+	}
+
+	CDEBUG(mask, "updates = %p magic = %x count = %d size = %zu\n", ourq,
+	       ourq->ourq_magic, ourq->ourq_count, total_size);
+}
+
 /**
- * Allocate an asynchronous request and initialize it with the given parameters.
+ * Allocate an osp request and initialize it with the given parameters.
  *
  * \param[in] obj		pointer to the operation target
  * \param[in] data		pointer to the data used by the interpreter
@@ -90,38 +205,38 @@ struct osp_async_request {
  * \retval			pointer to the asychronous request
  * \retval			NULL if the allocation failed
  */
-static struct osp_async_request *
-osp_async_request_init(struct osp_object *obj, void *data,
-		       osp_async_request_interpreter_t interpreter)
+static struct osp_update_callback *
+osp_update_callback_init(struct osp_object *obj, void *data,
+			 osp_update_interpreter_t interpreter)
 {
-	struct osp_async_request *oar;
+	struct osp_update_callback *ouc;
 
-	OBD_ALLOC_PTR(oar);
-	if (oar == NULL)
+	OBD_ALLOC_PTR(ouc);
+	if (ouc == NULL)
 		return NULL;
 
 	lu_object_get(osp2lu_obj(obj));
-	INIT_LIST_HEAD(&oar->oar_list);
-	oar->oar_obj = obj;
-	oar->oar_data = data;
-	oar->oar_interpreter = interpreter;
+	INIT_LIST_HEAD(&ouc->ouc_list);
+	ouc->ouc_obj = obj;
+	ouc->ouc_data = data;
+	ouc->ouc_interpreter = interpreter;
 
-	return oar;
+	return ouc;
 }
 
 /**
- * Destroy the asychronous request.
+ * Destroy the osp_update_callback.
  *
  * \param[in] env	pointer to the thread context
- * \param[in] oar	pointer to asychronous request
+ * \param[in] ouc	pointer to osp_update_callback
  */
-static void osp_async_request_fini(const struct lu_env *env,
-				   struct osp_async_request *oar)
+static void osp_update_callback_fini(const struct lu_env *env,
+				     struct osp_update_callback *ouc)
 {
-	LASSERT(list_empty(&oar->oar_list));
+	LASSERT(list_empty(&ouc->ouc_list));
 
-	lu_object_put(env, osp2lu_obj(oar->oar_obj));
-	OBD_FREE_PTR(oar);
+	lu_object_put(env, osp2lu_obj(ouc->ouc_obj));
+	OBD_FREE_PTR(ouc);
 }
 
 /**
@@ -138,15 +253,14 @@ static void osp_async_request_fini(const struct lu_env *env,
  * \retval		0 for success
  * \retval		negative error number on failure
  */
-static int osp_async_update_interpret(const struct lu_env *env,
-				      struct ptlrpc_request *req,
-				      void *arg, int rc)
+static int osp_update_interpret(const struct lu_env *env,
+				struct ptlrpc_request *req, void *arg, int rc)
 {
 	struct object_update_reply	*reply	= NULL;
-	struct osp_async_update_args	*oaua	= arg;
+	struct osp_update_args		*oaua	= arg;
 	struct dt_update_request	*dt_update = oaua->oaua_update;
-	struct osp_async_request	*oar;
-	struct osp_async_request	*next;
+	struct osp_update_callback	*ouc;
+	struct osp_update_callback	*next;
 	int				 count	= 0;
 	int				 index  = 0;
 	int				 rc1	= 0;
@@ -168,9 +282,9 @@ static int osp_async_update_interpret(const struct lu_env *env,
 		rc1 = rc;
 	}
 
-	list_for_each_entry_safe(oar, next, &dt_update->dur_cb_items,
-				 oar_list) {
-		list_del_init(&oar->oar_list);
+	list_for_each_entry_safe(ouc, next, &dt_update->dur_cb_items,
+				 ouc_list) {
+		list_del_init(&ouc->ouc_list);
 
 		/* The peer may only have handled some requests (indicated
 		 * by the 'count') in the packaged OUT RPC, we can only get
@@ -189,13 +303,18 @@ static int osp_async_update_interpret(const struct lu_env *env,
 				rc1 = -EINVAL;
 		}
 
-		oar->oar_interpreter(env, reply, req, oar->oar_obj,
-				       oar->oar_data, index, rc1);
-		osp_async_request_fini(env, oar);
+		if (ouc->ouc_interpreter != NULL)
+			ouc->ouc_interpreter(env, reply, req, ouc->ouc_obj,
+					     ouc->ouc_data, index, rc1);
+
+		osp_update_callback_fini(env, ouc);
 		index++;
 	}
 
-	out_destroy_update_req(dt_update);
+	if (oaua->oaua_count != NULL && atomic_dec_and_test(oaua->oaua_count))
+		wake_up_all(oaua->oaua_waitq);
+
+	dt_update_request_destroy(dt_update);
 
 	return 0;
 }
@@ -215,30 +334,33 @@ int osp_unplug_async_request(const struct lu_env *env,
 			     struct osp_device *osp,
 			     struct dt_update_request *update)
 {
-	struct osp_async_update_args	*args;
-	struct ptlrpc_request		*req = NULL;
-	int				 rc;
+	struct osp_update_args	*args;
+	struct ptlrpc_request	*req = NULL;
+	int			 rc;
 
-	rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
-				 update->dur_req, &req);
+	rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
+				 update->dur_buf.ub_req, &req);
 	if (rc != 0) {
-		struct osp_async_request *oar;
-		struct osp_async_request *next;
-
-		list_for_each_entry_safe(oar, next,
-					 &update->dur_cb_items, oar_list) {
-			list_del_init(&oar->oar_list);
-			oar->oar_interpreter(env, NULL, NULL, oar->oar_obj,
-					       oar->oar_data, 0, rc);
-			osp_async_request_fini(env, oar);
+		struct osp_update_callback *ouc;
+		struct osp_update_callback *next;
+
+		list_for_each_entry_safe(ouc, next,
+					 &update->dur_cb_items, ouc_list) {
+			list_del_init(&ouc->ouc_list);
+			if (ouc->ouc_interpreter != NULL)
+				ouc->ouc_interpreter(env, NULL, NULL,
+						     ouc->ouc_obj,
+						     ouc->ouc_data, 0, rc);
+			osp_update_callback_fini(env, ouc);
 		}
-		out_destroy_update_req(update);
+		dt_update_request_destroy(update);
 	} else {
-		LASSERT(list_empty(&update->dur_list));
-
 		args = ptlrpc_req_async_args(req);
 		args->oaua_update = update;
-		req->rq_interpret_reply = osp_async_update_interpret;
+		args->oaua_count = NULL;
+		args->oaua_waitq = NULL;
+		args->oaua_flow_control = false;
+		req->rq_interpret_reply = osp_update_interpret;
 		ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 	}
 
@@ -265,7 +387,7 @@ osp_find_or_create_async_update_request(struct osp_device *osp)
 	if (update != NULL)
 		return update;
 
-	update = out_create_update_req(&osp->opd_dt_dev);
+	update = dt_update_request_create(&osp->opd_dt_dev);
 	if (!IS_ERR(update))
 		osp->opd_async_requests = update;
 
@@ -273,6 +395,37 @@ osp_find_or_create_async_update_request(struct osp_device *osp)
 }
 
 /**
+ * Insert an osp_update_callback into the dt_update_request.
+ *
+ * Insert an osp_update_callback to the dt_update_request. Usually each update
+ * in the dt_update_request will have one correspondent callback, and these
+ * callbacks will be called in rq_interpret_reply.
+ *
+ * \param[in] env		pointer to the thread context
+ * \param[in] obj		pointer to the operation target object
+ * \param[in] data		pointer to the data used by the interpreter
+ * \param[in] interpreter	pointer to the interpreter function
+ *
+ * \retval			0 for success
+ * \retval			negative error number on failure
+ */
+int osp_insert_update_callback(const struct lu_env *env,
+			       struct dt_update_request *update,
+			       struct osp_object *obj, void *data,
+			       osp_update_interpreter_t interpreter)
+{
+	struct osp_update_callback  *ouc;
+
+	ouc = osp_update_callback_init(obj, data, interpreter);
+	if (ouc == NULL)
+		RETURN(-ENOMEM);
+
+	list_add_tail(&ouc->ouc_list, &update->dur_cb_items);
+
+	return 0;
+}
+
+/**
  * Insert an asynchronous idempotent request to the shared request queue that
  * is attached to the osp_device.
  *
@@ -287,8 +440,8 @@ osp_find_or_create_async_update_request(struct osp_device *osp)
  * \param[in] env		pointer to the thread context
  * \param[in] op		operation type, see 'enum update_type'
  * \param[in] obj		pointer to the operation target
- * \param[in] count		array size of the subsequent @lens and @bufs
- * \param[in] lens		buffer length array for the subsequent @bufs
+ * \param[in] count		array size of the subsequent \a lens and \a bufs
+ * \param[in] lens		buffer length array for the subsequent \a bufs
  * \param[in] bufs		the buffers to compose the request
  * \param[in] data		pointer to the data used by the interpreter
  * \param[in] interpreter	pointer to the interpreter function
@@ -296,28 +449,31 @@ osp_find_or_create_async_update_request(struct osp_device *osp)
  * \retval			0 for success
  * \retval			negative error number on failure
  */
-int osp_insert_async_request(const struct lu_env *env,
-			     int op, struct osp_object *obj, int count,
-			     int *lens, const char **bufs, void *data,
-			     osp_async_request_interpreter_t interpreter)
+int osp_insert_async_request(const struct lu_env *env, enum update_type op,
+			     struct osp_object *obj, int count,
+			     __u16 *lens, const void **bufs, void *data,
+			     osp_update_interpreter_t interpreter)
 {
-	struct osp_async_request     *oar;
 	struct osp_device	     *osp = lu2osp_dev(osp2lu_obj(obj)->lo_dev);
-	struct dt_update_request     *update;
-	int			      rc  = 0;
+	struct dt_update_request	*update;
+	struct object_update		*object_update;
+	size_t				max_update_size;
+	struct object_update_request	*ureq;
+	int				rc = 0;
 	ENTRY;
 
-	oar = osp_async_request_init(obj, data, interpreter);
-	if (oar == NULL)
-		RETURN(-ENOMEM);
-
 	update = osp_find_or_create_async_update_request(osp);
 	if (IS_ERR(update))
-		GOTO(out, rc = PTR_ERR(update));
+		RETURN(PTR_ERR(update));
 
 again:
-	rc = out_insert_update(env, update, op, lu_object_fid(osp2lu_obj(obj)),
-			       count, lens, bufs);
+	ureq = update->dur_buf.ub_req;
+	max_update_size = update->dur_buf.ub_req_size -
+			    object_update_request_size(ureq);
+
+	object_update = update_buffer_get_update(ureq, ureq->ourq_count);
+	rc = out_update_pack(env, object_update, max_update_size, op,
+			     lu_object_fid(osp2lu_obj(obj)), count, lens, bufs);
 	/* The queue is full. */
 	if (rc == -E2BIG) {
 		osp->opd_async_requests = NULL;
@@ -326,25 +482,55 @@ again:
 		rc = osp_unplug_async_request(env, osp, update);
 		mutex_lock(&osp->opd_async_requests_mutex);
 		if (rc != 0)
-			GOTO(out, rc);
+			RETURN(rc);
 
 		update = osp_find_or_create_async_update_request(osp);
 		if (IS_ERR(update))
-			GOTO(out, rc = PTR_ERR(update));
+			RETURN(PTR_ERR(update));
 
 		goto again;
+	} else {
+		if (rc < 0)
+			RETURN(rc);
+
+		ureq->ourq_count++;
 	}
 
-	if (rc == 0)
-		list_add_tail(&oar->oar_list, &update->dur_cb_items);
+	rc = osp_insert_update_callback(env, update, obj, data, interpreter);
 
-	GOTO(out, rc);
+	RETURN(rc);
+}
 
-out:
-	if (rc != 0)
-		osp_async_request_fini(env, oar);
+int osp_trans_update_request_create(struct thandle *th)
+{
+	struct osp_thandle		*oth = thandle_to_osp_thandle(th);
+	struct dt_update_request	*update;
 
-	return rc;
+	if (oth->ot_dur != NULL)
+		return 0;
+
+	update = dt_update_request_create(th->th_dev);
+	if (IS_ERR(update)) {
+		th->th_result = PTR_ERR(update);
+		return PTR_ERR(update);
+	}
+
+	if (dt2osp_dev(th->th_dev)->opd_connect_mdt)
+		update->dur_flags = UPDATE_FL_SYNC;
+
+	oth->ot_dur = update;
+	return 0;
+}
+
+static void osp_thandle_get(struct osp_thandle *oth)
+{
+	atomic_inc(&oth->ot_refcount);
+}
+
+static void osp_thandle_put(struct osp_thandle *oth)
+{
+	if (atomic_dec_and_test(&oth->ot_refcount))
+		OBD_FREE_PTR(oth);
 }
 
 /**
@@ -376,45 +562,204 @@ out:
  */
 struct thandle *osp_trans_create(const struct lu_env *env, struct dt_device *d)
 {
-	struct thandle		*th = NULL;
-	struct thandle_update	*tu = NULL;
-	int			 rc = 0;
+	struct osp_thandle		*oth;
+	struct thandle			*th = NULL;
+	ENTRY;
 
-	OBD_ALLOC_PTR(th);
-	if (unlikely(th == NULL))
-		GOTO(out, rc = -ENOMEM);
+	OBD_ALLOC_PTR(oth);
+	if (unlikely(oth == NULL))
+		RETURN(ERR_PTR(-ENOMEM));
 
+	th = &oth->ot_super;
 	th->th_dev = d;
 	th->th_tags = LCT_TX_HANDLE;
-	atomic_set(&th->th_refc, 1);
-	th->th_alloc_size = sizeof(*th);
 
-	OBD_ALLOC_PTR(tu);
-	if (tu == NULL)
-		GOTO(out, rc = -ENOMEM);
+	atomic_set(&oth->ot_refcount, 1);
+	INIT_LIST_HEAD(&oth->ot_dcb_list);
 
-	INIT_LIST_HEAD(&tu->tu_remote_update_list);
-	tu->tu_only_remote_trans = 1;
-	th->th_update = tu;
+	RETURN(th);
+}
 
-out:
+/**
+ * Prepare update request.
+ *
+ * Prepare OUT update ptlrpc request, and the request usually includes
+ * all of updates (stored in \param ureq) from one operation.
+ *
+ * \param[in] env      execution environment
+ * \param[in] imp      import on which ptlrpc request will be sent
+ * \param[in] ureq     hold all of updates which will be packed into the req
+ * \param[in] reqp     request to be created
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+int osp_prep_update_req(const struct lu_env *env, struct obd_import *imp,
+			const struct object_update_request *ureq,
+			struct ptlrpc_request **reqp)
+{
+	struct ptlrpc_request           *req;
+	struct object_update_request    *tmp;
+	int                             ureq_len;
+	int                             rc;
+	ENTRY;
+
+	object_update_request_dump(ureq, D_INFO);
+	req = ptlrpc_request_alloc(imp, &RQF_OUT_UPDATE);
+	if (req == NULL)
+		RETURN(-ENOMEM);
+
+	ureq_len = object_update_request_size(ureq);
+	req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE, RCL_CLIENT,
+			     ureq_len);
+
+	rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, OUT_UPDATE);
 	if (rc != 0) {
-		if (tu != NULL)
-			OBD_FREE_PTR(tu);
-		if (th != NULL)
-			OBD_FREE_PTR(th);
-		th = ERR_PTR(rc);
+		ptlrpc_req_finished(req);
+		RETURN(rc);
+	}
+
+	req_capsule_set_size(&req->rq_pill, &RMF_OUT_UPDATE_REPLY,
+			    RCL_SERVER, OUT_UPDATE_REPLY_SIZE);
+
+	tmp = req_capsule_client_get(&req->rq_pill, &RMF_OUT_UPDATE);
+	memcpy(tmp, ureq, ureq_len);
+
+	ptlrpc_request_set_replen(req);
+	req->rq_request_portal = OUT_PORTAL;
+	req->rq_reply_portal = OSC_REPLY_PORTAL;
+	*reqp = req;
+
+	RETURN(rc);
+}
+
+/**
+* Send update RPC.
+*
+* Send update request to the remote MDT synchronously.
+*
+* \param[in] env	execution environment
+* \param[in] imp	import on which ptlrpc request will be sent
+* \param[in] dt_update	hold all of updates which will be packed into the req
+* \param[in] reqp	request to be created
+*
+* \retval		0 if RPC succeeds.
+* \retval		negative errno if RPC fails.
+*/
+int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
+		    struct dt_update_request *dt_update,
+		    struct ptlrpc_request **reqp)
+{
+	struct obd_import	*imp = osp->opd_obd->u.cli.cl_import;
+	struct ptlrpc_request	*req = NULL;
+	int			rc;
+	ENTRY;
+
+	rc = osp_prep_update_req(env, imp, dt_update->dur_buf.ub_req, &req);
+	if (rc != 0)
+		RETURN(rc);
+
+	/* This will only be called with read-only update, and these updates
+	 * might be used to retrieve update log during recovery process, so
+	 * it will be allowed to send during recovery process */
+	req->rq_allow_replay = 1;
+
+	/* Note: some dt index api might return non-zero result here, like
+	 * osd_index_ea_lookup, so we should only check rc < 0 here */
+	rc = ptlrpc_queue_wait(req);
+	if (rc < 0) {
+		ptlrpc_req_finished(req);
+		dt_update->dur_rc = rc;
+		RETURN(rc);
+	}
+
+	if (reqp != NULL) {
+		*reqp = req;
+		RETURN(rc);
+	}
+
+	dt_update->dur_rc = rc;
+
+	ptlrpc_req_finished(req);
+
+	RETURN(rc);
+}
+
+/**
+ * Add commit callback to transaction.
+ *
+ * Add commit callback to the osp thandle, which will be called
+ * when the thandle is committed remotely.
+ *
+ * \param[in] th	the thandle
+ * \param[in] dcb	commit callback structure
+ *
+ * \retval		only return 0 for now.
+ */
+int osp_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
+{
+	struct osp_thandle *oth = thandle_to_osp_thandle(th);
+
+	LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
+	LASSERT(&dcb->dcb_func != NULL);
+	list_add(&dcb->dcb_linkage, &oth->ot_dcb_list);
+
+	return 0;
+}
+
+static void osp_trans_commit_cb(struct osp_thandle *oth, int result)
+{
+	struct dt_txn_commit_cb *dcb;
+	struct dt_txn_commit_cb *tmp;
+
+	LASSERT(atomic_read(&oth->ot_refcount) > 0);
+	/* call per-transaction callbacks if any */
+	list_for_each_entry_safe(dcb, tmp, &oth->ot_dcb_list,
+				 dcb_linkage) {
+		LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
+			 "commit callback entry: magic=%x name='%s'\n",
+			 dcb->dcb_magic, dcb->dcb_name);
+		list_del_init(&dcb->dcb_linkage);
+		dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
 	}
+}
+
+static void osp_request_commit_cb(struct ptlrpc_request *req)
+{
+	struct thandle *th = req->rq_cb_data;
+	struct osp_thandle *oth = thandle_to_osp_thandle(th);
+	__u64			last_committed_transno = 0;
+	int			result = req->rq_status;
+	ENTRY;
+
+	if (lustre_msg_get_last_committed(req->rq_repmsg))
+		last_committed_transno =
+			lustre_msg_get_last_committed(req->rq_repmsg);
 
-	return th;
+	if (last_committed_transno <
+		req->rq_import->imp_peer_committed_transno)
+		last_committed_transno =
+			req->rq_import->imp_peer_committed_transno;
+
+	CDEBUG(D_HA, "trans no "LPU64" committed transno "LPU64"\n",
+	       req->rq_transno, last_committed_transno);
+
+	/* If the transaction is not really committed, mark result = 1 */
+	if (req->rq_transno != 0 &&
+	    (req->rq_transno > last_committed_transno) && result == 0)
+		result = 1;
+
+	osp_trans_commit_cb(oth, result);
+	req->rq_committed = 1;
+	osp_thandle_put(oth);
+	EXIT;
 }
 
 /**
  * Trigger the request for remote updates.
  *
- * If the transaction is a remote transaction, then related remote updates
- * will be sent asynchronously; otherwise, the cross MDTs transaction will
- * be synchronized.
+ * If th_sync is set, then the request will be sent synchronously,
+ * otherwise, the RPC will be sent asynchronously.
  *
  * Please refer to osp_trans_create() for transaction type.
  *
@@ -422,64 +767,129 @@ out:
  * \param[in] osp		pointer to the OSP device
  * \param[in] dt_update		pointer to the dt_update_request
  * \param[in] th		pointer to the transaction handler
- * \param[in] flow_control	whether need to control the flow
+ * \param[out] sent		whether the RPC has been sent
  *
  * \retval			0 for success
  * \retval			negative error number on failure
  */
 static int osp_trans_trigger(const struct lu_env *env, struct osp_device *osp,
 			     struct dt_update_request *dt_update,
-			     struct thandle *th, bool flow_control)
+			     struct thandle *th, int *sent)
 {
-	struct thandle_update	*tu = th->th_update;
-	int			 rc = 0;
+	struct osp_update_args	*args;
+	struct ptlrpc_request	*req;
+	int	rc = 0;
+	ENTRY;
 
-	LASSERT(tu != NULL);
-
-	if (is_only_remote_trans(th)) {
-		struct osp_async_update_args	*args;
-		struct ptlrpc_request		*req;
-
-		list_del_init(&dt_update->dur_list);
-		rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
-					 dt_update->dur_req, &req);
-		if (rc == 0) {
-			args = ptlrpc_req_async_args(req);
-			args->oaua_update = dt_update;
-			args->oaua_flow_control = flow_control;
-			req->rq_interpret_reply =
-				osp_async_update_interpret;
-			ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
-		} else {
-			out_destroy_update_req(dt_update);
+	rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
+				 dt_update->dur_buf.ub_req, &req);
+	if (rc != 0)
+		RETURN(rc);
+
+	*sent = 1;
+	req->rq_interpret_reply = osp_update_interpret;
+	args = ptlrpc_req_async_args(req);
+	args->oaua_update = dt_update;
+
+	if (is_only_remote_trans(th) && !th->th_sync &&
+	    !th->th_wait_submit) {
+		args->oaua_flow_control = true;
+
+		if (!osp->opd_connect_mdt) {
+			down_read(&osp->opd_async_updates_rwsem);
+			args->oaua_count = &osp->opd_async_updates_count;
+			args->oaua_waitq = &osp->opd_syn_barrier_waitq;
+			up_read(&osp->opd_async_updates_rwsem);
+			atomic_inc(args->oaua_count);
 		}
+
+		ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
 	} else {
-		th->th_sync = 1;
-		rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import,
-				     dt_update, NULL);
+		struct osp_thandle *oth = thandle_to_osp_thandle(th);
+		struct lu_device *top_device;
+
+		/* If the transaction is created during MDT recoverying
+		 * process, it means this is an recovery update, we need
+		 * to let OSP send it anyway without checking recoverying
+		 * status, in case the other target is being recoveried
+		 * at the same time, and if we wait here for the import
+		 * to be recoveryed, it might cause deadlock */
+		top_device = osp->opd_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
+		if (top_device->ld_obd->obd_recovering)
+			req->rq_allow_replay = 1;
+
+		args->oaua_flow_control = false;
+		req->rq_commit_cb = osp_request_commit_cb;
+		req->rq_cb_data = th;
+		osp_thandle_get(oth); /* for commit callback */
+		osp_get_rpc_lock(osp);
+		rc = ptlrpc_queue_wait(req);
+		osp_put_rpc_lock(osp);
+		if (req->rq_transno == 0 && !req->rq_committed)
+			osp_thandle_put(oth);
+		else
+			oth->ot_dur = NULL;
+		ptlrpc_req_finished(req);
 	}
 
-	return rc;
+	RETURN(rc);
 }
 
 /**
- * The OSP layer dt_device_operations::dt_trans_start() interface
- * to start the transaction.
+ * Get local thandle for osp_thandle
  *
- * If the transaction is a remote transaction, then related remote
- * updates will be triggered in the osp_trans_stop(); otherwise the
- * transaction contains both local and remote update(s), then when
- * the OUT RPC will be triggered depends on the operation, and is
- * indicated by the dt_device::tu_sent_after_local_trans, for example:
+ * Get the local OSD thandle from the OSP thandle. Currently, there
+ * are a few OSP API (osp_object_create() and osp_sync_add()) needs
+ * to update the object on local OSD device.
+ *
+ * If the osp_thandle comes from normal stack (MDD->LOD->OSP), then
+ * we will get local thandle by thandle_get_sub_by_dt.
  *
- * 1) If it is remote create, it will send the remote req after local
- * transaction. i.e. create the object locally first, then insert the
- * remote name entry.
+ * If the osp_thandle is remote thandle (th_top == NULL, only used
+ * by LFSCK), then it will create a local thandle, and stop it in
+ * osp_trans_stop(). And this only happens on OSP for OST.
  *
- * 2) If it is remote unlink, it will send the remote req before the
- * local transaction, i.e. delete the name entry remotely first, then
- * destroy the local object.
+ * These are temporary solution, once OSP accessing OSD object is
+ * being fixed properly, this function should be removed. XXX
+ *
+ * \param[in] env		pointer to the thread context
+ * \param[in] th		pointer to the transaction handler
+ * \param[in] dt		pointer to the OSP device
+ *
+ * \retval			pointer to the local thandle
+ * \retval			ERR_PTR(errno) if it fails.
+ **/
+struct thandle *osp_get_storage_thandle(const struct lu_env *env,
+					struct thandle *th,
+					struct osp_device *osp)
+{
+	struct osp_thandle	*oth;
+	struct thandle		*local_th;
+
+	if (th->th_top != NULL)
+		return thandle_get_sub_by_dt(env, th->th_top,
+					     osp->opd_storage);
+
+	LASSERT(!osp->opd_connect_mdt);
+	oth = thandle_to_osp_thandle(th);
+	if (oth->ot_storage_th != NULL)
+		return oth->ot_storage_th;
+
+	local_th = dt_trans_create(env, osp->opd_storage);
+	if (IS_ERR(local_th))
+		return local_th;
+
+	oth->ot_storage_th = local_th;
+
+	return local_th;
+}
+
+/**
+ * The OSP layer dt_device_operations::dt_trans_start() interface
+ * to start the transaction.
  *
+ * If the transaction is a remote transaction, then related remote
+ * updates will be triggered in the osp_trans_stop().
  * Please refer to osp_trans_create() for transaction type.
  *
  * \param[in] env		pointer to the thread context
@@ -492,31 +902,20 @@ static int osp_trans_trigger(const struct lu_env *env, struct osp_device *osp,
 int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
 		    struct thandle *th)
 {
-	struct thandle_update		*tu = th->th_update;
-	struct dt_update_request	*dt_update;
-	int				 rc = 0;
+	struct osp_thandle	*oth = thandle_to_osp_thandle(th);
 
-	if (tu == NULL)
-		return rc;
-
-	/* Check whether there are updates related with this OSP */
-	dt_update = out_find_update(tu, dt);
-	if (dt_update == NULL)
-		return rc;
-
-	if (!is_only_remote_trans(th) && !tu->tu_sent_after_local_trans)
-		rc = osp_trans_trigger(env, dt2osp_dev(dt), dt_update, th,
-				       false);
-
-	return rc;
+	/* For remote thandle, if there are local thandle, start it here*/
+	if (is_only_remote_trans(th) && oth->ot_storage_th != NULL)
+		return dt_trans_start(env, oth->ot_storage_th->th_dev,
+				      oth->ot_storage_th);
+	return 0;
 }
 
 /**
  * The OSP layer dt_device_operations::dt_trans_stop() interface
  * to stop the transaction.
  *
- * If the transaction is a remote transaction, or the update handler
- * is marked as 'tu_sent_after_local_trans', then related remote
+ * If the transaction is a remote transaction, related remote
  * updates will be triggered here via osp_trans_trigger().
  *
  * For synchronous mode update or any failed update, the request
@@ -534,61 +933,82 @@ int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
 int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
 		   struct thandle *th)
 {
-	struct thandle_update		*tu = th->th_update;
-	struct dt_update_request	*dt_update;
-	int				 rc = 0;
-
-	LASSERT(tu != NULL);
-	LASSERT(tu != LP_POISON);
-
-	/* Check whether there are updates related with this OSP */
-	dt_update = out_find_update(tu, dt);
-	if (dt_update == NULL) {
-		if (!is_only_remote_trans(th))
-			return rc;
-		goto put;
+	struct osp_thandle	 *oth = thandle_to_osp_thandle(th);
+	struct dt_update_request *dt_update;
+	int			 rc = 0;
+	int			 sent = 0;
+	ENTRY;
+
+	/* For remote transaction, if there is local storage thandle,
+	 * stop it first */
+	if (oth->ot_storage_th != NULL && th->th_top == NULL) {
+		dt_trans_stop(env, oth->ot_storage_th->th_dev,
+			      oth->ot_storage_th);
+		oth->ot_storage_th = NULL;
 	}
 
-	if (dt_update->dur_req->ourq_count == 0) {
-		out_destroy_update_req(dt_update);
-		goto put;
+	dt_update = oth->ot_dur;
+	if (dt_update == NULL || th->th_result != 0) {
+		rc = th->th_result;
+		GOTO(out, rc);
 	}
 
-	if (is_only_remote_trans(th)) {
-		if (th->th_result == 0) {
-			struct osp_device *osp = dt2osp_dev(th->th_dev);
-			struct client_obd *cli = &osp->opd_obd->u.cli;
+	LASSERT(dt_update != LP_POISON);
 
-			rc = obd_get_request_slot(cli);
-			if (!osp->opd_imp_active || osp->opd_got_disconnected) {
-				if (rc == 0)
-					obd_put_request_slot(cli);
+	/* If there are no updates, destroy dt_update and thandle */
+	if (dt_update->dur_buf.ub_req == NULL ||
+	    dt_update->dur_buf.ub_req->ourq_count == 0)
+		GOTO(out, rc);
 
-				rc = -ENOTCONN;
-			}
+	if (is_only_remote_trans(th) && !th->th_sync &&
+	    !th->th_wait_submit) {
+		struct osp_device *osp = dt2osp_dev(th->th_dev);
+		struct client_obd *cli = &osp->opd_obd->u.cli;
 
-			if (rc != 0) {
-				out_destroy_update_req(dt_update);
-				goto put;
-			}
+		rc = obd_get_request_slot(cli);
+		if (rc != 0)
+			GOTO(out, rc);
 
-			rc = osp_trans_trigger(env, dt2osp_dev(dt),
-					       dt_update, th, true);
-			if (rc != 0)
-				obd_put_request_slot(cli);
-		} else {
-			rc = th->th_result;
-			out_destroy_update_req(dt_update);
+		if (!osp->opd_imp_active || !osp->opd_imp_connected) {
+			obd_put_request_slot(cli);
+			GOTO(out, rc = -ENOTCONN);
 		}
+
+		rc = osp_trans_trigger(env, dt2osp_dev(dt),
+				       dt_update, th, &sent);
+		if (rc != 0)
+			obd_put_request_slot(cli);
 	} else {
-		if (tu->tu_sent_after_local_trans)
-			rc = osp_trans_trigger(env, dt2osp_dev(dt),
-					       dt_update, th, false);
-		rc = dt_update->dur_rc;
-		out_destroy_update_req(dt_update);
+		rc = osp_trans_trigger(env, dt2osp_dev(dt), dt_update,
+				       th, &sent);
 	}
 
-put:
-	thandle_put(th);
-	return rc;
+out:
+	/* If RPC is triggered successfully, dt_update will be freed in
+	 * osp_update_interpreter() */
+	if (sent == 0) {
+		struct osp_update_callback *ouc;
+		struct osp_update_callback *next;
+
+		if (dt_update != NULL) {
+			list_for_each_entry_safe(ouc, next,
+						 &dt_update->dur_cb_items,
+						 ouc_list) {
+				list_del_init(&ouc->ouc_list);
+				if (ouc->ouc_interpreter != NULL)
+					ouc->ouc_interpreter(env, NULL, NULL,
+							   ouc->ouc_obj,
+							   ouc->ouc_data, 0,
+							   rc);
+				osp_update_callback_fini(env, ouc);
+			}
+		}
+		osp_trans_commit_cb(oth, rc);
+		dt_update_request_destroy(dt_update);
+		oth->ot_dur = NULL;
+	}
+
+	osp_thandle_put(oth);
+
+	RETURN(rc);
 }