X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_io.c;h=dfeb1028e6a1f73016f2f41eea3472feaa3a560c;hb=e609d30003560ba534e3da42d3240a081b00143e;hp=2548612ef3a2597023195d46ac57322ca41f045f;hpb=4f35c341f22b06b6c15a4af763e49b1e05f0dff9;p=fs%2Flustre-release.git

diff --git a/lustre/mdt/mdt_io.c b/lustre/mdt/mdt_io.c
index 2548612..dfeb102 100644
--- a/lustre/mdt/mdt_io.c
+++ b/lustre/mdt/mdt_io.c
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012, 2017 Intel Corporation.
+ * Copyright (c) 2017, Intel Corporation.
  */
 /*
  * lustre/mdt/mdt_io.c
@@ -33,86 +33,314 @@
 #include <dt_object.h>
 #include "mdt_internal.h"
 
-/* --------------- MDT grant code ---------------- */
+/* functions below are stubs for now, they will be implemented with
+ * grant support on MDT */
+static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
+				       char *jobid, long amount)
+{
+	return;
+}
 
-long mdt_grant_connect(const struct lu_env *env,
-		       struct obd_export *exp,
-		       u64 want, bool conservative)
+static inline void mdt_dom_read_lock(struct mdt_object *mo)
 {
-	struct mdt_device *mdt = mdt_exp2dev(exp);
-	u64 left;
-	long grant;
+	down_read(&mo->mot_dom_sem);
+}
 
-	ENTRY;
+static inline void mdt_dom_read_unlock(struct mdt_object *mo)
+{
+	up_read(&mo->mot_dom_sem);
+}
+
+static inline void mdt_dom_write_lock(struct mdt_object *mo)
+{
+	down_write(&mo->mot_dom_sem);
+}
+
+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;
 
-	dt_statfs(env, mdt->mdt_bottom, &mdt->mdt_osfs);
+	if (AT_OFF)
+		return obd_timeout / 2;
 
-	left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 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);
+}
 
-	grant = left;
+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;
 
-	CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %ld want: %llu left: %llu\n",
-	       exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
-	       exp, grant, want, left);
+	ENTRY;
 
-	return grant;
+	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);
+			RETURN_EXIT;
+		}
+	}
+	EXIT;
 }
 
-void mdt_grant_prepare_write(const struct lu_env *env,
-			     struct obd_export *exp, struct obdo *oa,
-			     struct niobuf_remote *rnb, int niocount)
+static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
+				   struct ldlm_lock *lock)
 {
-	struct mdt_device *mdt = mdt_exp2dev(exp);
-	u64 left;
+	struct obd_ioobj *ioo;
+	enum ldlm_mode mode;
+	__u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 
 	ENTRY;
 
-	left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
+	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;
 
-	/* grant more space back to the client if possible */
-	oa->o_grant = left;
+	if (!(lock->l_granted_mode & mode))
+		RETURN(0);
+
+	RETURN(1);
 }
-/* ---------------- end of MDT grant code ---------------- */
 
-/* functions below are stubs for now, they will be implemented with
- * grant support on MDT */
-static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
-				       char *jobid, long amount)
+static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
 {
-	return;
+	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);
 }
 
-void mdt_grant_prepare_read(const struct lu_env *env,
-			    struct obd_export *exp, struct obdo *oa)
+static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
 {
-	return;
+	mdt_rw_hpreq_check(req);
 }
 
-void mdt_grant_commit(struct obd_export *exp, unsigned long pending,
-		      int rc)
+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)
 {
-	return;
+	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 inline void mdt_dom_read_lock(struct mdt_object *mo)
+static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
+				      struct ldlm_lock *lock)
 {
-	down_read(&mo->mot_dom_sem);
+	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);
 }
 
-static inline void mdt_dom_read_unlock(struct mdt_object *mo)
+/**
+ * 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)
 {
-	up_read(&mo->mot_dom_sem);
+	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);
 }
 
-static inline void mdt_dom_write_lock(struct mdt_object *mo)
+/**
+ * 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)
 {
-	down_write(&mo->mot_dom_sem);
+	mdt_punch_hpreq_check(req);
 }
 
-static inline void mdt_dom_write_unlock(struct mdt_object *mo)
+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)
 {
-	up_write(&mo->mot_dom_sem);
+	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,
@@ -174,7 +402,7 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 
 	/* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
 	 * space back if possible */
-	mdt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
+	tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
 
 	mdt_dom_read_lock(mo);
 	if (!mdt_object_exists(mo)) {
@@ -191,8 +419,11 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 		if (unlikely(rc < 0))
 			GOTO(err, rc);
 		/* correct index for local buffers to continue with */
-		for (k = 0; k < rc; k++)
-			lnb[j+k].lnb_flags = rnb[i].rnb_flags;
+		for (k = 0; k < rc; k++) {
+			lnb[j + k].lnb_flags = rnb[i].rnb_flags;
+			if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
+				lnb[j + k].lnb_rc = -ENOSPC;
+		}
 		j += rc;
 		*nr_local += rc;
 		tot_bytes += rnb[i].rnb_len;
@@ -209,11 +440,11 @@ err:
 unlock:
 	mdt_dom_read_unlock(mo);
 	/* tgt_grant_prepare_write() was called, so we must commit */
-	mdt_grant_commit(exp, oa->o_grant_used, rc);
+	tgt_grant_commit(exp, oa->o_grant_used, rc);
 	/* let's still process incoming grant information packed in the oa,
 	 * but without enforcing grant since we won't proceed with the write.
 	 * Just like a read request actually. */
-	mdt_grant_prepare_read(env, exp, oa);
+	tgt_grant_prepare_read(env, exp, oa);
 	return rc;
 }
 
@@ -256,7 +487,7 @@ int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 				      objcount, obj, rnb, nr_local, lnb,
 				      jobid);
 	} else if (cmd == OBD_BRW_READ) {
-		mdt_grant_prepare_read(env, exp, oa);
+		tgt_grant_prepare_read(env, exp, oa);
 		rc = mdt_preprw_read(env, exp, mdt, mo, la,
 				     obj->ioo_bufcnt, rnb, nr_local, lnb,
 				     jobid);
@@ -343,6 +574,7 @@ retry:
 			GOTO(out_stop, rc);
 	}
 
+	tgt_vbr_obj_set(env, dob);
 	rc = dt_trans_start(env, dt, th);
 	if (rc)
 		GOTO(out_stop, rc);
@@ -368,6 +600,12 @@ out_stop:
 	if (rc == -ENOSPC)
 		th->th_sync = 1;
 
+
+	if (rc == 0 && granted > 0) {
+		if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
+			granted = 0;
+	}
+
 	th->th_result = rc;
 	dt_trans_stop(env, dt, th);
 	if (rc == -ENOSPC && retries++ < 3) {
@@ -379,7 +617,8 @@ out_stop:
 out:
 	dt_bufs_put(env, dob, lnb, niocount);
 	mdt_dom_read_unlock(mo);
-	mdt_grant_commit(exp, granted, old_rc);
+	if (granted > 0)
+		tgt_grant_commit(exp, granted, old_rc);
 	RETURN(rc);
 }
 
@@ -594,9 +833,18 @@ int mdt_punch_hdl(struct tgt_session_info *tsi)
 	if (IS_ERR(mo))
 		GOTO(out_unlock, rc = PTR_ERR(mo));
 
-	mdt_dom_write_lock(mo);
 	if (!mdt_object_exists(mo))
 		GOTO(out_put, rc = -ENOENT);
+
+	/* Shouldn't happen on dirs */
+	if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
+		rc = -EPERM;
+		CERROR("%s: Truncate on dir "DFID": rc = %d\n",
+		       exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
+		GOTO(out_put, rc);
+	}
+
+	mdt_dom_write_lock(mo);
 	dob = mdt_obj2dt(mo);
 
 	la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
@@ -617,7 +865,7 @@ 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;
@@ -645,7 +893,7 @@ 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);
 
@@ -708,8 +956,8 @@ static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
 
 	lock_res(res);
 	res_lvb = res->lr_lvb_data;
-	mb->mbo_size = res_lvb->lvb_size;
-	mb->mbo_blocks = res_lvb->lvb_blocks;
+	mb->mbo_dom_size = res_lvb->lvb_size;
+	mb->mbo_dom_blocks = res_lvb->lvb_blocks;
 	mb->mbo_mtime = res_lvb->lvb_mtime;
 	mb->mbo_ctime = res_lvb->lvb_ctime;
 	mb->mbo_atime = res_lvb->lvb_atime;
@@ -717,7 +965,7 @@ static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
 	CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
 
 	mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
-			 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
+			 OBD_MD_DOM_SIZE;
 	unlock_res(res);
 }
 
@@ -741,23 +989,14 @@ int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
 	fid_build_reg_res_name(fid, &resid);
 	res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
 				LDLM_IBITS, 1);
-	if (IS_ERR(res) || res->lr_lvb_data == NULL)
+	if (IS_ERR(res))
 		RETURN(-ENOENT);
 
-	/* if there is no DOM bit in the lock then glimpse is needed
-	 * to return valid size */
-	if (!dom_lock) {
-		rc = mdt_do_glimpse(env, mdt->mdt_namespace, res);
-		if (rc < 0)
-			GOTO(out, rc);
-	}
-
 	/* 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_lvb2body(res, mb);
-out:
 	ldlm_resource_putref(res);
 	RETURN(rc);
 }
@@ -819,7 +1058,7 @@ int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
 		__u64 tmpflags = 0;
 		enum ldlm_error err;
 
-		rc = policy(lock, &tmpflags, 0, &err, NULL);
+		rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
 		check_res_locked(res);
 	}
 	unlock_res(res);
@@ -946,7 +1185,7 @@ void mdt_dom_discard_data(struct mdt_thread_info *info,
 	struct ldlm_res_id *res_id = &info->mti_res_id;
 	struct lustre_handle dom_lh;
 	__u64 flags = LDLM_FL_AST_DISCARD_DATA;
-	__u64 rc = 0;
+	int rc = 0;
 
 	policy->l_inodebits.bits = MDS_INODELOCK_DOM;
 	policy->l_inodebits.try_bits = 0;
@@ -964,3 +1203,132 @@ void mdt_dom_discard_data(struct mdt_thread_info *info,
 		ldlm_lock_decref(&dom_lh, LCK_PW);
 }
 
+/* check if client has already DoM lock for given resource */
+bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
+			     const struct lu_fid *fid)
+{
+	struct mdt_device *mdt = info->mti_mdt;
+	union ldlm_policy_data *policy = &info->mti_policy;
+	struct ldlm_res_id *res_id = &info->mti_res_id;
+	struct lustre_handle lockh;
+	enum ldlm_mode mode;
+	struct ldlm_lock *lock;
+	bool rc;
+
+	policy->l_inodebits.bits = MDS_INODELOCK_DOM;
+	fid_build_reg_res_name(fid, res_id);
+
+	mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
+			       LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
+			       LCK_PW, &lockh, 0);
+
+	/* There is no other PW lock on this object; finished. */
+	if (mode == 0)
+		return false;
+
+	lock = ldlm_handle2lock(&lockh);
+	if (lock == 0)
+		return false;
+
+	/* check if lock from the same client */
+	rc = (lock->l_export->exp_handle.h_cookie ==
+	      info->mti_exp->exp_handle.h_cookie);
+	LDLM_LOCK_PUT(lock);
+	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);
+}