Whamcloud - gitweb
LU-4843 mdt: disallow old clients access striped dir
[fs/lustre-release.git] / lustre / mdt / mdt_reint.c
index 6c657da..8cb8cb7 100644 (file)
@@ -245,6 +245,79 @@ int mdt_lookup_version_check(struct mdt_thread_info *info,
 
 }
 
+/**
+ * mdt_remote_permission: Check whether the remote operation is permitted,
+ *
+ * Before we implement async cross-MDT updates (DNE phase 2). There are a few
+ * limitations here:
+ *
+ * 1.Only sysadmin can create remote directory and striped directory and
+ *   migrate directory now, unless
+ *   lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
+ * 2.Remote directory can only be created on MDT0, unless
+ *   lctl set_param mdt.*.enable_remote_dir = 1
+ * 3.Only new clients can access remote dir( >= 2.4) and striped dir(>= 2.6),
+ *   old client will return -ENOTSUPP.
+ *
+ * XXX these check are only needed for remote synchronization, once async
+ * update is supported, these check will be removed.
+ *
+ * param[in]info:      execution environment.
+ * param[in]parent:    the directory of this operation.
+ * param[in]child:     the child of this operation.
+ *
+ * retval      = 0 remote operation is allowed.
+ *              < 0 remote operation is denied.
+ */
+static int mdt_remote_permission(struct mdt_thread_info *info,
+                                struct mdt_object *parent,
+                                struct mdt_object *child)
+{
+       struct mdt_device       *mdt = info->mti_mdt;
+       struct lu_ucred         *uc  = mdt_ucred(info);
+       struct md_op_spec       *spec = &info->mti_spec;
+       struct lu_attr          *attr = &info->mti_attr.ma_attr;
+       struct obd_export       *exp = mdt_info_req(info)->rq_export;
+
+       /* Only check create remote directory, striped directory and
+        * migration */
+       if (mdt_object_remote(parent) == 0 && mdt_object_remote(child) == 0 &&
+           !(S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
+                                       spec->u.sp_ea.eadatalen != 0) &&
+           info->mti_rr.rr_opcode != REINT_MIGRATE)
+               return 0;
+
+       if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
+               if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
+                   mdt->mdt_enable_remote_dir_gid != -1)
+                       return -EPERM;
+       }
+
+       if (mdt->mdt_enable_remote_dir == 0) {
+               struct seq_server_site  *ss = mdt_seq_site(mdt);
+               struct lu_seq_range     range = { 0 };
+               int                     rc;
+
+               fld_range_set_type(&range, LU_SEQ_RANGE_MDT);
+               rc = fld_server_lookup(info->mti_env, ss->ss_server_fld,
+                                      fid_seq(mdt_object_fid(parent)), &range);
+               if (rc != 0)
+                       return rc;
+
+               if (range.lsr_index != 0)
+                       return -EPERM;
+       }
+
+       if (!mdt_is_dne_client(exp))
+               return -ENOTSUPP;
+
+       if (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
+           spec->u.sp_ea.eadatalen != 0 && !mdt_is_striped_client(exp))
+               return -ENOTSUPP;
+
+       return 0;
+}
+
 /*
  * VBR: we save three versions in reply:
  * 0 - parent. Check that parent version is the same during replay.
@@ -314,39 +387,11 @@ static int mdt_md_create(struct mdt_thread_info *info)
         if (likely(!IS_ERR(child))) {
                 struct md_object *next = mdt_object_child(parent);
 
-               if (mdt_object_remote(child)) {
-                       struct seq_server_site *ss;
-                       struct lu_ucred *uc  = mdt_ucred(info);
-
-                       if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
-                               if (uc->uc_gid !=
-                                   mdt->mdt_enable_remote_dir_gid &&
-                                   mdt->mdt_enable_remote_dir_gid != -1) {
-                                       CERROR("%s: Creating remote dir is only"
-                                              " permitted for administrator or"
-                                              " set mdt_enable_remote_dir_gid:"
-                                              " rc = %d\n",
-                                               mdt_obd_name(mdt), -EPERM);
-                                       GOTO(out_put_child, rc = -EPERM);
-                               }
-                       }
-
-                       ss = mdt_seq_site(mdt);
-                       if (ss->ss_node_id != 0 &&
-                           mdt->mdt_enable_remote_dir == 0) {
-                               CERROR("%s: remote dir is only permitted on"
-                                      " MDT0 or set_param"
-                                      " mdt.*.enable_remote_dir=1\n",
-                                      mdt_obd_name(mdt));
-                               GOTO(out_put_child, rc = -EPERM);
-                       }
-                       if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
-                               /* Return -EIO for old client */
-                               GOTO(out_put_child, rc = -EIO);
-                       }
+               rc = mdt_remote_permission(info, parent, child);
+               if (rc != 0)
+                       GOTO(out_put_child, rc);
 
-               }
-                ma->ma_need = MA_INODE;
+               ma->ma_need = MA_INODE;
                 ma->ma_valid = 0;
                 /* capa for cross-ref will be stored here */
                 ma->ma_capa = req_capsule_server_get(info->mti_pill,
@@ -399,6 +444,8 @@ put_parent:
 
 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
                             struct mdt_object *obj, __u64 ibits,
+                            struct mdt_lock_handle *s0_lh,
+                            struct mdt_object *s0_obj,
                             struct ldlm_enqueue_info *einfo)
 {
        ldlm_policy_data_t      *policy = &mti->mti_policy;
@@ -408,6 +455,12 @@ static int mdt_unlock_slaves(struct mdt_thread_info *mti,
        if (!S_ISDIR(obj->mot_header.loh_attr))
                RETURN(0);
 
+       /* Unlock stripe 0 */
+       if (s0_lh != NULL && lustre_handle_is_used(&s0_lh->mlh_reg_lh)) {
+               LASSERT(s0_obj != NULL);
+               mdt_object_unlock_put(mti, s0_obj, s0_lh, 1);
+       }
+
        memset(policy, 0, sizeof(*policy));
        policy->l_inodebits.bits = ibits;
 
@@ -422,15 +475,51 @@ static int mdt_unlock_slaves(struct mdt_thread_info *mti,
  **/
 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
                           ldlm_mode_t mode, __u64 ibits,
+                          struct mdt_lock_handle *s0_lh,
+                          struct mdt_object **s0_objp,
                           struct ldlm_enqueue_info *einfo)
 {
        ldlm_policy_data_t      *policy = &mti->mti_policy;
-       int                     rc;
+       struct lu_buf           *buf = &mti->mti_buf;
+       struct lmv_mds_md_v1    *lmv;
+       struct lu_fid           *fid = &mti->mti_tmp_fid1;
+       int rc;
        ENTRY;
 
        if (!S_ISDIR(obj->mot_header.loh_attr))
                RETURN(0);
 
+       buf->lb_buf = mti->mti_xattr_buf;
+       buf->lb_len = sizeof(mti->mti_xattr_buf);
+       rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), buf,
+                         XATTR_NAME_LMV);
+       if (rc == -ERANGE) {
+               rc = mdt_big_xattr_get(mti, obj, XATTR_NAME_LMV);
+               if (rc > 0) {
+                       buf->lb_buf = mti->mti_big_lmm;
+                       buf->lb_len = mti->mti_big_lmmsize;
+               }
+       }
+
+       if (rc == -ENODATA || rc == -ENOENT)
+               RETURN(0);
+
+       if (rc <= 0)
+               RETURN(rc);
+
+       lmv = buf->lb_buf;
+       if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
+               RETURN(-EINVAL);
+
+       /* Sigh, 0_stripe and master object are different
+        * object, though they are in the same MDT, to avoid
+        * adding osd_object_lock here, so we will enqueue the
+        * stripe0 lock in MDT0 for now */
+       fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[0]);
+       *s0_objp = mdt_object_find_lock(mti, fid, s0_lh, ibits);
+       if (IS_ERR(*s0_objp))
+               RETURN(PTR_ERR(*s0_objp));
+
        memset(einfo, 0, sizeof(*einfo));
        einfo->ei_type = LDLM_IBITS;
        einfo->ei_mode = mode;
@@ -448,12 +537,14 @@ static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
                  struct md_attr *ma, int flags)
 {
-        struct mdt_lock_handle  *lh;
-        int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
-        __u64 lockpart = MDS_INODELOCK_UPDATE;
+       struct mdt_lock_handle  *lh;
+       int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
+       __u64 lockpart = MDS_INODELOCK_UPDATE;
        struct ldlm_enqueue_info *einfo = &info->mti_einfo;
-        int rc;
-        ENTRY;
+       struct mdt_lock_handle  *s0_lh;
+       struct mdt_object       *s0_obj = NULL;
+       int rc;
+       ENTRY;
 
        /* attr shouldn't be set on remote object */
        LASSERT(!mdt_object_remote(mo));
@@ -472,7 +563,9 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
         if (rc != 0)
                 RETURN(rc);
 
-       rc = mdt_lock_slaves(info, mo, LCK_EX, lockpart, einfo);
+       s0_lh = &info->mti_lh[MDT_LH_LOCAL];
+       mdt_lock_reg_init(s0_lh, LCK_EX);
+       rc = mdt_lock_slaves(info, mo, LCK_PW, lockpart, s0_lh, &s0_obj, einfo);
        if (rc != 0)
                GOTO(out_unlock, rc);
 
@@ -511,7 +604,7 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 
         EXIT;
 out_unlock:
-       mdt_unlock_slaves(info, mo, lockpart, einfo);
+       mdt_unlock_slaves(info, mo, lockpart, s0_lh, s0_obj, einfo);
         mdt_object_unlock(info, mo, lh, rc);
         return rc;
 }
@@ -769,6 +862,8 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
         struct mdt_lock_handle  *parent_lh;
         struct mdt_lock_handle  *child_lh;
        struct ldlm_enqueue_info *einfo = &info->mti_einfo;
+       struct mdt_lock_handle  *s0_lh = NULL;
+       struct mdt_object       *s0_obj = NULL;
        int                     rc;
        int                     no_name = 0;
        ENTRY;
@@ -860,8 +955,8 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
                       PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 
                if (!mdt_is_dne_client(req->rq_export))
-                       /* Return -EIO for old client */
-                       GOTO(put_child, rc = -EIO);
+                       /* Return -ENOTSUPP for old client */
+                       GOTO(put_child, rc = -ENOTSUPP);
 
                if (info->mti_spec.sp_rm_entry) {
                        struct lu_ucred *uc  = mdt_ucred(info);
@@ -916,7 +1011,10 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
        ma->ma_need = MA_INODE;
        ma->ma_valid = 0;
 
-       rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, einfo);
+       s0_lh = &info->mti_lh[MDT_LH_LOCAL];
+       mdt_lock_reg_init(s0_lh, LCK_EX);
+       rc = mdt_lock_slaves(info, mc, LCK_EX, MDS_INODELOCK_UPDATE, s0_lh,
+                            &s0_obj, einfo);
        if (rc != 0)
                GOTO(unlock_child, rc);
 
@@ -961,13 +1059,8 @@ static int mdt_reint_unlink(struct mdt_thread_info *info,
         EXIT;
 
 unlock_child:
-       mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, einfo);
+       mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo);
        mdt_object_unlock(info, mc, child_lh, rc);
-
-       /* Since we do not need reply md striped dir info to client, so
-        * reset mti_big_lmm_used to avoid confusing mdt_fix_reply */
-       if (info->mti_big_lmm_used)
-               info->mti_big_lmm_used = 0;
 put_child:
        mdt_object_put(info->mti_env, mc);
 unlock_parent:
@@ -1195,17 +1288,18 @@ static void mdt_rename_unlock(struct lustre_handle *lh)
  * target. Source should not be ancestor of target dir. May be other rename
  * checks can be moved here later.
  */
-static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
+static int mdt_rename_sanity(struct mdt_thread_info *info,
+                            const struct lu_fid *dir_fid,
+                            const struct lu_fid *fid)
 {
-        struct mdt_reint_record *rr = &info->mti_rr;
-        struct lu_fid dst_fid = *rr->rr_fid2;
         struct mdt_object *dst;
+       struct lu_fid dst_fid = *dir_fid;
         int rc = 0;
         ENTRY;
 
        /* If the source and target are in the same directory, they can not
         * be parent/child relationship, so subdir check is not needed */
-       if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
+       if (lu_fid_eq(dir_fid, fid))
                return 0;
 
        do {
@@ -1217,7 +1311,13 @@ static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid)
                                           &dst_fid);
                        mdt_object_put(info->mti_env, dst);
                        if (rc != -EREMOTE && rc < 0) {
-                               CERROR("Failed mdo_is_subdir(), rc %d\n", rc);
+                               CERROR("%s: failed subdir check in "DFID" for "
+                                      DFID": rc = %d\n",
+                                      mdt_obd_name(info->mti_mdt),
+                                      PFID(dir_fid), PFID(fid), rc);
+                               /* Return EINVAL only if a parent is the @fid */
+                               if (rc == -EINVAL)
+                                       rc = -EIO;
                        } else {
                                /* check the found fid */
                                if (lu_fid_eq(&dst_fid, fid))
@@ -1358,6 +1458,7 @@ static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
 
        CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
               PNAME(&rr->rr_name), PFID(rr->rr_fid2));
+
        /* 1: lock the source dir. */
        msrcdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
        if (IS_ERR(msrcdir)) {
@@ -1412,6 +1513,10 @@ static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
                GOTO(out_put_child, rc = -EPERM);
        }
 
+       rc = mdt_remote_permission(info, msrcdir, mold);
+       if (rc != 0)
+               GOTO(out_put_child, rc);
+
        /* 3: iterate the linkea of the object and lock all of the objects */
        CFS_INIT_LIST_HEAD(&lock_list);
        rc = mdt_lock_objects_in_linkea(info, mold, msrcdir, &lock_list);
@@ -1440,7 +1545,7 @@ static int mdt_reint_migrate_internal(struct mdt_thread_info *info,
 
                lmv_le_to_cpu(ma->ma_lmv, ma->ma_lmv);
                lmm1 = &ma->ma_lmv->lmv_md_v1;
-               if (lmm1->lmv_magic != LMV_MAGIC_MIGRATE) {
+               if (!(lmm1->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) {
                        CERROR("%s: can not migrate striped dir "DFID
                               ": rc = %d\n", mdt_obd_name(info->mti_mdt),
                               PFID(mdt_object_fid(mold)), -EPERM);
@@ -1517,6 +1622,144 @@ out_put_parent:
        RETURN(rc);
 }
 
+static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
+                                               const struct lu_fid *fid,
+                                               int idx)
+{
+       struct mdt_object *dir;
+       int rc;
+       ENTRY;
+
+       dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
+       if (IS_ERR(dir))
+               RETURN(dir);
+
+       /* check early, the real version will be saved after locking */
+       rc = mdt_version_get_check(info, dir, idx);
+       if (rc)
+               GOTO(out_put, rc);
+
+       RETURN(dir);
+out_put:
+       mdt_object_put(info->mti_env, dir);
+       return ERR_PTR(rc);
+}
+
+static int mdt_object_lock_save(struct mdt_thread_info *info,
+                               struct mdt_object *dir,
+                               struct mdt_lock_handle *lh,
+                               int idx)
+{
+       int rc;
+
+       /* we lock the target dir if it is local */
+       rc = mdt_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
+                            MDT_LOCAL_LOCK);
+       if (rc != 0)
+               return rc;
+
+       /* get and save correct version after locking */
+       mdt_version_get_save(info, dir, idx);
+       return 0;
+}
+
+
+static int mdt_rename_parents_lock(struct mdt_thread_info *info,
+                                  struct mdt_object **srcp,
+                                  struct mdt_object **tgtp)
+{
+       struct mdt_reint_record *rr = &info->mti_rr;
+       const struct lu_fid     *fid_src = rr->rr_fid1;
+       const struct lu_fid     *fid_tgt = rr->rr_fid2;
+       struct mdt_lock_handle  *lh_src = &info->mti_lh[MDT_LH_PARENT];
+       struct mdt_lock_handle  *lh_tgt = &info->mti_lh[MDT_LH_CHILD];
+       struct mdt_object       *src;
+       struct mdt_object       *tgt;
+       int                      reverse = 0;
+       int                      rc;
+       ENTRY;
+
+       /* Check if the @src is not a child of the @tgt, otherwise a
+        * reverse locking must take place.
+        *
+        * Note: cannot be called after object_find, because if the object
+        * is destroyed in between it gets stuck in lu_object_find_at(),
+        * waiting for the last ref. */
+       rc = mdt_rename_sanity(info, fid_src, fid_tgt);
+       if (rc == -EINVAL)
+               reverse = 1;
+       else if (rc)
+               RETURN(rc);
+
+       /* find both parents. */
+       src = mdt_object_find_check(info, fid_src, 0);
+       if (IS_ERR(src))
+               RETURN(PTR_ERR(src));
+
+       OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
+
+       if (lu_fid_eq(fid_src, fid_tgt)) {
+               tgt = src;
+               mdt_object_get(info->mti_env, tgt);
+       } else {
+               tgt = mdt_object_find_check(info, fid_tgt, 1);
+               if (IS_ERR(tgt))
+                       GOTO(err_src_put, rc = PTR_ERR(tgt));
+
+               if (unlikely(mdt_object_remote(tgt))) {
+                       CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
+                              "on different MDTs\n", PFID(fid_src),
+                              PFID(fid_tgt));
+                       GOTO(err_tgt_put, rc = -EXDEV);
+               }
+       }
+
+       /* lock parents in the proper order. */
+       if (reverse) {
+               rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
+               if (rc)
+                       GOTO(err_tgt_put, rc);
+
+               OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
+
+               rc = mdt_object_lock_save(info, src, lh_src, 0);
+       } else {
+               rc = mdt_object_lock_save(info, src, lh_src, 0);
+               if (rc)
+                       GOTO(err_tgt_put, rc);
+
+               OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
+
+               if (tgt != src)
+                       rc = mdt_object_lock_save(info, tgt, lh_tgt, 1);
+               else if (lh_src->mlh_pdo_hash != lh_tgt->mlh_pdo_hash) {
+                       rc = mdt_pdir_hash_lock(info, lh_tgt, tgt,
+                                               MDS_INODELOCK_UPDATE);
+                       OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
+               }
+       }
+       if (rc)
+               GOTO(err_unlock, rc);
+
+       if (lu_object_is_dying(&tgt->mot_header))
+               GOTO(err_unlock, rc = -ENOENT);
+
+       *srcp = src;
+       *tgtp = tgt;
+       RETURN(0);
+
+err_unlock:
+       /* The order does not matter as the handle is checked inside,
+        * as well as not used handle. */
+       mdt_object_unlock(info, src, lh_src, rc);
+       mdt_object_unlock(info, tgt, lh_tgt, rc);
+err_tgt_put:
+       mdt_object_put(info->mti_env, tgt);
+err_src_put:
+       mdt_object_put(info->mti_env, src);
+       RETURN(rc);
+}
+
 /*
  * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
  * 2 - src child; 3 - tgt child.
@@ -1539,8 +1782,8 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info,
        struct mdt_reint_record *rr = &info->mti_rr;
        struct md_attr          *ma = &info->mti_attr;
        struct ptlrpc_request   *req = mdt_info_req(info);
-       struct mdt_object       *msrcdir;
-       struct mdt_object       *mtgtdir;
+       struct mdt_object       *msrcdir = NULL;
+       struct mdt_object       *mtgtdir = NULL;
        struct mdt_object       *mold;
        struct mdt_object       *mnew = NULL;
        struct mdt_lock_handle  *lh_srcdirp;
@@ -1556,78 +1799,39 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info,
                  PFID(rr->rr_fid1), PNAME(&rr->rr_name),
                  PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
 
-       /* step 1: lock the source dir. */
        lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
        mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
-       msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
-                                      MDS_INODELOCK_UPDATE);
-       if (IS_ERR(msrcdir))
-               RETURN(PTR_ERR(msrcdir));
-
-       rc = mdt_version_get_check_save(info, msrcdir, 0);
-       if (rc)
-               GOTO(out_unlock_source, rc);
-
-       /* step 2: find & lock the target dir. */
        lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
        mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
-       if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
-               mdt_object_get(info->mti_env, msrcdir);
-               mtgtdir = msrcdir;
-               if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
-                       rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
-                                        MDS_INODELOCK_UPDATE);
-                       if (rc != 0)
-                               GOTO(out_unlock_source, rc);
-                       OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
-               }
-       } else {
-               mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
-                                         rr->rr_fid2);
-               if (IS_ERR(mtgtdir))
-                       GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
 
-               /* check early, the real version will be saved after locking */
-               rc = mdt_version_get_check(info, mtgtdir, 1);
-               if (rc)
-                       GOTO(out_put_target, rc);
+       /* step 1&2: lock the source and target dirs. */
+       rc = mdt_rename_parents_lock(info, &msrcdir, &mtgtdir);
+       if (rc)
+               RETURN(rc);
 
-               if (unlikely(mdt_object_remote(mtgtdir))) {
-                       CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
-                              "on different MDTs\n", PFID(rr->rr_fid1),
-                              PFID(rr->rr_fid2));
-                       GOTO(out_put_target, rc = -EXDEV);
-               } else {
-                       if (likely(mdt_object_exists(mtgtdir))) {
-                               /* we lock the target dir if it is local */
-                               rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
-                                                    MDS_INODELOCK_UPDATE,
-                                                    MDT_LOCAL_LOCK);
-                               if (rc != 0)
-                                       GOTO(out_put_target, rc);
-                               /* get and save correct version after locking */
-                               mdt_version_get_save(info, mtgtdir, 1);
-                       } else {
-                               GOTO(out_put_target, rc = -ESTALE);
-                       }
-               }
-       }
+       OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
 
        /* step 3: find & lock the old object. */
        fid_zero(old_fid);
        rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
        if (rc != 0)
-               GOTO(out_unlock_target, rc);
+               GOTO(out_unlock_parents, rc);
 
        if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
-               GOTO(out_unlock_target, rc = -EINVAL);
+               GOTO(out_unlock_parents, rc = -EINVAL);
 
        if (!fid_is_md_operative(old_fid))
-               GOTO(out_unlock_target, rc = -EPERM);
+               GOTO(out_unlock_parents, rc = -EPERM);
 
        mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
        if (IS_ERR(mold))
-               GOTO(out_unlock_target, rc = PTR_ERR(mold));
+               GOTO(out_unlock_parents, rc = PTR_ERR(mold));
+
+       /* Check if @mtgtdir is subdir of @mold, before locking child
+        * to avoid reverse locking. */
+       rc = mdt_rename_sanity(info, rr->rr_fid2, old_fid);
+       if (rc)
+               GOTO(out_put_old, rc);
 
        tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
        /* save version after locking */
@@ -1675,6 +1879,12 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info,
                if (rc != 0)
                        GOTO(out_put_new, rc);
 
+               /* Check if @msrcdir is subdir of @mnew, before locking child
+                * to avoid reverse locking. */
+               rc = mdt_rename_sanity(info, rr->rr_fid1, new_fid);
+               if (rc)
+                       GOTO(out_unlock_old, rc);
+
                /* We used to acquire MDS_INODELOCK_FULL here but we
                 * can't do this now because a running HSM restore on
                 * the rename onto victim will hold the layout
@@ -1719,11 +1929,6 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info,
        mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
                       OBD_FAIL_MDS_REINT_RENAME_WRITE);
 
-       /* Check if @dst is subdir of @src. */
-       rc = mdt_rename_sanity(info, old_fid);
-       if (rc)
-               GOTO(out_unlock_new, rc);
-
        if (mnew != NULL)
                mutex_lock(&mnew->mot_lov_mutex);
 
@@ -1746,7 +1951,6 @@ static int mdt_reint_rename_internal(struct mdt_thread_info *info,
        }
 
        EXIT;
-out_unlock_new:
        if (mnew != NULL)
                mdt_object_unlock(info, mnew, lh_newp, rc);
 out_unlock_old:
@@ -1756,11 +1960,8 @@ out_put_new:
                mdt_object_put(info->mti_env, mnew);
 out_put_old:
        mdt_object_put(info->mti_env, mold);
-out_unlock_target:
-       mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
-out_put_target:
-       mdt_object_put(info->mti_env, mtgtdir);
-out_unlock_source:
+out_unlock_parents:
+       mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc);
        mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
        return rc;
 }