X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_io.c;h=1fc93b9f96d487975e97c9fce0f0cfe96124cdca;hp=bdbb3885799c3aafd771d8807628a0f15984f843;hb=d30f9e6b6c5df9dcf1510b5c2414600165d904a8;hpb=ccabce23bd9e366c345c852f565766a799f61238

diff --git a/lustre/mdt/mdt_io.c b/lustre/mdt/mdt_io.c
index bdbb388..1fc93b9 100644
--- a/lustre/mdt/mdt_io.c
+++ b/lustre/mdt/mdt_io.c
@@ -61,6 +61,320 @@ static inline void mdt_dom_write_unlock(struct mdt_object *mo)
 	up_write(&mo->mot_dom_sem);
 }
 
+/**
+ * Lock prolongation for Data-on-MDT.
+ * This is similar to OFD code but for DOM ibits lock.
+ */
+static inline time64_t prolong_timeout(struct ptlrpc_request *req)
+{
+	struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+	time64_t req_timeout;
+
+	if (AT_OFF)
+		return obd_timeout / 2;
+
+	req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
+	return max_t(time64_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
+		     req_timeout);
+}
+
+static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
+{
+	struct ldlm_resource *res;
+	struct ldlm_lock *lock;
+
+	ENTRY;
+
+	res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
+				&arg->lpa_resid, LDLM_EXTENT, 0);
+	if (IS_ERR(res)) {
+		CDEBUG(D_DLMTRACE,
+		       "Failed to get resource for resid %llu/%llu\n",
+		       arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
+		RETURN_EXIT;
+	}
+
+	lock_res(res);
+	list_for_each_entry(lock, &res->lr_granted, l_res_link) {
+		if (ldlm_has_dom(lock)) {
+			LDLM_DEBUG(lock, "DOM lock to prolong ");
+			ldlm_lock_prolong_one(lock, arg);
+			break;
+		}
+	}
+	unlock_res(res);
+	ldlm_resource_putref(res);
+
+	EXIT;
+}
+
+static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
+				 struct ldlm_prolong_args *data)
+{
+	struct obdo *oa = &tsi->tsi_ost_body->oa;
+	struct ldlm_lock *lock;
+
+	ENTRY;
+
+	data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
+	data->lpa_export = tsi->tsi_exp;
+	data->lpa_resid = tsi->tsi_resid;
+
+	CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
+	       tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
+
+	if (oa->o_valid & OBD_MD_FLHANDLE) {
+		/* mostly a request should be covered by only one lock, try
+		 * fast path. */
+		lock = ldlm_handle2lock(&oa->o_handle);
+		if (lock != NULL) {
+			LASSERT(lock->l_export == data->lpa_export);
+			ldlm_lock_prolong_one(lock, data);
+			lock->l_last_used = ktime_get();
+			LDLM_LOCK_PUT(lock);
+			if (data->lpa_locks_cnt > 0)
+				RETURN_EXIT;
+		}
+	}
+	mdt_dom_resource_prolong(data);
+	EXIT;
+}
+
+static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
+				   struct ldlm_lock *lock)
+{
+	struct obd_ioobj *ioo;
+	enum ldlm_mode mode;
+	__u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
+
+	ENTRY;
+
+	if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
+		RETURN(0);
+
+	ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+	LASSERT(ioo != NULL);
+
+	LASSERT(lock->l_resource != NULL);
+	if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
+		RETURN(0);
+
+	/* a bulk write can only hold a reference on a PW extent lock. */
+	mode = LCK_PW;
+	if (opc == OST_READ)
+		/* whereas a bulk read can be protected by either a PR or PW
+		 * extent lock */
+		mode |= LCK_PR;
+
+	if (!(lock->l_granted_mode & mode))
+		RETURN(0);
+
+	RETURN(1);
+}
+
+static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
+{
+	struct tgt_session_info *tsi;
+	struct obd_ioobj *ioo;
+	struct niobuf_remote *rnb;
+	int opc;
+	struct ldlm_prolong_args pa = { 0 };
+
+	ENTRY;
+
+	/* Don't use tgt_ses_info() to get session info, because lock_match()
+	 * can be called while request has no processing thread yet. */
+	tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+	/*
+	 * Use LASSERT below because malformed RPCs should have
+	 * been filtered out in tgt_hpreq_handler().
+	 */
+	opc = lustre_msg_get_opc(req->rq_reqmsg);
+	LASSERT(opc == OST_READ || opc == OST_WRITE);
+
+	ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+	LASSERT(ioo != NULL);
+
+	rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+	LASSERT(rnb != NULL);
+	LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
+
+	pa.lpa_mode = LCK_PW;
+	if (opc == OST_READ)
+		pa.lpa_mode |= LCK_PR;
+
+	DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID"\n",
+		  tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
+
+	mdt_prolong_dom_lock(tsi, &pa);
+
+	if (pa.lpa_blocks_cnt > 0) {
+		CDEBUG(D_DLMTRACE,
+		       "%s: refreshed %u locks timeout for req %p.\n",
+		       tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
+		RETURN(1);
+	}
+
+	RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
+}
+
+static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
+{
+	mdt_rw_hpreq_check(req);
+}
+
+static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
+	.hpreq_lock_match = mdt_rw_hpreq_lock_match,
+	.hpreq_check = mdt_rw_hpreq_check,
+	.hpreq_fini = mdt_rw_hpreq_fini
+};
+
+/**
+ * Assign high priority operations to an IO request.
+ *
+ * Check if the incoming request is a candidate for
+ * high-priority processing. If it is, assign it a high
+ * priority operations table.
+ *
+ * \param[in] tsi	target session environment for this request
+ */
+void mdt_hp_brw(struct tgt_session_info *tsi)
+{
+	struct niobuf_remote	*rnb;
+	struct obd_ioobj	*ioo;
+
+	ENTRY;
+
+	ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
+	LASSERT(ioo != NULL); /* must exist after request preprocessing */
+	if (ioo->ioo_bufcnt > 0) {
+		rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
+		LASSERT(rnb != NULL); /* must exist after preprocessing */
+
+		/* no high priority if server lock is needed */
+		if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
+		    (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
+		     MSG_REPLAY))
+			return;
+	}
+	tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
+}
+
+static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
+				      struct ldlm_lock *lock)
+{
+	struct tgt_session_info *tsi;
+	struct obdo *oa;
+
+	ENTRY;
+
+	/* Don't use tgt_ses_info() to get session info, because lock_match()
+	 * can be called while request has no processing thread yet. */
+	tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+
+	/*
+	 * Use LASSERT below because malformed RPCs should have
+	 * been filtered out in tgt_hpreq_handler().
+	 */
+	LASSERT(tsi->tsi_ost_body != NULL);
+	if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
+	    tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
+		RETURN(1);
+
+	oa = &tsi->tsi_ost_body->oa;
+
+	LASSERT(lock->l_resource != NULL);
+	if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
+		RETURN(0);
+
+	if (!(lock->l_granted_mode & LCK_PW))
+		RETURN(0);
+
+	RETURN(1);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
+ *
+ * High-priority queue request check for whether the given punch request
+ * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
+ * covered by an LDLM lock.
+ *
+
+ *
+ * \param[in] req	the incoming request
+ *
+ * \retval		1 if \a req is blocking an LDLM lock cancel
+ * \retval		0 if it is not
+ * \retval		-ESTALE if lock is not found
+ */
+static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
+{
+	struct tgt_session_info *tsi;
+	struct obdo *oa;
+	struct ldlm_prolong_args pa = { 0 };
+
+	ENTRY;
+
+	/* Don't use tgt_ses_info() to get session info, because lock_match()
+	 * can be called while request has no processing thread yet. */
+	tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
+	LASSERT(tsi != NULL);
+	oa = &tsi->tsi_ost_body->oa;
+
+	LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
+		  oa->o_flags & OBD_FL_SRVLOCK));
+
+	pa.lpa_mode = LCK_PW;
+
+	CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
+	       tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
+
+	mdt_prolong_dom_lock(tsi, &pa);
+
+
+	if (pa.lpa_blocks_cnt > 0) {
+		CDEBUG(D_DLMTRACE,
+		       "%s: refreshed %u locks timeout for req %p.\n",
+		       tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
+		RETURN(1);
+	}
+
+	RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
+}
+
+/**
+ * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
+ *
+ * Called after the request has been handled. It refreshes lock timeout again
+ * so that client has more time to send lock cancel RPC.
+ *
+ * \param[in] req	request which is being processed.
+ */
+static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
+{
+	mdt_punch_hpreq_check(req);
+}
+
+static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
+	.hpreq_lock_match = mdt_punch_hpreq_lock_match,
+	.hpreq_check = mdt_punch_hpreq_check,
+	.hpreq_fini = mdt_punch_hpreq_fini
+};
+
+void mdt_hp_punch(struct tgt_session_info *tsi)
+{
+	LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
+	/* no high-priority if server lock is needed */
+	if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
+	     tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
+	    tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
+	    lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
+		return;
+	tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
+}
+
 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
 			   struct mdt_device *mdt, struct mdt_object *mo,
 			   struct lu_attr *la, int niocount,
@@ -292,6 +606,7 @@ retry:
 			GOTO(out_stop, rc);
 	}
 
+	tgt_vbr_obj_set(env, dob);
 	rc = dt_trans_start(env, dt, th);
 	if (rc)
 		GOTO(out_stop, rc);
@@ -417,8 +732,15 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 					oa->o_flags = OBD_FL_NO_GRPQUOTA;
 			}
 
+			if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
+				if (oa->o_valid & OBD_MD_FLFLAGS)
+					oa->o_flags |= OBD_FL_NO_PRJQUOTA;
+				else
+					oa->o_flags = OBD_FL_NO_PRJQUOTA;
+			}
+
 			oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
-				       OBD_MD_FLGRPQUOTA;
+				       OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
 		}
 	} else if (cmd == OBD_BRW_READ) {
 		/* If oa != NULL then mdt_preprw_read updated the inode
@@ -582,7 +904,7 @@ out_put:
 	lu_object_put(tsi->tsi_env, &mo->mot_obj);
 out_unlock:
 	if (srvlock)
-		mdt_save_lock(info, &lh, LCK_PW, rc);
+		tgt_extent_unlock(&lh, LCK_PW);
 out:
 	mdt_thread_info_fini(info);
 	return rc;
@@ -610,7 +932,7 @@ int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
 
 	/* There can be only one write lock covering data, try to match it. */
 	policy.l_inodebits.bits = MDS_INODELOCK_DOM;
-	mode = ldlm_lock_match(ns, LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK,
+	mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
 			       &res->lr_name, LDLM_IBITS, &policy,
 			       LCK_PW, &lockh, 0);
 
@@ -711,7 +1033,7 @@ int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
 
 	/* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
 	if (dom_lock || !mdt_dom_lvb_is_valid(res))
-		mdt_dom_lvbo_update(res, NULL, NULL, false);
+		mdt_dom_lvbo_update(env, res, NULL, NULL, false);
 
 	mdt_lvb2body(res, mb);
 	ldlm_resource_putref(res);
@@ -772,7 +1094,7 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
 	if (flags & LDLM_FL_RESENT) {
 		rc = LDLM_ITER_CONTINUE;
 	} else {
-		__u64 tmpflags = 0;
+		__u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
 		enum ldlm_error err;
 
 		rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
@@ -809,7 +1131,7 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
 fill_mbo:
 	/* LVB can be without valid data in case of DOM */
 	if (!mdt_dom_lvb_is_valid(res))
-		mdt_dom_lvbo_update(res, lock, NULL, false);
+		mdt_dom_lvbo_update(mti->mti_env, res, lock, NULL, false);
 	mdt_lvb2body(res, mbo);
 	RETURN(rc);
 }
@@ -910,14 +1232,14 @@ void mdt_dom_discard_data(struct mdt_thread_info *info,
 
 	/* Tell the clients that the object is gone now and that they should
 	 * throw away any cached pages. */
-	rc = ldlm_cli_enqueue_local(mdt->mdt_namespace, res_id, LDLM_IBITS,
-				    policy, LCK_PW, &flags, ldlm_blocking_ast,
-				    ldlm_completion_ast, NULL, NULL, 0,
-				    LVB_T_NONE, NULL, &dom_lh);
+	rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, res_id,
+				    LDLM_IBITS, policy, LCK_PW, &flags,
+				    ldlm_blocking_ast, ldlm_completion_ast,
+				    NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
 
 	/* We only care about the side-effects, just drop the lock. */
 	if (rc == ELDLM_OK)
-		ldlm_lock_decref(&dom_lh, LCK_PW);
+		ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
 }
 
 /* check if client has already DoM lock for given resource */
@@ -954,3 +1276,281 @@ bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
 	return rc;
 }
 
+/**
+ * MDT request handler for OST_GETATTR RPC.
+ *
+ * This is data-specific request to get object and layout versions under
+ * IO lock. It is reliable only for Data-on-MDT files.
+ *
+ * \param[in] tsi target session environment for this request
+ *
+ * \retval 0 if successful
+ * \retval negative value on error
+ */
+int mdt_data_version_get(struct tgt_session_info *tsi)
+{
+	struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
+	struct mdt_device *mdt = mti->mti_mdt;
+	struct mdt_body *repbody;
+	struct mdt_object *mo = mti->mti_object;
+	struct lov_comp_md_v1 *comp;
+	struct lustre_handle lh = { 0 };
+	__u64 flags = 0;
+	__s64 version;
+	enum ldlm_mode lock_mode = LCK_PR;
+	bool srvlock;
+	int rc;
+
+	ENTRY;
+
+	req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
+	req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
+	rc = req_capsule_server_pack(tsi->tsi_pill);
+	if (unlikely(rc != 0))
+		RETURN(err_serious(rc));
+
+	repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
+	if (repbody == NULL)
+		RETURN(-ENOMEM);
+
+	srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
+		  tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
+
+	if (srvlock) {
+		if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
+			lock_mode = LCK_PW;
+
+		fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
+		rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
+				       &lh, lock_mode, &flags);
+		if (rc != 0)
+			RETURN(rc);
+	}
+
+	if (!mdt_object_exists(mo))
+		GOTO(out, rc = -ENOENT);
+	if (mdt_object_remote(mo))
+		GOTO(out, rc = -EREMOTE);
+	if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
+		GOTO(out, rc = -EBADF);
+
+	/* Get version first */
+	version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
+	if (version && version != -EOPNOTSUPP) {
+		repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
+		/* re-use mbo_ioepoch to transfer version */
+		repbody->mbo_version = version;
+	}
+
+	/* Read layout to get its version */
+	rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
+	if (rc == -ENODATA) /* File has no layout yet */
+		GOTO(out, rc = 0);
+	else if (rc < 0)
+		GOTO(out, rc);
+
+	comp = mti->mti_buf.lb_buf;
+	if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
+		CDEBUG(D_INFO, DFID" has no composite layout",
+		       PFID(&tsi->tsi_fid));
+		GOTO(out, rc = -ESTALE);
+	}
+
+	CDEBUG(D_INODE, DFID": layout version: %u\n",
+	       PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
+
+	repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
+	/* re-use mbo_rdev for that */
+	repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
+	rc = 0;
+out:
+	if (srvlock)
+		tgt_mdt_data_unlock(&lh, lock_mode);
+
+	repbody->mbo_valid |= OBD_MD_FLFLAGS;
+	repbody->mbo_flags = OBD_FL_FLUSH;
+	RETURN(rc);
+}
+
+/* read file data to the buffer */
+int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
+			 struct lustre_handle *lh)
+{
+	const struct lu_env *env = mti->mti_env;
+	struct tgt_session_info *tsi = tgt_ses_info(env);
+	struct req_capsule *pill = tsi->tsi_pill;
+	const struct lu_fid *fid;
+	struct ptlrpc_request *req = tgt_ses_req(tsi);
+	struct mdt_body *mbo;
+	struct dt_device *dt = mdt->mdt_bottom;
+	struct dt_object *mo;
+	void *buf;
+	struct niobuf_remote *rnb = NULL;
+	struct niobuf_local *lnb;
+	int rc;
+	int max_reply_len;
+	loff_t offset;
+	unsigned int len, copied = 0;
+	int lnbs, nr_local, i;
+	bool dom_lock = false;
+
+	ENTRY;
+
+	if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
+		/* There is no reply buffers for this field, this means that
+		 * client has no support for data in reply.
+		 */
+		RETURN(0);
+	}
+
+	mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
+
+	if (lustre_handle_is_used(lh)) {
+		struct ldlm_lock *lock;
+
+		lock = ldlm_handle2lock(lh);
+		if (lock) {
+			dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
+			LDLM_LOCK_PUT(lock);
+		}
+	}
+
+	/* return data along with open only along with DoM lock */
+	if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
+		RETURN(0);
+
+	if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
+		RETURN(0);
+
+	if (mbo->mbo_dom_size == 0)
+		RETURN(0);
+
+	/* check the maximum size available in reply */
+	max_reply_len =
+		req->rq_rqbd->rqbd_svcpt->scp_service->srv_max_reply_size;
+
+	CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d/%d\n",
+	       mbo->mbo_dom_size, max_reply_len, req->rq_reqmsg->lm_repsize,
+	       req->rq_replen);
+	len = req->rq_reqmsg->lm_repsize - req->rq_replen;
+	max_reply_len -= req->rq_replen;
+
+	/* NB: at this moment we have the following sizes:
+	 * - req->rq_replen: used data in reply
+	 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
+	 * - max_reply_len: maximum reply size allowed by protocol
+	 *
+	 * Ideal case when file size fits in allocated reply buffer,
+	 * that mean we can return whole data in reply. We can also fit more
+	 * data up to max_reply_size in total reply size, but this will cause
+	 * re-allocation on client and resend with larger buffer. This is still
+	 * faster than separate READ IO.
+	 * Third case if file is too big to fit even in maximum size, in that
+	 * case we return just tail to optimize possible append.
+	 *
+	 * At the moment the following strategy is used:
+	 * 1) try to fit into the buffer we have
+	 * 2) respond with bigger buffer so client will re-allocate it and
+	 *    resend (up to srv_max_reply_size value).
+	 * 3) return just file tail otherwise.
+	 */
+	if (mbo->mbo_dom_size <= len) {
+		/* can fit whole data */
+		len = mbo->mbo_dom_size;
+		offset = 0;
+	} else if (mbo->mbo_dom_size <= max_reply_len) {
+		/* It is worth to make this tunable ON/OFF because this will
+		 * cause buffer re-allocation and resend
+		 */
+		len = mbo->mbo_dom_size;
+		offset = 0;
+	} else {
+		int tail = mbo->mbo_dom_size % PAGE_SIZE;
+
+		/* no tail or tail can't fit in reply */
+		if (tail == 0 || len < tail)
+			RETURN(0);
+
+		len = tail;
+		offset = mbo->mbo_dom_size - len;
+	}
+	LASSERT((offset % PAGE_SIZE) == 0);
+	rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
+				     sizeof(*rnb) + len);
+	if (rc != 0) {
+		/* failed to grow data buffer, just exit */
+		GOTO(out, rc = -E2BIG);
+	}
+
+	/* re-take MDT_BODY buffer after the buffer growing above */
+	mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
+	fid = &mbo->mbo_fid1;
+	if (!fid_is_sane(fid))
+		RETURN(0);
+
+	rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
+	if (rnb == NULL)
+		GOTO(out, rc = -EPROTO);
+	buf = (char *)rnb + sizeof(*rnb);
+	rnb->rnb_len = len;
+	rnb->rnb_offset = offset;
+
+	mo = dt_locate(env, dt, fid);
+	if (IS_ERR(mo))
+		GOTO(out, rc = PTR_ERR(mo));
+	LASSERT(mo != NULL);
+
+	dt_read_lock(env, mo, 0);
+	if (!dt_object_exists(mo))
+		GOTO(unlock, rc = -ENOENT);
+
+	/* parse remote buffers to local buffers and prepare the latter */
+	lnbs = (len >> PAGE_SHIFT) + 1;
+	OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
+	if (lnb == NULL)
+		GOTO(unlock, rc = -ENOMEM);
+
+	rc = dt_bufs_get(env, mo, rnb, lnb, 0);
+	if (unlikely(rc < 0))
+		GOTO(free, rc);
+	LASSERT(rc <= lnbs);
+	nr_local = rc;
+	rc = dt_read_prep(env, mo, lnb, nr_local);
+	if (unlikely(rc))
+		GOTO(buf_put, rc);
+	/* copy data to the buffer finally */
+	for (i = 0; i < nr_local; i++) {
+		char *p = kmap(lnb[i].lnb_page);
+		long off;
+
+		LASSERT(lnb[i].lnb_page_offset == 0);
+		off = lnb[i].lnb_len & ~PAGE_MASK;
+		if (off > 0)
+			memset(p + off, 0, PAGE_SIZE - off);
+
+		memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
+		kunmap(lnb[i].lnb_page);
+		copied += lnb[i].lnb_len;
+		LASSERT(rc <= len);
+	}
+	CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
+	       len, offset);
+	if (copied < len)
+		CWARN("%s: read %i bytes for "DFID
+		      " but wanted %u, is size wrong?\n",
+		      tsi->tsi_exp->exp_obd->obd_name, copied,
+		      PFID(&tsi->tsi_fid), len);
+	EXIT;
+buf_put:
+	dt_bufs_put(env, mo, lnb, nr_local);
+free:
+	OBD_FREE(lnb, sizeof(*lnb) * lnbs);
+unlock:
+	dt_read_unlock(env, mo);
+	lu_object_put(env, &mo->do_lu);
+out:
+	if (rnb != NULL)
+		rnb->rnb_len = copied;
+	RETURN(0);
+}
+