Whamcloud - gitweb
LU-7408 target: declare write for reply data
[fs/lustre-release.git] / lustre / osp / osp_md_object.c
index 46de68b..69c815c 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2013, 2015, Intel Corporation.
  */
 /*
  * lustre/osp/osp_md_object.c
@@ -82,7 +82,7 @@ static int osp_object_create_interpreter(const struct lu_env *env,
                                         struct osp_object *obj,
                                         void *data, int index, int rc)
 {
-       if (rc != 0) {
+       if (rc != 0 && rc != -EEXIST) {
                obj->opo_obj.do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
                obj->opo_non_exist = 1;
        }
@@ -153,30 +153,6 @@ update_buffer_get_update(struct object_update_request *request,
        return ptr;
 }
 
-int osp_extend_update_buffer(const struct lu_env *env,
-                            struct osp_update_request *our)
-{
-       struct object_update_request *obj_update_req;
-       size_t new_size = our->our_req_size + OUT_UPDATE_BUFFER_SIZE_ADD;
-
-       /* enlarge object update request size */
-       if (new_size > OUT_UPDATE_BUFFER_SIZE_MAX)
-               return -E2BIG;
-
-       OBD_ALLOC_LARGE(obj_update_req, new_size);
-       if (obj_update_req == NULL)
-               return -ENOMEM;
-
-       memcpy(obj_update_req, our->our_req, our->our_req_size);
-
-       OBD_FREE_LARGE(our->our_req, our->our_req_size);
-
-       our->our_req = obj_update_req;
-       our->our_req_size = new_size;
-
-       return 0;
-}
-
 /**
  * Implementation of dt_object_operations::do_create
  *
@@ -887,7 +863,9 @@ static int osp_md_index_try(const struct lu_env *env,
  * 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.
+ * 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
@@ -902,7 +880,7 @@ static int osp_md_object_lock(const struct lu_env *env,
                              struct dt_object *dt,
                              struct lustre_handle *lh,
                              struct ldlm_enqueue_info *einfo,
-                             ldlm_policy_data_t *policy)
+                             union ldlm_policy_data *policy)
 {
        struct ldlm_res_id      *res_id;
        struct dt_device        *dt_dev = lu2dt_dev(dt->do_lu.lo_dev);
@@ -910,7 +888,7 @@ static int osp_md_object_lock(const struct lu_env *env,
        struct ptlrpc_request   *req;
        int                     rc = 0;
        __u64                   flags = 0;
-       ldlm_mode_t             mode;
+       enum ldlm_mode          mode;
 
        res_id = einfo->ei_res_id;
        LASSERT(res_id != NULL);
@@ -922,15 +900,25 @@ static int osp_md_object_lock(const struct lu_env *env,
        if (mode > 0)
                return ELDLM_OK;
 
+       if (einfo->ei_nonblock)
+               flags |= LDLM_FL_BLOCK_NOWAIT;
+
        req = ldlm_enqueue_pack(osp->opd_exp, 0);
        if (IS_ERR(req))
                RETURN(PTR_ERR(req));
 
        rc = ldlm_cli_enqueue(osp->opd_exp, &req, einfo, res_id,
-                             (const ldlm_policy_data_t *)policy,
+                             (const union ldlm_policy_data *)policy,
                              &flags, NULL, 0, LVB_T_NONE, lh, 0);
 
        ptlrpc_req_finished(req);
+       if (rc == ELDLM_OK) {
+               struct ldlm_lock *lock;
+
+               lock = __ldlm_handle2lock(lh, 0);
+               ldlm_set_cbpending(lock);
+               LDLM_LOCK_PUT(lock);
+       }
 
        return rc == ELDLM_OK ? 0 : -EIO;
 }
@@ -950,7 +938,7 @@ static int osp_md_object_lock(const struct lu_env *env,
 static int osp_md_object_unlock(const struct lu_env *env,
                                struct dt_object *dt,
                                struct ldlm_enqueue_info *einfo,
-                               ldlm_policy_data_t *policy)
+                               union ldlm_policy_data *policy)
 {
        struct lustre_handle    *lockh = einfo->ei_cbdata;
 
@@ -980,35 +968,6 @@ int osp_md_declare_object_destroy(const struct lu_env *env,
 }
 
 /**
- * Interpreter call for object destroy
- *
- * Object destroy interpreter, which will be called after destroying
- * 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 destroying object
- * \param[in] obj      object to be destroyed
- * \param[in] data     data used in this function.
- * \param[in] index    index(position) of destroy update in the whole
- *                      updates
- * \param[in] rc       update result on the remote MDT.
- *
- * \retval             only return 0 for now
- */
-static int osp_md_object_destroy_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)
-{
-       /* not needed in cache any more */
-       set_bit(LU_OBJECT_HEARD_BANSHEE,
-               &obj->opo_obj.do_lu.lo_header->loh_flags);
-       return 0;
-}
-
-/**
  * Implement OSP layer dt_object_operations::do_destroy() interface.
  *
  * Pack the destroy update into the RPC buffer, which will be sent
@@ -1043,8 +1002,10 @@ int osp_md_object_destroy(const struct lu_env *env, struct dt_object *dt,
        if (rc != 0)
                RETURN(rc);
 
+       set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
        rc = osp_insert_update_callback(env, update, dt2osp_obj(dt), NULL,
-                                       osp_md_object_destroy_interpreter);
+                                       NULL);
+
        RETURN(rc);
 }
 
@@ -1154,14 +1115,19 @@ static ssize_t osp_md_write(const struct lu_env *env, struct dt_object *dt,
 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;
-       struct osp_update_request   *update;
+       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 = NULL;
+       struct ptlrpc_request *req = NULL;
+       struct out_read_reply *orr;
+       struct ptlrpc_bulk_desc *desc;
        struct object_update_reply *reply;
-       struct out_read_reply      *orr;
-       struct ptlrpc_request      *req = NULL;
-       int                        rc;
+       __u32 left_size;
+       int nbufs;
+       int i;
+       int rc;
        ENTRY;
 
        /* Because it needs send the update buffer right away,
@@ -1169,23 +1135,62 @@ static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
         * update_remote list of the thandle.  */
        update = osp_update_request_create(dt_dev);
        if (IS_ERR(update))
-               RETURN(PTR_ERR(update));
+               GOTO(out, rc = PTR_ERR(update));
 
        rc = osp_update_rpc_pack(env, read, update, OUT_READ,
-                                lu_object_fid(&dt->do_lu), rbuf->lb_len, *pos);
+                                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);
        }
 
-       rc = osp_remote_sync(env, osp, update, &req);
+       rc = osp_prep_update_req(env, osp->opd_obd->u.cli.cl_import, update,
+                                &req);
+       if (rc != 0)
+               GOTO(out, 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,
@@ -1202,25 +1207,16 @@ static ssize_t osp_md_read(const struct lu_env *env, struct dt_object *dt,
 
        orr = lbuf->lb_buf;
        orr_le_to_cpu(orr, orr);
-
+       rc = orr->orr_size;
        *pos = orr->orr_offset;
-
-       if (orr->orr_size > rbuf->lb_len)
-               GOTO(out, rc = -EPROTO);
-
-       memcpy(rbuf->lb_buf, orr->orr_data, orr->orr_size);
-
-       CDEBUG(D_INFO, "%s: read "DFID" pos "LPU64" len %u\n",
-              osp->opd_obd->obd_name, PFID(lu_object_fid(&dt->do_lu)),
-              *pos, orr->orr_size);
-       GOTO(out, rc = (int)orr->orr_size);
 out:
        if (req != NULL)
                ptlrpc_req_finished(req);
 
-       osp_update_request_destroy(update);
+       if (update != NULL)
+               osp_update_request_destroy(update);
 
-       return rc;
+       RETURN(rc);
 }
 
 /* These body operation will be used to write symlinks during migration etc */