X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_object.c;h=63ab904b92fceb4ba0a8d07a8449c858234d07f6;hp=f57bcf98ca81c45bd5c317560f43af205246628b;hb=ec46feab365a10960f36da017b41066915b79c40;hpb=c0a455e6d035c3ce4f9d52395b0a77773b2fac93

diff --git a/lustre/osd-zfs/osd_object.c b/lustre/osd-zfs/osd_object.c
index f57bcf9..63ab904 100644
--- a/lustre/osd-zfs/osd_object.c
+++ b/lustre/osd-zfs/osd_object.c
@@ -23,7 +23,7 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2016, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -62,14 +62,12 @@
 #include <sys/txg.h>
 
 char *osd_obj_tag = "osd_object";
+static int osd_object_sync_delay_us = -1;
 
 static struct dt_object_operations osd_obj_ops;
 static struct lu_object_operations osd_lu_obj_ops;
-extern struct dt_body_operations osd_body_ops;
 static struct dt_object_operations osd_obj_otable_it_ops;
 
-extern struct kmem_cache *osd_object_kmem;
-
 static void
 osd_object_sa_fini(struct osd_object *obj)
 {
@@ -185,8 +183,8 @@ osd_object_sa_bulk_update(struct osd_object *obj, sa_bulk_attr_t *attrs,
 /*
  * Retrieve the attributes of a DMU object
  */
-int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
-			  struct osd_object *obj, struct lu_attr *la)
+static int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
+				 struct osd_object *obj, struct lu_attr *la)
 {
 	struct osa_attr	*osa = &osd_oti_get(env)->oti_osa;
 	sa_bulk_attr_t	*bulk = osd_oti_get(env)->oti_attr_bulk;
@@ -290,6 +288,7 @@ int __osd_obj2dnode(objset_t *os, uint64_t oid, dnode_t **dnp)
 	dbi = (dmu_buf_impl_t *)db;
 	DB_DNODE_ENTER(dbi);
 	*dnp = DB_DNODE(dbi);
+	DB_DNODE_EXIT(dbi);
 	LASSERT(*dnp != NULL);
 
 	return 0;
@@ -308,9 +307,26 @@ struct lu_object *osd_object_alloc(const struct lu_env *env,
 	OBD_SLAB_ALLOC_PTR_GFP(mo, osd_object_kmem, GFP_NOFS);
 	if (mo != NULL) {
 		struct lu_object *l;
+		struct lu_object_header *h;
+		struct osd_device *o = osd_dev(d);
 
 		l = &mo->oo_dt.do_lu;
-		dt_object_init(&mo->oo_dt, NULL, d);
+		if (unlikely(o->od_in_init)) {
+			OBD_ALLOC_PTR(h);
+			if (!h) {
+				OBD_FREE_PTR(mo);
+				return NULL;
+			}
+
+			lu_object_header_init(h);
+			lu_object_init(l, h, d);
+			lu_object_add_top(h, l);
+			mo->oo_header = h;
+		} else {
+			dt_object_init(&mo->oo_dt, NULL, d);
+			mo->oo_header = NULL;
+		}
+
 		mo->oo_dt.do_ops = &osd_obj_ops;
 		l->lo_ops = &osd_lu_obj_ops;
 		INIT_LIST_HEAD(&mo->oo_sa_linkage);
@@ -388,7 +404,7 @@ out:
 /*
  * Concurrency: shouldn't matter.
  */
-int osd_object_init0(const struct lu_env *env, struct osd_object *obj)
+static int osd_object_init0(const struct lu_env *env, struct osd_object *obj)
 {
 	struct osd_device	*osd = osd_obj2dev(obj);
 	const struct lu_fid	*fid = lu_object_fid(&obj->oo_dt.do_lu);
@@ -436,9 +452,10 @@ static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 	struct lu_buf		buf;
 	int			rc;
 	struct lustre_mdt_attrs	*lma;
+	const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
 	ENTRY;
 
-	CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
+	BUILD_BUG_ON(sizeof(info->oti_buf) < sizeof(*lma));
 	lma = (struct lustre_mdt_attrs *)info->oti_buf;
 	buf.lb_buf = lma;
 	buf.lb_len = sizeof(info->oti_buf);
@@ -452,12 +469,23 @@ static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 			CWARN("%s: unsupported incompat LMA feature(s) %#x for "
 			      "fid = "DFID"\n", osd_obj2dev(obj)->od_svname,
 			      lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
-			      PFID(lu_object_fid(&obj->oo_dt.do_lu)));
+			      PFID(rfid));
 			rc = -EOPNOTSUPP;
+		} else if (unlikely(!lu_fid_eq(rfid, &lma->lma_self_fid))) {
+			CERROR("%s: FID-in-LMA "DFID" does not match the "
+			      "object self-fid "DFID"\n",
+			      osd_obj2dev(obj)->od_svname,
+			      PFID(&lma->lma_self_fid), PFID(rfid));
+			rc = -EREMCHG;
 		} else {
+			struct osd_device *osd = osd_obj2dev(obj);
+
 			if (lma->lma_compat & LMAC_STRIPE_INFO &&
-			    osd_obj2dev(obj)->od_is_ost)
+			    osd->od_is_ost)
 				obj->oo_pfid_in_lma = 1;
+			if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
+			    osd->od_remote_parent_dir != ZFS_NO_OBJECT)
+				lu_object_set_agent_entry(&obj->oo_dt.do_lu);
 		}
 	} else if (rc == -ENODATA) {
 		/* haven't initialize LMA xattr */
@@ -506,8 +534,15 @@ static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 	struct osd_object *obj = osd_obj(l);
 	struct osd_device *osd = osd_obj2dev(obj);
 	const struct lu_fid *fid = lu_object_fid(l);
+	struct lustre_scrub *scrub = &osd->od_scrub;
+	struct osd_thread_info *info = osd_oti_get(env);
+	struct luz_direntry *zde = &info->oti_zde;
+	struct osd_idmap_cache *idc;
+	char *name = info->oti_str;
 	uint64_t oid;
 	int rc = 0;
+	int rc1;
+	bool remote = false;
 	ENTRY;
 
 	LASSERT(osd_invariant(obj));
@@ -515,10 +550,11 @@ static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 	if (fid_is_otable_it(&l->lo_header->loh_fid)) {
 		obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
 		l->lo_header->loh_attr |= LOHA_EXISTS;
-		RETURN(0);
+
+		GOTO(out, rc = 0);
 	}
 
-	if (conf != NULL && conf->loc_flags & LOC_F_NEW)
+	if (conf && conf->loc_flags & LOC_F_NEW)
 		GOTO(out, rc = 0);
 
 	if (unlikely(fid_is_acct(fid))) {
@@ -531,31 +567,117 @@ static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 		GOTO(out, rc = 0);
 	}
 
-	rc = osd_fid_lookup(env, osd, fid, &oid);
-	if (rc == 0) {
-		LASSERT(obj->oo_dn == NULL);
-		rc = __osd_obj2dnode(osd->od_os, oid, &obj->oo_dn);
-		/* EEXIST will be returned if object is being deleted in ZFS */
-		if (rc == -EEXIST) {
-			rc = 0;
-			GOTO(out, rc);
+	idc = osd_idc_find(env, osd, fid);
+	if (idc && !idc->oic_remote && idc->oic_dnode != ZFS_NO_OBJECT) {
+		oid = idc->oic_dnode;
+		goto zget;
+	}
+
+	rc = -ENOENT;
+	if (!list_empty(&osd->od_scrub.os_inconsistent_items))
+		rc = osd_oii_lookup(osd, fid, &oid);
+
+	if (rc)
+		rc = osd_fid_lookup(env, osd, fid, &oid);
+
+	if (rc == -ENOENT) {
+		if (likely(!(fid_is_norm(fid) || fid_is_igif(fid)) ||
+			   fid_is_on_ost(env, osd, fid) ||
+			   !zfs_test_bit(osd_oi_fid2idx(osd, fid),
+					 scrub->os_file.sf_oi_bitmap)))
+			GOTO(out, rc = 0);
+
+		rc = -EREMCHG;
+		goto trigger;
+	}
+
+	if (rc)
+		GOTO(out, rc);
+
+zget:
+	LASSERT(obj->oo_dn == NULL);
+
+	rc = __osd_obj2dnode(osd->od_os, oid, &obj->oo_dn);
+	/* EEXIST will be returned if object is being deleted in ZFS */
+	if (rc == -EEXIST)
+		GOTO(out, rc = 0);
+
+	if (rc) {
+		CERROR("%s: lookup "DFID"/%#llx failed: rc = %d\n",
+		       osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
+		GOTO(out, rc);
+	}
+
+	rc = osd_object_init0(env, obj);
+	if (rc)
+		GOTO(out, rc);
+
+	if (unlikely(obj->oo_header))
+		GOTO(out, rc = 0);
+
+	rc = osd_check_lma(env, obj);
+	if ((!rc && !remote) || (rc != -EREMCHG))
+		GOTO(out, rc);
+
+trigger:
+	/* We still have chance to get the valid dnode: for the object that is
+	 * referenced by remote name entry, the object on the local MDT will be
+	 * linked under the dir /REMOTE_PARENT_DIR with its FID string as name.
+	 *
+	 * During the OI scrub, if we cannot find the OI mapping, we may still
+	 * have change to map the FID to local OID via lookup the dir
+	 * /REMOTE_PARENT_DIR. */
+	if (!remote && !fid_is_on_ost(env, osd, fid)) {
+		osd_fid2str(name, fid, sizeof(info->oti_str));
+		rc = osd_zap_lookup(osd, osd->od_remote_parent_dir,
+				    NULL, name, 8, 3, (void *)zde);
+		if (!rc) {
+			oid = zde->lzd_reg.zde_dnode;
+			osd_dnode_rele(obj->oo_dn);
+			obj->oo_dn = NULL;
+			remote = true;
+			goto zget;
 		}
-		if (rc != 0) {
-			CERROR("%s: lookup "DFID"/%#llx failed: rc = %d\n",
-			       osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
-			GOTO(out, rc);
+	}
+
+	/* The case someone triggered the OI scrub already. */
+	if (thread_is_running(&scrub->os_thread)) {
+		if (!rc) {
+			LASSERT(remote);
+
+			lu_object_set_agent_entry(l);
+			osd_oii_insert(env, osd, fid, oid, false);
+		} else {
+			rc = -EINPROGRESS;
 		}
-		rc = osd_object_init0(env, obj);
-		if (rc != 0)
-			GOTO(out, rc);
 
-		rc = osd_check_lma(env, obj);
-		if (rc != 0)
-			GOTO(out, rc);
-	} else if (rc == -ENOENT) {
-		rc = 0;
+		GOTO(out, rc);
 	}
-	LASSERT(osd_invariant(obj));
+
+	/* The case NOT allow to trigger OI scrub automatically. */
+	if (osd->od_auto_scrub_interval == AS_NEVER)
+		GOTO(out, rc);
+
+	/* It is me to trigger the OI scrub. */
+	rc1 = osd_scrub_start(env, osd, SS_CLEAR_DRYRUN |
+			      SS_CLEAR_FAILOUT | SS_AUTO_FULL);
+	LCONSOLE_WARN("%s: trigger OI scrub by RPC for the "DFID": rc = %d\n",
+		      osd_name(osd), PFID(fid), rc1);
+	if (!rc) {
+		LASSERT(remote);
+
+		lu_object_set_agent_entry(l);
+		if (!rc1)
+			osd_oii_insert(env, osd, fid, oid, false);
+	} else {
+		if (!rc1)
+			rc = -EINPROGRESS;
+		else
+			rc = -EREMCHG;
+	}
+
+	GOTO(out, rc);
+
 out:
 	RETURN(rc);
 }
@@ -567,11 +689,16 @@ out:
 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 {
 	struct osd_object *obj = osd_obj(l);
+	struct lu_object_header *h = obj->oo_header;
 
 	LASSERT(osd_invariant(obj));
 
 	dt_object_fini(&obj->oo_dt);
 	OBD_SLAB_FREE_PTR(obj, osd_object_kmem);
+	if (unlikely(h)) {
+		lu_object_header_fini(h);
+		OBD_FREE_PTR(h);
+	}
 }
 
 static int
@@ -633,6 +760,8 @@ static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
 	oh = container_of0(th, struct osd_thandle, ot_super);
 	LASSERT(oh->ot_tx != NULL);
 
+	dmu_tx_mark_netfree(oh->ot_tx);
+
 	/* declare that we'll remove object from fid-dnode mapping */
 	zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0, &dn);
 	osd_tx_hold_zap(oh->ot_tx, zapid, dn, FALSE, NULL);
@@ -661,6 +790,11 @@ static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
 		osd_tx_hold_zap(oh->ot_tx, osd->od_unlinked->dn_object,
 				osd->od_unlinked, TRUE, NULL);
 
+	/* remove agent entry (if have) from remote parent */
+	if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
+		osd_tx_hold_zap(oh->ot_tx, osd->od_remote_parent_dir,
+				NULL, FALSE, NULL);
+
 	/* will help to find FID->ino when this object is being
 	 * added to PENDING/ */
 	osd_idc_find_and_init(env, osd, obj);
@@ -696,13 +830,6 @@ static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
 	/* remove obj ref from index dir (it depends) */
 	zapid = osd_get_name_n_idx(env, osd, fid, buf,
 				   sizeof(info->oti_str), &zdn);
-	rc = osd_zap_remove(osd, zapid, zdn, buf, oh->ot_tx);
-	if (rc) {
-		CERROR("%s: zap_remove(%s) failed: rc = %d\n",
-		       osd->od_svname, buf, rc);
-		GOTO(out, rc);
-	}
-
 	rc = osd_xattrs_destroy(env, obj, oh);
 	if (rc) {
 		CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
@@ -710,6 +837,12 @@ static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
 		GOTO(out, rc);
 	}
 
+	if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
+		rc = osd_delete_from_remote_parent(env, osd, obj, oh, true);
+		if (rc)
+			GOTO(out, rc);
+	}
+
 	oid = obj->oo_dn->dn_object;
 	if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
 		/* this may happen if the destroy wasn't declared
@@ -741,6 +874,17 @@ static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
 			       osd->od_svname, buf, oid, rc);
 	}
 
+	/* Remove the OI mapping after the destroy to handle the race with
+	 * OI scrub that may insert missed OI mapping during the interval. */
+	rc = osd_zap_remove(osd, zapid, zdn, buf, oh->ot_tx);
+	if (unlikely(rc == -ENOENT))
+		rc = 0;
+	if (rc)
+		CERROR("%s: zap_remove(%s) failed: rc = %d\n",
+		       osd->od_svname, buf, rc);
+
+	GOTO(out, rc);
+
 out:
 	/* not needed in the cache anymore */
 	set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
@@ -838,8 +982,7 @@ static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
 	return rc;
 }
 
-static int osd_attr_get(const struct lu_env *env,
-			struct dt_object *dt,
+static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
 			struct lu_attr *attr)
 {
 	struct osd_object	*obj = osd_dt_obj(dt);
@@ -860,9 +1003,16 @@ static int osd_attr_get(const struct lu_env *env,
 
 	read_lock(&obj->oo_attr_lock);
 	*attr = obj->oo_attr;
-	if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
+	if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
+		attr->la_valid |= LA_FLAGS;
 		attr->la_flags |= LUSTRE_ORPHAN_FL;
+	}
 	read_unlock(&obj->oo_attr_lock);
+	if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ORPHAN_FL)
+		CDEBUG(D_INFO, "%s: set orphan flag on "DFID" (%llx/%x)\n",
+		       osd_obj2dev(obj)->od_svname,
+		       PFID(lu_object_fid(&dt->do_lu)),
+		       attr->la_valid, obj->oo_lma_flags);
 
 	/* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
 	 * from within sa_object_size() can block on a mutex, so
@@ -893,7 +1043,7 @@ static inline int qsd_transfer(const struct lu_env *env,
 			       struct qsd_instance *qsd,
 			       struct lquota_trans *trans, int qtype,
 			       __u64 orig_id, __u64 new_id, __u64 bspace,
-			       struct lquota_id_info *qi)
+			       struct lquota_id_info *qi, bool ignore_edquot)
 {
 	int	rc;
 
@@ -910,7 +1060,7 @@ static inline int qsd_transfer(const struct lu_env *env,
 	qi->lqi_id.qid_uid = new_id;
 	qi->lqi_space      = 1;
 	rc = qsd_op_begin(env, qsd, trans, qi, NULL);
-	if (rc == -EDQUOT || rc == -EINPROGRESS)
+	if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
 		rc = 0;
 	if (rc)
 		return rc;
@@ -932,7 +1082,7 @@ static inline int qsd_transfer(const struct lu_env *env,
 	qi->lqi_id.qid_uid = new_id;
 	qi->lqi_space      = bspace;
 	rc = qsd_op_begin(env, qsd, trans, qi, NULL);
-	if (rc == -EDQUOT || rc == -EINPROGRESS)
+	if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
 		rc = 0;
 	if (rc)
 		return rc;
@@ -1002,16 +1152,20 @@ static int osd_declare_attr_set(const struct lu_env *env,
 
 	if (attr && (attr->la_valid & (LA_UID | LA_GID | LA_PROJID))) {
 		sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
-		bspace = toqb(bspace * blksize);
+		bspace = toqb(bspace * 512);
+
+		CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u,"
+		       "the quota space is %lld (%u)\n", osd->od_svname,
+		       attr->la_uid, attr->la_gid, bspace, blksize);
 	}
 
 	if (attr && attr->la_valid & LA_UID) {
 		/* quota enforcement for user */
 		if (attr->la_uid != obj->oo_attr.la_uid) {
-			rc = qsd_transfer(env, osd->od_quota_slave,
+			rc = qsd_transfer(env, osd_def_qsd(osd),
 					  &oh->ot_quota_trans, USRQUOTA,
 					  obj->oo_attr.la_uid, attr->la_uid,
-					  bspace, &info->oti_qi);
+					  bspace, &info->oti_qi, true);
 			if (rc)
 				GOTO(out, rc);
 		}
@@ -1019,41 +1173,45 @@ static int osd_declare_attr_set(const struct lu_env *env,
 	if (attr && attr->la_valid & LA_GID) {
 		/* quota enforcement for group */
 		if (attr->la_gid != obj->oo_attr.la_gid) {
-			rc = qsd_transfer(env, osd->od_quota_slave,
+			rc = qsd_transfer(env, osd_def_qsd(osd),
 					  &oh->ot_quota_trans, GRPQUOTA,
 					  obj->oo_attr.la_gid, attr->la_gid,
-					  bspace, &info->oti_qi);
+					  bspace, &info->oti_qi,
+					  !(attr->la_flags &
+							LUSTRE_SET_SYNC_FL));
 			if (rc)
 				GOTO(out, rc);
 		}
 	}
 #ifdef ZFS_PROJINHERIT
 	if (attr && attr->la_valid & LA_PROJID) {
-		if (!osd->od_projectused_dn)
-			GOTO(out, rc = -EOPNOTSUPP);
-
-		/* Usually, if project quota is upgradable for the device,
-		 * then the upgrade will be done before or when mount the
-		 * device. So when we come here, this project should have
-		 * project ID attribute already (that is zero by default).
-		 * Otherwise, there was something wrong during the former
-		 * upgrade, let's return failure to report that.
-		 *
-		 * Please note that, different from other attributes, you
-		 * can NOT simply set the project ID attribute under such
-		 * case, because adding (NOT change) project ID attribute
-		 * needs to change the object's attribute layout to match
-		 * zfs backend quota accounting requirement. */
-		if (unlikely(!obj->oo_with_projid))
-			GOTO(out, rc = -ENXIO);
-
 		/* quota enforcement for project */
 		if (attr->la_projid != obj->oo_attr.la_projid) {
-			rc = qsd_transfer(env, osd->od_quota_slave,
+			if (!osd->od_projectused_dn)
+				GOTO(out, rc = -EOPNOTSUPP);
+
+			/* Usually, if project quota is upgradable for the
+			 * device, then the upgrade will be done before or when
+			 * mount the device. So when we come here, this project
+			 * should have project ID attribute already (that is
+			 * zero by default).  Otherwise, there was something
+			 * wrong during the former upgrade, let's return failure
+			 * to report that.
+			 *
+			 * Please note that, different from other attributes,
+			 * you can NOT simply set the project ID attribute under
+			 * such case, because adding (NOT change) project ID
+			 * attribute needs to change the object's attribute
+			 * layout to match zfs backend quota accounting
+			 * requirement. */
+			if (unlikely(!obj->oo_with_projid))
+				GOTO(out, rc = -ENXIO);
+
+			rc = qsd_transfer(env, osd_def_qsd(osd),
 					  &oh->ot_quota_trans, PRJQUOTA,
 					  obj->oo_attr.la_projid,
 					  attr->la_projid, bspace,
-					  &info->oti_qi);
+					  &info->oti_qi, true);
 			if (rc)
 				GOTO(out, rc);
 		}
@@ -1099,6 +1257,26 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 	   transaction group. */
 	LASSERT(oh->ot_tx->tx_txg != 0);
 
+	if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) && !osd->od_is_ost) {
+		struct zpl_direntry *zde = &info->oti_zde.lzd_reg;
+		char *buf = info->oti_str;
+		dnode_t *zdn = NULL;
+		uint64_t zapid;
+
+		zapid = osd_get_name_n_idx(env, osd, lu_object_fid(&dt->do_lu),
+					   buf, sizeof(info->oti_str), &zdn);
+		rc = osd_zap_lookup(osd, zapid, zdn, buf, 8,
+				    sizeof(*zde) / 8, zde);
+		if (!rc) {
+			zde->zde_dnode -= 1;
+			rc = -zap_update(osd->od_os, zapid, buf, 8,
+					 sizeof(*zde) / 8, zde, oh->ot_tx);
+		}
+		if (rc > 0)
+			rc = 0;
+		GOTO(out, rc);
+	}
+
 	/* Only allow set size for regular file */
 	if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 		valid &= ~(LA_SIZE | LA_BLOCKS);
@@ -1118,16 +1296,22 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 	if (valid & LA_FLAGS) {
 		struct lustre_mdt_attrs *lma;
 		struct lu_buf buf;
+		int size = 0;
 
 		if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
 			LASSERT(!obj->oo_pfid_in_lma);
-			CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
+			BUILD_BUG_ON(sizeof(info->oti_buf) < sizeof(*lma));
 			lma = (struct lustre_mdt_attrs *)&info->oti_buf;
 			buf.lb_buf = lma;
 			buf.lb_len = sizeof(info->oti_buf);
-			rc = osd_xattr_get(env, &obj->oo_dt, &buf,
-					   XATTR_NAME_LMA);
-			if (rc > 0) {
+
+			/* Please do NOT call osd_xattr_get() directly, that
+			 * will cause recursive down_read() on oo_guard. */
+			rc = osd_xattr_get_internal(env, obj, &buf,
+						    XATTR_NAME_LMA, &size);
+			if (!rc && unlikely(size < sizeof(*lma))) {
+				rc = -EINVAL;
+			} else if (!rc) {
 				lma->lma_incompat =
 					le32_to_cpu(lma->lma_incompat);
 				lma->lma_incompat |=
@@ -1154,18 +1338,15 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 
 	if (valid & LA_PROJID) {
 #ifdef ZFS_PROJINHERIT
-		/* osd_declare_attr_set() must be called firstly.
-		 * If osd::od_projectused_dn is not set, then we
-		 * can not arrive at here. */
-		LASSERT(osd->od_projectused_dn);
-		LASSERT(obj->oo_with_projid);
+		if (osd->od_projectused_dn) {
+			LASSERT(obj->oo_with_projid);
 
-		osa->projid = obj->oo_attr.la_projid = la->la_projid;
-		SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PROJID(osd), NULL,
-				 &osa->projid, 8);
-#else
-		valid &= ~LA_PROJID;
+			osa->projid = obj->oo_attr.la_projid = la->la_projid;
+			SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PROJID(osd), NULL,
+					 &osa->projid, 8);
+		} else
 #endif
+			valid &= ~LA_PROJID;
 	}
 
 	if (valid & LA_ATIME) {
@@ -1212,7 +1393,7 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 		 * copy */
 		obj->oo_attr.la_flags = attrs_zfs2fs(osa->flags);
 #ifdef ZFS_PROJINHERIT
-		if (obj->oo_with_projid)
+		if (obj->oo_with_projid && osd->od_projectused_dn)
 			osa->flags |= ZFS_PROJID;
 #endif
 		SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL,
@@ -1302,6 +1483,7 @@ static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
 	switch (dof->dof_type) {
 		case DFT_DIR:
 			dt->do_index_ops = &osd_dir_ops;
+			/* fallthrough */
 		case DFT_INDEX:
 			/* for zap create */
 			dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, FALSE, NULL);
@@ -1337,13 +1519,13 @@ int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
 		    struct lu_attr *la, uint64_t parent,
 		    nvlist_t *xattr)
 {
-	sa_bulk_attr_t	*bulk = osd_oti_get(env)->oti_attr_bulk;
-	struct osa_attr	*osa = &osd_oti_get(env)->oti_osa;
-	uint64_t	 gen;
-	uint64_t	 crtime[2];
-	timestruc_t	 now;
-	int		 cnt;
-	int		 rc;
+	sa_bulk_attr_t *bulk = osd_oti_get(env)->oti_attr_bulk;
+	struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
+	uint64_t gen;
+	uint64_t crtime[2];
+	inode_timespec_t now;
+	int cnt;
+	int rc;
 	char *dxattr = NULL;
 	size_t sa_size;
 
@@ -1434,8 +1616,8 @@ int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
 	return rc;
 }
 
-static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
-			      uint64_t oid, dnode_t **dnp)
+int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
+		       uint64_t oid, dnode_t **dnp)
 {
 	dmu_tx_hold_t *txh;
 	int rc = 0;
@@ -1461,11 +1643,12 @@ static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
 		LASSERT(db);
 		LASSERT(dn->dn_handle);
 		DB_DNODE_ENTER(db);
-		if (refcount_add(&db->db_holds, osd_obj_tag) == 1) {
-			refcount_add(&dn->dn_holds, tag);
+		if (zfs_refcount_add(&db->db_holds, osd_obj_tag) == 1) {
+			zfs_refcount_add(&dn->dn_holds, osd_obj_tag);
 			atomic_inc_32(&dn->dn_dbufs_count);
 		}
 		*dnp = dn;
+		DB_DNODE_EXIT(db);
 		dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH);
 		break;
 	}
@@ -1633,6 +1816,14 @@ static dnode_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
 			       osd->od_svname, rc);
 			return ERR_PTR(rc);
 		}
+	} else if ((fid_is_llog(fid))) {
+		rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
+					       LLOG_MIN_CHUNK_SIZE, 0, oh->ot_tx);
+		if (unlikely(rc)) {
+			CERROR("%s: can't change blocksize: %d\n",
+			       osd->od_svname, rc);
+			return ERR_PTR(rc);
+		}
 	}
 
 	return dn;
@@ -1720,6 +1911,7 @@ static int osd_create(const struct lu_env *env, struct dt_object *dt,
 	dnode_t *dn = NULL, *zdn = NULL;
 	uint64_t		 zapid, parent = 0;
 	int			 rc;
+	__u32 compat = 0;
 
 	ENTRY;
 
@@ -1772,9 +1964,18 @@ static int osd_create(const struct lu_env *env, struct dt_object *dt,
 
 	zapid = osd_get_name_n_idx(env, osd, fid, buf,
 				   sizeof(info->oti_str), &zdn);
+	if (CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY) ||
+	    (osd->od_is_ost && OBD_FAIL_CHECK(OBD_FAIL_OSD_COMPAT_NO_ENTRY)))
+		goto skip_add;
+
+	if (osd->od_is_ost && OBD_FAIL_CHECK(OBD_FAIL_OSD_COMPAT_INVALID_ENTRY))
+		zde->zde_dnode++;
+
 	rc = osd_zap_add(osd, zapid, zdn, buf, 8, 1, zde, oh->ot_tx);
 	if (rc)
 		GOTO(out, rc);
+
+skip_add:
 	obj->oo_dn = dn;
 	/* Now add in all of the "SA" attributes */
 	rc = osd_sa_handle_get(obj);
@@ -1786,7 +1987,9 @@ static int osd_create(const struct lu_env *env, struct dt_object *dt,
 		GOTO(out, rc);
 
 	/* initialize LMA */
-	lustre_lma_init(lma, fid, 0, 0);
+	if (fid_is_idif(fid) || (fid_is_norm(fid) && osd->od_is_ost))
+		compat |= LMAC_FID_ON_OST;
+	lustre_lma_init(lma, fid, compat, 0);
 	lustre_lma_swab(lma);
 	rc = -nvlist_add_byte_array(obj->oo_sa_xattr, XATTR_NAME_LMA,
 				    (uchar_t *)lma, sizeof(*lma));
@@ -1823,6 +2026,7 @@ out:
 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
 			       struct thandle *th)
 {
+	osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
 	return osd_declare_attr_set(env, dt, NULL, th);
 }
 
@@ -1863,6 +2067,7 @@ out:
 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
 			       struct thandle *handle)
 {
+	osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
 	return osd_declare_attr_set(env, dt, NULL, handle);
 }
 
@@ -1906,14 +2111,25 @@ static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
 			   __u64 start, __u64 end)
 {
 	struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
+	struct dmu_buf_impl *db = osd_dt_obj(dt)->oo_dn->dn_dbuf;
+	uint64_t txg = 0;
 	ENTRY;
 
-	/* XXX: no other option than syncing the whole filesystem until we
-	 * support ZIL.  If the object tracked the txg that it was last
-	 * modified in, it could pass that txg here instead of "0".  Maybe
-	 * the changes are already committed, so no wait is needed at all? */
-	if (!osd->od_dt_dev.dd_rdonly)
-		txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
+	if (osd->od_dt_dev.dd_rdonly)
+		RETURN(0);
+
+	mutex_enter(&db->db_mtx);
+	if (db->db_last_dirty)
+		txg = db->db_last_dirty->dr_txg;
+	mutex_exit(&db->db_mtx);
+
+	if (txg) {
+		/* the object is dirty or being synced */
+		if (osd_object_sync_delay_us < 0)
+			txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
+		else
+			udelay(osd_object_sync_delay_us);
+	}
 
 	RETURN(0);
 }
@@ -1973,3 +2189,7 @@ static struct dt_object_operations osd_obj_otable_it_ops = {
 	.do_attr_get		= osd_otable_it_attr_get,
 	.do_index_try		= osd_index_try,
 };
+
+module_param(osd_object_sync_delay_us, int, 0644);
+MODULE_PARM_DESC(osd_object_sync_delay_us,
+		 "If zero or larger delay N usec instead of doing object sync");