X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_object.c;h=fa9ec89434911e8c6098847f168a66d21256cc5e;hp=6bf8185af44214a83dd448e0e890e3f29b2b425f;hb=3cbdb896b6b380ff7c843f9a7104456e7e80b347;hpb=745c19c70319915a55b71b81b4e89d68e3a4e272

diff --git a/lustre/osd-zfs/osd_object.c b/lustre/osd-zfs/osd_object.c
index 6bf8185..fa9ec894 100644
--- a/lustre/osd-zfs/osd_object.c
+++ b/lustre/osd-zfs/osd_object.c
@@ -26,10 +26,8 @@
 /*
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
- */
-/*
- * Copyright (c) 2012, 2013, Intel Corporation.
- * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -138,7 +136,7 @@ void osd_object_sa_dirty_rele(struct osd_thandle *oh)
 	down(&oh->ot_sa_lock);
 	while (!list_empty(&oh->ot_sa_list)) {
 		obj = list_entry(oh->ot_sa_list.next,
-				     struct osd_object, oo_sa_linkage);
+				 struct osd_object, oo_sa_linkage);
 		sa_spill_rele(obj->oo_sa_hdl);
 		write_lock(&obj->oo_attr_lock);
 		list_del_init(&obj->oo_sa_linkage);
@@ -233,6 +231,28 @@ int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
 	la->la_flags = attrs_zfs2fs(osa->flags);
 	la->la_size = osa->size;
 
+	/* Try to get extra flag from LMA. Right now, only LMAI_ORPHAN
+	 * flags is stored in LMA, and it is only for orphan directory */
+	if (S_ISDIR(la->la_mode) && dt_object_exists(&obj->oo_dt)) {
+		struct osd_thread_info *info = osd_oti_get(env);
+		struct lustre_mdt_attrs *lma;
+		struct lu_buf buf;
+
+		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) {
+			rc = 0;
+			lma->lma_incompat = le32_to_cpu(lma->lma_incompat);
+			obj->oo_lma_flags =
+				lma_to_lustre_flags(lma->lma_incompat);
+
+		} else if (rc == -ENODATA) {
+			rc = 0;
+		}
+	}
+
 	if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode)) {
 		rc = -sa_lookup(sa_hdl, SA_ZPL_RDEV(o), &osa->rdev, 8);
 		if (rc)
@@ -293,9 +313,11 @@ struct lu_object *osd_object_alloc(const struct lu_env *env,
 		mo->oo_dt.do_ops = &osd_obj_ops;
 		l->lo_ops = &osd_lu_obj_ops;
 		INIT_LIST_HEAD(&mo->oo_sa_linkage);
+		INIT_LIST_HEAD(&mo->oo_unlinked_linkage);
 		init_rwsem(&mo->oo_sem);
-		sema_init(&mo->oo_guard, 1);
+		init_rwsem(&mo->oo_guard);
 		rwlock_init(&mo->oo_attr_lock);
+		mo->oo_destroy = OSD_DESTROY_NONE;
 		return l;
 	} else {
 		return NULL;
@@ -308,7 +330,7 @@ struct lu_object *osd_object_alloc(const struct lu_env *env,
 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);
+	const struct lu_fid	*fid = lu_object_fid(&obj->oo_dt.do_lu);
 	int			 rc = 0;
 	ENTRY;
 
@@ -354,7 +376,7 @@ static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 	buf.lb_buf = lma;
 	buf.lb_len = sizeof(info->oti_buf);
 
-	rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA, BYPASS_CAPA);
+	rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 	if (rc > 0) {
 		rc = 0;
 		lustre_lma_swab(lma);
@@ -395,10 +417,18 @@ static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 		RETURN(0);
 	}
 
+	if (conf != NULL && conf->loc_flags & LOC_F_NEW)
+		GOTO(out, rc = 0);
+
 	rc = osd_fid_lookup(env, osd, lu_object_fid(l), &oid);
 	if (rc == 0) {
 		LASSERT(obj->oo_db == NULL);
 		rc = __osd_obj2dbuf(env, osd->od_os, oid, &obj->oo_db);
+		/* EEXIST will be returned if object is being deleted in ZFS */
+		if (rc == -EEXIST) {
+			rc = 0;
+			GOTO(out, rc);
+		}
 		if (rc != 0) {
 			CERROR("%s: lookup "DFID"/"LPX64" failed: rc = %d\n",
 			       osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
@@ -434,53 +464,45 @@ static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 	OBD_SLAB_FREE_PTR(obj, osd_object_kmem);
 }
 
-static void __osd_declare_object_destroy(const struct lu_env *env,
-					 struct osd_object *obj,
-					 struct osd_thandle *oh)
+static int
+osd_object_unlinked_add(struct osd_object *obj, struct osd_thandle *oh)
 {
-	struct osd_device	*osd = osd_obj2dev(obj);
-	dmu_buf_t		*db = obj->oo_db;
-	zap_attribute_t		*za = &osd_oti_get(env)->oti_za;
-	uint64_t		 oid = db->db_object, xid;
-	dmu_tx_t		*tx = oh->ot_tx;
-	zap_cursor_t		*zc;
-	int			 rc = 0;
+	int rc = -EBUSY;
 
-	dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
+	LASSERT(obj->oo_destroy == OSD_DESTROY_ASYNC);
 
-	/* zap holding xattrs */
-	if (obj->oo_xattr != ZFS_NO_OBJECT) {
-		oid = obj->oo_xattr;
-
-		dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END);
-
-		rc = osd_zap_cursor_init(&zc, osd->od_os, oid, 0);
-		if (rc)
-			goto out;
+	/* the object is supposed to be exclusively locked by
+	 * the caller (osd_object_destroy()), while the transaction
+	 * (oh) is per-thread and not shared */
+	if (likely(list_empty(&obj->oo_unlinked_linkage))) {
+		list_add(&obj->oo_unlinked_linkage, &oh->ot_unlinked_list);
+		rc = 0;
+	}
 
-		while ((rc = -zap_cursor_retrieve(zc, za)) == 0) {
-			BUG_ON(za->za_integer_length != sizeof(uint64_t));
-			BUG_ON(za->za_num_integers != 1);
+	return rc;
+}
 
-			rc = -zap_lookup(osd->od_os, obj->oo_xattr, za->za_name,
-					 sizeof(uint64_t), 1, &xid);
-			if (rc) {
-				CERROR("%s: xattr lookup failed: rc = %d\n",
-				       osd->od_svname, rc);
-				goto out_err;
-			}
-			dmu_tx_hold_free(tx, xid, 0, DMU_OBJECT_END);
+/* Default to max data size covered by a level-1 indirect block */
+static unsigned long osd_sync_destroy_max_size =
+	1UL << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT + SPA_MAXBLOCKSHIFT);
+module_param(osd_sync_destroy_max_size, ulong, 0444);
+MODULE_PARM_DESC(osd_sync_destroy_max_size, "Maximum object size to use synchronous destroy.");
 
-			zap_cursor_advance(zc);
-		}
-		if (rc == -ENOENT)
-			rc = 0;
-out_err:
-		osd_zap_cursor_fini(zc);
+static inline void
+osd_object_set_destroy_type(struct osd_object *obj)
+{
+	/*
+	 * Lock-less OST_WRITE can race with OST_DESTROY, so set destroy type
+	 * only once and use it consistently thereafter.
+	 */
+	down_write(&obj->oo_guard);
+	if (obj->oo_destroy == OSD_DESTROY_NONE) {
+		if (obj->oo_attr.la_size <= osd_sync_destroy_max_size)
+			obj->oo_destroy = OSD_DESTROY_SYNC;
+		else /* Larger objects are destroyed asynchronously */
+			obj->oo_destroy = OSD_DESTROY_ASYNC;
 	}
-out:
-	if (rc && tx->tx_err == 0)
-		tx->tx_err = -rc;
+	up_write(&obj->oo_guard);
 }
 
 static int osd_declare_object_destroy(const struct lu_env *env,
@@ -492,8 +514,8 @@ static int osd_declare_object_destroy(const struct lu_env *env,
 	struct osd_object	*obj = osd_dt_obj(dt);
 	struct osd_device	*osd = osd_obj2dev(obj);
 	struct osd_thandle	*oh;
-	uint64_t		 zapid;
 	int			 rc;
+	uint64_t		 zapid;
 	ENTRY;
 
 	LASSERT(th != NULL);
@@ -502,19 +524,18 @@ static int osd_declare_object_destroy(const struct lu_env *env,
 	oh = container_of0(th, struct osd_thandle, ot_super);
 	LASSERT(oh->ot_tx != NULL);
 
-	/* declare that we'll destroy the object */
-	__osd_declare_object_destroy(env, obj, oh);
-
 	/* declare that we'll remove object from fid-dnode mapping */
 	zapid = osd_get_name_n_idx(env, osd, fid, buf);
 	dmu_tx_hold_bonus(oh->ot_tx, zapid);
-	dmu_tx_hold_zap(oh->ot_tx, zapid, 0, buf);
+	dmu_tx_hold_zap(oh->ot_tx, zapid, FALSE, buf);
+
+	osd_declare_xattrs_destroy(env, obj, oh);
 
 	/* declare that we'll remove object from inode accounting ZAPs */
 	dmu_tx_hold_bonus(oh->ot_tx, osd->od_iusr_oid);
-	dmu_tx_hold_zap(oh->ot_tx, osd->od_iusr_oid, 0, buf);
+	dmu_tx_hold_zap(oh->ot_tx, osd->od_iusr_oid, FALSE, buf);
 	dmu_tx_hold_bonus(oh->ot_tx, osd->od_igrp_oid);
-	dmu_tx_hold_zap(oh->ot_tx, osd->od_igrp_oid, 0, buf);
+	dmu_tx_hold_zap(oh->ot_tx, osd->od_igrp_oid, FALSE, buf);
 
 	/* one less inode */
 	rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
@@ -525,68 +546,21 @@ static int osd_declare_object_destroy(const struct lu_env *env,
 	/* data to be truncated */
 	rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 			       obj->oo_attr.la_gid, 0, oh, true, NULL, false);
-	RETURN(rc);
-}
-
-/*
- * Delete a DMU object
- *
- * The transaction passed to this routine must have
- * dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END) called
- * and then assigned to a transaction group.
- *
- * This will release db and set it to NULL to prevent further dbuf releases.
- */
-static int __osd_object_destroy(const struct lu_env *env,
-				struct osd_object *obj,
-				dmu_tx_t *tx, void *tag)
-{
-	struct osd_device	*osd = osd_obj2dev(obj);
-	uint64_t		 xid;
-	zap_attribute_t		*za = &osd_oti_get(env)->oti_za;
-	zap_cursor_t		*zc;
-	int			 rc;
-
-	/* Assert that the transaction has been assigned to a
-	   transaction group. */
-	LASSERT(tx->tx_txg != 0);
+	if (rc)
+		RETURN(rc);
 
-	/* zap holding xattrs */
-	if (obj->oo_xattr != ZFS_NO_OBJECT) {
-		rc = osd_zap_cursor_init(&zc, osd->od_os, obj->oo_xattr, 0);
-		if (rc)
-			return rc;
-		while ((rc = -zap_cursor_retrieve(zc, za)) == 0) {
-			BUG_ON(za->za_integer_length != sizeof(uint64_t));
-			BUG_ON(za->za_num_integers != 1);
-
-			rc = -zap_lookup(osd->od_os, obj->oo_xattr, za->za_name,
-					 sizeof(uint64_t), 1, &xid);
-			if (rc) {
-				CERROR("%s: lookup xattr %s failed: rc = %d\n",
-				       osd->od_svname, za->za_name, rc);
-				continue;
-			}
-			rc = -dmu_object_free(osd->od_os, xid, tx);
-			if (rc)
-				CERROR("%s: fetch xattr %s failed: rc = %d\n",
-				       osd->od_svname, za->za_name, rc);
-			zap_cursor_advance(zc);
-		}
-		osd_zap_cursor_fini(zc);
+	osd_object_set_destroy_type(obj);
+	if (obj->oo_destroy == OSD_DESTROY_SYNC)
+		dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object,
+				 0, DMU_OBJECT_END);
+	else
+		dmu_tx_hold_zap(oh->ot_tx, osd->od_unlinkedid, TRUE, NULL);
 
-		rc = -dmu_object_free(osd->od_os, obj->oo_xattr, tx);
-		if (rc)
-			CERROR("%s: freeing xattr failed: rc = %d\n",
-			       osd->od_svname, rc);
-	}
-
-	return -dmu_object_free(osd->od_os, obj->oo_db->db_object, tx);
+	RETURN(0);
 }
 
 static int osd_object_destroy(const struct lu_env *env,
-			      struct dt_object *dt,
-			      struct thandle *th)
+			      struct dt_object *dt, struct thandle *th)
 {
 	char			*buf = osd_oti_get(env)->oti_str;
 	struct osd_object	*obj = osd_dt_obj(dt);
@@ -594,24 +568,33 @@ static int osd_object_destroy(const struct lu_env *env,
 	const struct lu_fid	*fid = lu_object_fid(&dt->do_lu);
 	struct osd_thandle	*oh;
 	int			 rc;
-	uint64_t		 zapid;
+	uint64_t		 oid, zapid;
 	ENTRY;
 
+	down_write(&obj->oo_guard);
+
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = -ENOENT);
+
 	LASSERT(obj->oo_db != NULL);
-	LASSERT(dt_object_exists(dt));
-	LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
 
 	oh = container_of0(th, struct osd_thandle, ot_super);
 	LASSERT(oh != NULL);
 	LASSERT(oh->ot_tx != NULL);
 
-	zapid = osd_get_name_n_idx(env, osd, fid, buf);
-
 	/* remove obj ref from index dir (it depends) */
+	zapid = osd_get_name_n_idx(env, osd, fid, buf);
 	rc = -zap_remove(osd->od_os, zapid, buf, oh->ot_tx);
 	if (rc) {
-		CERROR("%s: zap_remove() failed: rc = %d\n",
-		       osd->od_svname, 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",
+		       osd->od_svname, buf, rc);
 		GOTO(out, rc);
 	}
 
@@ -619,30 +602,52 @@ static int osd_object_destroy(const struct lu_env *env,
 	 * operation if something goes wrong while updating accounting, but we
 	 * still log an error message to notify the administrator */
 	rc = -zap_increment_int(osd->od_os, osd->od_iusr_oid,
-			obj->oo_attr.la_uid, -1, oh->ot_tx);
+				obj->oo_attr.la_uid, -1, oh->ot_tx);
 	if (rc)
 		CERROR("%s: failed to remove "DFID" from accounting ZAP for usr"
-			" %d: rc = %d\n", osd->od_svname, PFID(fid),
-			obj->oo_attr.la_uid, rc);
+		       " %d: rc = %d\n", osd->od_svname, PFID(fid),
+		       obj->oo_attr.la_uid, rc);
 	rc = -zap_increment_int(osd->od_os, osd->od_igrp_oid,
 				obj->oo_attr.la_gid, -1, oh->ot_tx);
 	if (rc)
 		CERROR("%s: failed to remove "DFID" from accounting ZAP for grp"
-			" %d: rc = %d\n", osd->od_svname, PFID(fid),
-			obj->oo_attr.la_gid, rc);
+		       " %d: rc = %d\n", osd->od_svname, PFID(fid),
+		       obj->oo_attr.la_gid, rc);
+
+	oid = obj->oo_db->db_object;
+	if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
+		/* this may happen if the destroy wasn't declared
+		 * e.g. when the object is created and then destroyed
+		 * in the same transaction - we don't need additional
+		 * space for destroy specifically */
+		LASSERT(obj->oo_attr.la_size <= osd_sync_destroy_max_size);
+		rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
+		if (rc)
+			CERROR("%s: failed to free %s "LPU64": rc = %d\n",
+			       osd->od_svname, buf, oid, rc);
+	} else if (obj->oo_destroy == OSD_DESTROY_SYNC) {
+		rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
+		if (rc)
+			CERROR("%s: failed to free %s "LPU64": rc = %d\n",
+			       osd->od_svname, buf, oid, rc);
+	} else { /* asynchronous destroy */
+		rc = osd_object_unlinked_add(obj, oh);
+		if (rc)
+			GOTO(out, rc);
 
-	/* kill object */
-	rc = __osd_object_destroy(env, obj, oh->ot_tx, osd_obj_tag);
-	if (rc) {
-		CERROR("%s: __osd_object_destroy() failed: rc = %d\n",
-		       osd->od_svname, rc);
-		GOTO(out, rc);
+		rc = -zap_add_int(osd->od_os, osd->od_unlinkedid,
+				  oid, oh->ot_tx);
+		if (rc)
+			CERROR("%s: zap_add_int() failed %s "LPU64": rc = %d\n",
+			       osd->od_svname, buf, oid, rc);
 	}
 
 out:
 	/* not needed in the cache anymore */
 	set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
-
+	if (rc == 0)
+		obj->oo_destroyed = 1;
+	up_write(&obj->oo_guard);
 	RETURN (0);
 }
 
@@ -688,7 +693,7 @@ static void osd_object_read_lock(const struct lu_env *env,
 
 	LASSERT(osd_invariant(obj));
 
-	down_read(&obj->oo_sem);
+	down_read_nested(&obj->oo_sem, role);
 }
 
 static void osd_object_write_lock(const struct lu_env *env,
@@ -698,7 +703,7 @@ static void osd_object_write_lock(const struct lu_env *env,
 
 	LASSERT(osd_invariant(obj));
 
-	down_write(&obj->oo_sem);
+	down_write_nested(&obj->oo_sem, role);
 }
 
 static void osd_object_read_unlock(const struct lu_env *env,
@@ -736,19 +741,25 @@ static int osd_object_write_locked(const struct lu_env *env,
 
 static int osd_attr_get(const struct lu_env *env,
 			struct dt_object *dt,
-			struct lu_attr *attr,
-			struct lustre_capa *capa)
+			struct lu_attr *attr)
 {
 	struct osd_object	*obj = osd_dt_obj(dt);
 	uint64_t		 blocks;
 	uint32_t		 blksize;
+	int			 rc = 0;
+
+	down_read(&obj->oo_guard);
+
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = -ENOENT);
 
-	LASSERT(dt_object_exists(dt));
 	LASSERT(osd_invariant(obj));
 	LASSERT(obj->oo_db);
 
 	read_lock(&obj->oo_attr_lock);
 	*attr = obj->oo_attr;
+	if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
+		attr->la_flags |= LUSTRE_ORPHAN_FL;
 	read_unlock(&obj->oo_attr_lock);
 
 	/* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
@@ -761,13 +772,16 @@ static int osd_attr_get(const struct lu_env *env,
 		attr->la_size = 512 * blocks;
 	/* Block size may be not set; suggest maximal I/O transfers. */
 	if (blksize == 0)
-		blksize = 1ULL << SPA_MAXBLOCKSHIFT;
+		blksize = osd_spa_maxblocksize(
+			dmu_objset_spa(osd_obj2dev(obj)->od_os));
 
 	attr->la_blksize = blksize;
 	attr->la_blocks = blocks;
 	attr->la_valid |= LA_BLOCKS | LA_BLKSIZE;
 
-	return 0;
+out:
+	up_read(&obj->oo_guard);
+	return rc;
 }
 
 /* Simple wrapper on top of qsd API which implement quota transfer for osd
@@ -843,28 +857,31 @@ static int osd_declare_attr_set(const struct lu_env *env,
 	struct osd_thandle	*oh;
 	uint64_t		 bspace;
 	uint32_t		 blksize;
-	int			 rc;
+	int			 rc = 0;
 	ENTRY;
 
-	if (!dt_object_exists(dt)) {
-		/* XXX: sanity check that object creation is declared */
-		RETURN(0);
-	}
 
 	LASSERT(handle != NULL);
 	LASSERT(osd_invariant(obj));
 
 	oh = container_of0(handle, struct osd_thandle, ot_super);
 
+	down_read(&obj->oo_guard);
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = 0);
+
 	LASSERT(obj->oo_sa_hdl != NULL);
 	LASSERT(oh->ot_tx != NULL);
 	dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
 	if (oh->ot_tx->tx_err != 0)
-		RETURN(-oh->ot_tx->tx_err);
+		GOTO(out, rc = -oh->ot_tx->tx_err);
 
 	sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
 	bspace = toqb(bspace * blksize);
 
+	__osd_xattr_declare_set(env, obj, sizeof(struct lustre_mdt_attrs),
+				XATTR_NAME_LMA, oh);
+
 	if (attr && attr->la_valid & LA_UID) {
 		/* account for user inode tracking ZAP update */
 		dmu_tx_hold_bonus(oh->ot_tx, osd->od_iusr_oid);
@@ -877,7 +894,7 @@ static int osd_declare_attr_set(const struct lu_env *env,
 					  obj->oo_attr.la_uid, attr->la_uid,
 					  bspace, &info->oti_qi);
 			if (rc)
-				RETURN(rc);
+				GOTO(out, rc);
 		}
 	}
 	if (attr && attr->la_valid & LA_GID) {
@@ -892,11 +909,13 @@ static int osd_declare_attr_set(const struct lu_env *env,
 					  obj->oo_attr.la_gid, attr->la_gid,
 					  bspace, &info->oti_qi);
 			if (rc)
-				RETURN(rc);
+				GOTO(out, rc);
 		}
 	}
 
-	RETURN(0);
+out:
+	up_read(&obj->oo_guard);
+	RETURN(rc);
 }
 
 /*
@@ -907,21 +926,25 @@ static int osd_declare_attr_set(const struct lu_env *env,
  * to a transaction group.
  */
 static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
-			const struct lu_attr *la, struct thandle *handle,
-			struct lustre_capa *capa)
+			const struct lu_attr *la, struct thandle *handle)
 {
+	struct osd_thread_info	*info = osd_oti_get(env);
 	struct osd_object	*obj = osd_dt_obj(dt);
 	struct osd_device	*osd = osd_obj2dev(obj);
 	struct osd_thandle	*oh;
-	struct osa_attr		*osa = &osd_oti_get(env)->oti_osa;
+	struct osa_attr		*osa = &info->oti_osa;
 	sa_bulk_attr_t		*bulk;
 	__u64			 valid = la->la_valid;
 	int			 cnt;
 	int			 rc = 0;
 
 	ENTRY;
+
+	down_read(&obj->oo_guard);
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = -ENOENT);
+
 	LASSERT(handle != NULL);
-	LASSERT(dt_object_exists(dt));
 	LASSERT(osd_invariant(obj));
 	LASSERT(obj->oo_sa_hdl);
 
@@ -934,12 +957,54 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 	if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 		valid &= ~(LA_SIZE | LA_BLOCKS);
 
+	if (valid & LA_CTIME && la->la_ctime == obj->oo_attr.la_ctime)
+		valid &= ~LA_CTIME;
+
+	if (valid & LA_MTIME && la->la_mtime == obj->oo_attr.la_mtime)
+		valid &= ~LA_MTIME;
+
+	if (valid & LA_ATIME && la->la_atime == obj->oo_attr.la_atime)
+		valid &= ~LA_ATIME;
+
 	if (valid == 0)
-		RETURN(0);
+		GOTO(out, rc = 0);
+
+	if (valid & LA_FLAGS) {
+		struct lustre_mdt_attrs *lma;
+		struct lu_buf buf;
+
+		if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
+			CLASSERT(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) {
+				lma->lma_incompat =
+					le32_to_cpu(lma->lma_incompat);
+				lma->lma_incompat |=
+					lustre_to_lma_flags(la->la_flags);
+				lma->lma_incompat =
+					cpu_to_le32(lma->lma_incompat);
+				buf.lb_buf = lma;
+				buf.lb_len = sizeof(*lma);
+				rc = osd_xattr_set_internal(env, obj, &buf,
+							    XATTR_NAME_LMA,
+							    LU_XATTR_REPLACE,
+							    oh);
+			}
+			if (rc < 0) {
+				CWARN("%s: failed to set LMA flags: rc = %d\n",
+				       osd->od_svname, rc);
+				RETURN(rc);
+			}
+		}
+	}
 
 	OBD_ALLOC(bulk, sizeof(sa_bulk_attr_t) * 10);
 	if (bulk == NULL)
-		RETURN(-ENOMEM);
+		GOTO(out, rc = -ENOMEM);
 
 	/* do both accounting updates outside oo_attr_lock below */
 	if ((valid & LA_UID) && (la->la_uid != obj->oo_attr.la_uid)) {
@@ -1037,6 +1102,8 @@ static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 	rc = osd_object_sa_bulk_update(obj, bulk, cnt, oh);
 
 	OBD_FREE(bulk, sizeof(sa_bulk_attr_t) * 10);
+out:
+	up_read(&obj->oo_guard);
 	RETURN(rc);
 }
 
@@ -1052,7 +1119,6 @@ static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
 {
 	LASSERT(ah);
 
-	memset(ah, 0, sizeof(*ah));
 	ah->dah_parent = parent;
 	ah->dah_mode = child_mode;
 }
@@ -1092,6 +1158,7 @@ static int osd_declare_object_create(const struct lu_env *env,
 
 	switch (dof->dof_type) {
 		case DFT_DIR:
+			dt->do_index_ops = &osd_dir_ops;
 		case DFT_INDEX:
 			/* for zap create */
 			dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, 1, NULL);
@@ -1206,19 +1273,28 @@ out:
  * dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT) called and then assigned
  * to a transaction group.
  */
-int __osd_object_create(const struct lu_env *env, struct osd_device *osd,
+int __osd_object_create(const struct lu_env *env, struct osd_object *obj,
 			dmu_buf_t **dbp, dmu_tx_t *tx, struct lu_attr *la,
 			uint64_t parent)
 {
-	uint64_t oid;
-	int	 rc;
+	uint64_t	     oid;
+	int		     rc;
+	struct osd_device   *osd = osd_obj2dev(obj);
+	const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
+	dmu_object_type_t    type = DMU_OT_PLAIN_FILE_CONTENTS;
 
 	/* Assert that the transaction has been assigned to a
 	   transaction group. */
 	LASSERT(tx->tx_txg != 0);
 
+	/* Use DMU_OTN_UINT8_METADATA for local objects so their data blocks
+	 * would get an additional ditto copy */
+	if (unlikely(S_ISREG(la->la_mode) &&
+		     fid_seq_is_local_file(fid_seq(fid))))
+		type = DMU_OTN_UINT8_METADATA;
+
 	/* Create a new DMU object. */
-	oid = dmu_object_alloc(osd->od_os, DMU_OT_PLAIN_FILE_CONTENTS, 0,
+	oid = dmu_object_alloc(osd->od_os, type, 0,
 			       DMU_OT_SA, DN_MAX_BONUSLEN, tx);
 	rc = -sa_buf_hold(osd->od_os, oid, osd_obj_tag, dbp);
 	LASSERTF(rc == 0, "sa_buf_hold "LPU64" failed: %d\n", oid, rc);
@@ -1276,7 +1352,7 @@ int __osd_zap_create(const struct lu_env *env, struct osd_device *osd,
 	return __osd_attr_init(env, osd, oid, tx, la, parent);
 }
 
-static dmu_buf_t *osd_mkidx(const struct lu_env *env, struct osd_device *osd,
+static dmu_buf_t *osd_mkidx(const struct lu_env *env, struct osd_object *obj,
 			    struct lu_attr *la, uint64_t parent,
 			    struct osd_thandle *oh)
 {
@@ -1288,14 +1364,14 @@ static dmu_buf_t *osd_mkidx(const struct lu_env *env, struct osd_device *osd,
 	 * We set ZAP_FLAG_UINT64_KEY to let ZFS know than we are going to use
 	 * binary keys */
 	LASSERT(S_ISREG(la->la_mode));
-	rc = __osd_zap_create(env, osd, &db, oh->ot_tx, la, parent,
+	rc = __osd_zap_create(env, osd_obj2dev(obj), &db, oh->ot_tx, la, parent,
 			      ZAP_FLAG_UINT64_KEY);
 	if (rc)
 		return ERR_PTR(rc);
 	return db;
 }
 
-static dmu_buf_t *osd_mkdir(const struct lu_env *env, struct osd_device *osd,
+static dmu_buf_t *osd_mkdir(const struct lu_env *env, struct osd_object *obj,
 			    struct lu_attr *la, uint64_t parent,
 			    struct osd_thandle *oh)
 {
@@ -1303,32 +1379,36 @@ static dmu_buf_t *osd_mkdir(const struct lu_env *env, struct osd_device *osd,
 	int	   rc;
 
 	LASSERT(S_ISDIR(la->la_mode));
-	rc = __osd_zap_create(env, osd, &db, oh->ot_tx, la, parent, 0);
+	rc = __osd_zap_create(env, osd_obj2dev(obj), &db,
+			      oh->ot_tx, la, parent, 0);
 	if (rc)
 		return ERR_PTR(rc);
 	return db;
 }
 
-static dmu_buf_t* osd_mkreg(const struct lu_env *env, struct osd_device *osd,
+static dmu_buf_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
 			    struct lu_attr *la, uint64_t parent,
 			    struct osd_thandle *oh)
 {
-	dmu_buf_t *db;
-	int	    rc;
+	dmu_buf_t	  *db;
+	int		   rc;
+	struct osd_device *osd = osd_obj2dev(obj);
 
 	LASSERT(S_ISREG(la->la_mode));
-	rc = __osd_object_create(env, osd, &db, oh->ot_tx, la, parent);
+	rc = __osd_object_create(env, obj, &db, oh->ot_tx, la, parent);
 	if (rc)
 		return ERR_PTR(rc);
 
 	/*
-	 * XXX: a hack, OST to use bigger blocksize. we need
-	 * a method in OSD API to control this from OFD/MDD
+	 * XXX: This heuristic is non-optimal.  It would be better to
+	 * increase the blocksize up to osd->od_max_blksz during the write.
+	 * This is exactly how the ZPL behaves and it ensures that the right
+	 * blocksize is selected based on the file size rather than the
+	 * making broad assumptions based on the osd type.
 	 */
 	if (!lu_device_is_md(osd2lu_dev(osd))) {
-		rc = -dmu_object_set_blocksize(osd->od_os,
-					       db->db_object,
-				128 << 10, 0, oh->ot_tx);
+		rc = -dmu_object_set_blocksize(osd->od_os, db->db_object,
+					       osd->od_max_blksz, 0, oh->ot_tx);
 		if (unlikely(rc)) {
 			CERROR("%s: can't change blocksize: %d\n",
 			       osd->od_svname, rc);
@@ -1339,7 +1419,7 @@ static dmu_buf_t* osd_mkreg(const struct lu_env *env, struct osd_device *osd,
 	return db;
 }
 
-static dmu_buf_t *osd_mksym(const struct lu_env *env, struct osd_device *osd,
+static dmu_buf_t *osd_mksym(const struct lu_env *env, struct osd_object *obj,
 			    struct lu_attr *la, uint64_t parent,
 			    struct osd_thandle *oh)
 {
@@ -1347,13 +1427,13 @@ static dmu_buf_t *osd_mksym(const struct lu_env *env, struct osd_device *osd,
 	int	   rc;
 
 	LASSERT(S_ISLNK(la->la_mode));
-	rc = __osd_object_create(env, osd, &db, oh->ot_tx, la, parent);
+	rc = __osd_object_create(env, obj, &db, oh->ot_tx, la, parent);
 	if (rc)
 		return ERR_PTR(rc);
 	return db;
 }
 
-static dmu_buf_t *osd_mknod(const struct lu_env *env, struct osd_device *osd,
+static dmu_buf_t *osd_mknod(const struct lu_env *env, struct osd_object *obj,
 			    struct lu_attr *la, uint64_t parent,
 			    struct osd_thandle *oh)
 {
@@ -1364,14 +1444,14 @@ static dmu_buf_t *osd_mknod(const struct lu_env *env, struct osd_device *osd,
 	if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode))
 		la->la_valid |= LA_RDEV;
 
-	rc = __osd_object_create(env, osd, &db, oh->ot_tx, la, parent);
+	rc = __osd_object_create(env, obj, &db, oh->ot_tx, la, parent);
 	if (rc)
 		return ERR_PTR(rc);
 	return db;
 }
 
 typedef dmu_buf_t *(*osd_obj_type_f)(const struct lu_env *env,
-				     struct osd_device *osd,
+				     struct osd_object *obj,
 				     struct lu_attr *la,
 				     uint64_t parent,
 				     struct osd_thandle *oh);
@@ -1420,7 +1500,7 @@ static inline int osd_init_lma(const struct lu_env *env, struct osd_object *obj,
 	buf.lb_len = sizeof(*lma);
 
 	rc = osd_xattr_set_internal(env, obj, &buf, XATTR_NAME_LMA,
-				    LU_XATTR_CREATE, oh, BYPASS_CAPA);
+				    LU_XATTR_CREATE, oh);
 
 	return rc;
 }
@@ -1449,10 +1529,12 @@ static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
 	/* concurrent create declarations should not see
 	 * the object inconsistent (db, attr, etc).
 	 * in regular cases acquisition should be cheap */
-	down(&obj->oo_guard);
+	down_write(&obj->oo_guard);
+
+	if (unlikely(dt_object_exists(dt)))
+		GOTO(out, rc = -EEXIST);
 
 	LASSERT(osd_invariant(obj));
-	LASSERT(!dt_object_exists(dt));
 	LASSERT(dof != NULL);
 
 	LASSERT(th != NULL);
@@ -1467,10 +1549,11 @@ static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
 	/* to follow ZFS on-disk format we need
 	 * to initialize parent dnode properly */
 	zapid = 0;
-	if (hint && hint->dah_parent)
+	if (hint != NULL && hint->dah_parent != NULL &&
+	    !dt_object_remote(hint->dah_parent))
 		zapid = osd_dt_obj(hint->dah_parent)->oo_db->db_object;
 
-	db = osd_create_type_f(dof->dof_type)(env, osd, attr, zapid, oh);
+	db = osd_create_type_f(dof->dof_type)(env, obj, attr, zapid, oh);
 	if (IS_ERR(db))
 		GOTO(out, rc = PTR_ERR(db));
 
@@ -1506,15 +1589,12 @@ static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
 	LASSERT(osd_invariant(obj));
 
 	rc = osd_init_lma(env, obj, fid, oh);
-	if (rc) {
+	if (rc != 0)
 		CERROR("%s: can not set LMA on "DFID": rc = %d\n",
 		       osd->od_svname, PFID(fid), rc);
-		/* ignore errors during LMA initialization */
-		rc = 0;
-	}
 
 out:
-	up(&obj->oo_guard);
+	up_write(&obj->oo_guard);
 	RETURN(rc);
 }
 
@@ -1540,8 +1620,11 @@ static int osd_object_ref_add(const struct lu_env *env,
 
 	ENTRY;
 
+	down_read(&obj->oo_guard);
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = -ENOENT);
+
 	LASSERT(osd_invariant(obj));
-	LASSERT(dt_object_exists(dt));
 	LASSERT(obj->oo_sa_hdl != NULL);
 
 	oh = container_of0(handle, struct osd_thandle, ot_super);
@@ -1551,7 +1634,10 @@ static int osd_object_ref_add(const struct lu_env *env,
 	write_unlock(&obj->oo_attr_lock);
 
 	rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
-	return rc;
+
+out:
+	up_read(&obj->oo_guard);
+	RETURN(rc);
 }
 
 static int osd_declare_object_ref_del(const struct lu_env *env,
@@ -1576,8 +1662,12 @@ static int osd_object_ref_del(const struct lu_env *env,
 
 	ENTRY;
 
+	down_read(&obj->oo_guard);
+
+	if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
+		GOTO(out, rc = -ENOENT);
+
 	LASSERT(osd_invariant(obj));
-	LASSERT(dt_object_exists(dt));
 	LASSERT(obj->oo_sa_hdl != NULL);
 
 	oh = container_of0(handle, struct osd_thandle, ot_super);
@@ -1588,146 +1678,10 @@ static int osd_object_ref_del(const struct lu_env *env,
 	write_unlock(&obj->oo_attr_lock);
 
 	rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
-	return rc;
-}
-
-static int capa_is_sane(const struct lu_env *env, struct osd_device *dev,
-			struct lustre_capa *capa, struct lustre_capa_key *keys)
-{
-	struct osd_thread_info	*oti = osd_oti_get(env);
-	struct obd_capa		*oc;
-	int			 i, rc = 0;
-	ENTRY;
-
-	oc = capa_lookup(dev->od_capa_hash, capa, 0);
-	if (oc) {
-		if (capa_is_expired(oc)) {
-			DEBUG_CAPA(D_ERROR, capa, "expired");
-			rc = -ESTALE;
-		}
-		capa_put(oc);
-		RETURN(rc);
-	}
-
-	spin_lock(&capa_lock);
-	for (i = 0; i < 2; i++) {
-		if (keys[i].lk_keyid == capa->lc_keyid) {
-			oti->oti_capa_key = keys[i];
-			break;
-		}
-	}
-	spin_unlock(&capa_lock);
-
-	if (i == 2) {
-		DEBUG_CAPA(D_ERROR, capa, "no matched capa key");
-		RETURN(-ESTALE);
-	}
 
-	rc = capa_hmac(oti->oti_capa.lc_hmac, capa, oti->oti_capa_key.lk_key);
-	if (rc)
-		RETURN(rc);
-	if (memcmp(oti->oti_capa.lc_hmac, capa->lc_hmac, sizeof(capa->lc_hmac)))
-	{
-		DEBUG_CAPA(D_ERROR, capa, "HMAC mismatch");
-		RETURN(-EACCES);
-	}
-
-	oc = capa_add(dev->od_capa_hash, capa);
-	capa_put(oc);
-
-	RETURN(0);
-}
-
-static int osd_object_auth(const struct lu_env *env, struct dt_object *dt,
-			   struct lustre_capa *capa, __u64 opc)
-{
-	const struct lu_fid	*fid = lu_object_fid(&dt->do_lu);
-	struct osd_device	*dev = osd_dev(dt->do_lu.lo_dev);
-	int			 rc;
-
-	if (!dev->od_fl_capa)
-		return 0;
-
-	if (capa == BYPASS_CAPA)
-		return 0;
-
-	if (!capa) {
-		CERROR("no capability is provided for fid "DFID"\n", PFID(fid));
-		return -EACCES;
-	}
-
-	if (!lu_fid_eq(fid, &capa->lc_fid)) {
-		DEBUG_CAPA(D_ERROR, capa, "fid "DFID" mismatch with",PFID(fid));
-		return -EACCES;
-	}
-
-	if (!capa_opc_supported(capa, opc)) {
-		DEBUG_CAPA(D_ERROR, capa, "opc "LPX64" not supported by", opc);
-		return -EACCES;
-	}
-
-	if ((rc = capa_is_sane(env, dev, capa, dev->od_capa_keys))) {
-		DEBUG_CAPA(D_ERROR, capa, "insane (rc %d)", rc);
-		return -EACCES;
-	}
-
-	return 0;
-}
-
-static struct obd_capa *osd_capa_get(const struct lu_env *env,
-				     struct dt_object *dt,
-				     struct lustre_capa *old,
-				     __u64 opc)
-{
-	struct osd_thread_info	*info = osd_oti_get(env);
-	const struct lu_fid	*fid = lu_object_fid(&dt->do_lu);
-	struct osd_object	*obj = osd_dt_obj(dt);
-	struct osd_device	*dev = osd_obj2dev(obj);
-	struct lustre_capa_key	*key = &info->oti_capa_key;
-	struct lustre_capa	*capa = &info->oti_capa;
-	struct obd_capa		*oc;
-	int			 rc;
-	ENTRY;
-
-	if (!dev->od_fl_capa)
-		RETURN(ERR_PTR(-ENOENT));
-
-	LASSERT(dt_object_exists(dt));
-	LASSERT(osd_invariant(obj));
-
-	/* renewal sanity check */
-	if (old && osd_object_auth(env, dt, old, opc))
-		RETURN(ERR_PTR(-EACCES));
-
-	capa->lc_fid = *fid;
-	capa->lc_opc = opc;
-	capa->lc_uid = 0;
-	capa->lc_flags = dev->od_capa_alg << 24;
-	capa->lc_timeout = dev->od_capa_timeout;
-	capa->lc_expiry = 0;
-
-	oc = capa_lookup(dev->od_capa_hash, capa, 1);
-	if (oc) {
-		LASSERT(!capa_is_expired(oc));
-		RETURN(oc);
-	}
-
-	spin_lock(&capa_lock);
-	*key = dev->od_capa_keys[1];
-	spin_unlock(&capa_lock);
-
-	capa->lc_keyid = key->lk_keyid;
-	capa->lc_expiry = cfs_time_current_sec() + dev->od_capa_timeout;
-
-	rc = capa_hmac(capa->lc_hmac, capa, key->lk_key);
-	if (rc) {
-		DEBUG_CAPA(D_ERROR, capa, "HMAC failed: %d for", rc);
-		LBUG();
-		RETURN(ERR_PTR(rc));
-	}
-
-	oc = capa_add(dev->od_capa_hash, capa);
-	RETURN(oc);
+out:
+	up_read(&obj->oo_guard);
+	RETURN(rc);
 }
 
 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
@@ -1770,7 +1724,6 @@ static struct dt_object_operations osd_obj_ops = {
 	.do_declare_xattr_del	= osd_declare_xattr_del,
 	.do_xattr_del		= osd_xattr_del,
 	.do_xattr_list		= osd_xattr_list,
-	.do_capa_get		= osd_capa_get,
 	.do_object_sync		= osd_object_sync,
 };
 
@@ -1785,8 +1738,7 @@ static struct lu_object_operations osd_lu_obj_ops = {
 
 static int osd_otable_it_attr_get(const struct lu_env *env,
 				struct dt_object *dt,
-				struct lu_attr *attr,
-				struct lustre_capa *capa)
+				struct lu_attr *attr)
 {
 	attr->la_valid = 0;
 	return 0;
@@ -1796,4 +1748,3 @@ static struct dt_object_operations osd_obj_otable_it_ops = {
         .do_attr_get    = osd_otable_it_attr_get,
         .do_index_try   = osd_index_try,
 };
-