X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_index.c;h=43779a9b8b7c1a2f0e35bed371c92a3b03ba389f;hp=cb865b5bcdadd8232722b6b67a9b8714ac01d3a1;hb=b69b7de30c3977cb69a741099218bc4a81752717;hpb=52953ae668710484a8396cdb2fe94016051496a6

diff --git a/lustre/osd-zfs/osd_index.c b/lustre/osd-zfs/osd_index.c
index cb865b5..43779a9 100644
--- a/lustre/osd-zfs/osd_index.c
+++ b/lustre/osd-zfs/osd_index.c
@@ -28,7 +28,7 @@
  * Use is subject to license terms.
  */
 /*
- * Copyright (c) 2011, 2012 Whamcloud, Inc.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  * Use is subject to license terms.
  */
 /*
@@ -41,14 +41,10 @@
  * 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>
@@ -71,10 +67,10 @@
 #include <sys/sa_impl.h>
 #include <sys/txg.h>
 
-static struct dt_it *osd_zap_it_init(const struct lu_env *env,
-				     struct dt_object *dt,
-				     __u32 unused,
-				     struct lustre_capa *capa)
+static struct dt_it *osd_index_it_init(const struct lu_env *env,
+				       struct dt_object *dt,
+				       __u32 unused,
+				       struct lustre_capa *capa)
 {
 	struct osd_thread_info  *info = osd_oti_get(env);
 	struct osd_zap_it       *it;
@@ -104,7 +100,7 @@ static struct dt_it *osd_zap_it_init(const struct lu_env *env,
 	RETURN((struct dt_it *)it);
 }
 
-static void osd_zap_it_fini(const struct lu_env *env, struct dt_it *di)
+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;
@@ -121,42 +117,8 @@ static void osd_zap_it_fini(const struct lu_env *env, struct dt_it *di)
 	EXIT;
 }
 
-/**
- *  Move Iterator to record specified by \a key
- *
- *  \param  di      osd iterator
- *  \param  key     key for index
- *
- *  \retval +ve  di points to record with least key not larger than key
- *  \retval  0   di points to exact matched key
- *  \retval -ve  failure
- */
-
-static int osd_zap_it_get(const struct lu_env *env,
-			  struct dt_it *di, const struct dt_key *key)
-{
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	struct osd_object *obj = it->ozi_obj;
-	struct osd_device *osd = osd_obj2dev(obj);
-	ENTRY;
-
-	LASSERT(it);
-	LASSERT(it->ozi_zc);
-
-	/* XXX: API is broken at the moment */
-	LASSERT(((const char *)key)[0] == '\0');
-
-	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);
-
-	it->ozi_reset = 1;
-
-	RETURN(+1);
-}
 
-static void osd_zap_it_put(const struct lu_env *env, struct dt_it *di)
+static void osd_index_it_put(const struct lu_env *env, struct dt_it *di)
 {
 	/* PBS: do nothing : ref are incremented at retrive and decreamented
 	 *      next/finish. */
@@ -166,96 +128,28 @@ 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;
+	int		 err;
 
 	if ((err = zap_cursor_retrieve(zc, za)))
 		return err;
 
-	if (key) {
-		if (strlen(za->za_name) > max)
-			return EOVERFLOW;
+	if (key)
 		strcpy(key, za->za_name);
-	}
 
 	return 0;
 }
 
-/**
- * to load a directory entry at a time and stored it in
- * iterator's in-memory data structure.
- *
- * \param di, struct osd_it_ea, iterator's in memory structure
- *
- * \retval +ve, iterator reached to end
- * \retval   0, iterator not reached to end
- * \retval -ve, on error
- */
-static int osd_zap_it_next(const struct lu_env *env, struct dt_it *di)
-{
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	int                rc;
-	ENTRY;
-
-	if (it->ozi_reset == 0)
-		zap_cursor_advance(it->ozi_zc);
-	it->ozi_reset = 0;
-
-	/*
-	 * According to current API we need to return error if its last entry.
-	 * zap_cursor_advance() does return any value. So we need to call
-	 * retrieve to check if there is any record.  We should make
-	 * changes to Iterator API to not return status for this API
-	 */
-	rc = -udmu_zap_cursor_retrieve_key(env, it->ozi_zc, NULL, NAME_MAX);
-	if (rc == -ENOENT) /* end of dir*/
-		RETURN(+1);
-
-	RETURN((rc));
-}
-
-static struct dt_key *osd_zap_it_key(const struct lu_env *env,
-					const struct dt_it *di)
-{
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	int                rc = 0;
-	ENTRY;
-
-	it->ozi_reset = 0;
-	rc = -udmu_zap_cursor_retrieve_key(env, it->ozi_zc, it->ozi_name,
-					   NAME_MAX + 1);
-	if (!rc)
-		RETURN((struct dt_key *)it->ozi_name);
-	else
-		RETURN(ERR_PTR(rc));
-}
-
-static int osd_zap_it_key_size(const struct lu_env *env, const struct dt_it *di)
-{
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	int                rc;
-	ENTRY;
-
-	it->ozi_reset = 0;
-	rc = -udmu_zap_cursor_retrieve_key(env, it->ozi_zc, it->ozi_name,
-					   NAME_MAX + 1);
-	if (!rc)
-		RETURN(strlen(it->ozi_name));
-	else
-		RETURN(rc);
-}
-
 /*
  * 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_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;
 
@@ -272,8 +166,8 @@ int udmu_zap_cursor_retrieve_value(const struct lu_env *env,
 	}
 
 	err = -zap_lookup(zc->zc_objset, zc->zc_zapobj,
-			  za->za_name, za->za_integer_length,
-			  buf_size, buf);
+			za->za_name, za->za_integer_length,
+			buf_size, buf);
 
 	if (!err)
 		*bytes_read = actual_size;
@@ -292,98 +186,166 @@ static inline void osd_it_append_attrs(struct lu_dirent *ent, __u32 attr,
 		len = (len + align) & ~align;
 
 		lt = (void *)ent->lde_name + len;
-		lt->lt_type = cpu_to_le16(CFS_DTTOIF(type));
+		lt->lt_type = cpu_to_le16(DTTOIF(type));
 		ent->lde_attrs |= LUDA_TYPE;
 	}
 
 	ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
 }
 
-static int osd_zap_it_rec(const struct lu_env *env, const struct dt_it *di,
-			  struct dt_rec *dtrec, __u32 attr)
+/*
+ * 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 luz_direntry *zde = &osd_oti_get(env)->oti_zde;
-	zap_attribute_t     *za = &osd_oti_get(env)->oti_za;
-	struct osd_zap_it   *it = (struct osd_zap_it *)di;
-	struct lu_dirent    *lde = (struct lu_dirent *)dtrec;
-	int                  rc, namelen;
+	struct lustre_mdt_attrs	*lma;
+	udmu_objset_t		*uos = &osd_obj2dev(osd_dt_obj(o))->od_objset;
+	struct lu_buf		 buf;
+	sa_handle_t		*sa_hdl;
+	nvlist_t		*nvbuf = NULL;
+	uchar_t			*value;
+	uint64_t		 dnode;
+	int			 rc, size;
 	ENTRY;
 
-	it->ozi_reset = 0;
-	LASSERT(lde);
-
-	lde->lde_hash = cpu_to_le64(udmu_zap_cursor_serialize(it->ozi_zc));
-
-	if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
-		GOTO(out, rc);
-
-	namelen = strlen(za->za_name);
-	if (namelen > NAME_MAX)
-		GOTO(out, rc = -EOVERFLOW);
-	strcpy(lde->lde_name, za->za_name);
-	lde->lde_namelen = cpu_to_le16(namelen);
+	/* 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);
 
-	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);
+	dnode = ZFS_NO_OBJECT;
+	rc = -sa_lookup(sa_hdl, SA_ZPL_PARENT(uos), &dnode, 8);
+	sa_handle_destroy(sa_hdl);
+	if (rc)
+		RETURN(rc);
 
-		GOTO(out, rc = -EIO);
+	/* now get EA buffer */
+	rc = __osd_xattr_load(uos, dnode, &nvbuf);
+	if (rc)
+		GOTO(regular, rc);
+
+	/* XXX: if we get that far.. should we cache the result? */
+
+	/* 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);
 	}
 
-	rc = -zap_lookup(it->ozi_zc->zc_objset, it->ozi_zc->zc_zapobj,
-			 za->za_name, za->za_integer_length, 3, zde);
+regular:
+	/* no LMA attribute in SA, let's try regular EA */
+
+	/* 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(out, rc);
 
-	lde->lde_fid = zde->lzd_fid;
-	lde->lde_attrs = LUDA_FID;
-
-	/* append lustre attributes */
-	osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
+	dnode = ZFS_NO_OBJECT;
+	rc = -sa_lookup(sa_hdl, SA_ZPL_XATTR(uos), &dnode, 8);
+	sa_handle_destroy(sa_hdl);
+	if (rc)
+		GOTO(out, rc);
 
-	lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
+	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,
+				   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) {
+		GOTO(out, rc);
+	} else {
+		GOTO(out, rc = -EIO);
+	}
 
 out:
+	if (nvbuf != NULL)
+		nvlist_free(nvbuf);
 	RETURN(rc);
 }
 
-static __u64 osd_zap_it_store(const struct lu_env *env, const struct dt_it *di)
+static int osd_find_parent_fid(const struct lu_env *env, struct dt_object *o,
+			       struct lu_fid *fid)
 {
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	struct link_ea_header  *leh;
+	struct link_ea_entry   *lee;
+	struct lu_buf		buf;
+	int			rc;
+	ENTRY;
 
-	it->ozi_reset = 0;
-	RETURN(udmu_zap_cursor_serialize(it->ozi_zc));
-}
+	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);
+	if (rc == -ERANGE) {
+		rc = osd_xattr_get(env, o, &LU_BUF_NULL,
+				   XATTR_NAME_LINK, BYPASS_CAPA);
+		if (rc < 0)
+			RETURN(rc);
+		LASSERT(rc > 0);
+		OBD_ALLOC(buf.lb_buf, rc);
+		if (buf.lb_buf == NULL)
+			RETURN(-ENOMEM);
+		buf.lb_len = rc;
+		rc = osd_xattr_get(env, o, &buf, XATTR_NAME_LINK, BYPASS_CAPA);
+	}
+	if (rc < 0)
+		GOTO(out, rc);
+	if (rc < sizeof(*leh) + sizeof(*lee))
+		GOTO(out, rc = -EINVAL);
+
+	leh = buf.lb_buf;
+	if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
+		leh->leh_magic = LINK_EA_MAGIC;
+		leh->leh_reccount = __swab32(leh->leh_reccount);
+		leh->leh_len = __swab64(leh->leh_len);
+	}
+	if (leh->leh_magic != LINK_EA_MAGIC)
+		GOTO(out, rc = -EINVAL);
+	if (leh->leh_reccount == 0)
+		GOTO(out, rc = -ENODATA);
 
-/*
- * return status :
- *  rc == 0 -> ok, proceed.
- *  rc >  0 -> end of directory.
- *  rc <  0 -> error.  ( EOVERFLOW  can be masked.)
- */
-static int osd_zap_it_load(const struct lu_env *env,
-			const struct dt_it *di, __u64 hash)
-{
-	struct osd_zap_it *it = (struct osd_zap_it *)di;
-	struct osd_object *obj = it->ozi_obj;
-	struct osd_device *osd = osd_obj2dev(obj);
-	int                rc;
-	ENTRY;
+	lee = (struct link_ea_entry *)(leh + 1);
+	fid_be_to_cpu(fid, (const struct lu_fid *)&lee->lee_parent_fid);
+	rc = 0;
 
-	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);
-	it->ozi_reset = 0;
+out:
+	if (buf.lb_buf != osd_oti_get(env)->oti_buf)
+		OBD_FREE(buf.lb_buf, buf.lb_len);
+
+#if 0
+	/* this block can be enabled for additional verification
+	 * it's trying to match FID from LinkEA vs. FID from LMA */
+	if (rc == 0) {
+		struct lu_fid fid2;
+		int rc2;
+		rc2 = osd_find_parent_by_dnode(env, o, &fid2);
+		if (rc2 == 0)
+			if (lu_fid_eq(fid, &fid2) == 0)
+				CERROR("wrong parent: "DFID" != "DFID"\n",
+				       PFID(fid), PFID(&fid2));
+	}
+#endif
 
-	/* same as osd_zap_it_next()*/
-	rc = -udmu_zap_cursor_retrieve_key(env, it->ozi_zc, NULL,
-					   NAME_MAX + 1);
-	if (rc == 0)
-		RETURN(+1);
-	else if (rc == -ENOENT) /* end of dir*/
-		RETURN(0);
+	/* no LinkEA is found, let's try to find the fid in parent's LMA */
+	if (unlikely(rc != 0))
+		rc = osd_find_parent_by_dnode(env, o, fid);
 
 	RETURN(rc);
 }
@@ -395,11 +357,23 @@ static int osd_dir_lookup(const struct lu_env *env, struct dt_object *dt,
 	struct osd_thread_info *oti = osd_oti_get(env);
 	struct osd_object  *obj = osd_dt_obj(dt);
 	struct osd_device  *osd = osd_obj2dev(obj);
+	char		   *name = (char *)key;
 	int                 rc;
 	ENTRY;
 
 	LASSERT(udmu_object_is_zap(obj->oo_db));
 
+	if (name[0] == '.') {
+		if (name[1] == 0) {
+			const struct lu_fid *f = lu_object_fid(&dt->do_lu);
+			memcpy(rec, f, sizeof(*f));
+			RETURN(1);
+		} else if (name[1] == '.' && name[2] == 0) {
+			rc = osd_find_parent_fid(env, dt, (struct lu_fid *)rec);
+			RETURN(rc == 0 ? 1 : rc);
+		}
+	}
+
 	rc = -zap_lookup(osd->od_objset.os, obj->oo_db->db_object,
 			 (char *)key, 8, sizeof(oti->oti_zde) / 8,
 			 (void *)&oti->oti_zde);
@@ -498,6 +472,50 @@ static inline void osd_object_put(const struct lu_env *env,
 	lu_object_put(env, &obj->oo_dt.do_lu);
 }
 
+static int osd_seq_exists(const struct lu_env *env, struct osd_device *osd,
+			  obd_seq 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;
+
+	LASSERT(ss != NULL);
+	LASSERT(ss->ss_server_fld != NULL);
+
+	rc = osd_fld_lookup(env, osd, seq, range);
+	if (rc != 0) {
+		CERROR("%s: Can not lookup fld for "LPX64"\n",
+		       osd_name(osd), seq);
+		RETURN(0);
+	}
+
+	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);
+}
+
 /**
  *      Inserts (key, value) pair in \a directory object.
  *
@@ -519,10 +537,12 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 	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;
+	struct osd_object   *child = NULL;
 	__u32                attr;
+	char		    *name = (char *)key;
 	int                  rc;
 	ENTRY;
 
@@ -532,36 +552,63 @@ static int osd_dir_insert(const struct lu_env *env, struct dt_object *dt,
 	LASSERT(dt_object_exists(dt));
 	LASSERT(osd_invariant(parent));
 
-	/*
-	 * zfs_readdir() generates ./.. on fly, but
-	 * we want own entries (.. at least) with a fid
-	 */
-#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 3, 58, 0)
-#warning "fix '.' and '..' handling"
-#endif
-
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
-	child = osd_object_find(env, dt, fid);
-	if (IS_ERR(child))
-		RETURN(PTR_ERR(child));
+	rc = osd_remote_fid(env, osd, fid);
+	if (rc < 0) {
+		CERROR("%s: Can not find object "DFID": rc = %d\n",
+		       osd->od_svname, PFID(fid), rc);
+		RETURN(rc);
+	}
 
-	LASSERT(child->oo_db);
+	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(rec1->rec_type & S_IFMT);
+	} else {
+		/*
+		 * To simulate old Orion setups with ./..  stored in the
+		 * directories
+		 */
+		/* Insert local entry */
+		child = osd_object_find(env, dt, fid);
+		if (IS_ERR(child))
+			RETURN(PTR_ERR(child));
+
+		LASSERT(child->oo_db);
+		if (name[0] == '.') {
+			if (name[1] == 0) {
+				/* do not store ".", instead generate it
+				 * during iteration */
+				GOTO(out, rc = 0);
+			} else if (name[1] == '.' && name[2] == 0) {
+				/* 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),
+						 &child->oo_db->db_object,
+						 8, oh);
+				GOTO(out, rc);
+			}
+		}
+		CLASSERT(sizeof(oti->oti_zde.lzd_reg) == 8);
+		CLASSERT(sizeof(oti->oti_zde) % 8 == 0);
+		attr = child->oo_dt.do_lu.lo_header ->loh_attr;
+		oti->oti_zde.lzd_reg.zde_type = IFTODT(attr & S_IFMT);
+		oti->oti_zde.lzd_reg.zde_dnode = child->oo_db->db_object;
+	}
 
-	CLASSERT(sizeof(oti->oti_zde.lzd_reg) == 8);
-	CLASSERT(sizeof(oti->oti_zde) % 8 == 0);
-	attr = child->oo_dt.do_lu.lo_header ->loh_attr;
-	oti->oti_zde.lzd_reg.zde_type = IFTODT(attr & S_IFMT);
-	oti->oti_zde.lzd_reg.zde_dnode = child->oo_db->db_object;
 	oti->oti_zde.lzd_fid = *fid;
-
 	/* Insert (key,oid) into ZAP */
 	rc = -zap_add(osd->od_objset.os, parent->oo_db->db_object,
 		      (char *)key, 8, sizeof(oti->oti_zde) / 8,
 		      (void *)&oti->oti_zde, oh->ot_tx);
 
-	osd_object_put(env, child);
+out:
+	if (child != NULL)
+		osd_object_put(env, child);
 
 	RETURN(rc);
 }
@@ -597,6 +644,7 @@ static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
 	struct osd_device *osd = osd_obj2dev(obj);
 	struct osd_thandle *oh;
 	dmu_buf_t *zap_db = obj->oo_db;
+	char	  *name = (char *)key;
 	int rc;
 	ENTRY;
 
@@ -606,70 +654,498 @@ static int osd_dir_delete(const struct lu_env *env, struct dt_object *dt,
 	LASSERT(th != NULL);
 	oh = container_of0(th, struct osd_thandle, ot_super);
 
+	/*
+	 * In Orion . and .. were stored in the directory (not generated upon
+	 * request as now). we preserve them for backward compatibility
+	 */
+	if (name[0] == '.') {
+		if (name[1] == 0) {
+			RETURN(0);
+		} else if (name[1] == '.' && name[2] == 0) {
+			RETURN(0);
+		}
+	}
+
 	/* Remove key from the ZAP */
 	rc = -zap_remove(osd->od_objset.os, zap_db->db_object,
 			 (char *) key, oh->ot_tx);
 
-	if (rc && rc != -ENOENT)
+#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);
 
 	RETURN(rc);
 }
 
-static struct dt_index_operations osd_dir_ops = {
-	.dio_lookup         = osd_dir_lookup,
-	.dio_declare_insert = osd_declare_dir_insert,
-	.dio_insert         = osd_dir_insert,
-	.dio_declare_delete = osd_declare_dir_delete,
-	.dio_delete         = osd_dir_delete,
-	.dio_it     = {
-		.init     = osd_zap_it_init,
-		.fini     = osd_zap_it_fini,
-		.get      = osd_zap_it_get,
-		.put      = osd_zap_it_put,
-		.next     = osd_zap_it_next,
-		.key      = osd_zap_it_key,
-		.key_size = osd_zap_it_key_size,
-		.rec      = osd_zap_it_rec,
-		.store    = osd_zap_it_store,
-		.load     = osd_zap_it_load
-	}
-};
-
-/*
- * Primitives for index files using binary keys.
- * XXX: only 64-bit keys are supported for now.
- */
-
-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)
+static struct dt_it *osd_dir_it_init(const struct lu_env *env,
+				     struct dt_object *dt,
+				     __u32 unused,
+				     struct lustre_capa *capa)
 {
-	struct osd_object *obj = osd_dt_obj(dt);
-	struct osd_device *osd = osd_obj2dev(obj);
-	int                rc;
-	ENTRY;
+	struct osd_zap_it *it;
 
-	rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
-				(const __u64 *)key, 1, 8, obj->oo_recsize,
-				(void *)rec);
-	RETURN(rc == 0 ? 1 : rc);
+	it = (struct osd_zap_it *)osd_index_it_init(env, dt, unused, capa);
+	if (!IS_ERR(it))
+		it->ozi_pos = 0;
+
+	RETURN((struct dt_it *)it);
 }
 
-static int osd_declare_index_insert(const struct lu_env *env,
-				    struct dt_object *dt,
-				    const struct dt_rec *rec,
-				    const struct dt_key *key,
-				    struct thandle *th)
+/**
+ *  Move Iterator to record specified by \a key
+ *
+ *  \param  di      osd iterator
+ *  \param  key     key for index
+ *
+ *  \retval +ve  di points to record with least key not larger than key
+ *  \retval  0   di points to exact matched key
+ *  \retval -ve  failure
+ */
+static int osd_dir_it_get(const struct lu_env *env,
+			  struct dt_it *di, const struct dt_key *key)
 {
-	struct osd_object  *obj = osd_dt_obj(dt);
-	struct osd_thandle *oh;
+	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;
 
-	LASSERT(th != NULL);
-	oh = container_of0(th, struct osd_thandle, ot_super);
+	LASSERT(it);
+	LASSERT(it->ozi_zc);
 
-	LASSERT(obj->oo_db);
+	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);
+
+	/* XXX: implementation of the API is broken at the moment */
+	LASSERT(((const char *)key)[0] == 0);
+
+	if (name[0] == 0) {
+		it->ozi_pos = 0;
+		RETURN(1);
+	}
+
+	if (name[0] == '.') {
+		if (name[1] == 0) {
+			it->ozi_pos = 1;
+			GOTO(out, rc = 1);
+		} else if (name[1] == '.' && name[2] == 0) {
+			it->ozi_pos = 2;
+			GOTO(out, rc = 1);
+		}
+	}
+
+	/* neither . nor .. - some real record */
+	it->ozi_pos = 3;
+	rc = +1;
+
+out:
+	RETURN(rc);
+}
+
+static void osd_dir_it_put(const struct lu_env *env, struct dt_it *di)
+{
+	/* PBS: do nothing : ref are incremented at retrive and decreamented
+	 *      next/finish. */
+}
+
+/*
+ * in Orion . and .. were stored in the directory, while ZPL
+ * and current osd-zfs generate them up on request. so, we
+ * need to ignore previously stored . and ..
+ */
+static int osd_index_retrieve_skip_dots(struct osd_zap_it *it,
+					zap_attribute_t *za)
+{
+	int rc, isdot;
+
+	do {
+		rc = -zap_cursor_retrieve(it->ozi_zc, za);
+
+		isdot = 0;
+		if (unlikely(rc == 0 && za->za_name[0] == '.')) {
+			if (za->za_name[1] == 0) {
+				isdot = 1;
+			} else if (za->za_name[1] == '.' &&
+				   za->za_name[2] == 0) {
+				isdot = 1;
+			}
+			if (unlikely(isdot))
+				zap_cursor_advance(it->ozi_zc);
+		}
+	} while (unlikely(rc == 0 && isdot));
+
+	return rc;
+}
+
+/**
+ * to load a directory entry at a time and stored it in
+ * iterator's in-memory data structure.
+ *
+ * \param di, struct osd_it_ea, iterator's in memory structure
+ *
+ * \retval +ve, iterator reached to end
+ * \retval   0, iterator not reached to end
+ * \retval -ve, on error
+ */
+static int osd_dir_it_next(const struct lu_env *env, struct dt_it *di)
+{
+	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	zap_attribute_t	  *za = &osd_oti_get(env)->oti_za;
+	int		   rc;
+
+	/* temp. storage should be enough for any key supported by ZFS */
+	CLASSERT(sizeof(za->za_name) <= sizeof(it->ozi_name));
+
+	/*
+	 * the first ->next() moves the cursor to .
+	 * the second ->next() moves the cursor to ..
+	 * then we get to the real records and have to verify any exist
+	 */
+	if (it->ozi_pos <= 2) {
+		it->ozi_pos++;
+		if (it->ozi_pos <=2)
+			RETURN(0);
+	}
+
+	zap_cursor_advance(it->ozi_zc);
+
+	/*
+	 * According to current API we need to return error if its last entry.
+	 * zap_cursor_advance() does not return any value. So we need to call
+	 * retrieve to check if there is any record.  We should make
+	 * changes to Iterator API to not return status for this API
+	 */
+	rc = osd_index_retrieve_skip_dots(it, za);
+
+	if (rc == -ENOENT) /* end of dir */
+		RETURN(+1);
+
+	RETURN(rc);
+}
+
+static struct dt_key *osd_dir_it_key(const struct lu_env *env,
+				     const struct dt_it *di)
+{
+	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	zap_attribute_t	  *za = &osd_oti_get(env)->oti_za;
+	int		   rc = 0;
+	ENTRY;
+
+	if (it->ozi_pos <= 1) {
+		it->ozi_pos = 1;
+		RETURN((struct dt_key *)".");
+	} else if (it->ozi_pos == 2) {
+		RETURN((struct dt_key *)"..");
+	}
+
+	if ((rc = -zap_cursor_retrieve(it->ozi_zc, za)))
+		RETURN(ERR_PTR(rc));
+
+	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);
+}
+
+static int osd_dir_it_key_size(const struct lu_env *env, const struct dt_it *di)
+{
+	struct osd_zap_it *it = (struct osd_zap_it *)di;
+	zap_attribute_t	  *za = &osd_oti_get(env)->oti_za;
+	int		   rc;
+	ENTRY;
+
+	if (it->ozi_pos <= 1) {
+		it->ozi_pos = 1;
+		RETURN(2);
+	} else if (it->ozi_pos == 2) {
+		RETURN(3);
+	}
+
+	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);
+}
+
+static int osd_dir_it_rec(const struct lu_env *env, const struct dt_it *di,
+			  struct dt_rec *dtrec, __u32 attr)
+{
+	struct osd_zap_it   *it = (struct osd_zap_it *)di;
+	struct lu_dirent    *lde = (struct lu_dirent *)dtrec;
+	struct luz_direntry *zde = &osd_oti_get(env)->oti_zde;
+	zap_attribute_t     *za = &osd_oti_get(env)->oti_za;
+	int		     rc, namelen;
+	ENTRY;
+
+	if (it->ozi_pos <= 1) {
+		lde->lde_hash = cpu_to_le64(1);
+		strcpy(lde->lde_name, ".");
+		lde->lde_namelen = cpu_to_le16(1);
+		lde->lde_fid = *lu_object_fid(&it->ozi_obj->oo_dt.do_lu);
+		lde->lde_attrs = LUDA_FID;
+		/* append lustre attributes */
+		osd_it_append_attrs(lde, attr, 1, IFTODT(S_IFDIR));
+		lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(1, attr));
+		it->ozi_pos = 1;
+		GOTO(out, rc = 0);
+
+	} else if (it->ozi_pos == 2) {
+		lde->lde_hash = cpu_to_le64(2);
+		strcpy(lde->lde_name, "..");
+		lde->lde_namelen = cpu_to_le16(2);
+		lde->lde_attrs = LUDA_FID;
+		/* append lustre attributes */
+		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);
+	}
+
+	LASSERT(lde);
+
+	rc = -zap_cursor_retrieve(it->ozi_zc, za);
+	if (unlikely(rc != 0))
+		GOTO(out, rc);
+
+	lde->lde_hash = cpu_to_le64(udmu_zap_cursor_serialize(it->ozi_zc));
+	namelen = strlen(za->za_name);
+	if (namelen > NAME_MAX)
+		GOTO(out, rc = -EOVERFLOW);
+	strcpy(lde->lde_name, za->za_name);
+	lde->lde_namelen = cpu_to_le16(namelen);
+
+	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);
+
+		GOTO(out, rc = -EIO);
+	}
+
+	rc = -zap_lookup(it->ozi_zc->zc_objset, it->ozi_zc->zc_zapobj,
+			 za->za_name, za->za_integer_length, 3, zde);
+	if (rc)
+		GOTO(out, rc);
+
+	lde->lde_fid = zde->lzd_fid;
+	lde->lde_attrs = LUDA_FID;
+
+	/* append lustre attributes */
+	osd_it_append_attrs(lde, attr, namelen, zde->lzd_reg.zde_type);
+
+	lde->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
+
+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;
+	int		     rc, namelen = 0;
+	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;
+	__u64		   pos;
+	ENTRY;
+
+	if (it->ozi_pos <= 2)
+		pos = it->ozi_pos;
+	else
+		pos = udmu_zap_cursor_serialize(it->ozi_zc);
+
+	RETURN(pos);
+}
+
+/*
+ * return status :
+ *  rc == 0 -> end of directory.
+ *  rc >  0 -> ok, proceed.
+ *  rc <  0 -> error.  ( EOVERFLOW  can be masked.)
+ */
+static int osd_dir_it_load(const struct lu_env *env,
+			const struct dt_it *di, __u64 hash)
+{
+	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);
+
+	if (hash <= 2) {
+		it->ozi_pos = hash;
+		rc = +1;
+	} else {
+		it->ozi_pos = 3;
+		/* to return whether the end has been reached */
+		rc = osd_index_retrieve_skip_dots(it, za);
+		if (rc == 0)
+			rc = +1;
+		else if (rc == -ENOENT)
+			rc = 0;
+	}
+
+	RETURN(rc);
+}
+
+static struct dt_index_operations osd_dir_ops = {
+	.dio_lookup         = osd_dir_lookup,
+	.dio_declare_insert = osd_declare_dir_insert,
+	.dio_insert         = osd_dir_insert,
+	.dio_declare_delete = osd_declare_dir_delete,
+	.dio_delete         = osd_dir_delete,
+	.dio_it     = {
+		.init     = osd_dir_it_init,
+		.fini     = osd_index_it_fini,
+		.get      = osd_dir_it_get,
+		.put      = osd_dir_it_put,
+		.next     = osd_dir_it_next,
+		.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
+	}
+};
+
+/*
+ * Primitives for index files using binary keys.
+ */
+
+/* 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 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 = osd_prepare_key_uint64(obj, k, key);
+
+	rc = -zap_lookup_uint64(osd->od_objset.os, obj->oo_db->db_object,
+				k, rc, obj->oo_recusize, obj->oo_recsize,
+				(void *)rec);
+	RETURN(rc == 0 ? 1 : rc);
+}
+
+static int osd_declare_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 osd_object  *obj = osd_dt_obj(dt);
+	struct osd_thandle *oh;
+	ENTRY;
+
+	LASSERT(th != NULL);
+	oh = container_of0(th, struct osd_thandle, ot_super);
+
+	LASSERT(obj->oo_db);
 
 	dmu_tx_hold_bonus(oh->ot_tx, obj->oo_db->db_object);
 
@@ -689,6 +1165,7 @@ static int osd_index_insert(const struct lu_env *env, struct dt_object *dt,
 	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;
 
@@ -699,9 +1176,11 @@ 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,
+			     k, rc, obj->oo_recusize, obj->oo_recsize,
 			     (void *)rec, oh->ot_tx);
 	RETURN(rc);
 }
@@ -733,6 +1212,7 @@ static int osd_index_delete(const struct lu_env *env, struct dt_object *dt,
 	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;
 
@@ -740,9 +1220,11 @@ 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);
+				k, rc, oh->ot_tx);
 	RETURN(rc);
 }
 
@@ -757,8 +1239,12 @@ 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));
@@ -796,6 +1282,7 @@ 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;
@@ -806,7 +1293,7 @@ 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);
 }
@@ -814,8 +1301,9 @@ 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,
@@ -825,6 +1313,7 @@ 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;
 
@@ -833,9 +1322,11 @@ static int osd_index_it_rec(const struct lu_env *env, const struct dt_it *di,
 	if (rc)
 		RETURN(rc);
 
+	rc = osd_prepare_key_uint64(obj, k, (const struct dt_key *)za->za_name);
+
 	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);
+				k, rc, obj->oo_recusize, obj->oo_recsize,
+				(void *)rec);
 	RETURN(rc);
 }
 
@@ -883,10 +1374,10 @@ static struct dt_index_operations osd_index_ops = {
 	.dio_declare_delete	= osd_declare_index_delete,
 	.dio_delete		= osd_index_delete,
 	.dio_it	= {
-		.init		= osd_zap_it_init,
-		.fini		= osd_zap_it_fini,
+		.init		= osd_index_it_init,
+		.fini		= osd_index_it_fini,
 		.get		= osd_index_it_get,
-		.put		= osd_zap_it_put,
+		.put		= osd_index_it_put,
 		.next		= osd_index_it_next,
 		.key		= osd_index_it_key,
 		.key_size	= osd_index_it_key_size,
@@ -896,6 +1387,216 @@ 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 lustre_capa *capa)
+{
+	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_objset.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;
+	udmu_objset_t		*uos = &dev->od_objset;
+	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(uos->os, &it->mit_prefetched_dnode,
+				      B_FALSE, 0);
+		if (unlikely(rc != 0))
+			break;
+
+		/* dmu_prefetch() was exported in 0.6.2, if you use with
+		 * an older release, just comment it out - this is an
+		 * optimization */
+		dmu_prefetch(uos->os, it->mit_prefetched_dnode, 0, 0);
+
+		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;
+	udmu_objset_t		*uos = &dev->od_objset;
+	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(uos->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(uos, 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)
 {
@@ -911,9 +1612,10 @@ int osd_index_try(const struct lu_env *env, struct dt_object *dt,
 	if (feat->dif_flags & DT_IND_RANGE)
 		RETURN(-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;
+		RETURN(0);
+	}
 
 	LASSERT(obj->oo_db != NULL);
 	if (likely(feat == &dt_directory_features)) {
@@ -931,10 +1633,9 @@ int osd_index_try(const struct lu_env *env, struct dt_object *dt,
 		if ((feat->dif_flags & ~DT_IND_UPDATE) != 0)
 			RETURN(-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))
+		if (feat->dif_keysize_max > ZAP_MAXNAMELEN)
+			RETURN(-E2BIG);
+		if (feat->dif_keysize_max != feat->dif_keysize_min)
 			RETURN(-EINVAL);
 
 		/* As for the record size, it should be a multiple of 8 bytes
@@ -942,14 +1643,20 @@ int osd_index_try(const struct lu_env *env, struct dt_object *dt,
 		 */
 		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)))
+		if (feat->dif_recsize_max != feat->dif_recsize_min)
 			RETURN(-EINVAL);
 
-		obj->oo_recsize = feat->dif_recsize_max / sizeof(__u64);
+		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);
 }
-