Whamcloud - gitweb
LU-13195 osp: invalidate object on write error
[fs/lustre-release.git] / lustre / osp / osp_md_object.c
index b66a66a..b607b78 100644 (file)
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2013, Intel Corporation.
+ * Copyright (c) 2013, 2017, Intel Corporation.
  */
 /*
  * lustre/osp/osp_md_object.c
  *
- * Lustre MDT Proxy Device
+ * OST/MDT proxy device (OSP) Metadata methods
+ *
+ * This file implements methods for remote MD object, which include
+ * dt_object_operations, dt_index_operations and dt_body_operations.
+ *
+ * If there are multiple MDTs in one filesystem, one operation might
+ * include modifications in several MDTs. In such cases, clients
+ * send the RPC to the master MDT, then the operation is decomposed into
+ * object updates which will be dispatched to OSD or OSP. The local updates
+ * go to local OSD and the remote updates go to OSP. In OSP, these remote
+ * object updates will be packed into an update RPC, sent to the remote MDT
+ * and handled by Object Update Target (OUT).
+ *
+ * In DNE phase I, because of missing complete recovery solution, updates
+ * will be executed in order and synchronously.
+ *     1. The transaction is created.
+ *     2. In transaction declare, it collects and packs remote
+ *        updates (in osp_md_declare_xxx()).
+ *     3. In transaction start, it sends these remote updates
+ *        to remote MDTs, which will execute these updates synchronously.
+ *     4. In transaction execute phase, the local updates will be executed
+ *        synchronously.
  *
  * Author: Di Wang <di.wang@intel.com>
  */
 
 #define DEBUG_SUBSYSTEM S_MDS
 
+#include <llog_swab.h>
 #include <lustre_log.h>
-#include <lustre_update.h>
 #include "osp_internal.h"
-static const char dot[] = ".";
-static const char dotdot[] = "..";
-
-static int osp_prep_update_req(const struct lu_env *env,
-                              struct osp_device *osp,
-                              struct update_buf *ubuf, int ubuf_len,
-                              struct ptlrpc_request **reqp)
-{
-       struct obd_import      *imp;
-       struct ptlrpc_request  *req;
-       struct update_buf      *tmp;
-       int                     rc;
-       ENTRY;
-
-       imp = osp->opd_obd->u.cli.cl_import;
-       LASSERT(imp);
-
-       req = ptlrpc_request_alloc(imp, &RQF_UPDATE_OBJ);
-       if (req == NULL)
-               RETURN(-ENOMEM);
-
-       req_capsule_set_size(&req->rq_pill, &RMF_UPDATE, RCL_CLIENT,
-                            UPDATE_BUFFER_SIZE);
-
-       rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, UPDATE_OBJ);
-       if (rc != 0) {
-               ptlrpc_req_finished(req);
-               RETURN(rc);
-       }
-
-       req_capsule_set_size(&req->rq_pill, &RMF_UPDATE_REPLY, RCL_SERVER,
-                            UPDATE_BUFFER_SIZE);
-
-       tmp = req_capsule_client_get(&req->rq_pill, &RMF_UPDATE);
-       memcpy(tmp, ubuf, ubuf_len);
-
-       ptlrpc_request_set_replen(req);
-
-       *reqp = req;
-
-       RETURN(rc);
-}
-
-static int osp_remote_sync(const struct lu_env *env, struct dt_device *dt,
-                          struct update_request *update,
-                          struct ptlrpc_request **reqp)
-{
-       struct osp_device       *osp = dt2osp_dev(dt);
-       struct ptlrpc_request   *req = NULL;
-       int                     rc;
-       ENTRY;
-
-       rc = osp_prep_update_req(env, osp, update->ur_buf, UPDATE_BUFFER_SIZE,
-                                &req);
-       if (rc)
-               RETURN(rc);
-
-       /* 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);
-               update->ur_rc = rc;
-               RETURN(rc);
-       }
-
-       if (reqp != NULL) {
-               *reqp = req;
-               RETURN(rc);
-       }
-
-       update->ur_rc = rc;
-
-       ptlrpc_req_finished(req);
 
-       RETURN(rc);
-}
+#define OUT_UPDATE_BUFFER_SIZE_ADD     4096
+#define OUT_UPDATE_BUFFER_SIZE_MAX     (256 * 4096)  /*  1M update size now */
 
 /**
- * Create a new update request for the device.
+ * Interpreter call for object creation
+ *
+ * Object creation interpreter, which will be called after creating
+ * the remote object to set flags and status.
+ *
+ * \param[in] env      execution environment
+ * \param[in] reply    update reply
+ * \param[in] req      ptlrpc update request for creating object
+ * \param[in] obj      object to be created
+ * \param[in] data     data used in this function.
+ * \param[in] index    index(position) of create update in the whole
+ *                      updates
+ * \param[in] rc       update result on the remote MDT.
+ *
+ * \retval             only return 0 for now
  */
-static struct update_request
-*osp_create_update_req(struct dt_device *dt)
+static int osp_create_interpreter(const struct lu_env *env,
+                                 struct object_update_reply *reply,
+                                 struct ptlrpc_request *req,
+                                 struct osp_object *obj,
+                                 void *data, int index, int rc)
 {
-       struct update_request *update;
-
-       OBD_ALLOC_PTR(update);
-       if (!update)
-               return ERR_PTR(-ENOMEM);
-
-       OBD_ALLOC_LARGE(update->ur_buf, UPDATE_BUFFER_SIZE);
-       if (update->ur_buf == NULL) {
-               OBD_FREE_PTR(update);
-               return ERR_PTR(-ENOMEM);
+       if (rc != 0 && rc != -EEXIST) {
+               obj->opo_obj.do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
+               obj->opo_non_exist = 1;
        }
 
-       CFS_INIT_LIST_HEAD(&update->ur_list);
-       update->ur_dt = dt;
-       update->ur_batchid = 0;
-       update->ur_buf->ub_magic = UPDATE_BUFFER_MAGIC;
-       update->ur_buf->ub_count = 0;
-
-       return update;
-}
-
-static void osp_destroy_update_req(struct update_request *update)
-{
-       if (update == NULL)
-               return;
-
-       cfs_list_del(&update->ur_list);
-       if (update->ur_buf != NULL)
-               OBD_FREE_LARGE(update->ur_buf, UPDATE_BUFFER_SIZE);
-
-       OBD_FREE_PTR(update);
-       return;
-}
-
-int osp_trans_stop(const struct lu_env *env, struct thandle *th)
-{
-       int rc = 0;
-
-       rc = th->th_current_request->ur_rc;
-       osp_destroy_update_req(th->th_current_request);
-       th->th_current_request = NULL;
+       /*
+        * invalidate opo cache for the object after the object is created, so
+        * attr_get will try to get attr from remote object.
+        */
+       osp_obj_invalidate_cache(obj);
 
-       return rc;
+       return 0;
 }
 
 /**
- * In DNE phase I, all remote updates will be packed into RPC (the format
- * description is in lustre_idl.h) during declare phase, all of updates
- * are attached to the transaction, one entry per OSP. Then in trans start,
- * LOD will walk through these entries and send these UPDATEs to the remote
- * MDT to be executed synchronously.
+ * Implementation of dt_object_operations::do_declare_create
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       remote object to be created
+ * \param[in] attr     attribute of the created object
+ * \param[in] hint     creation hint
+ * \param[in] dof      creation format information
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
  */
-int osp_trans_start(const struct lu_env *env, struct dt_device *dt,
-                   struct thandle *th)
+int osp_md_declare_create(const struct lu_env *env, struct dt_object *dt,
+                         struct lu_attr *attr, struct dt_allocation_hint *hint,
+                         struct dt_object_format *dof, struct thandle *th)
 {
-       struct update_request *update;
-       int rc = 0;
-
-       /* In phase I, if the transaction includes remote updates, the local
-        * update should be synchronized, so it will set th_sync = 1 */
-       update = th->th_current_request;
-       LASSERT(update != NULL && update->ur_dt == dt);
-       if (update->ur_buf->ub_count > 0) {
-               rc = osp_remote_sync(env, dt, update, NULL);
-               th->th_sync = 1;
-       }
-
-       RETURN(rc);
+       return osp_trans_update_request_create(th);
 }
 
-/**
- * Insert the update into the th_bufs for the device.
- */
-static int osp_insert_update(const struct lu_env *env,
-                            struct update_request *update, int op,
-                            struct lu_fid *fid, int count,
-                            int *lens, char **bufs)
+struct object_update *
+update_buffer_get_update(struct object_update_request *request,
+                        unsigned int index)
 {
-       struct update_buf    *ubuf = update->ur_buf;
-       struct update        *obj_update;
-       char                 *ptr;
-       int                   i;
-       int                   update_length;
-       int                   rc = 0;
-       ENTRY;
-
-       obj_update = (struct update *)((char *)ubuf +
-                     cfs_size_round(update_buf_size(ubuf)));
-
-       /* Check update size to make sure it can fit into the buffer */
-       update_length = cfs_size_round(offsetof(struct update,
-                                      u_bufs[0]));
-       for (i = 0; i < count; i++)
-               update_length += cfs_size_round(lens[i]);
-
-       if (cfs_size_round(update_buf_size(ubuf)) + update_length >
-           UPDATE_BUFFER_SIZE || ubuf->ub_count >= UPDATE_MAX_OPS) {
-               CERROR("%s: insert up %p, idx %d cnt %d len %lu: rc = %d\n",
-                       update->ur_dt->dd_lu_dev.ld_obd->obd_name, ubuf,
-                       update_length, ubuf->ub_count, update_buf_size(ubuf),
-                       -E2BIG);
-               RETURN(-E2BIG);
-       }
-
-       if (count > UPDATE_BUF_COUNT) {
-               CERROR("%s: Insert too much params %d "DFID" op %d: rc = %d\n",
-                       update->ur_dt->dd_lu_dev.ld_obd->obd_name, count,
-                       PFID(fid), op, -E2BIG);
-               RETURN(-E2BIG);
-       }
-
-       /* fill the update into the update buffer */
-       fid_cpu_to_le(&obj_update->u_fid, fid);
-       obj_update->u_type = cpu_to_le32(op);
-       obj_update->u_batchid = update->ur_batchid;
-       for (i = 0; i < count; i++)
-               obj_update->u_lens[i] = cpu_to_le32(lens[i]);
-
-       ptr = (char *)obj_update +
-                       cfs_size_round(offsetof(struct update, u_bufs[0]));
-       for (i = 0; i < count; i++)
-               LOGL(bufs[i], lens[i], ptr);
+       void    *ptr;
+       int     i;
 
-       ubuf->ub_count++;
+       if (index > request->ourq_count)
+               return NULL;
 
-       CDEBUG(D_INFO, "%s: %p "DFID" idx %d: op %d params %d:%lu\n",
-              update->ur_dt->dd_lu_dev.ld_obd->obd_name, ubuf, PFID(fid),
-              ubuf->ub_count, op, count, update_buf_size(ubuf));
+       ptr = &request->ourq_updates[0];
+       for (i = 0; i < index; i++)
+               ptr += object_update_size(ptr);
 
-       RETURN(rc);
-}
-
-static struct update_request
-*osp_find_update(struct thandle *th, struct dt_device *dt_dev)
-{
-       struct update_request   *update;
-
-       /* Because transaction api does not proivde the interface
-        * to transfer the update from LOD to OSP,  we need walk
-        * remote update list to find the update, this probably
-        * should move to LOD layer, when update can be part of
-        * the trancation api parameter. XXX */
-       cfs_list_for_each_entry(update, &th->th_remote_update_list, ur_list) {
-               if (update->ur_dt == dt_dev)
-                       return update;
-       }
-       return NULL;
-}
-
-static inline void osp_md_add_update_batchid(struct update_request *update)
-{
-       update->ur_batchid++;
+       return ptr;
 }
 
 /**
- * Find one loc in th_dev/dev_obj_update for the update,
- * Because only one thread can access this thandle, no need
- * lock now.
+ * Implementation of dt_object_operations::do_create
+ *
+ * It adds an OUT_CREATE sub-request into the OUT RPC that will be flushed
+ * when the transaction stop, and sets necessary flags for created object.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be created
+ * \param[in] attr     attribute of the created object
+ * \param[in] hint     creation hint
+ * \param[in] dof      creation format information
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing creation succeeds.
+ * \retval             negative errno if packing creation fails.
  */
-static struct update_request
-*osp_find_create_update_loc(struct thandle *th, struct dt_object *dt)
-{
-       struct dt_device        *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-       struct update_request   *update;
-       ENTRY;
-
-       update = osp_find_update(th, dt_dev);
-       if (update != NULL)
-               RETURN(update);
-
-       update = osp_create_update_req(dt_dev);
-       if (IS_ERR(update))
-               RETURN(update);
-
-       cfs_list_add_tail(&update->ur_list, &th->th_remote_update_list);
-
-       RETURN(update);
-}
-
-static int osp_get_attr_from_req(const struct lu_env *env,
-                                struct ptlrpc_request *req,
-                                struct lu_attr *attr, int index)
-{
-       struct update_reply     *reply;
-       struct obdo             *lobdo = &osp_env_info(env)->osi_obdo;
-       struct obdo             *wobdo;
-       int                     size;
-
-       LASSERT(attr != NULL);
-
-       reply = req_capsule_server_sized_get(&req->rq_pill, &RMF_UPDATE_REPLY,
-                                            UPDATE_BUFFER_SIZE);
-       if (reply == NULL || reply->ur_version != UPDATE_REPLY_V1)
-               return -EPROTO;
-
-       size = update_get_reply_buf(reply, (void **)&wobdo, index);
-       if (size != sizeof(struct obdo))
-               return -EPROTO;
-
-       obdo_le_to_cpu(wobdo, wobdo);
-       lustre_get_wire_obdo(NULL, lobdo, wobdo);
-       la_from_obdo(attr, lobdo, lobdo->o_valid);
-
-       return 0;
-}
-
-static int osp_md_declare_object_create(const struct lu_env *env,
-                                       struct dt_object *dt,
-                                       struct lu_attr *attr,
-                                       struct dt_allocation_hint *hint,
-                                       struct dt_object_format *dof,
-                                       struct thandle *th)
+int osp_md_create(const struct lu_env *env, struct dt_object *dt,
+                 struct lu_attr *attr, struct dt_allocation_hint *hint,
+                 struct dt_object_format *dof, struct thandle *th)
 {
-       struct osp_thread_info  *osi = osp_env_info(env);
-       struct update_request   *update;
-       struct lu_fid           *fid1;
-       int                     sizes[2] = {sizeof(struct obdo), 0};
-       char                    *bufs[2] = {NULL, NULL};
-       int                     buf_count;
-       int                     rc;
-
+       struct osp_update_request *update;
+       struct osp_object *obj = dt2osp_obj(dt);
+       int rc;
 
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
 
-       osi->osi_obdo.o_valid = 0;
-       LASSERT(S_ISDIR(attr->la_mode));
-       obdo_from_la(&osi->osi_obdo, attr, attr->la_valid);
-       lustre_set_wire_obdo(NULL, &osi->osi_obdo, &osi->osi_obdo);
-       obdo_cpu_to_le(&osi->osi_obdo, &osi->osi_obdo);
-
-       bufs[0] = (char *)&osi->osi_obdo;
-       buf_count = 1;
-       fid1 = (struct lu_fid *)lu_object_fid(&dt->do_lu);
-       if (hint->dah_parent) {
-               struct lu_fid *fid2;
-               struct lu_fid *tmp_fid = &osi->osi_fid;
-
-               fid2 = (struct lu_fid *)lu_object_fid(&hint->dah_parent->do_lu);
-               fid_cpu_to_le(tmp_fid, fid2);
-               sizes[1] = sizeof(*tmp_fid);
-               bufs[1] = (char *)tmp_fid;
-               buf_count++;
-       }
+       LASSERT(attr->la_valid & LA_TYPE);
+       rc = OSP_UPDATE_RPC_PACK(env, out_create_pack, update,
+                                lu_object_fid(&dt->do_lu), attr, hint, dof);
+       if (rc != 0)
+               GOTO(out, rc);
 
-       if (lu_object_exists(&dt->do_lu)) {
-               /* If the object already exists, we needs to destroy
-                * this orphan object first.
-                *
-                * The scenario might happen in this case
-                *
-                * 1. client send remote create to MDT0.
-                * 2. MDT0 send create update to MDT1.
-                * 3. MDT1 finished create synchronously.
-                * 4. MDT0 failed and reboot.
-                * 5. client resend remote create to MDT0.
-                * 6. MDT0 tries to resend create update to MDT1,
-                *    but find the object already exists
-                */
-               CDEBUG(D_HA, "%s: object "DFID" exists, destroy this orphan\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name, PFID(fid1));
-
-               rc = osp_insert_update(env, update, OBJ_REF_DEL, fid1, 0,
-                                      NULL, NULL);
-               if (rc != 0)
-                       GOTO(out, rc);
-
-               if (S_ISDIR(lu_object_attr(&dt->do_lu))) {
-                       /* decrease for ".." */
-                       rc = osp_insert_update(env, update, OBJ_REF_DEL, fid1,
-                                              0, NULL, NULL);
-                       if (rc != 0)
-                               GOTO(out, rc);
-               }
+       rc = osp_insert_update_callback(env, update, dt2osp_obj(dt), NULL,
+                                       osp_create_interpreter);
 
-               rc = osp_insert_update(env, update, OBJ_DESTROY, fid1, 0, NULL,
-                                      NULL);
-               if (rc != 0)
-                       GOTO(out, rc);
+       if (rc < 0)
+               GOTO(out, rc);
 
-               dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
-               /* Increase batchid to add this orphan object deletion
-                * to separate transaction */
-               osp_md_add_update_batchid(update);
-       }
+       dt->do_lu.lo_header->loh_attr |= LOHA_EXISTS | (attr->la_mode & S_IFMT);
+       dt2osp_obj(dt)->opo_non_exist = 0;
+       obj->opo_stale = 0;
 
-       rc = osp_insert_update(env, update, OBJ_CREATE, fid1, buf_count, sizes,
-                              bufs);
+       obj->opo_attr = *attr;
 out:
-       if (rc)
-               CERROR("%s: Insert update error: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name, rc);
-
        return rc;
 }
 
-static int osp_md_object_create(const struct lu_env *env, struct dt_object *dt,
-                               struct lu_attr *attr,
-                               struct dt_allocation_hint *hint,
-                               struct dt_object_format *dof,
-                               struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_declare_ref_del
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to decrease the reference count.
+ * \param[in] th       the transaction handle of refcount decrease.
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+static int osp_md_declare_ref_del(const struct lu_env *env,
+                                 struct dt_object *dt, struct thandle *th)
 {
-       struct osp_object  *obj = dt2osp_obj(dt);
-
-       CDEBUG(D_INFO, "create object "DFID"\n",
-              PFID(&dt->do_lu.lo_header->loh_fid));
-
-       /* Because the create update RPC will be sent during declare phase,
-        * if creation reaches here, it means the object has been created
-        * successfully */
-       dt->do_lu.lo_header->loh_attr |= LOHA_EXISTS | (attr->la_mode & S_IFMT);
-       obj->opo_empty = 1;
-
-       return 0;
+       return osp_trans_update_request_create(th);
 }
 
-static int osp_md_declare_object_ref_del(const struct lu_env *env,
-                                        struct dt_object *dt,
-                                        struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_ref_del
+ *
+ * Add an OUT_REF_DEL sub-request into the OUT RPC that will be
+ * flushed when the transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to decrease the reference count
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing ref_del succeeds.
+ * \retval             negative errno if packing fails.
+ */
+static int osp_md_ref_del(const struct lu_env *env, struct dt_object *dt,
+                         struct thandle *th)
 {
-       struct update_request   *update;
-       struct lu_fid           *fid;
-       int                     rc;
-
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                     (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
-
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
+       struct osp_update_request *update;
+       int rc;
 
-       rc = osp_insert_update(env, update, OBJ_REF_DEL, fid, 0, NULL, NULL);
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
 
+       rc = OSP_UPDATE_RPC_PACK(env, out_ref_del_pack, update,
+                                lu_object_fid(&dt->do_lu));
        return rc;
 }
 
-static int osp_md_object_ref_del(const struct lu_env *env,
-                                struct dt_object *dt,
-                                struct thandle *th)
-{
-       CDEBUG(D_INFO, "ref del object "DFID"\n",
-              PFID(&dt->do_lu.lo_header->loh_fid));
-
-       return 0;
-}
-
+/**
+ * Implementation of dt_object_operations::do_declare_ref_del
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object on which to increase the reference count.
+ * \param[in] th       the transaction handle.
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
 static int osp_md_declare_ref_add(const struct lu_env *env,
                                  struct dt_object *dt, struct thandle *th)
 {
-       struct update_request   *update;
-       struct lu_fid           *fid;
-       int                     rc;
-
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
-
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
-
-       rc = osp_insert_update(env, update, OBJ_REF_ADD, fid, 0, NULL, NULL);
-
-       return rc;
+       return osp_trans_update_request_create(th);
 }
 
-static int osp_md_object_ref_add(const struct lu_env *env,
-                                struct dt_object *dt,
-                                struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_ref_add
+ *
+ * Add an OUT_REF_ADD sub-request into the OUT RPC that will be flushed
+ * when the transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object on which to increase the reference count
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing ref_add succeeds.
+ * \retval             negative errno if packing fails.
+ */
+static int osp_md_ref_add(const struct lu_env *env, struct dt_object *dt,
+                         struct thandle *th)
 {
-       CDEBUG(D_INFO, "ref add object "DFID"\n",
-              PFID(&dt->do_lu.lo_header->loh_fid));
+       struct osp_update_request *update;
+       int rc;
 
-       return 0;
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
+
+       rc = OSP_UPDATE_RPC_PACK(env, out_ref_add_pack, update,
+                                lu_object_fid(&dt->do_lu));
+       return rc;
 }
 
+/**
+ * Implementation of dt_object_operations::do_ah_init
+ *
+ * Initialize the allocation hint for object creation, which is usually called
+ * before the creation, and these hints (parent and child mode) will be sent to
+ * the remote Object Update Target (OUT) and used in the object create process,
+ * same as OSD object creation.
+ *
+ * \param[in] env      execution environment
+ * \param[in] ah       the hint to be initialized
+ * \param[in] parent   the parent of the object
+ * \param[in] child    the object to be created
+ * \param[in] child_mode the mode of the created object
+ */
 static void osp_md_ah_init(const struct lu_env *env,
                           struct dt_allocation_hint *ah,
                           struct dt_object *parent,
@@ -513,639 +299,1024 @@ static void osp_md_ah_init(const struct lu_env *env,
 {
        LASSERT(ah);
 
-       memset(ah, 0, sizeof(*ah));
        ah->dah_parent = parent;
        ah->dah_mode = child_mode;
 }
 
-static int osp_md_declare_attr_set(const struct lu_env *env,
-                                  struct dt_object *dt,
-                                  const struct lu_attr *attr,
-                                  struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_declare_attr_get
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object on which to set attributes
+ * \param[in] attr     attributes to be set
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+int osp_md_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
+                           const struct lu_attr *attr, struct thandle *th)
 {
-       struct osp_thread_info *osi = osp_env_info(env);
-       struct update_request  *update;
-       struct lu_fid          *fid;
-       int                     size = sizeof(struct obdo);
-       char                   *buf;
-       int                     rc;
-
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
-
-       osi->osi_obdo.o_valid = 0;
-       LASSERT(!(attr->la_valid & (LA_MODE | LA_TYPE)));
-       obdo_from_la(&osi->osi_obdo, (struct lu_attr *)attr,
-                    attr->la_valid);
-       lustre_set_wire_obdo(NULL, &osi->osi_obdo, &osi->osi_obdo);
-       obdo_cpu_to_le(&osi->osi_obdo, &osi->osi_obdo);
+       return osp_trans_update_request_create(th);
+}
 
-       buf = (char *)&osi->osi_obdo;
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
+/**
+ * Implementation of dt_object_operations::do_attr_set
+ *
+ * Set attributes to the specified remote object.
+ *
+ * Add the OUT_ATTR_SET sub-request into the OUT RPC that will be flushed
+ * when the transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to set attributes
+ * \param[in] attr     attributes to be set
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing attr_set succeeds.
+ * \retval             negative errno if packing fails.
+ */
+int osp_md_attr_set(const struct lu_env *env, struct dt_object *dt,
+                   const struct lu_attr *attr, struct thandle *th)
+{
+       struct osp_update_request *update;
+       int rc;
 
-       rc = osp_insert_update(env, update, OBJ_ATTR_SET, fid, 1, &size, &buf);
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
 
+       rc = OSP_UPDATE_RPC_PACK(env, out_attr_set_pack, update,
+                                lu_object_fid(&dt->do_lu), attr);
        return rc;
 }
 
-static int osp_md_attr_set(const struct lu_env *env, struct dt_object *dt,
-                          const struct lu_attr *attr, struct thandle *th,
-                          struct lustre_capa *capa)
+/**
+ * Implementation of dt_object_operations::do_read_lock
+ *
+ * osp_md_{read,write}_lock() will only lock the remote object in the
+ * local cache, which uses the semaphore (opo_sem) inside the osp_object to
+ * lock the object. Note: it will not lock the object in the whole cluster,
+ * which relies on the LDLM lock.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be locked
+ * \param[in] role     lock role from MDD layer, see dt_object_role().
+ */
+static void osp_md_read_lock(const struct lu_env *env, struct dt_object *dt,
+                            unsigned role)
 {
-       CDEBUG(D_INFO, "attr set object "DFID"\n",
-              PFID(&dt->do_lu.lo_header->loh_fid));
+       struct osp_object  *obj = dt2osp_obj(dt);
+
+       LASSERT(obj->opo_owner != env);
+       down_read_nested(&obj->opo_sem, role);
 
-       RETURN(0);
+       LASSERT(obj->opo_owner == NULL);
 }
 
-static int osp_md_declare_xattr_set(const struct lu_env *env,
-                                   struct dt_object *dt,
-                                   const struct lu_buf *buf,
-                                   const char *name, int flag,
-                                   struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_write_lock
+ *
+ * Lock the remote object in write mode.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be locked
+ * \param[in] role     lock role from MDD layer, see dt_object_role().
+ */
+static void osp_md_write_lock(const struct lu_env *env, struct dt_object *dt,
+                             unsigned role)
 {
-       struct update_request   *update;
-       struct lu_fid           *fid;
-       int                     sizes[3] = {strlen(name), buf->lb_len,
-                                           sizeof(int)};
-       char                    *bufs[3] = {(char *)name, (char *)buf->lb_buf };
-       int                     rc;
+       struct osp_object *obj = dt2osp_obj(dt);
 
-       LASSERT(buf->lb_len > 0 && buf->lb_buf != NULL);
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
+       down_write_nested(&obj->opo_sem, role);
+
+       LASSERT(obj->opo_owner == NULL);
+       obj->opo_owner = env;
+}
 
-       flag = cpu_to_le32(flag);
-       bufs[2] = (char *)&flag;
+/**
+ * Implementation of dt_object_operations::do_read_unlock
+ *
+ * Unlock the read lock of remote object.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be unlocked
+ */
+static void osp_md_read_unlock(const struct lu_env *env, struct dt_object *dt)
+{
+       struct osp_object *obj = dt2osp_obj(dt);
 
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
-       rc = osp_insert_update(env, update, OBJ_XATTR_SET, fid,
-                              ARRAY_SIZE(sizes), sizes, bufs);
+       up_read(&obj->opo_sem);
+}
 
-       return rc;
+/**
+ * Implementation of dt_object_operations::do_write_unlock
+ *
+ * Unlock the write lock of remote object.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be unlocked
+ */
+static void osp_md_write_unlock(const struct lu_env *env, struct dt_object *dt)
+{
+       struct osp_object *obj = dt2osp_obj(dt);
+
+       LASSERT(obj->opo_owner == env);
+       obj->opo_owner = NULL;
+       up_write(&obj->opo_sem);
 }
 
-static int osp_md_xattr_set(const struct lu_env *env, struct dt_object *dt,
-                           const struct lu_buf *buf, const char *name, int fl,
-                           struct thandle *th, struct lustre_capa *capa)
+/**
+ * Implementation of dt_object_operations::do_write_locked
+ *
+ * Test if the object is locked in write mode.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be tested
+ */
+static int osp_md_write_locked(const struct lu_env *env, struct dt_object *dt)
 {
-       CDEBUG(D_INFO, "xattr %s set object "DFID"\n", name,
-              PFID(&dt->do_lu.lo_header->loh_fid));
+       struct osp_object *obj = dt2osp_obj(dt);
 
-       return 0;
+       return obj->opo_owner == env;
 }
 
-static int osp_md_xattr_get(const struct lu_env *env, struct dt_object *dt,
-                           struct lu_buf *buf, const char *name,
-                           struct lustre_capa *capa)
+/**
+ * Implementation of dt_index_operations::dio_lookup
+ *
+ * Look up record by key under a remote index object. It packs lookup update
+ * into RPC, sends to the remote OUT and waits for the lookup result.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       index object to lookup
+ * \param[out] rec     record in which to return lookup result
+ * \param[in] key      key of index which will be looked up
+ *
+ * \retval             1 if the lookup succeeds.
+ * \retval              negative errno if the lookup fails.
+ */
+static int osp_md_index_lookup(const struct lu_env *env, struct dt_object *dt,
+                              struct dt_rec *rec, const struct dt_key *key)
 {
-       struct dt_device        *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-       struct update_request   *update = NULL;
-       struct ptlrpc_request   *req = NULL;
-       int                     rc;
-       int                     buf_len;
-       int                     size;
-       struct update_reply     *reply;
-       void                    *ea_buf;
+       struct lu_buf           *lbuf   = &osp_env_info(env)->osi_lb2;
+       struct osp_device       *osp    = lu2osp_dev(dt->do_lu.lo_dev);
+       struct dt_device        *dt_dev = &osp->opd_dt_dev;
+       struct osp_update_request   *update;
+       struct object_update_reply *reply;
+       struct ptlrpc_request      *req = NULL;
+       struct lu_fid              *fid;
+       int                        rc;
        ENTRY;
 
        /* Because it needs send the update buffer right away,
         * just create an update buffer, instead of attaching the
         * update_remote list of the thandle.
         */
-       update = osp_create_update_req(dt_dev);
+       update = osp_update_request_create(dt_dev);
        if (IS_ERR(update))
                RETURN(PTR_ERR(update));
 
-       LASSERT(name != NULL);
-       buf_len = strlen(name);
-       rc = osp_insert_update(env, update, OBJ_XATTR_GET,
-                              (struct lu_fid *)lu_object_fid(&dt->do_lu),
-                              1, &buf_len, (char **)&name);
+       rc = OSP_UPDATE_RPC_PACK(env, out_index_lookup_pack, update,
+                                lu_object_fid(&dt->do_lu), rec, key);
        if (rc != 0) {
                CERROR("%s: Insert update error: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name, rc);
+                      dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
                GOTO(out, rc);
        }
-       dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
 
-       rc = osp_remote_sync(env, dt_dev, update, &req);
-       if (rc != 0)
+       rc = osp_remote_sync(env, osp, update, &req);
+       if (rc < 0)
                GOTO(out, rc);
 
-       reply = req_capsule_server_sized_get(&req->rq_pill, &RMF_UPDATE_REPLY,
-                                           UPDATE_BUFFER_SIZE);
-       if (reply->ur_version != UPDATE_REPLY_V1) {
+       reply = req_capsule_server_sized_get(&req->rq_pill,
+                                            &RMF_OUT_UPDATE_REPLY,
+                                            OUT_UPDATE_REPLY_SIZE);
+       if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
                CERROR("%s: Wrong version %x expected %x: rc = %d\n",
                       dt_dev->dd_lu_dev.ld_obd->obd_name,
-                      reply->ur_version, UPDATE_REPLY_V1, -EPROTO);
+                      reply->ourp_magic, UPDATE_REPLY_MAGIC, -EPROTO);
                GOTO(out, rc = -EPROTO);
        }
 
-       size = update_get_reply_buf(reply, &ea_buf, 0);
-       if (size < 0)
-               GOTO(out, rc = size);
+       rc = object_update_result_data_get(reply, lbuf, 0);
+       if (rc < 0)
+               GOTO(out, rc);
 
-       LASSERT(size > 0 && size < PAGE_CACHE_SIZE);
-       LASSERT(ea_buf != NULL);
+       if (lbuf->lb_len != sizeof(*fid)) {
+               CERROR("%s: lookup "DFID" %s wrong size %d\n",
+                      dt_dev->dd_lu_dev.ld_obd->obd_name,
+                      PFID(lu_object_fid(&dt->do_lu)), (char *)key,
+                      (int)lbuf->lb_len);
+               GOTO(out, rc = -EINVAL);
+       }
+
+       fid = lbuf->lb_buf;
+       if (ptlrpc_rep_need_swab(req))
+               lustre_swab_lu_fid(fid);
+       if (!fid_is_sane(fid)) {
+               CERROR("%s: lookup "DFID" %s invalid fid "DFID"\n",
+                      dt_dev->dd_lu_dev.ld_obd->obd_name,
+                      PFID(lu_object_fid(&dt->do_lu)), (char *)key, PFID(fid));
+               GOTO(out, rc = -EINVAL);
+       }
+
+       memcpy(rec, fid, sizeof(*fid));
+
+       GOTO(out, rc = 1);
 
-       rc = size;
-       if (buf->lb_buf != NULL)
-               memcpy(buf->lb_buf, ea_buf, size);
 out:
        if (req != NULL)
                ptlrpc_req_finished(req);
 
-       osp_destroy_update_req(update);
+       osp_update_request_destroy(env, update);
 
-       RETURN(rc);
+       return rc;
 }
 
-static void osp_md_object_read_lock(const struct lu_env *env,
-                                   struct dt_object *dt, unsigned role)
+/**
+ * Implementation of dt_index_operations::dio_declare_insert
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object for which to insert index
+ * \param[in] rec      record of the index which will be inserted
+ * \param[in] key      key of the index which will be inserted
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+static int osp_md_declare_index_insert(const struct lu_env *env,
+                                      struct dt_object *dt,
+                                      const struct dt_rec *rec,
+                                      const struct dt_key *key,
+                                      struct thandle *th)
 {
-       struct osp_object  *obj = dt2osp_obj(dt);
+       return osp_trans_update_request_create(th);
+}
 
-       LASSERT(obj->opo_owner != env);
-       down_read_nested(&obj->opo_sem, role);
+/**
+ * Implementation of dt_index_operations::dio_insert
+ *
+ * Add an OUT_INDEX_INSERT sub-request into the OUT RPC that will
+ * be flushed when the transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object for which to insert index
+ * \param[in] rec      record of the index to be inserted
+ * \param[in] key      key of the index to be inserted
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing index insert succeeds.
+ * \retval             negative errno if packing fails.
+ */
+static int osp_md_index_insert(const struct lu_env *env, struct dt_object *dt,
+                              const struct dt_rec *rec,
+                              const struct dt_key *key, struct thandle *th)
+{
+       struct osp_update_request *update;
+       int rc;
 
-       LASSERT(obj->opo_owner == NULL);
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
+
+       rc = OSP_UPDATE_RPC_PACK(env, out_index_insert_pack, update,
+                                lu_object_fid(&dt->do_lu), rec, key);
+       return rc;
 }
 
-static void osp_md_object_write_lock(const struct lu_env *env,
-                                    struct dt_object *dt, unsigned role)
+/**
+ * Implementation of dt_index_operations::dio_declare_delete
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object for which to delete index
+ * \param[in] key      key of the index
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+static int osp_md_declare_index_delete(const struct lu_env *env,
+                                      struct dt_object *dt,
+                                      const struct dt_key *key,
+                                      struct thandle *th)
 {
-       struct osp_object *obj = dt2osp_obj(dt);
+       return osp_trans_update_request_create(th);
+}
 
-       down_write_nested(&obj->opo_sem, role);
+/**
+ * Implementation of dt_index_operations::dio_delete
+ *
+ * Add an OUT_INDEX_DELETE sub-request into the OUT RPC that will
+ * be flushed when the transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object for which to delete index
+ * \param[in] key      key of the index which will be deleted
+ * \param[in] th       the transaction handle
+ *
+ * \retval             0 if packing index delete succeeds.
+ * \retval             negative errno if packing fails.
+ */
+static int osp_md_index_delete(const struct lu_env *env,
+                              struct dt_object *dt,
+                              const struct dt_key *key,
+                              struct thandle *th)
+{
+       struct osp_update_request *update;
+       int                      rc;
 
-       LASSERT(obj->opo_owner == NULL);
-       obj->opo_owner = env;
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
+
+       rc = OSP_UPDATE_RPC_PACK(env, out_index_delete_pack, update,
+                                lu_object_fid(&dt->do_lu), key);
+
+       return rc;
 }
 
-static void osp_md_object_read_unlock(const struct lu_env *env,
-                                     struct dt_object *dt)
+/**
+ * Implementation of dt_index_operations::dio_it.next
+ *
+ * Advance the pointer of the iterator to the next entry. It shares a similar
+ * internal implementation with osp_orphan_it_next(), which is being used for
+ * remote orphan index object. This method will be used for remote directory.
+ *
+ * \param[in] env      execution environment
+ * \param[in] di       iterator of this iteration
+ *
+ * \retval             0 if the pointer is advanced successfully.
+ * \retval             1 if it reaches to the end of the index object.
+ * \retval             negative errno if the pointer cannot be advanced.
+ */
+static int osp_md_index_it_next(const struct lu_env *env, struct dt_it *di)
 {
-       struct osp_object *obj = dt2osp_obj(dt);
+       struct osp_it           *it = (struct osp_it *)di;
+       struct lu_idxpage       *idxpage;
+       struct lu_dirent        *ent = (struct lu_dirent *)it->ooi_ent;
+       int                     rc;
+       ENTRY;
 
-       up_read(&obj->opo_sem);
+again:
+       idxpage = it->ooi_cur_idxpage;
+       if (idxpage != NULL) {
+               if (idxpage->lip_nr == 0)
+                       RETURN(1);
+
+               it->ooi_pos_ent++;
+               if (ent == NULL) {
+                       it->ooi_ent =
+                             (struct lu_dirent *)idxpage->lip_entries;
+                       RETURN(0);
+               } else if (le16_to_cpu(ent->lde_reclen) != 0 &&
+                          it->ooi_pos_ent < idxpage->lip_nr) {
+                       ent = (struct lu_dirent *)(((char *)ent) +
+                                       le16_to_cpu(ent->lde_reclen));
+                       it->ooi_ent = ent;
+                       RETURN(0);
+               } else {
+                       it->ooi_ent = NULL;
+               }
+       }
+
+       rc = osp_it_next_page(env, di);
+       if (rc == 0)
+               goto again;
+
+       RETURN(rc);
 }
 
-static void osp_md_object_write_unlock(const struct lu_env *env,
-                                      struct dt_object *dt)
+/**
+ * Implementation of dt_index_operations::dio_it.key
+ *
+ * Get the key at current iterator poisiton. These iteration methods
+ * (dio_it) will only be used for iterating the remote directory, so
+ * the key is the name of the directory entry.
+ *
+ * \param[in] env      execution environment
+ * \param[in] di       iterator of this iteration
+ *
+ * \retval             name of the current entry
+ */
+static struct dt_key *osp_it_key(const struct lu_env *env,
+                                const struct dt_it *di)
 {
-       struct osp_object *obj = dt2osp_obj(dt);
+       struct osp_it           *it = (struct osp_it *)di;
+       struct lu_dirent        *ent = (struct lu_dirent *)it->ooi_ent;
 
-       LASSERT(obj->opo_owner == env);
-       obj->opo_owner = NULL;
-       up_write(&obj->opo_sem);
+       return (struct dt_key *)ent->lde_name;
 }
 
-static int osp_md_object_write_locked(const struct lu_env *env,
-                                     struct dt_object *dt)
+/**
+ * Implementation of dt_index_operations::dio_it.key_size
+ *
+ * Get the key size at current iterator poisiton. These iteration methods
+ * (dio_it) will only be used for iterating the remote directory, so the key
+ * size is the name size of the directory entry.
+ *
+ * \param[in] env      execution environment
+ * \param[in] di       iterator of this iteration
+ *
+ * \retval             name size of the current entry
+ */
+
+static int osp_it_key_size(const struct lu_env *env, const struct dt_it *di)
 {
-       struct osp_object *obj = dt2osp_obj(dt);
+       struct osp_it           *it = (struct osp_it *)di;
+       struct lu_dirent        *ent = (struct lu_dirent *)it->ooi_ent;
 
-       return obj->opo_owner == env;
+       return (int)le16_to_cpu(ent->lde_namelen);
 }
 
-static int osp_md_index_lookup(const struct lu_env *env, struct dt_object *dt,
-                              struct dt_rec *rec, const struct dt_key *key,
-                              struct lustre_capa *capa)
+/**
+ * Implementation of dt_index_operations::dio_it.rec
+ *
+ * Get the record at current iterator position. These iteration methods
+ * (dio_it) will only be used for iterating the remote directory, so it
+ * uses lu_dirent_calc_size() to calculate the record size.
+ *
+ * \param[in] env      execution environment
+ * \param[in] di       iterator of this iteration
+ * \param[out] rec     the record to be returned
+ * \param[in] attr     attributes of the index object, so it knows
+ *                      how to pack the entry.
+ *
+ * \retval             only return 0 for now
+ */
+static int osp_md_index_it_rec(const struct lu_env *env, const struct dt_it *di,
+                              struct dt_rec *rec, __u32 attr)
 {
-       struct dt_device        *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-       struct update_request   *update;
-       struct ptlrpc_request   *req = NULL;
-       int                     size = strlen((char *)key) + 1;
-       char                    *name = (char *)key;
-       int                     rc;
-       struct update_reply     *reply;
-       struct lu_fid           *fid;
+       struct osp_it           *it = (struct osp_it *)di;
+       struct lu_dirent        *ent = (struct lu_dirent *)it->ooi_ent;
+       size_t                  reclen;
+
+       reclen = lu_dirent_calc_size(le16_to_cpu(ent->lde_namelen), attr);
+       memcpy(rec, ent, reclen);
+       return 0;
+}
+
+/**
+ * Implementation of dt_index_operations::dio_it.load
+ *
+ * Locate the iteration cursor to the specified position (cookie).
+ *
+ * \param[in] env      pointer to the thread context
+ * \param[in] di       pointer to the iteration structure
+ * \param[in] hash     the specified position
+ *
+ * \retval             positive number for locating to the exactly position
+ *                     or the next
+ * \retval             0 for arriving at the end of the iteration
+ * \retval             negative error number on failure
+ */
+static int osp_it_load(const struct lu_env *env, const struct dt_it *di,
+                      __u64 hash)
+{
+       struct osp_it   *it     = (struct osp_it *)di;
+       int              rc;
+
+       it->ooi_next = hash;
+       rc = osp_md_index_it_next(env, (struct dt_it *)di);
+       if (rc == 1)
+               return 0;
+
+       if (rc == 0)
+               return 1;
+
+       return rc;
+}
+
+const struct dt_index_operations osp_md_index_ops = {
+       .dio_lookup         = osp_md_index_lookup,
+       .dio_declare_insert = osp_md_declare_index_insert,
+       .dio_insert         = osp_md_index_insert,
+       .dio_declare_delete = osp_md_declare_index_delete,
+       .dio_delete         = osp_md_index_delete,
+       .dio_it     = {
+               .init     = osp_it_init,
+               .fini     = osp_it_fini,
+               .get      = osp_it_get,
+               .put      = osp_it_put,
+               .next     = osp_md_index_it_next,
+               .key      = osp_it_key,
+               .key_size = osp_it_key_size,
+               .rec      = osp_md_index_it_rec,
+               .store    = osp_it_store,
+               .load     = osp_it_load,
+               .key_rec  = osp_it_key_rec,
+       }
+};
+
+/**
+ * Implement OSP layer dt_object_operations::do_xattr_list() interface.
+ *
+ * List extended attribute from the specified MDT/OST object, result is not
+ * cached because this is called by directory migration only.
+ *
+ * \param[in] env      pointer to the thread context
+ * \param[in] dt       pointer to the OSP layer dt_object
+ * \param[out] buf     pointer to the lu_buf to hold the extended attribute
+ *
+ * \retval             positive bytes used/required in the buffer
+ * \retval             negative error number on failure
+ */
+static int osp_md_xattr_list(const struct lu_env *env, struct dt_object *dt,
+                            const struct lu_buf *buf)
+{
+       struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+       struct osp_object *obj = dt2osp_obj(dt);
+       struct dt_device *dev = &osp->opd_dt_dev;
+       struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
+       struct osp_update_request *update = NULL;
+       struct ptlrpc_request *req = NULL;
+       struct object_update_reply *reply;
+       const char *dname  = dt->do_lu.lo_dev->ld_obd->obd_name;
+       int rc = 0;
 
        ENTRY;
 
-       /* Because it needs send the update buffer right away,
-        * just create an update buffer, instead of attaching the
-        * update_remote list of the thandle.
-        */
-       update = osp_create_update_req(dt_dev);
+       LASSERT(buf);
+
+       if (unlikely(obj->opo_non_exist))
+               RETURN(-ENOENT);
+
+       update = osp_update_request_create(dev);
        if (IS_ERR(update))
                RETURN(PTR_ERR(update));
 
-       rc = osp_insert_update(env, update, OBJ_INDEX_LOOKUP,
-                              (struct lu_fid *)lu_object_fid(&dt->do_lu),
-                              1, &size, (char **)&name);
+       rc = OSP_UPDATE_RPC_PACK(env, out_xattr_list_pack, update,
+                                lu_object_fid(&dt->do_lu), buf->lb_len);
        if (rc) {
-               CERROR("%s: Insert update error: rc = %d\n",
-                      dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
+               CERROR("%s: Insert update error "DFID": rc = %d\n",
+                      dname, PFID(lu_object_fid(&dt->do_lu)), rc);
                GOTO(out, rc);
        }
 
-       rc = osp_remote_sync(env, dt_dev, update, &req);
-       if (rc < 0)
-               GOTO(out, rc);
-
-       reply = req_capsule_server_sized_get(&req->rq_pill, &RMF_UPDATE_REPLY,
-                                            UPDATE_BUFFER_SIZE);
-       if (reply->ur_version != UPDATE_REPLY_V1) {
-               CERROR("%s: Wrong version %x expected %x: rc = %d\n",
-                      dt_dev->dd_lu_dev.ld_obd->obd_name,
-                      reply->ur_version, UPDATE_REPLY_V1, -EPROTO);
-               GOTO(out, rc = -EPROTO);
-       }
-
-       rc = update_get_reply_result(reply, NULL, 0);
+       rc = osp_remote_sync(env, osp, update, &req);
        if (rc < 0) {
-               CERROR("%s: wrong version lookup "DFID" %s: rc = %d\n",
-                      dt_dev->dd_lu_dev.ld_obd->obd_name,
-                      PFID(lu_object_fid(&dt->do_lu)), (char *)key, rc);
+               if (rc == -ENOENT) {
+                       dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
+                       obj->opo_non_exist = 1;
+               }
                GOTO(out, rc);
        }
 
-       size = update_get_reply_buf(reply, (void **)&fid, 0);
-       if (size < 0)
-               GOTO(out, rc = size);
-
-       if (size != sizeof(struct lu_fid)) {
-               CERROR("%s: lookup "DFID" %s wrong size %d: rc = %d\n",
-                      dt_dev->dd_lu_dev.ld_obd->obd_name,
-                      PFID(lu_object_fid(&dt->do_lu)), (char *)key, size, rc);
-               GOTO(out, rc = -EINVAL);
-       }
+       reply = req_capsule_server_sized_get(&req->rq_pill,
+                                            &RMF_OUT_UPDATE_REPLY,
+                                            OUT_UPDATE_REPLY_SIZE);
+       if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
+               DEBUG_REQ(D_ERROR, req,
+                         "%s: Wrong version %x expected %x "DFID": rc = %d",
+                         dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
+                         PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
 
-       fid_le_to_cpu(fid, fid);
-       if (!fid_is_sane(fid)) {
-               CERROR("%s: lookup "DFID" %s invalid fid "DFID": rc = %d\n",
-                      dt_dev->dd_lu_dev.ld_obd->obd_name,
-                      PFID(lu_object_fid(&dt->do_lu)), (char *)key, PFID(fid),
-                      rc);
-               GOTO(out, rc = -EINVAL);
+               GOTO(out, rc = -EPROTO);
        }
-       memcpy(rec, fid, sizeof(*fid));
-out:
-       if (req != NULL)
-               ptlrpc_req_finished(req);
 
-       osp_destroy_update_req(update);
-
-       RETURN(rc);
-}
+       rc = object_update_result_data_get(reply, rbuf, 0);
+       if (rc < 0)
+               GOTO(out, rc);
 
-static int osp_md_declare_insert(const struct lu_env *env,
-                                struct dt_object *dt,
-                                const struct dt_rec *rec,
-                                const struct dt_key *key,
-                                struct thandle *th)
-{
-       struct update_request   *update;
-       struct lu_fid           *fid;
-       struct lu_fid           *rec_fid = (struct lu_fid *)rec;
-       int                     size[2] = {strlen((char *)key) + 1,
-                                                 sizeof(*rec_fid)};
-       char                    *bufs[2] = {(char *)key, (char *)rec_fid};
-       int                     rc;
+       if (!buf->lb_buf)
+               GOTO(out, rc);
 
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
+       if (unlikely(buf->lb_len < rbuf->lb_len))
+               GOTO(out, rc = -ERANGE);
 
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
+       memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
+       EXIT;
 
-       CDEBUG(D_INFO, "%s: insert index of "DFID" %s: "DFID"\n",
-              dt->do_lu.lo_dev->ld_obd->obd_name,
-              PFID(fid), (char *)key, PFID(rec_fid));
+out:
+       if (req)
+               ptlrpc_req_finished(req);
 
-       fid_cpu_to_le(rec_fid, rec_fid);
+       if (update && !IS_ERR(update))
+               osp_update_request_destroy(env, update);
 
-       rc = osp_insert_update(env, update, OBJ_INDEX_INSERT, fid,
-                              ARRAY_SIZE(size), size, bufs);
        return rc;
 }
 
-static int osp_md_index_insert(const struct lu_env *env,
-                              struct dt_object *dt,
-                              const struct dt_rec *rec,
-                              const struct dt_key *key,
-                              struct thandle *th,
-                              struct lustre_capa *capa,
-                              int ignore_quota)
+/**
+ * Implementation of dt_object_operations::do_index_try
+ *
+ * Try to initialize the index API pointer for the given object. This
+ * is the entry point of the index API, i.e. we must call this method
+ * to initialize the index object before calling other index methods.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       index object to be initialized
+ * \param[in] feat     the index feature of the object
+ *
+ * \retval             0 if the initialization succeeds.
+ * \retval              negative errno if the initialization fails.
+ */
+static int osp_md_index_try(const struct lu_env *env,
+                           struct dt_object *dt,
+                           const struct dt_index_features *feat)
 {
+       dt->do_index_ops = &osp_md_index_ops;
        return 0;
 }
 
-static int osp_md_declare_delete(const struct lu_env *env,
-                                struct dt_object *dt,
-                                const struct dt_key *key,
-                                struct thandle *th)
+/**
+ * Implementation of dt_object_operations::do_object_lock
+ *
+ * Enqueue a lock (by ldlm_cli_enqueue()) of remote object on the remote MDT,
+ * which will lock the object in the global namespace. And because the
+ * cross-MDT locks are relatively rare compared with normal local MDT operation,
+ * let's release it right away, instead of putting it into the LRU list.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be locked
+ * \param[out] lh      lock handle
+ * \param[in] einfo    enqueue information
+ * \param[in] policy   lock policy
+ *
+ * \retval             ELDLM_OK if locking the object succeeds.
+ * \retval             negative errno if locking fails.
+ */
+static int osp_md_object_lock(const struct lu_env *env,
+                             struct dt_object *dt,
+                             struct lustre_handle *lh,
+                             struct ldlm_enqueue_info *einfo,
+                             union ldlm_policy_data *policy)
 {
-       struct update_request *update;
-       struct lu_fid *fid;
-       int size = strlen((char *)key) + 1;
-       char *buf = (char *)key;
-       int rc;
+       struct ldlm_res_id      *res_id;
+       struct osp_device       *osp = dt2osp_dev(lu2dt_dev(dt->do_lu.lo_dev));
+       struct ptlrpc_request   *req;
+       int                     rc = 0;
+       __u64                   flags = LDLM_FL_NO_LRU;
+       ENTRY;
 
-       update = osp_find_create_update_loc(th, dt);
-       if (IS_ERR(update)) {
-               CERROR("%s: Get OSP update buf failed: rc = %d\n",
-                      dt->do_lu.lo_dev->ld_obd->obd_name,
-                      (int)PTR_ERR(update));
-               return PTR_ERR(update);
-       }
+       res_id = einfo->ei_res_id;
+       LASSERT(res_id != NULL);
 
-       fid = (struct lu_fid *)lu_object_fid(&dt->do_lu);
+       if (einfo->ei_mode & (LCK_EX | LCK_PW))
+               flags |= LDLM_FL_COS_INCOMPAT;
 
-       rc = osp_insert_update(env, update, OBJ_INDEX_DELETE, fid, 1, &size,
-                              &buf);
+       req = ldlm_enqueue_pack(osp->opd_exp, 0);
+       if (IS_ERR(req))
+               RETURN(PTR_ERR(req));
 
-       return rc;
-}
+       osp_set_req_replay(osp, req);
+       rc = ldlm_cli_enqueue(osp->opd_exp, &req, einfo, res_id,
+                             (const union ldlm_policy_data *)policy,
+                             &flags, NULL, 0, LVB_T_NONE, lh, 0);
 
-static int osp_md_index_delete(const struct lu_env *env,
-                              struct dt_object *dt,
-                              const struct dt_key *key,
-                              struct thandle *th,
-                              struct lustre_capa *capa)
-{
-       CDEBUG(D_INFO, "index delete "DFID" %s\n",
-              PFID(&dt->do_lu.lo_header->loh_fid), (char *)key);
+       ptlrpc_req_finished(req);
 
-       return 0;
+       RETURN(rc == ELDLM_OK ? 0 : -EIO);
 }
 
 /**
- * Creates or initializes iterator context.
- *
- * Note: for OSP, these index iterate api is only used to check
- * whether the directory is empty now (see mdd_dir_is_empty).
- * Since dir_empty will be return by OBJ_ATTR_GET(see osp_md_attr_get/
- * out_attr_get). So the implementation of these iterator is simplied
- * to make mdd_dir_is_empty happy. The real iterator should be
- * implemented, if we need it one day.
+ * Implementation of dt_object_operations::do_object_unlock
+ *
+ * Cancel a lock of a remote object.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be unlocked
+ * \param[in] einfo    lock enqueue information
+ * \param[in] policy   lock policy
+ *
+ * \retval             Only return 0 for now.
  */
-static struct dt_it *osp_it_init(const struct lu_env *env,
-                                struct dt_object *dt,
-                                __u32 attr,
-                               struct lustre_capa *capa)
+static int osp_md_object_unlock(const struct lu_env *env,
+                               struct dt_object *dt,
+                               struct ldlm_enqueue_info *einfo,
+                               union ldlm_policy_data *policy)
 {
-       lu_object_get(&dt->do_lu);
-       return (struct dt_it *)dt;
-}
+       struct lustre_handle    *lockh = einfo->ei_cbdata;
 
-static void osp_it_fini(const struct lu_env *env, struct dt_it *di)
-{
-       struct dt_object *dt = (struct dt_object *)di;
-       lu_object_put(env, &dt->do_lu);
-}
+       /* unlock finally */
+       ldlm_lock_decref(lockh, einfo->ei_mode);
 
-static int osp_it_get(const struct lu_env *env,
-                     struct dt_it *di, const struct dt_key *key)
-{
-       return 1;
+       return 0;
 }
 
-static void osp_it_put(const struct lu_env *env, struct dt_it *di)
+/**
+ * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
+ *
+ * Create the dt_update_request to track the update for this OSP
+ * in the transaction.
+ *
+ * \param[in] env      pointer to the thread context
+ * \param[in] dt       pointer to the OSP layer dt_object to be destroyed
+ * \param[in] th       pointer to the transaction handler
+ *
+ * \retval             0 for success
+ * \retval             negative error number on failure
+ */
+int osp_md_declare_destroy(const struct lu_env *env, struct dt_object *dt,
+                          struct thandle *th)
 {
-       return;
+       return osp_trans_update_request_create(th);
 }
 
-static int osp_it_next(const struct lu_env *env, struct dt_it *di)
+/**
+ * Implement OSP layer dt_object_operations::do_destroy() interface.
+ *
+ * Pack the destroy update into the RPC buffer, which will be sent
+ * to the remote MDT during transaction stop.
+ *
+ * It also marks the object as non-cached.
+ *
+ * \param[in] env      pointer to the thread context
+ * \param[in] dt       pointer to the OSP layer dt_object to be destroyed
+ * \param[in] th       pointer to the transaction handler
+ *
+ * \retval             0 for success
+ * \retval             negative error number on failure
+ */
+int osp_md_destroy(const struct lu_env *env, struct dt_object *dt,
+                  struct thandle *th)
 {
-       struct dt_object *dt = (struct dt_object *)di;
        struct osp_object *o = dt2osp_obj(dt);
+       struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+       struct osp_update_request *update;
+       int rc = 0;
 
-       if (o->opo_empty)
-               return 1;
+       ENTRY;
+       o->opo_non_exist = 1;
 
-       return 0;
-}
+       LASSERT(osp->opd_connect_mdt);
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
 
-static struct dt_key *osp_it_key(const struct lu_env *env,
-                                const struct dt_it *di)
-{
-       LBUG();
-       return NULL;
-}
+       rc = OSP_UPDATE_RPC_PACK(env, out_destroy_pack, update,
+                                lu_object_fid(&dt->do_lu));
+       if (rc != 0)
+               RETURN(rc);
 
-static int osp_it_key_size(const struct lu_env *env, const struct dt_it *di)
-{
-       LBUG();
-       return 0;
-}
+       set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
+       rc = osp_insert_update_callback(env, update, dt2osp_obj(dt), NULL,
+                                       NULL);
 
-static int osp_it_rec(const struct lu_env *env, const struct dt_it *di,
-                     struct dt_rec *lde, __u32 attr)
-{
-       LBUG();
-       return 0;
+       RETURN(rc);
 }
 
-static __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di)
+struct dt_object_operations osp_md_obj_ops = {
+       .do_read_lock         = osp_md_read_lock,
+       .do_write_lock        = osp_md_write_lock,
+       .do_read_unlock       = osp_md_read_unlock,
+       .do_write_unlock      = osp_md_write_unlock,
+       .do_write_locked      = osp_md_write_locked,
+       .do_declare_create    = osp_md_declare_create,
+       .do_create            = osp_md_create,
+       .do_declare_ref_add   = osp_md_declare_ref_add,
+       .do_ref_add           = osp_md_ref_add,
+       .do_declare_ref_del   = osp_md_declare_ref_del,
+       .do_ref_del           = osp_md_ref_del,
+       .do_declare_destroy   = osp_md_declare_destroy,
+       .do_destroy           = osp_md_destroy,
+       .do_ah_init           = osp_md_ah_init,
+       .do_attr_get          = osp_attr_get,
+       .do_declare_attr_set  = osp_md_declare_attr_set,
+       .do_attr_set          = osp_md_attr_set,
+       .do_xattr_get         = osp_xattr_get,
+       .do_xattr_list        = osp_md_xattr_list,
+       .do_declare_xattr_set = osp_declare_xattr_set,
+       .do_xattr_set         = osp_xattr_set,
+       .do_declare_xattr_del = osp_declare_xattr_del,
+       .do_xattr_del         = osp_xattr_del,
+       .do_index_try         = osp_md_index_try,
+       .do_object_lock       = osp_md_object_lock,
+       .do_object_unlock     = osp_md_object_unlock,
+       .do_invalidate        = osp_invalidate,
+};
+
+/**
+ * Implementation of dt_body_operations::dbo_declare_write
+ *
+ * Create the osp_update_request to track the update for this OSP
+ * in the transaction.
+  *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be written
+ * \param[in] buf      buffer to write which includes an embedded size field
+ * \param[in] pos      offet in the object to start writing at
+ * \param[in] th       transaction handle
+ *
+ * \retval             0 if preparation succeeds.
+ * \retval             negative errno if preparation fails.
+ */
+static ssize_t osp_md_declare_write(const struct lu_env *env,
+                                   struct dt_object *dt,
+                                   const struct lu_buf *buf,
+                                   loff_t pos, struct thandle *th)
 {
-       LBUG();
+       struct osp_device *osp = dt2osp_dev(th->th_dev);
+       int rc;
+
+       rc = osp_trans_update_request_create(th);
+       if (rc != 0)
+               return rc;
+
+       if (osp->opd_update == NULL)
+               return 0;
+
+       if (dt2osp_obj(dt)->opo_stale)
+               return -ESTALE;
+
        return 0;
 }
 
-static int osp_it_load(const struct lu_env *env, const struct dt_it *di,
-                      __u64 hash)
+static int osp_write_interpreter(const struct lu_env *env,
+                                 struct object_update_reply *reply,
+                                 struct ptlrpc_request *req,
+                                 struct osp_object *obj,
+                                 void *data, int index, int rc)
 {
-       LBUG();
+       if (rc) {
+               CDEBUG(D_HA, "error "DFID": rc = %d\n",
+                      PFID(lu_object_fid(&obj->opo_obj.do_lu)), rc);
+               spin_lock(&obj->opo_lock);
+               obj->opo_attr.la_valid = 0;
+               obj->opo_stale = 1;
+               spin_unlock(&obj->opo_lock);
+       }
        return 0;
 }
 
-static int osp_it_key_rec(const struct lu_env *env, const struct dt_it *di,
-                         void *key_rec)
+/**
+ * Implementation of dt_body_operations::dbo_write
+ *
+ * Pack the write object update into the RPC buffer, which will be sent
+ * to the remote MDT during transaction stop.
+ *
+ * \param[in] env      execution environment
+ * \param[in] dt       object to be written
+ * \param[in] buf      buffer to write which includes an embedded size field
+ * \param[in] pos      offet in the object to start writing at
+ * \param[in] th       transaction handle
+ *
+ * \retval             the buffer size in bytes if packing succeeds.
+ * \retval             negative errno if packing fails.
+ */
+static ssize_t osp_md_write(const struct lu_env *env, struct dt_object *dt,
+                           const struct lu_buf *buf, loff_t *pos,
+                           struct thandle *th)
 {
-       LBUG();
-       return 0;
-}
+       struct osp_object         *obj = dt2osp_obj(dt);
+       struct osp_update_request  *update;
+       struct osp_thandle        *oth = thandle_to_osp_thandle(th);
+       ssize_t                   rc;
+       ENTRY;
 
-static const struct dt_index_operations osp_md_index_ops = {
-       .dio_lookup         = osp_md_index_lookup,
-       .dio_declare_insert = osp_md_declare_insert,
-       .dio_insert         = osp_md_index_insert,
-       .dio_declare_delete = osp_md_declare_delete,
-       .dio_delete         = osp_md_index_delete,
-       .dio_it     = {
-               .init     = osp_it_init,
-               .fini     = osp_it_fini,
-               .get      = osp_it_get,
-               .put      = osp_it_put,
-               .next     = osp_it_next,
-               .key      = osp_it_key,
-               .key_size = osp_it_key_size,
-               .rec      = osp_it_rec,
-               .store    = osp_it_store,
-               .load     = osp_it_load,
-               .key_rec  = osp_it_key_rec,
+       update = thandle_to_osp_update_request(th);
+       LASSERT(update != NULL);
+
+       CDEBUG(D_INFO, "write "DFID" offset = %llu length = %zu\n",
+              PFID(lu_object_fid(&dt->do_lu)), *pos, buf->lb_len);
+
+       rc = OSP_UPDATE_RPC_PACK(env, out_write_pack, update,
+                                lu_object_fid(&dt->do_lu), buf, *pos);
+       if (rc < 0)
+               RETURN(rc);
+
+       rc = osp_check_and_set_rpc_version(oth, obj);
+       if (rc < 0)
+               RETURN(rc);
+
+       /* to be able to invalidate object's state in case of an error */
+       rc = osp_insert_update_callback(env, update, obj, NULL,
+                       osp_write_interpreter);
+       if (rc < 0)
+               RETURN(rc);
+
+       /* XXX: how about the write error happened later? */
+       *pos += buf->lb_len;
+
+       if (obj->opo_attr.la_valid & LA_SIZE && obj->opo_attr.la_size < *pos)
+               obj->opo_attr.la_size = *pos;
+
+       spin_lock(&obj->opo_lock);
+       if (list_empty(&obj->opo_invalidate_cb_list)) {
+               lu_object_get(&obj->opo_obj.do_lu);
+
+               list_add_tail(&obj->opo_invalidate_cb_list,
+                             &update->our_invalidate_cb_list);
        }
-};
+       spin_unlock(&obj->opo_lock);
 
-static int osp_md_index_try(const struct lu_env *env,
-                           struct dt_object *dt,
-                           const struct dt_index_features *feat)
+       RETURN(buf->lb_len);
+}
+
+static inline void orr_le_to_cpu(struct out_read_reply *orr_dst,
+                                const struct out_read_reply *orr_src)
 {
-       dt->do_index_ops = &osp_md_index_ops;
-       return 0;
+       orr_dst->orr_size = le32_to_cpu(orr_src->orr_size);
+       orr_dst->orr_padding = le32_to_cpu(orr_src->orr_padding);
+       orr_dst->orr_offset = le64_to_cpu(orr_dst->orr_offset);
 }
 
-static int osp_md_attr_get(const struct lu_env *env,
-                          struct dt_object *dt, struct lu_attr *attr,
-                          struct lustre_capa *capa)
+
+
+static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
+                          struct lu_buf *rbuf, loff_t *pos)
 {
-       struct osp_object     *obj = dt2osp_obj(dt);
-       struct dt_device      *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-       struct update_request *update = NULL;
+       struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
+       struct dt_device *dt_dev        = &osp->opd_dt_dev;
+       struct lu_buf *lbuf = &osp_env_info(env)->osi_lb2;
+       char *ptr = rbuf->lb_buf;
+       struct osp_update_request *update;
        struct ptlrpc_request *req = NULL;
+       struct out_read_reply *orr;
+       struct ptlrpc_bulk_desc *desc;
+       struct object_update_reply *reply;
+       int pages;
        int rc;
        ENTRY;
 
        /* Because it needs send the update buffer right away,
         * just create an update buffer, instead of attaching the
-        * update_remote list of the thandle.
-        */
-       update = osp_create_update_req(dt_dev);
+        * update_remote list of the thandle.  */
+       update = osp_update_request_create(dt_dev);
        if (IS_ERR(update))
                RETURN(PTR_ERR(update));
 
-       rc = osp_insert_update(env, update, OBJ_ATTR_GET,
-                              (struct lu_fid *)lu_object_fid(&dt->do_lu),
-                              0, NULL, NULL);
-       if (rc) {
-               CERROR("%s: Insert update error: rc = %d\n",
+       rc = OSP_UPDATE_RPC_PACK(env, out_read_pack, update,
+                                lu_object_fid(&dt->do_lu),
+                                rbuf->lb_len, *pos);
+       if (rc != 0) {
+               CERROR("%s: cannot insert update: rc = %d\n",
                       dt_dev->dd_lu_dev.ld_obd->obd_name, rc);
-               GOTO(out, rc);
+               GOTO(out_update, rc);
        }
-       dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-
-       rc = osp_remote_sync(env, dt_dev, update, &req);
-       if (rc < 0)
-               GOTO(out, rc);
-
-       rc = osp_get_attr_from_req(env, req, attr, 0);
-       if (rc)
-               GOTO(out, rc);
-
-       if (attr->la_flags == 1)
-               obj->opo_empty = 0;
-       else
-               obj->opo_empty = 1;
-out:
-       if (req != NULL)
-               ptlrpc_req_finished(req);
-
-       osp_destroy_update_req(update);
 
-       RETURN(rc);
-}
-
-static int osp_md_declare_object_destroy(const struct lu_env *env,
-                                        struct dt_object *dt,
-                                        struct thandle *th)
-{
-       struct osp_object  *o = dt2osp_obj(dt);
-       int                 rc = 0;
-       ENTRY;
+       CDEBUG(D_INFO, "%s "DFID" read offset %llu size %zu\n",
+              dt_dev->dd_lu_dev.ld_obd->obd_name,
+              PFID(lu_object_fid(&dt->do_lu)), *pos, rbuf->lb_len);
+       rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import, update,
+                                &req);
+       if (rc != 0)
+               GOTO(out_update, rc);
 
-       /*
-        * track objects to be destroyed via llog
-        */
-       rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
+       /* First *and* last might be partial pages, hence +1 */
+       pages = DIV_ROUND_UP(rbuf->lb_len, PAGE_SIZE) + 1;
 
-       RETURN(rc);
-}
+       /* allocate bulk descriptor */
+       desc = ptlrpc_prep_bulk_imp(req, pages, 1,
+                                   PTLRPC_BULK_PUT_SINK,
+                                   MDS_BULK_PORTAL,
+                                   &ptlrpc_bulk_kiov_nopin_ops);
+       if (desc == NULL)
+               GOTO(out, rc = -ENOMEM);
 
-static int osp_md_object_destroy(const struct lu_env *env,
-                                struct dt_object *dt, struct thandle *th)
-{
-       struct osp_object  *o = dt2osp_obj(dt);
-       int                 rc = 0;
-       ENTRY;
+       desc->bd_frag_ops->add_iov_frag(desc, ptr, rbuf->lb_len);
 
-       /*
-        * once transaction is committed put proper command on
-        * the queue going to our OST
-        */
-       rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
+       osp_set_req_replay(osp, req);
+       req->rq_bulk_read = 1;
+       /* send request to master and wait for RPC to complete */
+       rc = ptlrpc_queue_wait(req);
+       if (rc != 0)
+               GOTO(out, rc);
 
-       /* not needed in cache any more */
-       set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
+       rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
+                                         req->rq_bulk->bd_nob_transferred);
+       if (rc < 0)
+               GOTO(out, rc);
 
-       RETURN(rc);
-}
+       reply = req_capsule_server_sized_get(&req->rq_pill,
+                                            &RMF_OUT_UPDATE_REPLY,
+                                            OUT_UPDATE_REPLY_SIZE);
 
-static int osp_md_object_lock(const struct lu_env *env,
-                             struct dt_object *dt,
-                             struct lustre_handle *lh,
-                             struct ldlm_enqueue_info *einfo,
-                             void *policy)
-{
-       struct osp_thread_info *info = osp_env_info(env);
-       struct ldlm_res_id     *res_id = &info->osi_resid;
-       struct dt_device       *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
-       struct osp_device      *osp = dt2osp_dev(dt_dev);
-       struct ptlrpc_request  *req = NULL;
-       int                     rc = 0;
-       __u64                   flags = 0;
-       ldlm_mode_t             mode;
-
-       fid_build_reg_res_name(lu_object_fid(&dt->do_lu), res_id);
-
-       mode = ldlm_lock_match(osp->opd_obd->obd_namespace,
-                              LDLM_FL_BLOCK_GRANTED, res_id,
-                              einfo->ei_type,
-                              (ldlm_policy_data_t *)policy,
-                              einfo->ei_mode, lh, 0);
-       if (mode > 0)
-               return ELDLM_OK;
+       if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
+               CERROR("%s: invalid update reply magic %x expected %x:"
+                      " rc = %d\n", dt_dev->dd_lu_dev.ld_obd->obd_name,
+                      reply->ourp_magic, UPDATE_REPLY_MAGIC, -EPROTO);
+               GOTO(out, rc = -EPROTO);
+       }
 
-       req = ldlm_enqueue_pack(osp->opd_exp, 0);
-       if (IS_ERR(req))
-               RETURN(PTR_ERR(req));
+       rc = object_update_result_data_get(reply, lbuf, 0);
+       if (rc < 0)
+               GOTO(out, rc);
 
-       rc = ldlm_cli_enqueue(osp->opd_exp, &req, einfo, res_id,
-                             (const ldlm_policy_data_t *)policy,
-                             &flags, NULL, 0, LVB_T_NONE, lh, 0);
+       if (lbuf->lb_len < sizeof(*orr))
+               GOTO(out, rc = -EPROTO);
 
+       orr = lbuf->lb_buf;
+       orr_le_to_cpu(orr, orr);
+       rc = orr->orr_size;
+       *pos = orr->orr_offset;
+out:
        ptlrpc_req_finished(req);
 
-       return rc == ELDLM_OK ? 0 : -EIO;
+out_update:
+       osp_update_request_destroy(env, update);
+
+       RETURN(rc);
 }
 
-struct dt_object_operations osp_md_obj_ops = {
-       .do_read_lock         = osp_md_object_read_lock,
-       .do_write_lock        = osp_md_object_write_lock,
-       .do_read_unlock       = osp_md_object_read_unlock,
-       .do_write_unlock      = osp_md_object_write_unlock,
-       .do_write_locked      = osp_md_object_write_locked,
-       .do_declare_create    = osp_md_declare_object_create,
-       .do_create            = osp_md_object_create,
-       .do_declare_ref_add   = osp_md_declare_ref_add,
-       .do_ref_add           = osp_md_object_ref_add,
-       .do_declare_ref_del   = osp_md_declare_object_ref_del,
-       .do_ref_del           = osp_md_object_ref_del,
-       .do_declare_destroy   = osp_md_declare_object_destroy,
-       .do_destroy           = osp_md_object_destroy,
-       .do_ah_init           = osp_md_ah_init,
-       .do_attr_get          = osp_md_attr_get,
-       .do_declare_attr_set  = osp_md_declare_attr_set,
-       .do_attr_set          = osp_md_attr_set,
-       .do_declare_xattr_set = osp_md_declare_xattr_set,
-       .do_xattr_set         = osp_md_xattr_set,
-       .do_xattr_get         = osp_md_xattr_get,
-       .do_index_try         = osp_md_index_try,
-       .do_object_lock       = osp_md_object_lock,
+/* These body operation will be used to write symlinks during migration etc */
+struct dt_body_operations osp_md_body_ops = {
+       .dbo_declare_write      = osp_md_declare_write,
+       .dbo_write              = osp_md_write,
+       .dbo_read               = osp_md_read,
 };