+ .do_object_unlock = osp_md_object_unlock,
+};
+
+/**
+ * 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)
+{
+ 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;
+}
+
+/**
+ * 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
+ * \param[in] ignore_quota quota enforcement for this write
+ *
+ * \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, int ignore_quota)
+{
+ 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;
+
+ update = thandle_to_osp_update_request(th);
+ LASSERT(update != NULL);
+
+ CDEBUG(D_INFO, "write "DFID" offset = "LPU64" length = %zu\n",
+ PFID(lu_object_fid(&dt->do_lu)), *pos, buf->lb_len);
+
+ rc = osp_update_rpc_pack(env, write, update, OUT_WRITE,
+ 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);
+
+ /* XXX: how about the write error happened later? */
+ *pos += buf->lb_len;
+
+ if (obj->opo_ooa != NULL &&
+ obj->opo_ooa->ooa_attr.la_valid & LA_SIZE &&
+ obj->opo_ooa->ooa_attr.la_size < *pos)
+ obj->opo_ooa->ooa_attr.la_size = *pos;
+
+ RETURN(buf->lb_len);
+}
+
+static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
+ struct lu_buf *rbuf, loff_t *pos)
+{
+ 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;
+ __u32 left_size;
+ int nbufs;
+ int i;
+ 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_update_request_create(dt_dev);
+ if (IS_ERR(update))
+ RETURN(PTR_ERR(update));
+
+ rc = osp_update_rpc_pack(env, read, update, OUT_READ,
+ 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_update, rc);
+ }
+
+ 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);
+
+ nbufs = (rbuf->lb_len + OUT_BULK_BUFFER_SIZE - 1) /
+ OUT_BULK_BUFFER_SIZE;
+ /* allocate bulk descriptor */
+ desc = ptlrpc_prep_bulk_imp(req, nbufs, 1,
+ PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KVEC,
+ MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
+ if (desc == NULL)
+ GOTO(out, rc = -ENOMEM);
+
+ /* split the buffer into small chunk size */
+ left_size = rbuf->lb_len;
+ for (i = 0; i < nbufs; i++) {
+ int read_size;
+
+ read_size = left_size > OUT_BULK_BUFFER_SIZE ?
+ OUT_BULK_BUFFER_SIZE : left_size;
+ desc->bd_frag_ops->add_iov_frag(desc, ptr, read_size);
+
+ ptr += read_size;
+ }
+
+ /* 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;
+ 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);
+
+ rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
+ req->rq_bulk->bd_nob_transferred);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ 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: 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);
+ }
+
+ rc = object_update_result_data_get(reply, lbuf, 0);
+ if (rc < 0)
+ GOTO(out, rc);
+
+ 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);
+
+out_update:
+ osp_update_request_destroy(update);
+
+ RETURN(rc);
+}
+
+/* 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,