X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_index.c;h=8c1ffbf02a5737893527b630a24012ff7595d22a;hb=e3d507eec50fc1ff79acf2a9f93d52d698c887d7;hp=b68833f8613777ba16ddc5b32ee6800c8912bba9;hpb=98060d83459ba10409f295898f0ec917f938b4d3;p=fs%2Flustre-release.git

diff --git a/lustre/osd-zfs/osd_index.c b/lustre/osd-zfs/osd_index.c
index b68833f..8c1ffbf 100644
--- a/lustre/osd-zfs/osd_index.c
+++ b/lustre/osd-zfs/osd_index.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/
@@ -41,14 +39,10 @@
  * Author: Mike Pershin <tappro@whamcloud.com>
  */
 
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
 #define DEBUG_SUBSYSTEM S_OSD
 
 #include <lustre_ver.h>
 #include <libcfs/libcfs.h>
-#include <lustre_fsfilt.h>
 #include <obd_support.h>
 #include <lustre_net.h>
 #include <obd.h>
@@ -71,33 +65,120 @@
 #include <sys/sa_impl.h>
 #include <sys/txg.h>
 
+static inline int osd_object_is_zap(dmu_buf_t *db)
+{
+	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *) db;
+	dnode_t *dn;
+	int rc;
+
+	DB_DNODE_ENTER(dbi);
+	dn = DB_DNODE(dbi);
+	rc = (dn->dn_type == DMU_OT_DIRECTORY_CONTENTS ||
+			dn->dn_type == DMU_OT_USERGROUP_USED);
+	DB_DNODE_EXIT(dbi);
+
+	return rc;
+}
+
+/* We don't actually have direct access to the zap_hashbits() function
+ * so just pretend like we do for now.  If this ever breaks we can look at
+ * it at that time. */
+#define zap_hashbits(zc) 48
+/*
+ * ZFS hash format:
+ * | cd (16 bits) | hash (48 bits) |
+ * we need it in other form:
+ * |0| hash (48 bit) | cd (15 bit) |
+ * to be a full 64-bit ordered hash so that Lustre readdir can use it to merge
+ * the readdir hashes from multiple directory stripes uniformly on the client.
+ * Another point is sign bit, the hash range should be in [0, 2^63-1] because
+ * loff_t (for llseek) needs to be a positive value.  This means the "cd" field
+ * should only be the low 15 bits.
+ */
+uint64_t osd_zap_cursor_serialize(zap_cursor_t *zc)
+{
+	uint64_t zfs_hash = zap_cursor_serialize(zc) & (~0ULL >> 1);
+
+	return (zfs_hash >> zap_hashbits(zc)) |
+		(zfs_hash << (63 - zap_hashbits(zc)));
+}
+
+void osd_zap_cursor_init_serialized(zap_cursor_t *zc, struct objset *os,
+				    uint64_t id, uint64_t dirhash)
+{
+	uint64_t zfs_hash = ((dirhash << zap_hashbits(zc)) & (~0ULL >> 1)) |
+		(dirhash >> (63 - zap_hashbits(zc)));
+
+	zap_cursor_init_serialized(zc, os, id, zfs_hash);
+}
+
+int osd_zap_cursor_init(zap_cursor_t **zc, struct objset *os,
+			uint64_t id, uint64_t dirhash)
+{
+	zap_cursor_t *t;
+
+	OBD_ALLOC_PTR(t);
+	if (unlikely(t == NULL))
+		return -ENOMEM;
+
+	osd_zap_cursor_init_serialized(t, os, id, dirhash);
+	*zc = t;
+
+	return 0;
+}
+
+void osd_zap_cursor_fini(zap_cursor_t *zc)
+{
+	zap_cursor_fini(zc);
+	OBD_FREE_PTR(zc);
+}
+
+static inline void osd_obj_cursor_init_serialized(zap_cursor_t *zc,
+						 struct osd_object *o,
+						 uint64_t dirhash)
+{
+	struct osd_device *d = osd_obj2dev(o);
+	osd_zap_cursor_init_serialized(zc, d->od_os,
+				       o->oo_db->db_object, dirhash);
+}
+
+static inline int osd_obj_cursor_init(zap_cursor_t **zc, struct osd_object *o,
+			uint64_t dirhash)
+{
+	struct osd_device *d = osd_obj2dev(o);
+	return osd_zap_cursor_init(zc, d->od_os, o->oo_db->db_object, dirhash);
+}
+
 static struct dt_it *osd_index_it_init(const struct lu_env *env,
 				       struct dt_object *dt,
-				       __u32 unused,
-				       struct lustre_capa *capa)
+				       __u32 unused)
 {
 	struct osd_thread_info  *info = osd_oti_get(env);
 	struct osd_zap_it       *it;
 	struct osd_object       *obj = osd_dt_obj(dt);
-	struct osd_device       *osd = osd_obj2dev(obj);
 	struct lu_object        *lo  = &dt->do_lu;
+	int			 rc;
 	ENTRY;
 
-	/* XXX: check capa ? */
+	if (obj->oo_destroyed)
+		RETURN(ERR_PTR(-ENOENT));
 
 	LASSERT(lu_object_exists(lo));
 	LASSERT(obj->oo_db);
-	LASSERT(udmu_object_is_zap(obj->oo_db));
+	LASSERT(osd_object_is_zap(obj->oo_db));
 	LASSERT(info);
 
-	it = &info->oti_it_zap;
-
-	if (udmu_zap_cursor_init(&it->ozi_zc, &osd->od_objset,
-				 obj->oo_db->db_object, 0))
+	OBD_SLAB_ALLOC_PTR_GFP(it, osd_zapit_cachep, GFP_NOFS);
+	if (it == NULL)
 		RETURN(ERR_PTR(-ENOMEM));
 
+	rc = osd_obj_cursor_init(&it->ozi_zc, obj, 0);
+	if (rc != 0) {
+		OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
+		RETURN(ERR_PTR(rc));
+	}
+
 	it->ozi_obj   = obj;
-	it->ozi_capa  = capa;
 	it->ozi_reset = 1;
 	lu_object_get(lo);
 
@@ -106,8 +187,8 @@ static struct dt_it *osd_index_it_init(const struct lu_env *env,
 
 static void osd_index_it_fini(const struct lu_env *env, struct dt_it *di)
 {
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	struct osd_object *obj;
+	struct osd_zap_it	*it	= (struct osd_zap_it *)di;
+	struct osd_object	*obj;
 	ENTRY;
 
 	LASSERT(it);
@@ -115,8 +196,9 @@ static void osd_index_it_fini(const struct lu_env *env, struct dt_it *di)
 
 	obj = it->ozi_obj;
 
-	udmu_zap_cursor_fini(it->ozi_zc);
+	osd_zap_cursor_fini(it->ozi_zc);
 	lu_object_put(env, &obj->oo_dt.do_lu);
+	OBD_SLAB_FREE_PTR(it, osd_zapit_cachep);
 
 	EXIT;
 }
@@ -128,57 +210,6 @@ static void osd_index_it_put(const struct lu_env *env, struct dt_it *di)
 	 *      next/finish. */
 }
 
-int udmu_zap_cursor_retrieve_key(const struct lu_env *env,
-				 zap_cursor_t *zc, char *key, int max)
-{
-	zap_attribute_t *za = &osd_oti_get(env)->oti_za;
-	int		 err;
-
-	if ((err = zap_cursor_retrieve(zc, za)))
-		return err;
-
-	if (key)
-		strcpy(key, za->za_name);
-
-	return 0;
-}
-
-/*
- * zap_cursor_retrieve read from current record.
- * to read bytes we need to call zap_lookup explicitly.
- */
-int udmu_zap_cursor_retrieve_value(const struct lu_env *env,
-		zap_cursor_t *zc,  char *buf,
-		int buf_size, int *bytes_read)
-{
-	zap_attribute_t *za = &osd_oti_get(env)->oti_za;
-	int err, actual_size;
-
-	if ((err = zap_cursor_retrieve(zc, za)))
-		return err;
-
-	if (za->za_integer_length <= 0)
-		return (ERANGE);
-
-	actual_size = za->za_integer_length * za->za_num_integers;
-
-	if (actual_size > buf_size) {
-		actual_size = buf_size;
-		buf_size = actual_size / za->za_integer_length;
-	} else {
-		buf_size = za->za_num_integers;
-	}
-
-	err = -zap_lookup(zc->zc_objset, zc->zc_zapobj,
-			za->za_name, za->za_integer_length,
-			buf_size, buf);
-
-	if (!err)
-		*bytes_read = actual_size;
-
-	return err;
-}
-
 static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
 				       int len, __u16 type)
 {
@@ -197,90 +228,121 @@ static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
 	ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
 }
 
-/*
- * as we don't know FID, we can't use LU object, so this function
- * partially duplicate __osd_xattr_get() which is built around
- * LU-object and uses it to cache data like regular EA dnode, etc
+/**
+ * Get the object's FID from its LMA EA.
+ *
+ * \param[in] env	pointer to the thread context
+ * \param[in] osd	pointer to the OSD device
+ * \param[in] oid	the object's local identifier
+ * \param[out] fid	the buffer to hold the object's FID
+ *
+ * \retval		0 for success
+ * \retval		negative error number on failure
  */
-static int osd_find_parent_by_dnode(const struct lu_env *env,
-				    struct dt_object *o,
-				    struct lu_fid *fid)
+static int osd_get_fid_by_oid(const struct lu_env *env, struct osd_device *osd,
+			      uint64_t oid, struct lu_fid *fid)
 {
-	struct lustre_mdt_attrs	*lma;
-	udmu_objset_t		*uos = &osd_obj2dev(osd_dt_obj(o))->od_objset;
+	struct objset		*os	  = osd->od_os;
+	struct osd_thread_info	*oti	  = osd_oti_get(env);
+	struct lustre_mdt_attrs	*lma	  =
+			(struct lustre_mdt_attrs *)oti->oti_buf;
 	struct lu_buf		 buf;
-	sa_handle_t		*sa_hdl;
-	nvlist_t		*nvbuf = NULL;
-	uchar_t			*value;
-	uint64_t		 dnode;
-	int			 rc, size;
+	nvlist_t		*sa_xattr = NULL;
+	sa_handle_t		*sa_hdl   = NULL;
+	uchar_t			*nv_value = NULL;
+	uint64_t		 xattr	  = ZFS_NO_OBJECT;
+	int			 size	  = 0;
+	int			 rc;
 	ENTRY;
 
-	/* first of all, get parent dnode from own attributes */
-	LASSERT(osd_dt_obj(o)->oo_db);
-	rc = -sa_handle_get(uos->os, osd_dt_obj(o)->oo_db->db_object,
-			    NULL, SA_HDL_PRIVATE, &sa_hdl);
-	if (rc)
-		RETURN(rc);
+	rc = __osd_xattr_load(osd, oid, &sa_xattr);
+	if (rc == -ENOENT)
+		goto regular;
 
-	dnode = ZFS_NO_OBJECT;
-	rc = -sa_lookup(sa_hdl, SA_ZPL_PARENT(uos), &dnode, 8);
-	sa_handle_destroy(sa_hdl);
-	if (rc)
-		RETURN(rc);
+	if (rc != 0)
+		GOTO(out, rc);
 
-	/* now get EA buffer */
-	rc = __osd_xattr_load(uos, dnode, &nvbuf);
-	if (rc)
-		GOTO(regular, rc);
+	rc = -nvlist_lookup_byte_array(sa_xattr, XATTR_NAME_LMA, &nv_value,
+				       &size);
+	if (rc == -ENOENT)
+		goto regular;
 
-	/* XXX: if we get that far.. should we cache the result? */
+	if (rc != 0)
+		GOTO(out, rc);
 
-	/* try to find LMA attribute */
-	LASSERT(nvbuf != NULL);
-	rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &value, &size);
-	if (rc == 0 && size >= sizeof(*lma)) {
-		lma = (struct lustre_mdt_attrs *)value;
-		lustre_lma_swab(lma);
-		*fid = lma->lma_self_fid;
-		GOTO(out, rc = 0);
-	}
+	if (unlikely(size > sizeof(oti->oti_buf)))
+		GOTO(out, rc = -ERANGE);
 
-regular:
-	/* no LMA attribute in SA, let's try regular EA */
+	memcpy(lma, nv_value, size);
 
-	/* first of all, get parent dnode storing regular EA */
-	rc = -sa_handle_get(uos->os, dnode, NULL, SA_HDL_PRIVATE, &sa_hdl);
-	if (rc)
+	goto found;
+
+regular:
+	rc = -sa_handle_get(os, oid, NULL, SA_HDL_PRIVATE, &sa_hdl);
+	if (rc != 0)
 		GOTO(out, rc);
 
-	dnode = ZFS_NO_OBJECT;
-	rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(uos), &dnode, 8);
+	rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(osd), &xattr, 8);
 	sa_handle_destroy(sa_hdl);
-	if (rc)
+	if (rc != 0)
 		GOTO(out, rc);
 
-	CLASSERT(sizeof(*lma) <= sizeof(osd_oti_get(env)->oti_buf));
-	buf.lb_buf = osd_oti_get(env)->oti_buf;
-	buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
-
-	/* now try to find LMA */
-	rc = __osd_xattr_get_large(env, uos, dnode, &buf,
+	buf.lb_buf = lma;
+	buf.lb_len = sizeof(oti->oti_buf);
+	rc = __osd_xattr_get_large(env, osd, xattr, &buf,
 				   XATTR_NAME_LMA, &size);
-	if (rc == 0 && size >= sizeof(*lma)) {
-		lma = buf.lb_buf;
-		lustre_lma_swab(lma);
-		*fid = lma->lma_self_fid;
-		GOTO(out, rc = 0);
-	} else if (rc < 0) {
+	if (rc != 0)
 		GOTO(out, rc);
-	} else {
+
+found:
+	if (size < sizeof(*lma))
 		GOTO(out, rc = -EIO);
+
+	lustre_lma_swab(lma);
+	if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
+		     CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
+		CWARN("%s: unsupported incompat LMA feature(s) %#x for "
+		      "oid = "LPX64"\n", osd->od_svname,
+		      lma->lma_incompat & ~LMA_INCOMPAT_SUPP, oid);
+		GOTO(out, rc = -EOPNOTSUPP);
+	} else {
+		*fid = lma->lma_self_fid;
+		GOTO(out, rc = 0);
 	}
 
 out:
-	if (nvbuf != NULL)
-		nvlist_free(nvbuf);
+	if (sa_xattr != NULL)
+		nvlist_free(sa_xattr);
+	return rc;
+}
+
+/*
+ * As we don't know FID, we can't use LU object, so this function
+ * partially duplicate __osd_xattr_get() which is built around
+ * LU-object and uses it to cache data like regular EA dnode, etc
+ */
+static int osd_find_parent_by_dnode(const struct lu_env *env,
+				    struct dt_object *o,
+				    struct lu_fid *fid)
+{
+	struct osd_device	*osd = osd_obj2dev(osd_dt_obj(o));
+	sa_handle_t		*sa_hdl;
+	uint64_t		 dnode = ZFS_NO_OBJECT;
+	int			 rc;
+	ENTRY;
+
+	/* first of all, get parent dnode from own attributes */
+	LASSERT(osd_dt_obj(o)->oo_db);
+	rc = -sa_handle_get(osd->od_os, osd_dt_obj(o)->oo_db->db_object,
+			    NULL, SA_HDL_PRIVATE, &sa_hdl);
+	if (rc != 0)
+		RETURN(rc);
+
+	rc = -sa_lookup(sa_hdl, SA_ZPL_PARENT(osd), &dnode, 8);
+	sa_handle_destroy(sa_hdl);
+	if (rc == 0)
+		rc = osd_get_fid_by_oid(env, osd, dnode, fid);
+
 	RETURN(rc);
 }
 
@@ -296,10 +358,9 @@ static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
 	buf.lb_buf = osd_oti_get(env)->oti_buf;
 	buf.lb_len = sizeof(osd_oti_get(env)->oti_buf);
 
-	rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK, BYPASS_CAPA);
+	rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
 	if (rc == -ERANGE) {
-		rc = osd_xattr_get(env, o, &LU_BUF_NULL,
-				   XATTR_NAME_LINK, BYPASS_CAPA);
+		rc = osd_xattr_get(env, o, &LU_BUF_NULL, XATTR_NAME_LINK);
 		if (rc < 0)
 			RETURN(rc);
 		LASSERT(rc > 0);
@@ -307,7 +368,7 @@ static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
 		if (buf.lb_buf == NULL)
 			RETURN(-ENOMEM);
 		buf.lb_len = rc;
-		rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK, BYPASS_CAPA);
+		rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK);
 	}
 	if (rc < 0)
 		GOTO(out, rc);
@@ -355,8 +416,7 @@ out:
 }
 
 static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
-			  struct dt_rec *rec, const struct dt_key *key,
-			  struct lustre_capa *capa)
+			  struct dt_rec *rec, const struct dt_key *key)
 {
 	struct osd_thread_info *oti = osd_oti_get(env);
 	struct osd_object  *obj = osd_dt_obj(dt);
@@ -365,7 +425,7 @@ static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
 	int                 rc;
 	ENTRY;
 
-	LASSERT(udmu_object_is_zap(obj->oo_db));
+	LASSERT(osd_object_is_zap(obj->oo_db));
 
 	if (name[0] == '.') {
 		if (name[1] == 0) {
@@ -378,12 +438,22 @@ static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
 		}
 	}
 
-	rc = -zap_lookup(osd->od_objset.os, obj->oo_db->db_object,
+	memset(&oti->oti_zde.lzd_fid, 0, sizeof(struct lu_fid));
+	rc = -zap_lookup(osd->od_os, obj->oo_db->db_object,
 			 (char *)key, 8, sizeof(oti->oti_zde) / 8,
 			 (void *)&oti->oti_zde);
-	memcpy(rec, &oti->oti_zde.lzd_fid, sizeof(struct lu_fid));
+	if (rc != 0)
+		RETURN(rc);
 
-	RETURN(rc == 0 ? 1 : rc);
+	if (likely(fid_is_sane(&oti->oti_zde.lzd_fid))) {
+		memcpy(rec, &oti->oti_zde.lzd_fid, sizeof(struct lu_fid));
+		RETURN(1);
+	}
+
+	rc = osd_get_fid_by_oid(env, osd, oti->oti_zde.lzd_reg.zde_dnode,
+				(struct lu_fid *)rec);
+
+	RETURN(rc == 0 ? 1 : (rc == -ENOENT ? -ENODATA : rc));
 }
 
 static int osd_declare_dir_insert(const struct lu_env *env,
@@ -394,16 +464,20 @@ static int osd_declare_dir_insert(const struct lu_env *env,
 {
 	struct osd_object  *obj = osd_dt_obj(dt);
 	struct osd_thandle *oh;
+	uint64_t object;
 	ENTRY;
 
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
-	LASSERT(obj->oo_db);
-	LASSERT(udmu_object_is_zap(obj->oo_db));
+	/* This is for inserting dot/dotdot for new created dir. */
+	if (obj->oo_db == NULL)
+		object = DMU_NEW_OBJECT;
+	else
+		object = obj->oo_db->db_object;
 
-	dmu_tx_hold_bonus(oh->ot_tx, obj->oo_db->db_object);
-	dmu_tx_hold_zap(oh->ot_tx, obj->oo_db->db_object, TRUE, (char *)key);
+	dmu_tx_hold_bonus(oh->ot_tx, object);
+	dmu_tx_hold_zap(oh->ot_tx, object, TRUE, (char *)key);
 
 	RETURN(0);
 }
@@ -445,7 +519,7 @@ struct osd_object *osd_object_find(const struct lu_env *env,
 			child = osd_obj(lo);
 		else
 			LU_OBJECT_DEBUG(D_ERROR, env, luch,
-					"%s: object can't be located "DFID"\n",
+					"%s: object can't be located "DFID,
 					osd_dev(ludev)->od_svname, PFID(fid));
 
 		if (child == NULL) {
@@ -456,7 +530,7 @@ struct osd_object *osd_object_find(const struct lu_env *env,
 		}
 	} else {
 		LU_OBJECT_DEBUG(D_ERROR, env, luch,
-				"%s: lu_object does not exists "DFID"\n",
+				"%s: lu_object does not exists "DFID,
 				osd_dev(ludev)->od_svname, PFID(fid));
 		lu_object_put(env, luch);
 		child = ERR_PTR(-ENOENT);
@@ -476,25 +550,49 @@ static inline void osd_object_put(const struct lu_env *env,
 	lu_object_put(env, &obj->oo_dt.do_lu);
 }
 
-static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
-			  struct lu_fid *fid)
+static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
+			  u64 seq)
 {
 	struct lu_seq_range	*range = &osd_oti_get(env)->oti_seq_range;
 	struct seq_server_site	*ss = osd_seq_site(osd);
 	int			rc;
 	ENTRY;
 
-	if (!fid_is_norm(fid) && !fid_is_root(fid))
-		RETURN(0);
+	LASSERT(ss != NULL);
+	LASSERT(ss->ss_server_fld != NULL);
 
-	rc = osd_fld_lookup(env, osd, fid, range);
+	rc = osd_fld_lookup(env, osd, seq, range);
 	if (rc != 0) {
-		CERROR("%s: Can not lookup fld for "DFID"\n",
-		       osd_name(osd), PFID(fid));
-		RETURN(rc);
+		if (rc != -ENOENT)
+			CERROR("%s: Can not lookup fld for "LPX64"\n",
+			       osd_name(osd), seq);
+		RETURN(0);
 	}
 
-	RETURN(ss->ss_node_id != range->lsr_index);
+	RETURN(ss->ss_node_id == range->lsr_index);
+}
+
+static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
+			  const struct lu_fid *fid)
+{
+	struct seq_server_site	*ss = osd_seq_site(osd);
+	ENTRY;
+
+	/* FID seqs not in FLDB, must be local seq */
+	if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
+		RETURN(0);
+
+	/* If FLD is not being initialized yet, it only happens during the
+	 * initialization, likely during mgs initialization, and we assume
+	 * this is local FID. */
+	if (ss == NULL || ss->ss_server_fld == NULL)
+		RETURN(0);
+
+	/* Only check the local FLDB here */
+	if (osd_seq_exists(env, osd, fid_seq(fid)))
+		RETURN(0);
+
+	RETURN(1);
 }
 
 /**
@@ -504,7 +602,6 @@ static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
  *      \param  key     key for index
  *      \param  rec     record reference
  *      \param  th      transaction handler
- *      \param  capa    capability descriptor
  *      \param  ignore_quota update should not affect quota
  *
  *      \retval  0  success
@@ -512,13 +609,13 @@ static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
  */
 static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 			  const struct dt_rec *rec, const struct dt_key *key,
-			  struct thandle *th, struct lustre_capa *capa,
-			  int ignore_quota)
+			  struct thandle *th, int ignore_quota)
 {
 	struct osd_thread_info *oti = osd_oti_get(env);
 	struct osd_object   *parent = osd_dt_obj(dt);
 	struct osd_device   *osd = osd_obj2dev(parent);
-	struct lu_fid       *fid = (struct lu_fid *)rec;
+	struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
+	const struct lu_fid *fid = rec1->rec_fid;
 	struct osd_thandle  *oh;
 	struct osd_object   *child = NULL;
 	__u32                attr;
@@ -527,7 +624,7 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 	ENTRY;
 
 	LASSERT(parent->oo_db);
-	LASSERT(udmu_object_is_zap(parent->oo_db));
+	LASSERT(osd_object_is_zap(parent->oo_db));
 
 	LASSERT(dt_object_exists(dt));
 	LASSERT(osd_invariant(parent));
@@ -545,7 +642,7 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 	if (unlikely(rc == 1)) {
 		/* Insert remote entry */
 		memset(&oti->oti_zde.lzd_reg, 0, sizeof(oti->oti_zde.lzd_reg));
-		oti->oti_zde.lzd_reg.zde_type = IFTODT(S_IFDIR & S_IFMT);
+		oti->oti_zde.lzd_reg.zde_type = IFTODT(rec1->rec_type & S_IFMT);
 	} else {
 		/*
 		 * To simulate old Orion setups with ./..  stored in the
@@ -563,13 +660,25 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 				 * during iteration */
 				GOTO(out, rc = 0);
 			} else if (name[1] == '.' && name[2] == 0) {
+				if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
+					struct lu_fid tfid = *fid;
+
+					osd_object_put(env, child);
+					tfid.f_oid--;
+					child = osd_object_find(env, dt, &tfid);
+					if (IS_ERR(child))
+						RETURN(PTR_ERR(child));
+
+					LASSERT(child->oo_db);
+				}
+
 				/* update parent dnode in the child.
 				 * later it will be used to generate ".." */
-				udmu_objset_t *uos = &osd->od_objset;
 				rc = osd_object_sa_update(parent,
-						 SA_ZPL_PARENT(uos),
+						 SA_ZPL_PARENT(osd),
 						 &child->oo_db->db_object,
 						 8, oh);
+
 				GOTO(out, rc);
 			}
 		}
@@ -582,9 +691,15 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 
 	oti->oti_zde.lzd_fid = *fid;
 	/* Insert (key,oid) into ZAP */
-	rc = -zap_add(osd->od_objset.os, parent->oo_db->db_object,
+	rc = -zap_add(osd->od_os, parent->oo_db->db_object,
 		      (char *)key, 8, sizeof(oti->oti_zde) / 8,
 		      (void *)&oti->oti_zde, oh->ot_tx);
+	if (unlikely(rc == -EEXIST &&
+		     name[0] == '.' && name[1] == '.' && name[2] == 0))
+		/* Update (key,oid) in ZAP */
+		rc = -zap_update(osd->od_os, parent->oo_db->db_object,
+				(char *)key, 8, sizeof(oti->oti_zde) / 8,
+				(void *)&oti->oti_zde, oh->ot_tx);
 
 out:
 	if (child != NULL)
@@ -598,8 +713,9 @@ static int osd_declare_dir_delete(const struct lu_env *env,
 				  const struct dt_key *key,
 				  struct thandle *th)
 {
-	struct osd_object *obj = osd_dt_obj(dt);
+	struct osd_object  *obj = osd_dt_obj(dt);
 	struct osd_thandle *oh;
+	uint64_t	    dnode;
 	ENTRY;
 
 	LASSERT(dt_object_exists(dt));
@@ -608,17 +724,20 @@ static int osd_declare_dir_delete(const struct lu_env *env,
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
-	LASSERT(obj->oo_db);
-	LASSERT(udmu_object_is_zap(obj->oo_db));
-
-	dmu_tx_hold_zap(oh->ot_tx, obj->oo_db->db_object, TRUE, (char *)key);
+	if (dt_object_exists(dt)) {
+		LASSERT(obj->oo_db);
+		LASSERT(osd_object_is_zap(obj->oo_db));
+		dnode = obj->oo_db->db_object;
+	} else {
+		dnode = DMU_NEW_OBJECT;
+	}
+	dmu_tx_hold_zap(oh->ot_tx, dnode, TRUE, (char *)key);
 
 	RETURN(0);
 }
 
 static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
-			  const struct dt_key *key, struct thandle *th,
-			  struct lustre_capa *capa)
+			  const struct dt_key *key, struct thandle *th)
 {
 	struct osd_object *obj = osd_dt_obj(dt);
 	struct osd_device *osd = osd_obj2dev(obj);
@@ -628,8 +747,8 @@ static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
 	int rc;
 	ENTRY;
 
-	LASSERT(obj->oo_db);
-	LASSERT(udmu_object_is_zap(obj->oo_db));
+	LASSERT(zap_db);
+	LASSERT(osd_object_is_zap(zap_db));
 
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
@@ -647,14 +766,9 @@ static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
 	}
 
 	/* Remove key from the ZAP */
-	rc = -zap_remove(osd->od_objset.os, zap_db->db_object,
+	rc = -zap_remove(osd->od_os, zap_db->db_object,
 			 (char *) key, oh->ot_tx);
 
-#if LUSTRE_VERSION_CODE <= OBD_OCD_VERSION(2, 4, 53, 0)
-	if (unlikely(rc == -ENOENT && name[0] == '.' &&
-	    (name[1] == 0 || (name[1] == '.' && name[2] == 0))))
-		rc = 0;
-#endif
 	if (unlikely(rc && rc != -ENOENT))
 		CERROR("%s: zap_remove failed: rc = %d\n", osd->od_svname, rc);
 
@@ -663,12 +777,11 @@ static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
 
 static struct dt_it *osd_dir_it_init(const struct lu_env *env,
 				     struct dt_object *dt,
-				     __u32 unused,
-				     struct lustre_capa *capa)
+				     __u32 unused)
 {
 	struct osd_zap_it *it;
 
-	it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused, capa);
+	it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused);
 	if (!IS_ERR(it))
 		it->ozi_pos = 0;
 
@@ -690,7 +803,6 @@ static int osd_dir_it_get(const struct lu_env *env,
 {
 	struct osd_zap_it *it = (struct osd_zap_it *)di;
 	struct osd_object *obj = it->ozi_obj;
-	struct osd_device *osd = osd_obj2dev(obj);
 	char		  *name = (char *)key;
 	int		   rc;
 	ENTRY;
@@ -698,11 +810,9 @@ static int osd_dir_it_get(const struct lu_env *env,
 	LASSERT(it);
 	LASSERT(it->ozi_zc);
 
-	udmu_zap_cursor_fini(it->ozi_zc);
-
-	if (udmu_zap_cursor_init(&it->ozi_zc, &osd->od_objset,
-				 obj->oo_db->db_object, 0))
-		RETURN(-ENOMEM);
+	/* reset the cursor */
+	zap_cursor_fini(it->ozi_zc);
+	osd_obj_cursor_init_serialized(it->ozi_zc, obj, 0);
 
 	/* XXX: implementation of the API is broken at the moment */
 	LASSERT(((const char *)key)[0] == 0);
@@ -781,6 +891,8 @@ static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
 	zap_attribute_t	  *za = &osd_oti_get(env)->oti_za;
 	int		   rc;
 
+	ENTRY;
+
 	/* temp. storage should be enough for any key supported by ZFS */
 	CLASSERT(sizeof(za->za_name) <= sizeof(it->ozi_name));
 
@@ -793,9 +905,10 @@ static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
 		it->ozi_pos++;
 		if (it->ozi_pos <=2)
 			RETURN(0);
-	}
 
-	zap_cursor_advance(it->ozi_zc);
+	} else {
+		zap_cursor_advance(it->ozi_zc);
+	}
 
 	/*
 	 * According to current API we need to return error if its last entry.
@@ -831,19 +944,6 @@ static struct dt_key *osd_dir_it_key(const struct lu_env *env,
 
 	strcpy(it->ozi_name, za->za_name);
 
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
-	if (za->za_name[0] == '.') {
-		if (za->za_name[1] == 0 || (za->za_name[1] == '.' &&
-		    za->za_name[2] == 0)) {
-			/* we should not get onto . and ..
-			 * stored in the directory. ->next() and
-			 * other methods should prevent this
-			 */
-			LBUG();
-		}
-	}
-#endif
-
 	RETURN((struct dt_key *)it->ozi_name);
 }
 
@@ -864,18 +964,6 @@ static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
 	if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)) == 0)
 		rc = strlen(za->za_name);
 
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
-	if (rc == 0 && za->za_name[0] == '.') {
-		if (za->za_name[1] == 0 || (za->za_name[1] == '.' &&
-		    za->za_name[2] == 0)) {
-			/* we should not get onto . and ..
-			 * stored in the directory. ->next() and
-			 * other methods should prevent this
-			 */
-			LBUG();
-		}
-	}
-#endif
 	RETURN(rc);
 }
 
@@ -910,23 +998,20 @@ static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
 		osd_it_append_attrs(lde, attr, 2, IFTODT(S_IFDIR));
 		lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(2, attr));
 		rc = osd_find_parent_fid(env, &it->ozi_obj->oo_dt, &lde->lde_fid);
-		/*
-		 * early Orion code was not setting LinkEA, so it's possible
-		 * some setups still have objects with no LinkEA set.
-		 * but at that time .. was a real record in the directory
-		 * so we should try to lookup .. in ZAP
-		 */
-		if (rc != -ENOENT)
-			GOTO(out, rc);
+
+		/* ENOENT happens at the root of filesystem so ignore it */
+		if (rc == -ENOENT)
+			rc = 0;
+		GOTO(out, rc);
 	}
 
 	LASSERT(lde);
 
-	lde->lde_hash = cpu_to_le64(udmu_zap_cursor_serialize(it->ozi_zc));
-
-	if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
+	rc = -zap_cursor_retrieve(it->ozi_zc, za);
+	if (unlikely(rc != 0))
 		GOTO(out, rc);
 
+	lde->lde_hash = cpu_to_le64(osd_zap_cursor_serialize(it->ozi_zc));
 	namelen = strlen(za->za_name);
 	if (namelen > NAME_MAX)
 		GOTO(out, rc = -EOVERFLOW);
@@ -958,6 +1043,45 @@ out:
 	RETURN(rc);
 }
 
+static int osd_dir_it_rec_size(const struct lu_env *env, const struct dt_it *di,
+			       __u32 attr)
+{
+	struct osd_zap_it   *it = (struct osd_zap_it *)di;
+	zap_attribute_t     *za = &osd_oti_get(env)->oti_za;
+	size_t		     namelen = 0;
+	int		     rc;
+	ENTRY;
+
+	if (it->ozi_pos <= 1)
+		namelen = 1;
+	else if (it->ozi_pos == 2)
+		namelen = 2;
+
+	if (namelen > 0) {
+		rc = lu_dirent_calc_size(namelen, attr);
+		RETURN(rc);
+	}
+
+	rc = -zap_cursor_retrieve(it->ozi_zc, za);
+	if (unlikely(rc != 0))
+		RETURN(rc);
+
+	if (za->za_integer_length != 8 || za->za_num_integers < 3) {
+		CERROR("%s: unsupported direntry format: %d %d\n",
+		       osd_obj2dev(it->ozi_obj)->od_svname,
+		       za->za_integer_length, (int)za->za_num_integers);
+		RETURN(-EIO);
+	}
+
+	namelen = strlen(za->za_name);
+	if (namelen > NAME_MAX)
+		RETURN(-EOVERFLOW);
+
+	rc = lu_dirent_calc_size(namelen, attr);
+
+	RETURN(rc);
+}
+
 static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
 {
 	struct osd_zap_it *it = (struct osd_zap_it *)di;
@@ -967,7 +1091,7 @@ static __u64 osd_dir_it_store(const struct lu_env *env, const struct dt_it *di)
 	if (it->ozi_pos <= 2)
 		pos = it->ozi_pos;
 	else
-		pos = udmu_zap_cursor_serialize(it->ozi_zc);
+		pos = osd_zap_cursor_serialize(it->ozi_zc);
 
 	RETURN(pos);
 }
@@ -983,15 +1107,13 @@ static int osd_dir_it_load(const struct lu_env *env,
 {
 	struct osd_zap_it *it = (struct osd_zap_it *)di;
 	struct osd_object *obj = it->ozi_obj;
-	struct osd_device *osd = osd_obj2dev(obj);
 	zap_attribute_t   *za = &osd_oti_get(env)->oti_za;
 	int		   rc;
 	ENTRY;
 
-	udmu_zap_cursor_fini(it->ozi_zc);
-	if (udmu_zap_cursor_init(&it->ozi_zc, &osd->od_objset,
-				 obj->oo_db->db_object, hash))
-		RETURN(-ENOMEM);
+	/* reset the cursor */
+	zap_cursor_fini(it->ozi_zc);
+	osd_obj_cursor_init_serialized(it->ozi_zc, obj, hash);
 
 	if (hash <= 2) {
 		it->ozi_pos = hash;
@@ -1009,7 +1131,7 @@ static int osd_dir_it_load(const struct lu_env *env,
 	RETURN(rc);
 }
 
-static struct dt_index_operations osd_dir_ops = {
+struct dt_index_operations osd_dir_ops = {
 	.dio_lookup         = osd_dir_lookup,
 	.dio_declare_insert = osd_declare_dir_insert,
 	.dio_insert         = osd_dir_insert,
@@ -1024,6 +1146,7 @@ static struct dt_index_operations osd_dir_ops = {
 		.key      = osd_dir_it_key,
 		.key_size = osd_dir_it_key_size,
 		.rec      = osd_dir_it_rec,
+		.rec_size = osd_dir_it_rec_size,
 		.store    = osd_dir_it_store,
 		.load     = osd_dir_it_load
 	}
@@ -1031,20 +1154,43 @@ static struct dt_index_operations osd_dir_ops = {
 
 /*
  * Primitives for index files using binary keys.
- * XXX: only 64-bit keys are supported for now.
  */
 
+/* key integer_size is 8 */
+static int osd_prepare_key_uint64(struct osd_object *o, __u64 *dst,
+				  const struct dt_key *src)
+{
+	int size;
+
+	LASSERT(dst);
+	LASSERT(src);
+
+	/* align keysize to 64bit */
+	size = (o->oo_keysize + sizeof(__u64) - 1) / sizeof(__u64);
+	size *= sizeof(__u64);
+
+	LASSERT(size <= MAXNAMELEN);
+
+	if (unlikely(size > o->oo_keysize))
+		memset(dst + o->oo_keysize, 0, size - o->oo_keysize);
+	memcpy(dst, (const char *)src, o->oo_keysize);
+
+	return (size/sizeof(__u64));
+}
+
 static int osd_index_lookup(const struct lu_env *env, struct dt_object *dt,
-			struct dt_rec *rec, const struct dt_key *key,
-			struct lustre_capa *capa)
+			struct dt_rec *rec, const struct dt_key *key)
 {
 	struct osd_object *obj = osd_dt_obj(dt);
 	struct osd_device *osd = osd_obj2dev(obj);
+	__u64		  *k = osd_oti_get(env)->oti_key64;
 	int                rc;
 	ENTRY;
 
-	rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
-				(const __u64 *)key, 1, 8, obj->oo_recsize,
+	rc = osd_prepare_key_uint64(obj, k, key);
+
+	rc = -zap_lookup_uint64(osd->od_os, obj->oo_db->db_object,
+				k, rc, obj->oo_recusize, obj->oo_recsize,
 				(void *)rec);
 	RETURN(rc == 0 ? 1 : rc);
 }
@@ -1076,12 +1222,12 @@ static int osd_declare_index_insert(const struct lu_env *env,
 
 static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
 			    const struct dt_rec *rec, const struct dt_key *key,
-			    struct thandle *th, struct lustre_capa *capa,
-			    int ignore_quota)
+			    struct thandle *th, int ignore_quota)
 {
 	struct osd_object  *obj = osd_dt_obj(dt);
 	struct osd_device  *osd = osd_obj2dev(obj);
 	struct osd_thandle *oh;
+	__u64		   *k = osd_oti_get(env)->oti_key64;
 	int                 rc;
 	ENTRY;
 
@@ -1092,9 +1238,11 @@ static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
 
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
+	rc = osd_prepare_key_uint64(obj, k, key);
+
 	/* Insert (key,oid) into ZAP */
-	rc = -zap_add_uint64(osd->od_objset.os, obj->oo_db->db_object,
-			     (const __u64 *)key, 1, 8, obj->oo_recsize,
+	rc = -zap_add_uint64(osd->od_os, obj->oo_db->db_object,
+			     k, rc, obj->oo_recusize, obj->oo_recsize,
 			     (void *)rec, oh->ot_tx);
 	RETURN(rc);
 }
@@ -1120,12 +1268,12 @@ static int osd_declare_index_delete(const struct lu_env *env,
 }
 
 static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
-			    const struct dt_key *key, struct thandle *th,
-			    struct lustre_capa *capa)
+			    const struct dt_key *key, struct thandle *th)
 {
 	struct osd_object  *obj = osd_dt_obj(dt);
 	struct osd_device  *osd = osd_obj2dev(obj);
 	struct osd_thandle *oh;
+	__u64		   *k = osd_oti_get(env)->oti_key64;
 	int                 rc;
 	ENTRY;
 
@@ -1133,9 +1281,11 @@ static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
+	rc = osd_prepare_key_uint64(obj, k, key);
+
 	/* Remove binary key from the ZAP */
-	rc = -zap_remove_uint64(osd->od_objset.os, obj->oo_db->db_object,
-				(const __u64 *)key, 1, oh->ot_tx);
+	rc = -zap_remove_uint64(osd->od_os, obj->oo_db->db_object,
+				k, rc, oh->ot_tx);
 	RETURN(rc);
 }
 
@@ -1150,12 +1300,15 @@ static int osd_index_it_get(const struct lu_env *env, struct dt_it *di,
 	LASSERT(it);
 	LASSERT(it->ozi_zc);
 
-	/* XXX: API is broken at the moment */
-	LASSERT(*((const __u64 *)key) == 0);
+	/*
+	 * XXX: we need a binary version of zap_cursor_move_to_key()
+	 *	to implement this API */
+	if (*((const __u64 *)key) != 0)
+		CERROR("NOT IMPLEMETED YET (move to "LPX64")\n",
+		       *((__u64 *)key));
 
 	zap_cursor_fini(it->ozi_zc);
-	memset(it->ozi_zc, 0, sizeof(*it->ozi_zc));
-	zap_cursor_init(it->ozi_zc, osd->od_objset.os, obj->oo_db->db_object);
+	zap_cursor_init(it->ozi_zc, osd->od_os, obj->oo_db->db_object);
 	it->ozi_reset = 1;
 
 	RETURN(+1);
@@ -1189,6 +1342,7 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
 				       const struct dt_it *di)
 {
 	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	struct osd_object *obj = it->ozi_obj;
 	zap_attribute_t   *za = &osd_oti_get(env)->oti_za;
 	int                rc = 0;
 	ENTRY;
@@ -1199,7 +1353,7 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
 		RETURN(ERR_PTR(rc));
 
 	/* the binary key is stored in the name */
-	it->ozi_key = *((__u64 *)za->za_name);
+	memcpy(&it->ozi_key, za->za_name, obj->oo_keysize);
 
 	RETURN((struct dt_key *)&it->ozi_key);
 }
@@ -1207,8 +1361,9 @@ static struct dt_key *osd_index_it_key(const struct lu_env *env,
 static int osd_index_it_key_size(const struct lu_env *env,
 				const struct dt_it *di)
 {
-	/* we only support 64-bit binary keys for the time being */
-	RETURN(sizeof(__u64));
+	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	struct osd_object *obj = it->ozi_obj;
+	RETURN(obj->oo_keysize);
 }
 
 static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
@@ -1218,6 +1373,7 @@ static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
 	struct osd_zap_it *it = (struct osd_zap_it *)di;
 	struct osd_object *obj = it->ozi_obj;
 	struct osd_device *osd = osd_obj2dev(obj);
+	__u64		  *k = osd_oti_get(env)->oti_key64;
 	int                rc;
 	ENTRY;
 
@@ -1226,9 +1382,11 @@ static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
 	if (rc)
 		RETURN(rc);
 
-	rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
-				(const __u64 *)za->za_name, 1, 8,
-				obj->oo_recsize, (void *)rec);
+	rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
+
+	rc = -zap_lookup_uint64(osd->od_os, obj->oo_db->db_object,
+				k, rc, obj->oo_recusize, obj->oo_recsize,
+				(void *)rec);
 	RETURN(rc);
 }
 
@@ -1251,12 +1409,9 @@ static int osd_index_it_load(const struct lu_env *env, const struct dt_it *di,
 	int                rc;
 	ENTRY;
 
-	/* close the current cursor */
+	/* reset the cursor */
 	zap_cursor_fini(it->ozi_zc);
-
-	/* create a new one starting at hash */
-	memset(it->ozi_zc, 0, sizeof(*it->ozi_zc));
-	zap_cursor_init_serialized(it->ozi_zc, osd->od_objset.os,
+	zap_cursor_init_serialized(it->ozi_zc, osd->od_os,
 				   obj->oo_db->db_object, hash);
 	it->ozi_reset = 0;
 
@@ -1289,60 +1444,274 @@ static struct dt_index_operations osd_index_ops = {
 	}
 };
 
+struct osd_metadnode_it {
+	struct osd_device       *mit_dev;
+	__u64			 mit_pos;
+	struct lu_fid		 mit_fid;
+	int			 mit_prefetched;
+	__u64			 mit_prefetched_dnode;
+};
+
+static struct dt_it *osd_zfs_otable_it_init(const struct lu_env *env,
+					    struct dt_object *dt, __u32 attr)
+{
+	struct osd_device	*dev   = osd_dev(dt->do_lu.lo_dev);
+	struct osd_metadnode_it *it;
+	ENTRY;
+
+	OBD_ALLOC_PTR(it);
+	if (unlikely(it == NULL))
+		RETURN(ERR_PTR(-ENOMEM));
+
+	it->mit_dev = dev;
+
+	/* XXX: dmu_object_next() does NOT find dnodes allocated
+	 *	in the current non-committed txg, so we force txg
+	 *	commit to find all existing dnodes ... */
+	txg_wait_synced(dmu_objset_pool(dev->od_os), 0ULL);
+
+	RETURN((struct dt_it *)it);
+}
+
+static void osd_zfs_otable_it_fini(const struct lu_env *env, struct dt_it *di)
+{
+	struct osd_metadnode_it *it  = (struct osd_metadnode_it *)di;
+
+	OBD_FREE_PTR(it);
+}
+
+static int osd_zfs_otable_it_get(const struct lu_env *env,
+				 struct dt_it *di, const struct dt_key *key)
+{
+	return 0;
+}
+
+static void osd_zfs_otable_it_put(const struct lu_env *env, struct dt_it *di)
+{
+}
+
+#define OTABLE_PREFETCH		256
+
+static void osd_zfs_otable_prefetch(const struct lu_env *env,
+				    struct osd_metadnode_it *it)
+{
+	struct osd_device	*dev = it->mit_dev;
+	int			 rc;
+
+	/* can go negative on the very first access to the iterator
+	 * or if some non-Lustre objects were found */
+	if (unlikely(it->mit_prefetched < 0))
+		it->mit_prefetched = 0;
+
+	if (it->mit_prefetched >= (OTABLE_PREFETCH >> 1))
+		return;
+
+	if (it->mit_prefetched_dnode == 0)
+		it->mit_prefetched_dnode = it->mit_pos;
+
+	while (it->mit_prefetched < OTABLE_PREFETCH) {
+		rc = -dmu_object_next(dev->od_os, &it->mit_prefetched_dnode,
+				      B_FALSE, 0);
+		if (unlikely(rc != 0))
+			break;
+
+		osd_dmu_prefetch(dev->od_os, it->mit_prefetched_dnode,
+				 0, 0, 0, ZIO_PRIORITY_ASYNC_READ);
+
+		it->mit_prefetched++;
+	}
+}
+
+static int osd_zfs_otable_it_next(const struct lu_env *env, struct dt_it *di)
+{
+	struct osd_metadnode_it *it  = (struct osd_metadnode_it *)di;
+	struct lustre_mdt_attrs *lma;
+	struct osd_device	*dev = it->mit_dev;
+	nvlist_t		*nvbuf = NULL;
+	uchar_t			*v;
+	__u64			 dnode;
+	int			 rc, s;
+
+	memset(&it->mit_fid, 0, sizeof(it->mit_fid));
+
+	dnode = it->mit_pos;
+	do {
+		rc = -dmu_object_next(dev->od_os, &it->mit_pos, B_FALSE, 0);
+		if (unlikely(rc != 0))
+			GOTO(out, rc = 1);
+		it->mit_prefetched--;
+
+		/* LMA is required for this to be a Lustre object.
+		 * If there is no xattr skip it. */
+		rc = __osd_xattr_load(dev, it->mit_pos, &nvbuf);
+		if (unlikely(rc != 0))
+			continue;
+
+		LASSERT(nvbuf != NULL);
+		rc = -nvlist_lookup_byte_array(nvbuf, XATTR_NAME_LMA, &v, &s);
+		if (likely(rc == 0)) {
+			/* Lustre object */
+			lma = (struct lustre_mdt_attrs *)v;
+			lustre_lma_swab(lma);
+			it->mit_fid = lma->lma_self_fid;
+			nvlist_free(nvbuf);
+			break;
+		} else {
+			/* not a Lustre object, try next one */
+			nvlist_free(nvbuf);
+		}
+
+	} while (1);
+
+
+	/* we aren't prefetching in the above loop because the number of
+	 * non-Lustre objects is very small and we will be repeating very
+	 * rare. in case we want to use this to iterate over non-Lustre
+	 * objects (i.e. when we convert regular ZFS in Lustre) it makes
+	 * sense to initiate prefetching in the loop */
+
+	/* 0 - there are more items, +1 - the end */
+	if (likely(rc == 0))
+		osd_zfs_otable_prefetch(env, it);
+
+	CDEBUG(D_OTHER, "advance: %llu -> %llu "DFID": %d\n", dnode,
+	       it->mit_pos, PFID(&it->mit_fid), rc);
+
+out:
+	return rc;
+}
+
+static struct dt_key *osd_zfs_otable_it_key(const struct lu_env *env,
+					    const struct dt_it *di)
+{
+	return NULL;
+}
+
+static int osd_zfs_otable_it_key_size(const struct lu_env *env,
+				      const struct dt_it *di)
+{
+	return sizeof(__u64);
+}
+
+static int osd_zfs_otable_it_rec(const struct lu_env *env,
+				 const struct dt_it *di,
+				 struct dt_rec *rec, __u32 attr)
+{
+	struct osd_metadnode_it *it  = (struct osd_metadnode_it *)di;
+	struct lu_fid *fid = (struct lu_fid *)rec;
+	ENTRY;
+
+	*fid = it->mit_fid;
+
+	RETURN(0);
+}
+
+
+static __u64 osd_zfs_otable_it_store(const struct lu_env *env,
+				     const struct dt_it *di)
+{
+	struct osd_metadnode_it *it  = (struct osd_metadnode_it *)di;
+
+	return it->mit_pos;
+}
+
+static int osd_zfs_otable_it_load(const struct lu_env *env,
+				  const struct dt_it *di, __u64 hash)
+{
+	struct osd_metadnode_it *it  = (struct osd_metadnode_it *)di;
+
+	it->mit_pos = hash;
+	it->mit_prefetched = 0;
+	it->mit_prefetched_dnode = 0;
+
+	return osd_zfs_otable_it_next(env, (struct dt_it *)di);
+}
+
+static int osd_zfs_otable_it_key_rec(const struct lu_env *env,
+				     const struct dt_it *di, void *key_rec)
+{
+	return 0;
+}
+
+const struct dt_index_operations osd_zfs_otable_ops = {
+	.dio_it = {
+		.init     = osd_zfs_otable_it_init,
+		.fini     = osd_zfs_otable_it_fini,
+		.get      = osd_zfs_otable_it_get,
+		.put	  = osd_zfs_otable_it_put,
+		.next     = osd_zfs_otable_it_next,
+		.key	  = osd_zfs_otable_it_key,
+		.key_size = osd_zfs_otable_it_key_size,
+		.rec      = osd_zfs_otable_it_rec,
+		.store    = osd_zfs_otable_it_store,
+		.load     = osd_zfs_otable_it_load,
+		.key_rec  = osd_zfs_otable_it_key_rec,
+	}
+};
+
 int osd_index_try(const struct lu_env *env, struct dt_object *dt,
 		const struct dt_index_features *feat)
 {
 	struct osd_object *obj = osd_dt_obj(dt);
+	int rc = 0;
 	ENTRY;
 
-	LASSERT(dt_object_exists(dt));
+	down_read(&obj->oo_guard);
 
 	/*
 	 * XXX: implement support for fixed-size keys sorted with natural
 	 *      numerical way (not using internal hash value)
 	 */
 	if (feat->dif_flags & DT_IND_RANGE)
-		RETURN(-ERANGE);
+		GOTO(out, rc = -ERANGE);
 
-	if (unlikely(feat == &dt_otable_features))
-		/* do not support oi scrub yet. */
-		RETURN(-ENOTSUPP);
+	if (unlikely(feat == &dt_otable_features)) {
+		dt->do_index_ops = &osd_zfs_otable_ops;
+		GOTO(out, rc = 0);
+	}
 
-	LASSERT(obj->oo_db != NULL);
+	LASSERT(!dt_object_exists(dt) || obj->oo_db != NULL);
 	if (likely(feat == &dt_directory_features)) {
-		if (udmu_object_is_zap(obj->oo_db))
+		if (!dt_object_exists(dt) || osd_object_is_zap(obj->oo_db))
 			dt->do_index_ops = &osd_dir_ops;
 		else
-			RETURN(-ENOTDIR);
+			GOTO(out, rc = -ENOTDIR);
 	} else if (unlikely(feat == &dt_acct_features)) {
 		LASSERT(fid_is_acct(lu_object_fid(&dt->do_lu)));
 		dt->do_index_ops = &osd_acct_index_ops;
-	} else if (udmu_object_is_zap(obj->oo_db) &&
-		   dt->do_index_ops == NULL) {
+	} else if (dt->do_index_ops == NULL) {
 		/* For index file, we don't support variable key & record sizes
 		 * and the key has to be unique */
 		if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
-			RETURN(-EINVAL);
+			GOTO(out, rc = -EINVAL);
 
-		/* Although the zap_*_uint64() primitives support large keys, we
-		 * limit ourselves to 64-bit keys for now */
-		if (feat->dif_keysize_max != sizeof(__u64) ||
-		    feat->dif_keysize_min != sizeof(__u64))
-			RETURN(-EINVAL);
+		if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
+			GOTO(out, rc = -E2BIG);
+		if (feat->dif_keysize_max != feat->dif_keysize_min)
+			GOTO(out, rc = -EINVAL);
 
 		/* As for the record size, it should be a multiple of 8 bytes
 		 * and smaller than the maximum value length supported by ZAP.
 		 */
 		if (feat->dif_recsize_max > ZAP_MAXVALUELEN)
-			RETURN(-E2BIG);
-		if (feat->dif_recsize_max != feat->dif_recsize_min ||
-		    (feat->dif_recsize_max & (sizeof(__u64) - 1)))
-			RETURN(-EINVAL);
-
-		obj->oo_recsize = feat->dif_recsize_max / sizeof(__u64);
+			GOTO(out, rc = -E2BIG);
+		if (feat->dif_recsize_max != feat->dif_recsize_min)
+			GOTO(out, rc = -EINVAL);
+
+		obj->oo_keysize = feat->dif_keysize_max;
+		obj->oo_recsize = feat->dif_recsize_max;
+		obj->oo_recusize = 1;
+
+		/* ZFS prefers to work with array of 64bits */
+		if ((obj->oo_recsize & 7) == 0) {
+			obj->oo_recsize >>= 3;
+			obj->oo_recusize = 8;
+		}
 		dt->do_index_ops = &osd_index_ops;
 	}
 
-	RETURN(0);
-}
+out:
+	up_read(&obj->oo_guard);
 
+	RETURN(rc);
+}