X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_io.c;h=04cfe5d47bed2e2413befaf3d28208a5b1084c46;hp=637f5eebf529e2895ae286005111c761646b8375;hb=cd8fb1e8d300c0622200d6b25d187d8d0472e79f;hpb=9e338ee25609d9bbcc359c37012e6868dd6be5d3

diff --git a/lustre/mdt/mdt_io.c b/lustre/mdt/mdt_io.c
index 637f5ee..04cfe5d 100644
--- a/lustre/mdt/mdt_io.c
+++ b/lustre/mdt/mdt_io.c
@@ -35,12 +35,6 @@
 
 /* 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;
-}
-
 static inline void mdt_dom_read_lock(struct mdt_object *mo)
 {
 	down_read(&mo->mot_dom_sem);
@@ -61,21 +55,34 @@ 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)
+static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
 {
-	struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
-	time64_t req_timeout;
+	struct ldlm_resource *res;
+	struct ldlm_lock *lock;
 
-	if (AT_OFF)
-		return obd_timeout / 2;
+	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;
+	}
 
-	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);
+	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,
@@ -102,9 +109,11 @@ static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
 			ldlm_lock_prolong_one(lock, data);
 			lock->l_last_used = ktime_get();
 			LDLM_LOCK_PUT(lock);
-			RETURN_EXIT;
+			if (data->lpa_locks_cnt > 0)
+				RETURN_EXIT;
 		}
 	}
+	mdt_dom_resource_prolong(data);
 	EXIT;
 }
 
@@ -172,14 +181,14 @@ static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
 	if (opc == OST_READ)
 		pa.lpa_mode |= LCK_PR;
 
-	DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID"\n",
+	DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
 		  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",
+		       "%s: refreshed %u locks timeout for req %p\n",
 		       tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
 		RETURN(1);
 	}
@@ -349,24 +358,54 @@ static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
 			   struct niobuf_remote *rnb, int *nr_local,
 			   struct niobuf_local *lnb, char *jobid)
 {
+	struct tgt_session_info *tsi = tgt_ses_info(env);
+	struct ptlrpc_request *req = tgt_ses_req(tsi);
 	struct dt_object *dob;
 	int i, j, rc, tot_bytes = 0;
+	int maxlnb = *nr_local;
+	int level;
 
 	ENTRY;
 
 	mdt_dom_read_lock(mo);
-	if (!mdt_object_exists(mo))
-		GOTO(unlock, rc = -ENOENT);
+	*nr_local = 0;
+	/* the only valid case when READ can find object is missing or stale
+	 * when export is just evicted and open files are closed forcefully
+	 * on server while client's READ can be in progress.
+	 * This should not happen on healthy export, object can't be missing
+	 * or dying because both states means it was finally destroyed.
+	 */
+	level = exp->exp_failed ? D_INFO : D_ERROR;
+	if (!mdt_object_exists(mo)) {
+		CDEBUG_LIMIT(level,
+			     "%s: READ IO to missing obj "DFID": rc = %d\n",
+			     exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+			     -ENOENT);
+		/* return 0 and continue with empty commit to skip such READ
+		 * without more BRW errors.
+		 */
+		RETURN(0);
+	}
+	if (lu_object_is_dying(&mo->mot_header)) {
+		CDEBUG_LIMIT(level,
+			     "%s: READ IO to stale obj "DFID": rc = %d\n",
+			     exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+			     -ESTALE);
+		/* return 0 and continue with empty commit to skip such READ
+		 * without more BRW errors.
+		 */
+		RETURN(0);
+	}
 
 	dob = mdt_obj2dt(mo);
 	/* parse remote buffers to local buffers and prepare the latter */
-	*nr_local = 0;
 	for (i = 0, j = 0; i < niocount; i++) {
-		rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
+		rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
 		if (unlikely(rc < 0))
 			GOTO(buf_put, rc);
 		/* correct index for local buffers to continue with */
 		j += rc;
+		maxlnb -= rc;
 		*nr_local += rc;
 		tot_bytes += rnb[i].rnb_len;
 	}
@@ -379,11 +418,10 @@ static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
 	if (unlikely(rc))
 		GOTO(buf_put, rc);
 
-	mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
+	mdt_counter_incr(req, LPROC_MDT_IO_READ, tot_bytes);
 	RETURN(0);
 buf_put:
 	dt_bufs_put(env, dob, lnb, *nr_local);
-unlock:
 	mdt_dom_read_unlock(mo);
 	return rc;
 }
@@ -395,8 +433,11 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 			    struct niobuf_remote *rnb, int *nr_local,
 			    struct niobuf_local *lnb, char *jobid)
 {
+	struct tgt_session_info *tsi = tgt_ses_info(env);
+	struct ptlrpc_request *req = tgt_ses_req(tsi);
 	struct dt_object *dob;
 	int i, j, k, rc = 0, tot_bytes = 0;
+	int maxlnb = *nr_local;
 
 	ENTRY;
 
@@ -405,17 +446,36 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 	tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
 
 	mdt_dom_read_lock(mo);
+	*nr_local = 0;
+	/* don't report error in cases with failed export */
 	if (!mdt_object_exists(mo)) {
-		CDEBUG(D_ERROR, "%s: BRW to missing obj "DFID"\n",
-		       exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)));
-		GOTO(unlock, rc = -ENOENT);
+		int level = exp->exp_failed ? D_INFO : D_ERROR;
+
+		rc = -ENOENT;
+		CDEBUG_LIMIT(level,
+			     "%s: WRITE IO to missing obj "DFID": rc = %d\n",
+			     exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
+			     rc);
+		/* exit with no data written, note nr_local = 0 above */
+		GOTO(unlock, rc);
+	}
+	if (lu_object_is_dying(&mo->mot_header)) {
+		/* This is possible race between object destroy followed by
+		 * discard BL AST and client cache flushing. Object is
+		 * referenced until discard finish.
+		 */
+		CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
+		       PFID(mdt_object_fid(mo)));
+		/* Note: continue with no error here to don't cause BRW errors
+		 * but skip transaction in commitrw silently so no data is
+		 * written.
+		 */
 	}
 
 	dob = mdt_obj2dt(mo);
 	/* parse remote buffers to local buffers and prepare the latter */
-	*nr_local = 0;
 	for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
-		rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
+		rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
 		if (unlikely(rc < 0))
 			GOTO(err, rc);
 		/* correct index for local buffers to continue with */
@@ -425,6 +485,7 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 				lnb[j + k].lnb_rc = -ENOSPC;
 		}
 		j += rc;
+		maxlnb -= rc;
 		*nr_local += rc;
 		tot_bytes += rnb[i].rnb_len;
 	}
@@ -433,7 +494,7 @@ static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
 	if (likely(rc))
 		GOTO(err, rc);
 
-	mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
+	mdt_counter_incr(req, LPROC_MDT_IO_WRITE, tot_bytes);
 	RETURN(0);
 err:
 	dt_bufs_put(env, dob, lnb, *nr_local);
@@ -514,11 +575,10 @@ static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
 
 	ENTRY;
 
-	LASSERT(niocount > 0);
-
 	dob = mdt_obj2dt(mo);
 
-	dt_bufs_put(env, dob, lnb, niocount);
+	if (niocount)
+		dt_bufs_put(env, dob, lnb, niocount);
 
 	mdt_dom_read_unlock(mo);
 	RETURN(rc);
@@ -548,6 +608,20 @@ static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 retry:
 	if (!dt_object_exists(dob))
 		GOTO(out, rc = -ENOENT);
+	if (lu_object_is_dying(&mo->mot_header)) {
+		/* Commit to stale object can be just skipped silently. */
+		CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
+			PFID(mdt_object_fid(mo)));
+		GOTO(out, rc = 0);
+	}
+
+	if (niocount == 0) {
+		rc = -EPROTO;
+		DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
+			  "%s: commit with no pages for "DFID": rc = %d\n",
+			  exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
+		GOTO(out, rc);
+	}
 
 	th = dt_trans_create(env, dt);
 	if (IS_ERR(th))
@@ -574,12 +648,13 @@ retry:
 			GOTO(out_stop, rc);
 	}
 
+	tgt_vbr_obj_set(env, dob);
 	rc = dt_trans_start(env, dt, th);
 	if (rc)
 		GOTO(out_stop, rc);
 
 	dt_write_lock(env, dob, 0);
-	rc = dt_write_commit(env, dob, lnb, niocount, th);
+	rc = dt_write_commit(env, dob, lnb, niocount, th, 0);
 	if (rc)
 		GOTO(unlock, rc);
 
@@ -652,12 +727,6 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 	__u64 valid;
 	int rc = 0;
 
-	if (npages == 0) {
-		CERROR("%s: no pages to commit\n",
-		       exp->exp_obd->obd_name);
-		rc = -EPROTO;
-	}
-
 	LASSERT(mo);
 
 	if (cmd == OBD_BRW_WRITE) {
@@ -669,7 +738,10 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 		 * doesn't already exist so we can store the reservation handle
 		 * there. */
 		valid = OBD_MD_FLUID | OBD_MD_FLGID;
-		valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
+		if (tgt_fmd_check(exp, mdt_object_fid(mo),
+				  mdt_info_req(info)->rq_xid))
+			valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
+				 OBD_MD_FLCTIME;
 
 		la_from_obdo(la, oa, valid);
 
@@ -699,8 +771,15 @@ 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
@@ -714,7 +793,6 @@ int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 	} else {
 		rc = -EPROTO;
 	}
-	/* this put is pair to object_get in ofd_preprw_write */
 	mdt_thread_info_fini(info);
 	RETURN(rc);
 }
@@ -769,6 +847,7 @@ stop:
 int mdt_punch_hdl(struct tgt_session_info *tsi)
 {
 	const struct obdo *oa = &tsi->tsi_ost_body->oa;
+	struct ptlrpc_request *req = tgt_ses_req(tsi);
 	struct ost_body *repbody;
 	struct mdt_thread_info *info;
 	struct lu_attr *la;
@@ -779,6 +858,7 @@ int mdt_punch_hdl(struct tgt_session_info *tsi)
 	struct dt_object *dob;
 	__u64 flags = 0;
 	struct lustre_handle lh = { 0, };
+	ktime_t kstart = ktime_get();
 	__u64 start, end;
 	int rc;
 	bool srvlock;
@@ -786,7 +866,7 @@ int mdt_punch_hdl(struct tgt_session_info *tsi)
 	ENTRY;
 
 	/* check that we do support OBD_CONNECT_TRUNCLOCK. */
-	CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
+	BUILD_BUG_ON(!(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK));
 
 	if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 	    (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
@@ -850,6 +930,11 @@ int mdt_punch_hdl(struct tgt_session_info *tsi)
 	la->la_size = start;
 	la->la_valid |= LA_SIZE;
 
+	/* MDT supports FMD for Data-on-MDT needs */
+	if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
+		tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
+			       tgt_ses_req(tsi)->rq_xid);
+
 	rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
 			      start, end, la);
 	mdt_dom_write_unlock(mo);
@@ -857,8 +942,8 @@ int mdt_punch_hdl(struct tgt_session_info *tsi)
 		GOTO(out_put, rc);
 
 	mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
-	mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
-			    tsi->tsi_jobid, 1);
+	mdt_counter_incr(req, LPROC_MDT_IO_PUNCH,
+			 ktime_us_delta(ktime_get(), kstart));
 	EXIT;
 out_put:
 	lu_object_put(tsi->tsi_env, &mo->mot_obj);
@@ -885,7 +970,7 @@ int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
 	enum ldlm_mode mode;
 	struct ldlm_lock *lock;
 	struct ldlm_glimpse_work *gl_work;
-	struct list_head gl_list;
+	LIST_HEAD(gl_list);
 	int rc;
 
 	ENTRY;
@@ -926,7 +1011,6 @@ int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
 	gl_work->gl_lock = LDLM_LOCK_GET(lock);
 	/* The glimpse callback is sent to one single IO lock. As a result,
 	 * the gl_work list is just composed of one element */
-	INIT_LIST_HEAD(&gl_list);
 	list_add_tail(&gl_work->gl_list, &gl_list);
 	/* There is actually no need for a glimpse descriptor when glimpsing
 	 * IO locks */
@@ -1054,7 +1138,7 @@ 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);
@@ -1176,32 +1260,6 @@ out:
 	RETURN(rc);
 }
 
-void mdt_dom_discard_data(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 dom_lh;
-	__u64 flags = LDLM_FL_AST_DISCARD_DATA;
-	int rc = 0;
-
-	policy->l_inodebits.bits = MDS_INODELOCK_DOM;
-	policy->l_inodebits.try_bits = 0;
-	fid_build_reg_res_name(fid, res_id);
-
-	/* 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);
-
-	/* We only care about the side-effects, just drop the lock. */
-	if (rc == ELDLM_OK)
-		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)
@@ -1236,3 +1294,400 @@ 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;
+	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 (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
+		RETURN(0);
+
+	if (!mbo->mbo_dom_size)
+		RETURN(0);
+
+	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);
+
+	CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
+	       mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
+	len = req->rq_reqmsg->lm_repsize - 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
+	 *
+	 * 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) 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 <
+		   mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) {
+		int tail, pgbits;
+
+		/* File tail offset must be aligned with larger page size
+		 * between client and server, so the maximum page size is
+		 * used here to align offset.
+		 *
+		 * NB: DOM feature was introduced when server supports pagebits
+		 * already, so it should be always non-zero value. Report error
+		 * if it is not for some reason.
+		 */
+		if (!req->rq_export->exp_target_data.ted_pagebits) {
+			CERROR("%s: client page bits are not saved on server\n",
+			       mdt_obd_name(mdt));
+			RETURN(0);
+		}
+		pgbits = max_t(int, PAGE_SHIFT,
+			       req->rq_export->exp_target_data.ted_pagebits);
+		tail = mbo->mbo_dom_size % (1 << pgbits);
+
+		/* no partial tail or tail can't fit in reply */
+		if (tail == 0 || len < tail)
+			RETURN(0);
+
+		len = tail;
+		offset = mbo->mbo_dom_size - len;
+	} else {
+		/* DOM stripe is fully written, so don't expect its tail
+		 * will be used by append.
+		 */
+		RETURN(0);
+	}
+
+	LASSERT((offset & ~PAGE_MASK) == 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 and NIOBUF_INLINE buffers after the buffer grow */
+	mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
+	fid = &mbo->mbo_fid1;
+	if (!fid_is_sane(fid))
+		GOTO(out, rc = -EINVAL);
+
+	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_rnb, 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_PTR_ARRAY(lnb, lnbs);
+	if (lnb == NULL)
+		GOTO(unlock, rc = -ENOMEM);
+
+	rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 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);
+		/* Ignore partially copied data */
+		copied = 0;
+	}
+	EXIT;
+buf_put:
+	dt_bufs_put(env, mo, lnb, nr_local);
+free:
+	OBD_FREE_PTR_ARRAY(lnb, lnbs);
+unlock:
+	dt_read_unlock(env, mo);
+	lu_object_put(env, &mo->do_lu);
+out_rnb:
+	rnb->rnb_len = copied;
+out:
+	/* Don't fail OPEN request if read-on-open is failed, but drop
+	 * a message in log about the error.
+	 */
+	if (rc)
+		CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
+
+	RETURN(0);
+}
+
+/**
+ * Completion AST for DOM discard locks:
+ *
+ * CP AST an DOM discard lock is called always right after enqueue or from
+ * reprocess if lock was blocked, in the latest case l_ast_data is set to
+ * the mdt_object which is kept while there are pending locks on it.
+ */
+int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
+{
+	struct mdt_object *mo;
+	struct lustre_handle dom_lh;
+	struct lu_env *env;
+
+	ENTRY;
+
+	/* l_ast_data is set when lock was not granted immediately
+	 * in mdt_dom_discard_data() below but put into waiting list,
+	 * so this CP callback means we are finished and corresponding
+	 * MDT object should be released finally as well as lock itself.
+	 */
+	lock_res_and_lock(lock);
+	if (!lock->l_ast_data) {
+		unlock_res_and_lock(lock);
+		RETURN(0);
+	}
+
+	mo = lock->l_ast_data;
+	lock->l_ast_data = NULL;
+	unlock_res_and_lock(lock);
+
+	ldlm_lock2handle(lock, &dom_lh);
+	ldlm_lock_decref(&dom_lh, LCK_PW);
+
+	env = lu_env_find();
+	LASSERT(env);
+	mdt_object_put(env, mo);
+
+	RETURN(0);
+}
+
+void mdt_dom_discard_data(struct mdt_thread_info *info,
+			  struct mdt_object *mo)
+{
+	struct ptlrpc_request *req = mdt_info_req(info);
+	struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
+	union ldlm_policy_data policy;
+	struct ldlm_res_id res_id;
+	struct lustre_handle dom_lh;
+	struct ldlm_lock *lock;
+	__u64 flags = LDLM_FL_AST_DISCARD_DATA;
+	int rc = 0;
+	bool old_client;
+
+	ENTRY;
+
+	if (req && req_is_replay(req))
+		RETURN_EXIT;
+
+	policy.l_inodebits.bits = MDS_INODELOCK_DOM;
+	policy.l_inodebits.try_bits = 0;
+	fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
+
+	/* Keep blocking version of discard for an old client to avoid
+	 * crashes on non-patched clients. LU-11359.
+	 */
+	old_client = req && !(exp_connect_flags2(req->rq_export) &
+			      OBD_CONNECT2_ASYNC_DISCARD);
+
+	/* Tell the clients that the object is gone now and that they should
+	 * throw away any cached pages. */
+	rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
+				    LDLM_IBITS, &policy, LCK_PW, &flags,
+				    ldlm_blocking_ast, old_client ?
+				    ldlm_completion_ast :
+				    ldlm_dom_discard_cp_ast,
+				    NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
+	if (rc != ELDLM_OK) {
+		CDEBUG(D_DLMTRACE,
+		       "Failed to issue discard lock, rc = %d\n", rc);
+		RETURN_EXIT;
+	}
+
+	lock = ldlm_handle2lock(&dom_lh);
+	lock_res_and_lock(lock);
+	/* if lock is not granted then there are BL ASTs in progress and
+	 * lock will be granted in result of reprocessing with CP callback
+	 * notifying about that. The mdt object has to be kept until that and
+	 * it is saved in l_ast_data of the lock. Lock reference is kept too
+	 * until that to prevent it from canceling.
+	 */
+	if (!is_granted_or_cancelled_nolock(lock)) {
+		mdt_object_get(info->mti_env, mo);
+		lock->l_ast_data = mo;
+		unlock_res_and_lock(lock);
+	} else {
+		unlock_res_and_lock(lock);
+		ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
+	}
+	LDLM_LOCK_PUT(lock);
+
+	RETURN_EXIT;
+}