Whamcloud - gitweb
LU-3335 scrub: purge inconsistenct objects after OI scrub
[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_scrub.c
index 1c0c78f..72ed5b8 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012, Intel Corporation.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  */
 /*
  * lustre/osd-ldiskfs/osd_scrub.c
@@ -35,9 +35,6 @@
  * Author: Fan Yong <yong.fan@whamcloud.com>
  */
 
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
 #define DEBUG_SUBSYSTEM S_MDS
 
 #include <lustre/lustre_idl.h>
@@ -50,6 +47,8 @@
 
 #define HALF_SEC       (CFS_HZ >> 1)
 
+#define OSD_OTABLE_MAX_HASH            0x00000000ffffffffULL
+
 #define SCRUB_NEXT_BREAK       1 /* exit current loop and process next group */
 #define SCRUB_NEXT_CONTINUE    2 /* skip current object and process next bit */
 #define SCRUB_NEXT_EXIT        3 /* exit all the loops */
@@ -77,6 +76,13 @@ static inline int osd_scrub_has_window(struct osd_scrub *scrub,
        return scrub->os_pos_current < ooc->ooc_pos_preload + SCRUB_WINDOW_SIZE;
 }
 
+/**
+ * update/insert/delete the specified OI mapping (@fid @id) according to the ops
+ *
+ * \retval   1, changed nothing
+ * \retval   0, changed successfully
+ * \retval -ve, on error
+ */
 static int osd_scrub_refresh_mapping(struct osd_thread_info *info,
                                     struct osd_device *dev,
                                     const struct lu_fid *fid,
@@ -91,7 +97,8 @@ static int osd_scrub_refresh_mapping(struct osd_thread_info *info,
        ENTRY;
 
        fid_cpu_to_be(oi_fid, fid);
-       osd_id_pack(oi_id, id);
+       if (id != NULL)
+               osd_id_pack(oi_id, id);
        jh = ldiskfs_journal_start_sb(osd_sb(dev),
                                      osd_dto_credits_noquota[ops]);
        if (IS_ERR(jh)) {
@@ -110,13 +117,19 @@ static int osd_scrub_refresh_mapping(struct osd_thread_info *info,
                RETURN(-ENOMEM);
        }
 
-       if (ops == DTO_INDEX_UPDATE) {
+       switch (ops) {
+       case DTO_INDEX_UPDATE:
                rc = iam_update(jh, bag, (const struct iam_key *)oi_fid,
                                (struct iam_rec *)oi_id, ipd);
-       } else {
+               if (unlikely(rc == -ENOENT)) {
+                       /* Some unlink thread may removed the OI mapping. */
+                       rc = 1;
+               }
+               break;
+       case DTO_INDEX_INSERT:
                rc = iam_insert(jh, bag, (const struct iam_key *)oi_fid,
                                (struct iam_rec *)oi_id, ipd);
-               if (rc == -EEXIST) {
+               if (unlikely(rc == -EEXIST)) {
                        rc = 1;
                        /* XXX: There are trouble things when adding OI
                         *      mapping for IGIF object, which may cause
@@ -150,6 +163,18 @@ static int osd_scrub_refresh_mapping(struct osd_thread_info *info,
                         *
                         *      Anyway, it is rare, only exists in theory. */
                }
+               break;
+       case DTO_INDEX_DELETE:
+               rc = iam_delete(jh, bag, (const struct iam_key *)oi_fid, ipd);
+               if (rc == -ENOENT) {
+                       /* It is normal that the unlink thread has removed the
+                        * OI mapping already. */
+                       rc = 1;
+               }
+               break;
+       default:
+               LASSERTF(0, "Unexpected ops %d\n", ops);
+               break;
        }
        osd_ipd_put(info->oti_env, bag, ipd);
        ldiskfs_journal_stop(jh);
@@ -415,7 +440,8 @@ osd_scrub_check_update(struct osd_thread_info *info, struct osd_device *dev,
        if (fid_is_igif(fid))
                sf->sf_items_igif++;
 
-       if (val == SCRUB_NEXT_NOLMA && !dev->od_handle_nolma)
+       if ((val == SCRUB_NEXT_NOLMA) &&
+           (!dev->od_handle_nolma || OBD_FAIL_CHECK(OBD_FAIL_FID_NOLMA)))
                GOTO(out, rc = 0);
 
        if ((oii != NULL && oii->oii_insert) || (val == SCRUB_NEXT_NOLMA))
@@ -441,10 +467,8 @@ iget:
                        GOTO(out, rc);
                }
 
-               /* Prevent the inode to be unlinked during OI scrub. */
-               mutex_lock(&inode->i_mutex);
+               /* Check whether the inode to be unlinked during OI scrub. */
                if (unlikely(inode->i_nlink == 0)) {
-                       mutex_unlock(&inode->i_mutex);
                        iput(inode);
                        GOTO(out, rc = 0);
                }
@@ -452,11 +476,17 @@ iget:
                ops = DTO_INDEX_INSERT;
                idx = osd_oi_fid2idx(dev, fid);
                if (val == SCRUB_NEXT_NOLMA) {
-                       rc = osd_ea_fid_set(info, inode, fid);
+                       sf->sf_flags |= SF_UPGRADE;
+                       scrub->os_full_speed = 1;
+                       rc = osd_ea_fid_set(info, inode, fid, 0);
                        if (rc != 0)
                                GOTO(out, rc);
+
+                       if (!(sf->sf_flags & SF_INCONSISTENT))
+                               dev->od_igif_inoi = 0;
                } else {
-                       sf->sf_flags |= SF_RECREATED | SF_INCONSISTENT;
+                       sf->sf_flags |= SF_RECREATED;
+                       scrub->os_full_speed = 1;
                        if (unlikely(!ldiskfs_test_bit(idx, sf->sf_oi_bitmap)))
                                ldiskfs_set_bit(idx, sf->sf_oi_bitmap);
                }
@@ -464,6 +494,21 @@ iget:
                GOTO(out, rc = 0);
        } else {
                sf->sf_flags |= SF_INCONSISTENT;
+               scrub->os_full_speed = 1;
+
+               /* XXX: If the device is restored from file-level backup, then
+                *      some IGIFs may have been already in OI files, and some
+                *      may be not yet. Means upgrading from 1.8 may be partly
+                *      processed, but some clients may hold some immobilized
+                *      IGIFs, and use them to access related objects. Under
+                *      such case, OSD does not know whether an given IGIF has
+                *      been processed or to be processed, and it also cannot
+                *      generate local ino#/gen# directly from the immobilized
+                *      IGIF because of the backup/restore. Then force OSD to
+                *      lookup the given IGIF in OI files, and if no entry,
+                *      then ask the client to retry after upgrading completed.
+                *      No better choice. */
+               dev->od_igif_inoi = 1;
        }
 
        rc = osd_scrub_refresh_mapping(info, dev, fid, lid, ops);
@@ -472,6 +517,9 @@ iget:
                        sf->sf_items_updated_prior++;
                else
                        sf->sf_items_updated++;
+
+               /* The target has been changed, need to be re-loaded. */
+               lu_object_purge(info->oti_env, osd2lu_dev(dev), fid);
        }
 
        GOTO(out, rc);
@@ -487,7 +535,11 @@ out:
        }
 
        if (ops == DTO_INDEX_INSERT) {
-               mutex_unlock(&inode->i_mutex);
+               /* There may be conflict unlink during the OI scrub,
+                * if happend, then remove the new added OI mapping. */
+               if (unlikely(inode->i_nlink == 0))
+                       osd_scrub_refresh_mapping(info, dev, fid, lid,
+                                                 DTO_INDEX_DELETE);
                iput(inode);
        }
        up_write(&scrub->os_rwsem);
@@ -542,6 +594,10 @@ static void osd_scrub_post(struct osd_scrub *scrub, int result)
        }
        sf->sf_time_last_checkpoint = cfs_time_current_sec();
        if (result > 0) {
+               struct osd_device *dev =
+                       container_of0(scrub, struct osd_device, od_scrub);
+
+               dev->od_igif_inoi = 1;
                sf->sf_status = SS_COMPLETED;
                memset(sf->sf_oi_bitmap, 0, SCRUB_OI_BITMAP_SIZE);
                sf->sf_flags &= ~(SF_RECREATED | SF_INCONSISTENT |
@@ -642,12 +698,12 @@ static int osd_iit_iget(struct osd_thread_info *info, struct osd_device *dev,
 
        rc = osd_get_lma(info, inode, &info->oti_obj_dentry, lma);
        if (rc == 0) {
-               if (!scrub) {
-                       if (!fid_is_client_visible(&lma->lma_self_fid))
-                               rc = SCRUB_NEXT_CONTINUE;
-                       else
-                               *fid = lma->lma_self_fid;
-               }
+               if (fid_is_llog(&lma->lma_self_fid) ||
+                   (!scrub && fid_is_internal(&lma->lma_self_fid)) ||
+                   (scrub && (lma->lma_incompat & LMAI_AGENT)))
+                       rc = SCRUB_NEXT_CONTINUE;
+               else
+                       *fid = lma->lma_self_fid;
        } else if (rc == -ENODATA) {
                lu_igif_build(fid, inode->i_ino, inode->i_generation);
                if (scrub)
@@ -746,6 +802,21 @@ static int osd_preload_next(struct osd_thread_info *info,
        return rc;
 }
 
+static inline int
+osd_scrub_wakeup(struct osd_scrub *scrub, struct osd_otable_it *it)
+{
+       spin_lock(&scrub->os_lock);
+       if (osd_scrub_has_window(scrub, &it->ooi_cache) ||
+           !cfs_list_empty(&scrub->os_inconsistent_items) ||
+           it->ooi_waiting || !thread_is_running(&scrub->os_thread))
+               scrub->os_waiting = 0;
+       else
+               scrub->os_waiting = 1;
+       spin_unlock(&scrub->os_lock);
+
+       return !scrub->os_waiting;
+}
+
 static int osd_scrub_exec(struct osd_thread_info *info, struct osd_device *dev,
                          struct osd_iit_param *param,
                          struct osd_idmap_cache *oic, int *noslot, int rc)
@@ -789,28 +860,27 @@ static int osd_scrub_exec(struct osd_thread_info *info, struct osd_device *dev,
 
 next:
        scrub->os_pos_current = param->gbase + ++(param->offset);
+
+wait:
        if (it != NULL && it->ooi_waiting &&
            ooc->ooc_pos_preload < scrub->os_pos_current) {
+               spin_lock(&scrub->os_lock);
                it->ooi_waiting = 0;
                cfs_waitq_broadcast(&thread->t_ctl_waitq);
+               spin_unlock(&scrub->os_lock);
        }
 
        if (scrub->os_full_speed || rc == SCRUB_NEXT_CONTINUE)
                return 0;
 
-wait:
        if (osd_scrub_has_window(scrub, ooc)) {
                *noslot = 0;
                return 0;
        }
 
-       scrub->os_waiting = 1;
        l_wait_event(thread->t_ctl_waitq,
-                    osd_scrub_has_window(scrub, ooc) ||
-                    !cfs_list_empty(&scrub->os_inconsistent_items) ||
-                    !thread_is_running(thread),
+                    osd_scrub_wakeup(scrub, it),
                     &lwi);
-       scrub->os_waiting = 0;
 
        if (osd_scrub_has_window(scrub, ooc))
                *noslot = 0;
@@ -837,7 +907,7 @@ static int osd_preload_exec(struct osd_thread_info *info,
 #define SCRUB_IT_CRASH 2
 
 static int osd_inode_iteration(struct osd_thread_info *info,
-                              struct osd_device *dev, __u32 max, int preload)
+                              struct osd_device *dev, __u32 max, bool preload)
 {
        osd_iit_next_policy   next;
        osd_iit_exec_policy   exec;
@@ -849,7 +919,7 @@ static int osd_inode_iteration(struct osd_thread_info *info,
        int                   rc;
        ENTRY;
 
-       if (preload == 0) {
+       if (!preload) {
                struct osd_scrub *scrub = &dev->od_scrub;
 
                next = osd_scrub_next;
@@ -925,7 +995,7 @@ static int osd_otable_it_preload(const struct lu_env *env,
        ENTRY;
 
        rc = osd_inode_iteration(osd_oti_get(env), dev,
-                                OSD_OTABLE_IT_CACHE_SIZE, 1);
+                                OSD_OTABLE_IT_CACHE_SIZE, true);
        if (rc == SCRUB_IT_ALL)
                it->ooi_all_cached = 1;
 
@@ -951,7 +1021,6 @@ static int osd_scrub_main(void *args)
        int                   rc;
        ENTRY;
 
-       cfs_daemonize("OI_scrub");
        rc = lu_env_init(&env, LCT_DT_THREAD);
        if (rc != 0) {
                CERROR("%.16s: OI scrub, fail to init env, rc = %d\n",
@@ -977,14 +1046,13 @@ static int osd_scrub_main(void *args)
                if (unlikely(!thread_is_running(thread)))
                        GOTO(post, rc = 0);
 
-               LASSERT(scrub->os_pos_current >= ooc->ooc_pos_preload);
                scrub->os_pos_current = ooc->ooc_pos_preload;
        }
 
        CDEBUG(D_LFSCK, "OI scrub: flags = 0x%x, pos = %u\n",
               scrub->os_start_flags, scrub->os_pos_current);
 
-       rc = osd_inode_iteration(osd_oti_get(&env), dev, ~0U, 0);
+       rc = osd_inode_iteration(osd_oti_get(&env), dev, ~0U, false);
        if (unlikely(rc == SCRUB_IT_CRASH))
                GOTO(out, rc = -EINVAL);
        GOTO(post, rc);
@@ -1062,16 +1130,12 @@ static const struct osd_lf_map osd_lf_maps[] = {
                osd_ios_general_scan, osd_ios_varfid_fill },
 
        /* PENDING */
-       { "PENDING", { FID_SEQ_LOCAL_FILE, MDD_ORPHAN_OID, 0 }, 0, NULL, NULL },
+       { "PENDING", { 0, 0, 0 }, 0, NULL, NULL },
 
        /* ROOT */
-       { "ROOT", { FID_SEQ_LOCAL_FILE, MDD_ROOT_INDEX_OID, 0 },
+       { "ROOT", { FID_SEQ_ROOT, 1, 0 },
                OLF_SCAN_SUBITEMS | OLF_HIDE_FID, osd_ios_ROOT_scan, NULL },
 
-       /* capa_keys */
-       { CAPA_KEYS, { FID_SEQ_LOCAL_FILE, MDD_CAPA_KEYS_OID, 0 }, 0,
-               NULL, NULL },
-
        /* changelog_catalog */
        { CHANGELOG_CATALOG, { 0, 0, 0 }, 0, NULL, NULL },
 
@@ -1087,8 +1151,7 @@ static const struct osd_lf_map osd_lf_maps[] = {
                NULL, NULL },
 
        /* lfsck_bookmark */
-       { "lfsck_bookmark", { FID_SEQ_LOCAL_FILE, LFSCK_BOOKMARK_OID, 0 }, 0,
-               NULL, NULL },
+       { "lfsck_bookmark", { 0, 0, 0 }, 0, NULL, NULL },
 
        /* lov_objid */
        { LOV_OBJID, { FID_SEQ_LOCAL_FILE, MDD_LOV_OBJ_OID, 0 }, OLF_SHOW_NAME,
@@ -1106,10 +1169,6 @@ static const struct osd_lf_map osd_lf_maps[] = {
        { QSD_DIR, { 0, 0, 0 }, OLF_SCAN_SUBITEMS,
                osd_ios_general_scan, osd_ios_varfid_fill },
 
-       /* seq-200000003-lastid */
-       { "seq-200000003-lastid", { FID_SEQ_LOCAL_NAME, 1, 0 }, 0,
-               NULL, NULL },
-
        /* seq_ctl */
        { "seq_ctl", { FID_SEQ_LOCAL_FILE, FID_SEQ_CTL_OID, 0 },
                OLF_SHOW_NAME, NULL, NULL },
@@ -1118,14 +1177,13 @@ static const struct osd_lf_map osd_lf_maps[] = {
        { "seq_srv", { FID_SEQ_LOCAL_FILE, FID_SEQ_SRV_OID, 0 },
                OLF_SHOW_NAME, NULL, NULL },
 
-       /* LAST_GROUP */
-       { "LAST_GROUP", { FID_SEQ_LOCAL_FILE, OFD_LAST_GROUP_OID, 0 },
-               OLF_SHOW_NAME, NULL, NULL },
-
        /* health_check */
        { HEALTH_CHECK, { FID_SEQ_LOCAL_FILE, OFD_HEALTH_CHECK_OID, 0 },
                OLF_SHOW_NAME, NULL, NULL },
 
+       /* lfsck_namespace */
+       { "lfsck_namespace", { 0, 0, 0 }, 0, NULL, NULL },
+
        /* OBJECTS, upgrade from old device */
        { OBJECTS, { 0, 0, 0 }, OLF_SCAN_SUBITEMS, osd_ios_OBJECTS_scan, NULL },
 
@@ -1135,6 +1193,10 @@ static const struct osd_lf_map osd_lf_maps[] = {
        /* lquota_v2.group, upgrade from old device */
        { "lquota_v2.group", { 0, 0, 0 }, 0, NULL, NULL },
 
+       /* LAST_GROUP, upgrade from old device */
+       { "LAST_GROUP", { FID_SEQ_LOCAL_FILE, OFD_LAST_GROUP_OID, 0 },
+               OLF_SHOW_NAME, NULL, NULL },
+
        { NULL, { 0, 0, 0 }, 0, NULL, NULL }
 };
 
@@ -1239,7 +1301,7 @@ osd_ios_scan_one(struct osd_thread_info *info, struct osd_device *dev,
                        lu_igif_build(&tfid, inode->i_ino, inode->i_generation);
                else
                        tfid = *fid;
-               rc = osd_ea_fid_set(info, inode, &tfid);
+               rc = osd_ea_fid_set(info, inode, &tfid, 0);
                if (rc != 0)
                        RETURN(rc);
        } else {
@@ -1393,7 +1455,31 @@ osd_ios_ROOT_scan(struct osd_thread_info *info, struct osd_device *dev,
                        }
                }
        } else {
-               rc = osd_ios_scan_one(info, dev, child->d_inode, NULL, 0);
+               /* For lustre-2.x (x <= 3), the ".lustre" has NO FID-in-LMA,
+                * so the client will get IGIF for the ".lustre" object when
+                * the MDT restart.
+                *
+                * From the OI scrub view, when the MDT upgrade to Lustre-2.4,
+                * it does not know whether there are some old clients cached
+                * the ".lustre" IGIF during the upgrading. Two choices:
+                *
+                * 1) Generate IGIF-in-LMA and IGIF-in-OI for the ".lustre".
+                *    It will allow the old connected clients to access the
+                *    ".lustre" with cached IGIF. But it will cause others
+                *    on the MDT failed to check "fid_is_dot_lustre()".
+                *
+                * 2) Use fixed FID {FID_SEQ_DOT_LUSTRE, FID_OID_DOT_LUSTRE, 0}
+                *    for ".lustre" in spite of whether there are some clients
+                *    cached the ".lustre" IGIF or not. It enables the check
+                *    "fid_is_dot_lustre()" on the MDT, although it will cause
+                *    that the old connected clients cannot access the ".lustre"
+                *    with the cached IGIF.
+                *
+                * Usually, it is rare case for the old connected clients
+                * to access the ".lustre" with cached IGIF. So we prefer
+                * to the solution 2). */
+               rc = osd_ios_scan_one(info, dev, child->d_inode,
+                                     &LU_DOT_LUSTRE_FID, 0);
                dput(child);
        }
 
@@ -1436,11 +1522,12 @@ osd_ios_OBJECTS_scan(struct osd_thread_info *info, struct osd_device *dev,
 static int osd_initial_OI_scrub(struct osd_thread_info *info,
                                struct osd_device *dev)
 {
-       struct osd_ios_item *item    = NULL;
-       scandir_t            scandir = osd_ios_general_scan;
-       filldir_t            filldir = osd_ios_root_fill;
-       struct dentry       *dentry  = osd_sb(dev)->s_root;
-       int                  rc;
+       struct osd_ios_item     *item    = NULL;
+       scandir_t                scandir = osd_ios_general_scan;
+       filldir_t                filldir = osd_ios_root_fill;
+       struct dentry           *dentry  = osd_sb(dev)->s_root;
+       const struct osd_lf_map *map     = osd_lf_maps;
+       int                      rc;
        ENTRY;
 
        while (1) {
@@ -1474,7 +1561,32 @@ static int osd_initial_OI_scrub(struct osd_thread_info *info,
                OBD_FREE_PTR(item);
        }
 
-       RETURN(rc);
+       if (rc != 0)
+               RETURN(rc);
+
+       /* There maybe the case that the object has been removed, but its OI
+        * mapping is still in the OI file, such as the "CATALOGS" after MDT
+        * file-level backup/restore. So here cleanup the stale OI mappings. */
+       while (map->olm_name != NULL) {
+               struct dentry *child;
+
+               if (fid_is_zero(&map->olm_fid)) {
+                       map++;
+                       continue;
+               }
+
+               child = osd_ios_lookup_one_len(map->olm_name,
+                                              osd_sb(dev)->s_root,
+                                              strlen(map->olm_name));
+               if (!IS_ERR(child))
+                       dput(child);
+               else if (PTR_ERR(child) == -ENOENT)
+                       osd_scrub_refresh_mapping(info, dev, &map->olm_fid,
+                                                 NULL, DTO_INDEX_DELETE);
+               map++;
+       }
+
+       RETURN(0);
 }
 
 char *osd_lf_fid2name(const struct lu_fid *fid)
@@ -1526,8 +1638,8 @@ again:
 
        scrub->os_start_flags = flags;
        thread_set_flags(thread, 0);
-       rc = cfs_create_thread(osd_scrub_main, dev, 0);
-       if (rc < 0) {
+       rc = PTR_ERR(kthread_run(osd_scrub_main, dev, "OI_scrub"));
+       if (IS_ERR_VALUE(rc)) {
                CERROR("%.16s: cannot start iteration thread, rc = %d\n",
                       LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, rc);
                RETURN(rc);
@@ -1656,14 +1768,33 @@ int osd_scrub_setup(const struct lu_env *env, struct osd_device *dev)
                RETURN(rc);
 
        rc = osd_initial_OI_scrub(info, dev);
-       if (rc == 0 && !dev->od_noscrub &&
-           ((sf->sf_status == SS_PAUSED) ||
-            (sf->sf_status == SS_CRASHED &&
-             sf->sf_flags & (SF_RECREATED | SF_INCONSISTENT | SF_UPGRADE |
-                             SF_AUTO)) ||
-            (sf->sf_status == SS_INIT &&
-             sf->sf_flags & (SF_RECREATED | SF_INCONSISTENT | SF_UPGRADE))))
-               rc = osd_scrub_start(dev);
+       if (rc == 0) {
+               if ((sf->sf_flags & SF_UPGRADE) &&
+                  !(sf->sf_flags & SF_INCONSISTENT))
+                       /* The 'od_igif_inoi' will be set after the
+                        * upgrading completed, needs NOT remount. */
+                       dev->od_igif_inoi = 0;
+               else
+                       /* The 'od_igif_inoi' will be set under the
+                        * following cases:
+                        * 1) new created system, or
+                        * 2) restored from file-level backup, or
+                        * 3) the upgrading completed.
+                        *
+                        * The 'od_igif_inoi' may be cleared by OI scrub
+                        * later if found that the system is upgrading. */
+                       dev->od_igif_inoi = 1;
+
+               if (!dev->od_noscrub &&
+                   ((sf->sf_status == SS_PAUSED) ||
+                    (sf->sf_status == SS_CRASHED &&
+                     sf->sf_flags & (SF_RECREATED | SF_INCONSISTENT |
+                                     SF_UPGRADE | SF_AUTO)) ||
+                    (sf->sf_status == SS_INIT &&
+                     sf->sf_flags & (SF_RECREATED | SF_INCONSISTENT |
+                                     SF_UPGRADE))))
+                       rc = osd_scrub_start(dev);
+       }
 
        RETURN(rc);
 }
@@ -1724,16 +1855,14 @@ static struct dt_it *osd_otable_it_init(const struct lu_env *env,
        }
 
        rc = do_osd_scrub_start(dev, start);
-       if (rc == -EALREADY) {
-               it->ooi_cache.ooc_pos_preload = scrub->os_pos_current - 1;
-       } else if (rc < 0) {
+       if (rc < 0 && rc != -EALREADY) {
                dev->od_otable_it = NULL;
                OBD_FREE_PTR(it);
-               GOTO(out, it = ERR_PTR(-EALREADY));
-       } else {
-               it->ooi_cache.ooc_pos_preload = scrub->os_pos_current;
+               GOTO(out, it = ERR_PTR(rc));
        }
 
+       it->ooi_cache.ooc_pos_preload = scrub->os_pos_current;
+
        GOTO(out, it);
 
 out:
@@ -1756,50 +1885,29 @@ static void osd_otable_it_fini(const struct lu_env *env, struct dt_it *di)
        OBD_FREE_PTR(it);
 }
 
-/**
- * XXX: Temporary used to notify otable iteration to be paused.
- */
-static void osd_otable_it_put(const struct lu_env *env, struct dt_it *di)
-{
-       struct osd_device *dev = ((struct osd_otable_it *)di)->ooi_dev;
-
-       /* od_otable_mutex: prevent curcurrent init/fini */
-       mutex_lock(&dev->od_otable_mutex);
-       dev->od_scrub.os_paused = 1;
-       mutex_unlock(&dev->od_otable_mutex);
-}
-
-/**
- * Set the OSD layer iteration start position as the specified key.
- *
- * The LFSCK out of OSD layer does not know the detail of the key, so if there
- * are several keys, they cannot be compared out of OSD, so call "::get()" for
- * each key, and OSD will select the smallest one by itself.
- */
 static int osd_otable_it_get(const struct lu_env *env,
                             struct dt_it *di, const struct dt_key *key)
 {
-       struct osd_otable_it    *it  = (struct osd_otable_it *)di;
-       struct osd_otable_cache *ooc = &it->ooi_cache;
-       const char              *str = (const char *)key;
-       __u32                    ino;
-       ENTRY;
-
-       /* Forbid to set iteration position after iteration started. */
-       if (it->ooi_user_ready)
-               RETURN(-EPERM);
-
-       if (str[0] == '\0')
-               RETURN(-EINVAL);
+       return 0;
+}
 
-       if (sscanf(str, "%u", &ino) <= 0)
-               RETURN(-EINVAL);
+static void osd_otable_it_put(const struct lu_env *env, struct dt_it *di)
+{
+}
 
-       /* Skip the one that has been processed last time. */
-       if (ooc->ooc_pos_preload > ++ino)
-               ooc->ooc_pos_preload = ino;
+static inline int
+osd_otable_it_wakeup(struct osd_scrub *scrub, struct osd_otable_it *it)
+{
+       spin_lock(&scrub->os_lock);
+       if (it->ooi_cache.ooc_pos_preload < scrub->os_pos_current ||
+           scrub->os_waiting ||
+           !thread_is_running(&scrub->os_thread))
+               it->ooi_waiting = 0;
+       else
+               it->ooi_waiting = 1;
+       spin_unlock(&scrub->os_lock);
 
-       RETURN(0);
+       return !it->ooi_waiting;
 }
 
 static int osd_otable_it_next(const struct lu_env *env, struct dt_it *di)
@@ -1833,12 +1941,17 @@ again:
                RETURN(1);
        }
 
-       it->ooi_waiting = 1;
-       l_wait_event(thread->t_ctl_waitq,
-                    ooc->ooc_pos_preload < scrub->os_pos_current ||
-                    !thread_is_running(thread),
-                    &lwi);
-       it->ooi_waiting = 0;
+       if (scrub->os_waiting && osd_scrub_has_window(scrub, ooc)) {
+               spin_lock(&scrub->os_lock);
+               scrub->os_waiting = 0;
+               cfs_waitq_broadcast(&scrub->os_thread.t_ctl_waitq);
+               spin_unlock(&scrub->os_lock);
+       }
+
+       if (it->ooi_cache.ooc_pos_preload >= scrub->os_pos_current)
+               l_wait_event(thread->t_ctl_waitq,
+                            osd_otable_it_wakeup(scrub, it),
+                            &lwi);
 
        if (!thread_is_running(thread) && !it->ooi_used_outside)
                RETURN(1);
@@ -1853,18 +1966,13 @@ again:
 static struct dt_key *osd_otable_it_key(const struct lu_env *env,
                                        const struct dt_it *di)
 {
-       struct osd_otable_it    *it  = (struct osd_otable_it *)di;
-       struct osd_otable_cache *ooc = &it->ooi_cache;
-
-       sprintf(it->ooi_key, "%u",
-               ooc->ooc_cache[ooc->ooc_consumer_idx].oic_lid.oii_ino);
-       return (struct dt_key *)it->ooi_key;
+       return NULL;
 }
 
 static int osd_otable_it_key_size(const struct lu_env *env,
                                  const struct dt_it *di)
 {
-       return sizeof(((struct osd_otable_it *)di)->ooi_key);
+       return sizeof(__u64);
 }
 
 static int osd_otable_it_rec(const struct lu_env *env, const struct dt_it *di,
@@ -1874,9 +1982,33 @@ static int osd_otable_it_rec(const struct lu_env *env, const struct dt_it *di,
        struct osd_otable_cache *ooc = &it->ooi_cache;
 
        *(struct lu_fid *)rec = ooc->ooc_cache[ooc->ooc_consumer_idx].oic_fid;
+
+       /* Filter out Invald FID already. */
+       LASSERTF(fid_is_sane((struct lu_fid *)rec),
+                "Invalid FID "DFID", p_idx = %d, c_idx = %d\n",
+                PFID((struct lu_fid *)rec),
+                ooc->ooc_producer_idx, ooc->ooc_consumer_idx);
+
        return 0;
 }
 
+static __u64 osd_otable_it_store(const struct lu_env *env,
+                                const struct dt_it *di)
+{
+       struct osd_otable_it    *it  = (struct osd_otable_it *)di;
+       struct osd_otable_cache *ooc = &it->ooi_cache;
+       __u64                    hash;
+
+       if (it->ooi_user_ready && ooc->ooc_consumer_idx != -1)
+               hash = ooc->ooc_cache[ooc->ooc_consumer_idx].oic_lid.oii_ino;
+       else
+               hash = ooc->ooc_pos_preload;
+       return hash;
+}
+
+/**
+ * Set the OSD layer iteration start position as the specified hash.
+ */
 static int osd_otable_it_load(const struct lu_env *env,
                              const struct dt_it *di, __u64 hash)
 {
@@ -1884,31 +2016,49 @@ static int osd_otable_it_load(const struct lu_env *env,
        struct osd_device       *dev   = it->ooi_dev;
        struct osd_otable_cache *ooc   = &it->ooi_cache;
        struct osd_scrub        *scrub = &dev->od_scrub;
+       int                      rc;
+       ENTRY;
 
+       /* Forbid to set iteration position after iteration started. */
        if (it->ooi_user_ready)
-               return 0;
+               RETURN(-EPERM);
 
+       if (hash > OSD_OTABLE_MAX_HASH)
+               hash = OSD_OTABLE_MAX_HASH;
+
+       ooc->ooc_pos_preload = hash;
        if (ooc->ooc_pos_preload <= LDISKFS_FIRST_INO(osd_sb(dev)))
                ooc->ooc_pos_preload = LDISKFS_FIRST_INO(osd_sb(dev)) + 1;
+
        it->ooi_user_ready = 1;
        if (!scrub->os_full_speed)
                cfs_waitq_broadcast(&scrub->os_thread.t_ctl_waitq);
 
        /* Unplug OSD layer iteration by the first next() call. */
-       return osd_otable_it_next(env, (struct dt_it *)it);
+       rc = osd_otable_it_next(env, (struct dt_it *)it);
+
+       RETURN(rc);
+}
+
+static int osd_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_otable_ops = {
        .dio_it = {
                .init     = osd_otable_it_init,
                .fini     = osd_otable_it_fini,
-               .put      = osd_otable_it_put,
                .get      = osd_otable_it_get,
+               .put      = osd_otable_it_put,
                .next     = osd_otable_it_next,
-               .key      = osd_otable_it_key,
+               .key      = osd_otable_it_key,
                .key_size = osd_otable_it_key_size,
                .rec      = osd_otable_it_rec,
+               .store    = osd_otable_it_store,
                .load     = osd_otable_it_load,
+               .key_rec  = osd_otable_it_key_rec,
        }
 };